repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/crunet.py
""" CRU-Net, implemented in Gluon. Original paper: 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. """ __all__ = ['CRUNet', 'crunet56', 'crunet116'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import pre_conv1x1_block, pre_conv3x3_block from .resnet import ResInitBlock from .preresnet import PreResActivation def cru_conv3x3(in_channels, out_channels, strides=1, padding=1, groups=1, use_bias=False, conv_params=None): """ CRU-Net specific convolution 3x3 layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2D( channels=out_channels, kernel_size=3, strides=strides, padding=padding, groups=groups, use_bias=use_bias, in_channels=in_channels, params=conv_params) class CRUConvBlock(HybridBlock): """ CRU-Net specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. groups : int, default 1 Number of groups. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. It's used by PreResNet. conv_params : ParameterDict, default None Weights for the convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, groups=1, bn_use_global_stats=False, return_preact=False, conv_params=None, **kwargs): super(CRUConvBlock, self).__init__(**kwargs) self.return_preact = return_preact with self.name_scope(): self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, groups=groups, use_bias=False, in_channels=in_channels, params=conv_params) def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) if self.return_preact: x_pre_activ = x x = self.conv(x) if self.return_preact: return x, x_pre_activ else: return x def cru_conv1x1_block(in_channels, out_channels, strides=1, bn_use_global_stats=False, return_preact=False, conv_params=None): """ 1x1 version of the CRU-Net specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. conv_params : ParameterDict, default None Weights for the convolution layer. """ return CRUConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, bn_use_global_stats=bn_use_global_stats, return_preact=return_preact, conv_params=conv_params) class ResBottleneck(HybridBlock): """ Pre-ResNeXt bottleneck block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, **kwargs): super(ResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // 4 D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D with self.name_scope(): self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=group_width, bn_use_global_stats=bn_use_global_stats) self.conv2 = pre_conv3x3_block( in_channels=group_width, out_channels=group_width, strides=strides, groups=cardinality, bn_use_global_stats=bn_use_global_stats) self.conv3 = pre_conv1x1_block( in_channels=group_width, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class CRUBottleneck(HybridBlock): """ CRU-Net bottleneck block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. group_width: int Group width parameter. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. conv1_params : ParameterDict, default None Weights for the convolution layer #1. conv2_params : ParameterDict, default None Weights for the convolution layer #2. """ def __init__(self, in_channels, out_channels, strides, group_width, bn_use_global_stats, conv1_params=None, conv2_params=None, **kwargs): super(CRUBottleneck, self).__init__(**kwargs) with self.name_scope(): self.conv1 = cru_conv1x1_block( in_channels=in_channels, out_channels=group_width, bn_use_global_stats=bn_use_global_stats, conv_params=conv1_params) self.conv2 = cru_conv3x3( in_channels=group_width, out_channels=group_width, strides=strides, groups=group_width, conv_params=conv2_params) self.conv3 = pre_conv1x1_block( in_channels=group_width, out_channels=group_width, bn_use_global_stats=bn_use_global_stats) self.conv4 = pre_conv1x1_block( in_channels=group_width, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.conv4(x) return x class ResUnit(HybridBlock): """ CRU-Net residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, **kwargs): super(ResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = pre_conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity return x class CRUUnit(HybridBlock): """ CRU-Net collective residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. group_width: int Group width parameter. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. conv1_params : ParameterDict, default None Weights for the convolution layer #1. conv2_params : ParameterDict, default None Weights for the convolution layer #2. """ def __init__(self, in_channels, out_channels, strides, group_width, bn_use_global_stats, conv1_params=None, conv2_params=None, **kwargs): super(CRUUnit, self).__init__(**kwargs) assert (strides == 1) or ((conv1_params is None) and (conv2_params is None)) self.resize_input = (in_channels != out_channels) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if self.resize_input: self.input_conv = pre_conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.body = CRUBottleneck( in_channels=out_channels, out_channels=out_channels, strides=strides, group_width=group_width, bn_use_global_stats=bn_use_global_stats, conv1_params=conv1_params, conv2_params=conv2_params) if self.resize_identity: self.identity_conv = cru_conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, conv_params=self.input_conv.conv.params) def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x if self.resize_input: x = self.input_conv(x) x = self.body(x) x = x + identity return x class CRUNet(HybridBlock): """ CRU-Net model from 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. group_widths: list of int List of group width parameters. refresh_steps: list of int List of refresh step parameters. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, group_widths, refresh_steps, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(CRUNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) group_width = group_widths[i] refresh_step = refresh_steps[i] with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 if group_width != 0: if ((refresh_step == 0) and (j == 0)) or ((refresh_step != 0) and (j % refresh_step == 0)): conv1_params = None conv2_params = None unit = CRUUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, group_width=group_width, bn_use_global_stats=bn_use_global_stats, conv1_params=conv1_params, conv2_params=conv2_params) if conv1_params is None: conv1_params = unit.body.conv1.conv.params conv2_params = unit.body.conv2.params stage.add(unit) else: stage.add(ResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_crunet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create CRU-Net model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ cardinality = 32 bottleneck_width = 4 if blocks == 56: layers = [3, 4, 6, 3] group_widths = [0, 0, 640, 0] refresh_steps = [0, 0, 0, 0] elif blocks == 116: layers = [3, 6, 18, 3] group_widths = [0, 352, 704, 0] refresh_steps = [0, 0, 6, 0] else: raise ValueError("Unsupported CRU-Net with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = CRUNet( channels=channels, init_block_channels=init_block_channels, cardinality=cardinality, bottleneck_width=bottleneck_width, group_widths=group_widths, refresh_steps=refresh_steps, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def crunet56(**kwargs): """ CRU-Net-56 model from 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_crunet(blocks=56, model_name="crunet56", **kwargs) def crunet116(**kwargs): """ CRU-Net-116 model from 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_crunet(blocks=116, model_name="crunet116", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ crunet56, crunet116, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != crunet56 or weight_count == 25609384) assert (model != crunet116 or weight_count == 43656136) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
21,164
32.436019
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/fbnet.py
""" FBNet for ImageNet-1K, implemented in Gluon. Original paper: 'FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search,' https://arxiv.org/abs/1812.03443. """ __all__ = ['FBNet', 'fbnet_cb'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block class FBNetUnit(HybridBlock): """ FBNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the second convolution layer. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. use_kernel3 : bool Whether to use 3x3 (instead of 5x5) kernel. exp_factor : int Expansion factor for each unit. activation : str, default 'relu' Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, strides, bn_epsilon, bn_use_global_stats, use_kernel3, exp_factor, activation="relu", **kwargs): super(FBNetUnit, self).__init__(**kwargs) assert (exp_factor >= 1) self.residual = (in_channels == out_channels) and (strides == 1) self.use_exp_conv = True mid_channels = exp_factor * in_channels with self.name_scope(): if self.use_exp_conv: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, activation=activation) if use_kernel3: self.conv1 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, activation=activation) else: self.conv1 = dwconv5x5_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, activation=activation) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): if self.residual: identity = x if self.use_exp_conv: x = self.exp_conv(x) x = self.conv1(x) x = self.conv2(x) if self.residual: x = x + identity return x class FBNetInitBlock(HybridBlock): """ FBNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(FBNetInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=2, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.conv2 = FBNetUnit( in_channels=out_channels, out_channels=out_channels, strides=1, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, use_kernel3=True, exp_factor=1) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class FBNet(HybridBlock): """ FBNet model from 'FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search,' https://arxiv.org/abs/1812.03443. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. final_block_channels : int Number of output channels for the final block of the feature extractor. kernels3 : list of list of int/bool Using 3x3 (instead of 5x5) kernel for each unit. exp_factors : list of list of int Expansion factor for each unit. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, kernels3, exp_factors, bn_epsilon=1e-5, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(FBNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(FBNetInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) else 1 use_kernel3 = kernels3[i][j] == 1 exp_factor = exp_factors[i][j] stage.add(FBNetUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, use_kernel3=use_kernel3, exp_factor=exp_factor)) in_channels = out_channels self.features.add(stage) self.features.add(conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) in_channels = final_block_channels self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_fbnet(version, bn_epsilon=1e-5, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create FBNet model with specific parameters. Parameters: ---------- version : str Version of MobileNetV3 ('a', 'b' or 'c'). bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if version == "c": init_block_channels = 16 final_block_channels = 1984 channels = [[24, 24, 24], [32, 32, 32, 32], [64, 64, 64, 64, 112, 112, 112, 112], [184, 184, 184, 184, 352]] kernels3 = [[1, 1, 1], [0, 0, 0, 1], [0, 0, 0, 0, 0, 0, 0, 0], [0, 0, 0, 0, 1]] exp_factors = [[6, 1, 1], [6, 3, 6, 6], [6, 3, 6, 6, 6, 6, 6, 3], [6, 6, 6, 6, 6]] else: raise ValueError("Unsupported FBNet version {}".format(version)) net = FBNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, kernels3=kernels3, exp_factors=exp_factors, bn_epsilon=bn_epsilon, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def fbnet_cb(**kwargs): """ FBNet-Cb model (bn_epsilon=1e-3) from 'FBNet: Hardware-Aware Efficient ConvNet Design via Differentiable Neural Architecture Search,' https://arxiv.org/abs/1812.03443. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fbnet(version="c", bn_epsilon=1e-3, model_name="fbnet_cb", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ fbnet_cb, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fbnet_cb or weight_count == 5572200) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
11,752
33.567647
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/visemenet.py
""" VisemeNet for speech-driven facial animation, implemented in Gluon. Original paper: 'VisemeNet: Audio-Driven Animator-Centric Speech Animation,' https://arxiv.org/abs/1805.09488. """ __all__ = ['VisemeNet', 'visemenet20'] import os from mxnet import cpu from mxnet.gluon import nn, rnn, HybridBlock from .common import DenseBlock class VisemeDenseBranch(HybridBlock): """ VisemeNet dense branch. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of middle/output channels. """ def __init__(self, in_channels, out_channels_list, **kwargs): super(VisemeDenseBranch, self).__init__(**kwargs) with self.name_scope(): self.branch = nn.HybridSequential(prefix="") with self.branch.name_scope(): for out_channels in out_channels_list[:-1]: self.branch.add(DenseBlock( in_channels=in_channels, out_channels=out_channels, use_bias=True, use_bn=True)) in_channels = out_channels self.final_fc = nn.Dense( units=out_channels_list[-1], in_units=in_channels) def hybrid_forward(self, F, x): x = self.branch(x) y = self.final_fc(x) return y, x class VisemeRnnBranch(HybridBlock): """ VisemeNet RNN branch. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of middle/output channels. rnn_num_layers : int Number of RNN layers. dropout_rate : float Dropout rate. """ def __init__(self, in_channels, out_channels_list, rnn_num_layers, dropout_rate, **kwargs): super(VisemeRnnBranch, self).__init__(**kwargs) with self.name_scope(): self.rnn = rnn.LSTM( hidden_size=out_channels_list[0], num_layers=rnn_num_layers, dropout=dropout_rate, input_size=in_channels) self.fc_branch = VisemeDenseBranch( in_channels=out_channels_list[0], out_channels_list=out_channels_list[1:]) def hybrid_forward(self, F, x): x = self.rnn(x) x = x[:, -1, :] y, _ = self.fc_branch(x) return y class VisemeNet(HybridBlock): """ VisemeNet model from 'VisemeNet: Audio-Driven Animator-Centric Speech Animation,' https://arxiv.org/abs/1805.09488. Parameters: ---------- audio_features : int, default 195 Number of audio features (characters/sounds). audio_window_size : int, default 8 Size of audio window (for time related audio features). stage2_window_size : int, default 64 Size of window for stage #2. num_face_ids : int, default 76 Number of face IDs. num_landmarks : int, default 76 Number of landmarks. num_phonemes : int, default 21 Number of phonemes. num_visemes : int, default 20 Number of visemes. dropout_rate : float, default 0.5 Dropout rate for RNNs. """ def __init__(self, audio_features=195, audio_window_size=8, stage2_window_size=64, num_face_ids=76, num_landmarks=76, num_phonemes=21, num_visemes=20, dropout_rate=0.5, **kwargs): super(VisemeNet, self).__init__(**kwargs) stage1_rnn_hidden_size = 256 stage1_fc_mid_channels = 256 stage2_rnn_in_features = (audio_features + num_landmarks + stage1_fc_mid_channels) * \ stage2_window_size // audio_window_size self.audio_window_size = audio_window_size self.stage2_window_size = stage2_window_size with self.name_scope(): self.stage1_rnn = rnn.LSTM( hidden_size=stage1_rnn_hidden_size, num_layers=3, dropout=dropout_rate, input_size=audio_features) self.lm_branch = VisemeDenseBranch( in_channels=(stage1_rnn_hidden_size + num_face_ids), out_channels_list=[stage1_fc_mid_channels, num_landmarks]) self.ph_branch = VisemeDenseBranch( in_channels=(stage1_rnn_hidden_size + num_face_ids), out_channels_list=[stage1_fc_mid_channels, num_phonemes]) self.cls_branch = VisemeRnnBranch( in_channels=stage2_rnn_in_features, out_channels_list=[256, 200, num_visemes], rnn_num_layers=1, dropout_rate=dropout_rate) self.reg_branch = VisemeRnnBranch( in_channels=stage2_rnn_in_features, out_channels_list=[256, 200, 100, num_visemes], rnn_num_layers=3, dropout_rate=dropout_rate) self.jali_branch = VisemeRnnBranch( in_channels=stage2_rnn_in_features, out_channels_list=[128, 200, 2], rnn_num_layers=3, dropout_rate=dropout_rate) def hybrid_forward(self, F, x, pid): y = self.stage1_rnn(x) y = y[:, -1, :] y = F.concat(y, pid, dim=1) lm, _ = self.lm_branch(y) lm += pid ph, ph1 = self.ph_branch(y) z = F.concat(lm, ph1, dim=1) z2 = F.concat(z, x[:, self.audio_window_size // 2, :], dim=1) n_net2_input = z2.shape[1] z2 = F.concat( F.zeros((self.stage2_window_size // 2, n_net2_input)), z2, dim=0) z = F.stack( *[z2[i:i + self.stage2_window_size].reshape( (self.audio_window_size, n_net2_input * self.stage2_window_size // self.audio_window_size)) for i in range(z2.shape[0] - self.stage2_window_size)], axis=0) cls = self.cls_branch(z) reg = self.reg_branch(z) jali = self.jali_branch(z) return cls, reg, jali def get_visemenet(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create VisemeNet model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ net = VisemeNet( **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def visemenet20(**kwargs): """ VisemeNet model for 20 visemes (without co-articulation rules) from 'VisemeNet: Audio-Driven Animator-Centric Speech Animation,' https://arxiv.org/abs/1805.09488. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_visemenet(model_name="visemenet20", **kwargs) def _calc_width(net): import numpy as np net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) return weight_count def _test(): import mxnet as mx pretrained = False models = [ visemenet20, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != visemenet20 or weight_count == 14574303) batch = 34 audio_window_size = 8 audio_features = 195 num_face_ids = 76 num_visemes = 20 x = mx.nd.random.normal(shape=(batch, audio_window_size, audio_features), ctx=ctx) pid = mx.nd.full(shape=(batch, num_face_ids), val=3, ctx=ctx) y1, y2, y3 = net(x, pid) assert (y1.shape[0] == y2.shape[0] == y3.shape[0]) assert (y1.shape[1] == y2.shape[1] == num_visemes) assert (y3.shape[1] == 2) if __name__ == "__main__": _test()
9,324
31.155172
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/fractalnet_cifar.py
""" FractalNet for CIFAR, implemented in Gluon. Original paper: 'FractalNet: Ultra-Deep Neural Networks without Residuals,' https://arxiv.org/abs/1605.07648. """ __all__ = ['CIFARFractalNet', 'fractalnet_cifar10', 'fractalnet_cifar100'] import os import numpy as np import mxnet as mx from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import ParametricSequential class DropConvBlock(HybridBlock): """ Convolution block with Batch normalization, ReLU activation, and Dropout layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, use_bias=False, bn_use_global_stats=False, dropout_prob=0.0, **kwargs): super(DropConvBlock, self).__init__(**kwargs) self.use_dropout = (dropout_prob != 0.0) with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, use_bias=use_bias, in_channels=in_channels) self.bn = nn.BatchNorm( in_channels=out_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_prob) def hybrid_forward(self, F, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) if self.use_dropout: x = self.dropout(x) return x def drop_conv3x3_block(in_channels, out_channels, strides=1, padding=1, use_bias=False, bn_use_global_stats=False, dropout_prob=0.0, **kwargs): """ 3x3 version of the convolution block with dropout. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. """ return DropConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, use_bias=use_bias, bn_use_global_stats=bn_use_global_stats, dropout_prob=dropout_prob, **kwargs) class FractalBlock(HybridBlock): """ FractalNet block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. num_columns : int Number of columns in each block. loc_drop_prob : float Local drop path probability. dropout_prob : float Probability of dropout. """ def __init__(self, in_channels, out_channels, num_columns, loc_drop_prob, dropout_prob, **kwargs): super(FractalBlock, self).__init__(**kwargs) assert (num_columns >= 1) self.num_columns = num_columns self.loc_drop_prob = loc_drop_prob with self.name_scope(): self.blocks = nn.HybridSequential(prefix="") depth = 2 ** (num_columns - 1) for i in range(depth): level_block_i = nn.HybridSequential(prefix='block{}_'.format(i + 1)) for j in range(self.num_columns): column_step_j = 2 ** j if (i + 1) % column_step_j == 0: in_channels_ij = in_channels if (i + 1 == column_step_j) else out_channels level_block_i.add(drop_conv3x3_block( in_channels=in_channels_ij, out_channels=out_channels, dropout_prob=dropout_prob)) self.blocks.add(level_block_i) @staticmethod def calc_drop_mask(batch_size, glob_num_columns, curr_num_columns, max_num_columns, loc_drop_prob): """ Calculate drop path mask. Parameters: ---------- batch_size : int Size of batch. glob_num_columns : int Number of columns in global drop path mask. curr_num_columns : int Number of active columns in the current level of block. max_num_columns : int Number of columns for all network. loc_drop_prob : float Local drop path probability. Returns: ------- np.array Resulted mask. """ glob_batch_size = glob_num_columns.shape[0] glob_drop_mask = np.zeros((curr_num_columns, glob_batch_size), dtype=np.float32) glob_drop_num_columns = glob_num_columns - (max_num_columns - curr_num_columns) glob_drop_indices = np.where(glob_drop_num_columns >= 0)[0] glob_drop_mask[glob_drop_num_columns[glob_drop_indices], glob_drop_indices] = 1.0 loc_batch_size = batch_size - glob_batch_size loc_drop_mask = np.random.binomial( n=1, p=(1.0 - loc_drop_prob), size=(curr_num_columns, loc_batch_size)).astype(np.float32) alive_count = loc_drop_mask.sum(axis=0) dead_indices = np.where(alive_count == 0.0)[0] loc_drop_mask[np.random.randint(0, curr_num_columns, size=dead_indices.shape), dead_indices] = 1.0 drop_mask = np.concatenate((glob_drop_mask, loc_drop_mask), axis=1) return drop_mask @staticmethod def join_outs(F, raw_outs, glob_num_columns, num_columns, loc_drop_prob, training): """ Join outputs for current level of block. Parameters: ---------- F : namespace Symbol or NDArray namespace. raw_outs : list of Tensor Current outputs from active columns. glob_num_columns : int Number of columns in global drop path mask. num_columns : int Number of columns for all network. loc_drop_prob : float Local drop path probability. training : bool Whether training mode for network. Returns: ------- NDArray Joined output. """ curr_num_columns = len(raw_outs) out = F.stack(*raw_outs, axis=0) assert (out.shape[0] == curr_num_columns) if training: batch_size = out.shape[1] batch_mask = FractalBlock.calc_drop_mask( batch_size=batch_size, glob_num_columns=glob_num_columns, curr_num_columns=curr_num_columns, max_num_columns=num_columns, loc_drop_prob=loc_drop_prob) batch_mask = mx.nd.array(batch_mask, ctx=out.context) assert (batch_mask.shape[0] == curr_num_columns) assert (batch_mask.shape[1] == batch_size) batch_mask = batch_mask.expand_dims(2).expand_dims(3).expand_dims(4) masked_out = out * batch_mask num_alive = batch_mask.sum(axis=0).asnumpy() num_alive[num_alive == 0.0] = 1.0 num_alive = mx.nd.array(num_alive, ctx=out.context) out = masked_out.sum(axis=0) / num_alive else: out = out.mean(axis=0) return out def hybrid_forward(self, F, x, glob_num_columns): outs = [x] * self.num_columns for level_block_i in self.blocks._children.values(): outs_i = [] for j, block_ij in enumerate(level_block_i._children.values()): input_i = outs[j] outs_i.append(block_ij(input_i)) joined_out = FractalBlock.join_outs( F=F, raw_outs=outs_i[::-1], glob_num_columns=glob_num_columns, num_columns=self.num_columns, loc_drop_prob=self.loc_drop_prob, training=mx.autograd.is_training()) len_level_block_i = len(level_block_i._children.values()) for j in range(len_level_block_i): outs[j] = joined_out return outs[0] class FractalUnit(HybridBlock): """ FractalNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. num_columns : int Number of columns in each block. loc_drop_prob : float Local drop path probability. dropout_prob : float Probability of dropout. """ def __init__(self, in_channels, out_channels, num_columns, loc_drop_prob, dropout_prob, **kwargs): super(FractalUnit, self).__init__(**kwargs) with self.name_scope(): self.block = FractalBlock( in_channels=in_channels, out_channels=out_channels, num_columns=num_columns, loc_drop_prob=loc_drop_prob, dropout_prob=dropout_prob) self.pool = nn.MaxPool2D( pool_size=2, strides=2) def hybrid_forward(self, F, x, glob_num_columns): x = self.block(x, glob_num_columns) x = self.pool(x) return x class CIFARFractalNet(HybridBlock): """ FractalNet model for CIFAR from 'FractalNet: Ultra-Deep Neural Networks without Residuals,' https://arxiv.org/abs/1605.07648. Parameters: ---------- channels : list of int Number of output channels for each unit. num_columns : int Number of columns in each block. dropout_probs : list of float Probability of dropout in each block. loc_drop_prob : float Local drop path probability. glob_drop_ratio : float Global drop part fraction. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, num_columns, dropout_probs, loc_drop_prob, glob_drop_ratio, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARFractalNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes self.glob_drop_ratio = glob_drop_ratio self.num_columns = num_columns with self.name_scope(): self.features = ParametricSequential(prefix="") for i, out_channels in enumerate(channels): dropout_prob = dropout_probs[i] self.features.add(FractalUnit( in_channels=in_channels, out_channels=out_channels, num_columns=num_columns, loc_drop_prob=loc_drop_prob, dropout_prob=dropout_prob)) in_channels = out_channels self.output = nn.Dense( units=classes, in_units=in_channels) def hybrid_forward(self, F, x): glob_batch_size = int(x.shape[0] * self.glob_drop_ratio) glob_num_columns = np.random.randint(0, self.num_columns, size=(glob_batch_size,)) x = self.features(x, glob_num_columns) x = self.output(x) return x def get_fractalnet_cifar(num_classes, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create WRN model for CIFAR with specific parameters. Parameters: ---------- num_classes : int Number of classification classes. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ dropout_probs = (0.0, 0.1, 0.2, 0.3, 0.4) channels = [64 * (2 ** (i if i != len(dropout_probs) - 1 else i - 1)) for i in range(len(dropout_probs))] num_columns = 3 loc_drop_prob = 0.15 glob_drop_ratio = 0.5 net = CIFARFractalNet( channels=channels, num_columns=num_columns, dropout_probs=dropout_probs, loc_drop_prob=loc_drop_prob, glob_drop_ratio=glob_drop_ratio, classes=num_classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def fractalnet_cifar10(num_classes=10, **kwargs): """ FractalNet model for CIFAR-10 from 'FractalNet: Ultra-Deep Neural Networks without Residuals,' https://arxiv.org/abs/1605.07648. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fractalnet_cifar(num_classes=num_classes, model_name="fractalnet_cifar10", **kwargs) def fractalnet_cifar100(num_classes=100, **kwargs): """ FractalNet model for CIFAR-100 from 'FractalNet: Ultra-Deep Neural Networks without Residuals,' https://arxiv.org/abs/1605.07648. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fractalnet_cifar(num_classes=num_classes, model_name="fractalnet_cifar100", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (fractalnet_cifar10, 10), (fractalnet_cifar100, 100), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fractalnet_cifar10 or weight_count == 33724618) assert (model != fractalnet_cifar100 or weight_count == 33770788) x = mx.nd.zeros((14, 3, 32, 32), ctx=ctx) y = net(x) # with mx.autograd.record(): # y = net(x) # y.backward() assert (y.shape == (14, classes)) if __name__ == "__main__": _test()
17,294
32.195777
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/mobilenetv3.py
""" MobileNetV3 for ImageNet-1K, implemented in Gluon. Original paper: 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. """ __all__ = ['MobileNetV3', 'mobilenetv3_small_w7d20', 'mobilenetv3_small_wd2', 'mobilenetv3_small_w3d4', 'mobilenetv3_small_w1', 'mobilenetv3_small_w5d4', 'mobilenetv3_large_w7d20', 'mobilenetv3_large_wd2', 'mobilenetv3_large_w3d4', 'mobilenetv3_large_w1', 'mobilenetv3_large_w5d4'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import round_channels, conv1x1, conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block, SEBlock,\ HSwish class MobileNetV3Unit(HybridBlock): """ MobileNetV3 unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. exp_channels : int Number of middle (expanded) channels. strides : int or tuple/list of 2 int Strides of the second convolution layer. use_kernel3 : bool Whether to use 3x3 (instead of 5x5) kernel. activation : str Activation function or name of activation function. use_se : bool Whether to use SE-module. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. """ def __init__(self, in_channels, out_channels, exp_channels, strides, use_kernel3, activation, use_se, bn_use_global_stats=False, **kwargs): super(MobileNetV3Unit, self).__init__(**kwargs) assert (exp_channels >= out_channels) self.residual = (in_channels == out_channels) and (strides == 1) self.use_se = use_se self.use_exp_conv = exp_channels != out_channels mid_channels = exp_channels with self.name_scope(): if self.use_exp_conv: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, activation=activation) if use_kernel3: self.conv1 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=activation) else: self.conv1 = dwconv5x5_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=activation) if self.use_se: self.se = SEBlock( channels=mid_channels, reduction=4, round_mid=True, out_activation="hsigmoid") self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): if self.residual: identity = x if self.use_exp_conv: x = self.exp_conv(x) x = self.conv1(x) if self.use_se: x = self.se(x) x = self.conv2(x) if self.residual: x = x + identity return x class MobileNetV3FinalBlock(HybridBlock): """ MobileNetV3 final block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. use_se : bool Whether to use SE-module. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. """ def __init__(self, in_channels, out_channels, use_se, bn_use_global_stats=False, **kwargs): super(MobileNetV3FinalBlock, self).__init__(**kwargs) self.use_se = use_se with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation="hswish") if self.use_se: self.se = SEBlock( channels=out_channels, reduction=4, round_mid=True, out_activation="hsigmoid") def hybrid_forward(self, F, x): x = self.conv(x) if self.use_se: x = self.se(x) return x class MobileNetV3Classifier(HybridBlock): """ MobileNetV3 classifier. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of middle channels. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, mid_channels, dropout_rate, **kwargs): super(MobileNetV3Classifier, self).__init__(**kwargs) self.use_dropout = (dropout_rate != 0.0) with self.name_scope(): self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels) self.activ = HSwish() if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, use_bias=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.activ(x) if self.use_dropout: x = self.dropout(x) x = self.conv2(x) return x class MobileNetV3(HybridBlock): """ MobileNetV3 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- channels : list of list of int Number of output channels for each unit. exp_channels : list of list of int Number of middle (expanded) channels for each unit. init_block_channels : int Number of output channels for the initial unit. final_block_channels : int Number of output channels for the final block of the feature extractor. classifier_mid_channels : int Number of middle channels for classifier. kernels3 : list of list of int/bool Using 3x3 (instead of 5x5) kernel for each unit. use_relu : list of list of int/bool Using ReLU activation flag for each unit. use_se : list of list of int/bool Using SE-block flag for each unit. first_stride : bool Whether to use stride for the first stage. final_use_se : bool Whether to use SE-module in the final block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, exp_channels, init_block_channels, final_block_channels, classifier_mid_channels, kernels3, use_relu, use_se, first_stride, final_use_se, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(MobileNetV3, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, strides=2, bn_use_global_stats=bn_use_global_stats, activation="hswish")) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): exp_channels_ij = exp_channels[i][j] strides = 2 if (j == 0) and ((i != 0) or first_stride) else 1 use_kernel3 = kernels3[i][j] == 1 activation = "relu" if use_relu[i][j] == 1 else "hswish" use_se_flag = use_se[i][j] == 1 stage.add(MobileNetV3Unit( in_channels=in_channels, out_channels=out_channels, exp_channels=exp_channels_ij, use_kernel3=use_kernel3, strides=strides, activation=activation, use_se=use_se_flag, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(MobileNetV3FinalBlock( in_channels=in_channels, out_channels=final_block_channels, use_se=final_use_se, bn_use_global_stats=bn_use_global_stats)) in_channels = final_block_channels self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(MobileNetV3Classifier( in_channels=in_channels, out_channels=classes, mid_channels=classifier_mid_channels, dropout_rate=0.2)) self.output.add(nn.Flatten()) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_mobilenetv3(version, width_scale, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create MobileNetV3 model with specific parameters. Parameters: ---------- version : str Version of MobileNetV3 ('small' or 'large'). width_scale : float Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if version == "small": init_block_channels = 16 channels = [[16], [24, 24], [40, 40, 40, 48, 48], [96, 96, 96]] exp_channels = [[16], [72, 88], [96, 240, 240, 120, 144], [288, 576, 576]] kernels3 = [[1], [1, 1], [0, 0, 0, 0, 0], [0, 0, 0]] use_relu = [[1], [1, 1], [0, 0, 0, 0, 0], [0, 0, 0]] use_se = [[1], [0, 0], [1, 1, 1, 1, 1], [1, 1, 1]] first_stride = True final_block_channels = 576 elif version == "large": init_block_channels = 16 channels = [[16], [24, 24], [40, 40, 40], [80, 80, 80, 80, 112, 112], [160, 160, 160]] exp_channels = [[16], [64, 72], [72, 120, 120], [240, 200, 184, 184, 480, 672], [672, 960, 960]] kernels3 = [[1], [1, 1], [0, 0, 0], [1, 1, 1, 1, 1, 1], [0, 0, 0]] use_relu = [[1], [1, 1], [1, 1, 1], [0, 0, 0, 0, 0, 0], [0, 0, 0]] use_se = [[0], [0, 0], [1, 1, 1], [0, 0, 0, 0, 1, 1], [1, 1, 1]] first_stride = False final_block_channels = 960 else: raise ValueError("Unsupported MobileNetV3 version {}".format(version)) final_use_se = False classifier_mid_channels = 1280 if width_scale != 1.0: channels = [[round_channels(cij * width_scale) for cij in ci] for ci in channels] exp_channels = [[round_channels(cij * width_scale) for cij in ci] for ci in exp_channels] init_block_channels = round_channels(init_block_channels * width_scale) if width_scale > 1.0: final_block_channels = round_channels(final_block_channels * width_scale) net = MobileNetV3( channels=channels, exp_channels=exp_channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, classifier_mid_channels=classifier_mid_channels, kernels3=kernels3, use_relu=use_relu, use_se=use_se, first_stride=first_stride, final_use_se=final_use_se, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def mobilenetv3_small_w7d20(**kwargs): """ MobileNetV3 Small 224/0.35 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="small", width_scale=0.35, model_name="mobilenetv3_small_w7d20", **kwargs) def mobilenetv3_small_wd2(**kwargs): """ MobileNetV3 Small 224/0.5 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="small", width_scale=0.5, model_name="mobilenetv3_small_wd2", **kwargs) def mobilenetv3_small_w3d4(**kwargs): """ MobileNetV3 Small 224/0.75 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="small", width_scale=0.75, model_name="mobilenetv3_small_w3d4", **kwargs) def mobilenetv3_small_w1(**kwargs): """ MobileNetV3 Small 224/1.0 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="small", width_scale=1.0, model_name="mobilenetv3_small_w1", **kwargs) def mobilenetv3_small_w5d4(**kwargs): """ MobileNetV3 Small 224/1.25 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="small", width_scale=1.25, model_name="mobilenetv3_small_w5d4", **kwargs) def mobilenetv3_large_w7d20(**kwargs): """ MobileNetV3 Small 224/0.35 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="large", width_scale=0.35, model_name="mobilenetv3_small_w7d20", **kwargs) def mobilenetv3_large_wd2(**kwargs): """ MobileNetV3 Large 224/0.5 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="large", width_scale=0.5, model_name="mobilenetv3_large_wd2", **kwargs) def mobilenetv3_large_w3d4(**kwargs): """ MobileNetV3 Large 224/0.75 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="large", width_scale=0.75, model_name="mobilenetv3_large_w3d4", **kwargs) def mobilenetv3_large_w1(**kwargs): """ MobileNetV3 Large 224/1.0 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="large", width_scale=1.0, model_name="mobilenetv3_large_w1", **kwargs) def mobilenetv3_large_w5d4(**kwargs): """ MobileNetV3 Large 224/1.25 model from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="large", width_scale=1.25, model_name="mobilenetv3_large_w5d4", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ mobilenetv3_small_w7d20, mobilenetv3_small_wd2, mobilenetv3_small_w3d4, mobilenetv3_small_w1, mobilenetv3_small_w5d4, mobilenetv3_large_w7d20, mobilenetv3_large_wd2, mobilenetv3_large_w3d4, mobilenetv3_large_w1, mobilenetv3_large_w5d4, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mobilenetv3_small_w7d20 or weight_count == 2159600) assert (model != mobilenetv3_small_wd2 or weight_count == 2288976) assert (model != mobilenetv3_small_w3d4 or weight_count == 2581312) assert (model != mobilenetv3_small_w1 or weight_count == 2945288) assert (model != mobilenetv3_small_w5d4 or weight_count == 3643632) assert (model != mobilenetv3_large_w7d20 or weight_count == 2943080) assert (model != mobilenetv3_large_wd2 or weight_count == 3334896) assert (model != mobilenetv3_large_w3d4 or weight_count == 4263496) assert (model != mobilenetv3_large_w1 or weight_count == 5481752) assert (model != mobilenetv3_large_w5d4 or weight_count == 7459144) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
21,706
35.238731
118
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/diaresnet.py
""" DIA-ResNet for ImageNet-1K, implemented in Gluon. Original paper: 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. """ __all__ = ['DIAResNet', 'diaresnet10', 'diaresnet12', 'diaresnet14', 'diaresnetbc14b', 'diaresnet16', 'diaresnet18', 'diaresnet26', 'diaresnetbc26b', 'diaresnet34', 'diaresnetbc38b', 'diaresnet50', 'diaresnet50b', 'diaresnet101', 'diaresnet101b', 'diaresnet152', 'diaresnet152b', 'diaresnet200', 'diaresnet200b', 'DIAAttention', 'DIAResUnit'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, DualPathSequential from .resnet import ResBlock, ResBottleneck, ResInitBlock class FirstLSTMAmp(HybridBlock): """ First LSTM amplifier branch. Parameters: ---------- in_units : int Number of input channels. units : int Number of output channels. """ def __init__(self, in_units, units, **kwargs): super(FirstLSTMAmp, self).__init__(**kwargs) mid_units = in_units // 4 with self.name_scope(): self.fc1 = nn.Dense( units=mid_units, in_units=in_units) self.activ = nn.Activation("relu") self.fc2 = nn.Dense( units=units, in_units=mid_units) def hybrid_forward(self, F, x): x = self.fc1(x) x = self.activ(x) x = self.fc2(x) return x class DIALSTMCell(HybridBlock): """ DIA-LSTM cell. Parameters: ---------- in_x_features : int Number of x input channels. in_h_features : int Number of h input channels. num_layers : int Number of amplifiers. dropout_rate : float, default 0.1 Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_x_features, in_h_features, num_layers, dropout_rate=0.1, **kwargs): super(DIALSTMCell, self).__init__(**kwargs) self.num_layers = num_layers out_features = 4 * in_h_features with self.name_scope(): self.x_amps = nn.HybridSequential(prefix="") self.h_amps = nn.HybridSequential(prefix="") for i in range(num_layers): amp_class = FirstLSTMAmp if i == 0 else nn.Dense self.x_amps.add(amp_class( in_units=in_x_features, units=out_features)) self.h_amps.add(amp_class( in_units=in_h_features, units=out_features)) in_x_features = in_h_features self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x, h, c): hy = [] cy = [] for i in range(self.num_layers): hx_i = h[i] cx_i = c[i] gates = self.x_amps[i](x) + self.h_amps[i](hx_i) i_gate, f_gate, c_gate, o_gate = F.split(gates, axis=1, num_outputs=4) i_gate = F.sigmoid(i_gate) f_gate = F.sigmoid(f_gate) c_gate = F.tanh(c_gate) o_gate = F.sigmoid(o_gate) cy_i = (f_gate * cx_i) + (i_gate * c_gate) hy_i = o_gate * F.sigmoid(cy_i) cy.append(cy_i) hy.append(hy_i) x = self.dropout(hy_i) return hy, cy class DIAAttention(HybridBlock): """ DIA-Net attention module. Parameters: ---------- in_x_features : int Number of x input channels. in_h_features : int Number of h input channels. num_layers : int, default 1 Number of amplifiers. """ def __init__(self, in_x_features, in_h_features, num_layers=1, **kwargs): super(DIAAttention, self).__init__(**kwargs) self.num_layers = num_layers with self.name_scope(): self.lstm = DIALSTMCell( in_x_features=in_x_features, in_h_features=in_h_features, num_layers=num_layers) def hybrid_forward(self, F, x, hc=None): w = F.contrib.AdaptiveAvgPooling2D(x, output_size=1) w = w.flatten() if hc is None: h = [F.zeros_like(w)] * self.num_layers c = [F.zeros_like(w)] * self.num_layers else: h, c = hc h, c = self.lstm(w, h, c) w = h[self.num_layers - 1].expand_dims(axis=-1).expand_dims(axis=-1) x = F.broadcast_mul(x, w) return x, (h, c) class DIAResUnit(HybridBlock): """ DIA-ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for the second convolution layer in bottleneck. dilation : int or tuple/list of 2 int, default 1 Dilation value for the second convolution layer in bottleneck. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool, default True Whether to use a bottleneck or simple block in units. conv1_stride : bool, default False Whether to use stride in the first or the second convolution layer of the block. attention : nn.Module, default None Attention module. """ def __init__(self, in_channels, out_channels, strides, padding=1, dilation=1, bn_use_global_stats=False, bottleneck=True, conv1_stride=False, attention=None, **kwargs): super(DIAResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, padding=padding, dilation=dilation, bn_use_global_stats=bn_use_global_stats, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") self.attention = attention def hybrid_forward(self, F, x, hc=None): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x, hc = self.attention(x, hc) x = x + identity x = self.activ(x) return x, hc class DIAResNet(HybridBlock): """ DIA-ResNet model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(DIAResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential( return_two=False, prefix="stage{}_".format(i + 1)) attention = DIAAttention( in_x_features=channels_per_stage[0], in_h_features=channels_per_stage[0]) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(DIAResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=conv1_stride, attention=attention)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_diaresnet(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DIA-ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. width_scale : float, default 1.0 Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif (blocks == 38) and bottleneck: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] else: raise ValueError("Unsupported DIA-ResNet with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1.0: channels = [[int(cij * width_scale) if (i != len(channels) - 1) or (j != len(ci) - 1) else cij for j, cij in enumerate(ci)] for i, ci in enumerate(channels)] init_block_channels = int(init_block_channels * width_scale) net = DIAResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def diaresnet10(**kwargs): """ DIA-ResNet-10 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=10, model_name="diaresnet10", **kwargs) def diaresnet12(**kwargs): """ DIA-ResNet-12 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=12, model_name="diaresnet12", **kwargs) def diaresnet14(**kwargs): """ DIA-ResNet-14 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=14, model_name="diaresnet14", **kwargs) def diaresnetbc14b(**kwargs): """ DIA-ResNet-BC-14b model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=14, bottleneck=True, conv1_stride=False, model_name="diaresnetbc14b", **kwargs) def diaresnet16(**kwargs): """ DIA-ResNet-16 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=16, model_name="diaresnet16", **kwargs) def diaresnet18(**kwargs): """ DIA-ResNet-18 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=18, model_name="diaresnet18", **kwargs) def diaresnet26(**kwargs): """ DIA-ResNet-26 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=26, bottleneck=False, model_name="diaresnet26", **kwargs) def diaresnetbc26b(**kwargs): """ DIA-ResNet-BC-26b model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="diaresnetbc26b", **kwargs) def diaresnet34(**kwargs): """ DIA-ResNet-34 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=34, model_name="diaresnet34", **kwargs) def diaresnetbc38b(**kwargs): """ DIA-ResNet-BC-38b model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=38, bottleneck=True, conv1_stride=False, model_name="diaresnetbc38b", **kwargs) def diaresnet50(**kwargs): """ DIA-ResNet-50 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=50, model_name="diaresnet50", **kwargs) def diaresnet50b(**kwargs): """ DIA-ResNet-50 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=50, conv1_stride=False, model_name="diaresnet50b", **kwargs) def diaresnet101(**kwargs): """ DIA-ResNet-101 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=101, model_name="diaresnet101", **kwargs) def diaresnet101b(**kwargs): """ DIA-ResNet-101 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=101, conv1_stride=False, model_name="diaresnet101b", **kwargs) def diaresnet152(**kwargs): """ DIA-ResNet-152 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=152, model_name="diaresnet152", **kwargs) def diaresnet152b(**kwargs): """ DIA-ResNet-152 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=152, conv1_stride=False, model_name="diaresnet152b", **kwargs) def diaresnet200(**kwargs): """ DIA-ResNet-200 model 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=200, model_name="diaresnet200", **kwargs) def diaresnet200b(**kwargs): """ DIA-ResNet-200 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=200, conv1_stride=False, model_name="diaresnet200b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ diaresnet10, diaresnet12, diaresnet14, diaresnetbc14b, diaresnet16, diaresnet18, diaresnet26, diaresnetbc26b, diaresnet34, diaresnetbc38b, diaresnet50, diaresnet50b, diaresnet101, diaresnet101b, diaresnet152, diaresnet152b, diaresnet200, diaresnet200b, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != diaresnet10 or weight_count == 6297352) assert (model != diaresnet12 or weight_count == 6371336) assert (model != diaresnet14 or weight_count == 6666760) assert (model != diaresnetbc14b or weight_count == 24023976) assert (model != diaresnet16 or weight_count == 7847432) assert (model != diaresnet18 or weight_count == 12568072) assert (model != diaresnet26 or weight_count == 18838792) assert (model != diaresnetbc26b or weight_count == 29954216) assert (model != diaresnet34 or weight_count == 22676232) assert (model != diaresnetbc38b or weight_count == 35884456) assert (model != diaresnet50 or weight_count == 39516072) assert (model != diaresnet50b or weight_count == 39516072) assert (model != diaresnet101 or weight_count == 58508200) assert (model != diaresnet101b or weight_count == 58508200) assert (model != diaresnet152 or weight_count == 74151848) assert (model != diaresnet152b or weight_count == 74151848) assert (model != diaresnet200 or weight_count == 78632872) assert (model != diaresnet200b or weight_count == 78632872) x = mx.nd.zeros((14, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (14, 1000)) if __name__ == "__main__": _test()
27,098
33.565051
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/lffd.py
""" LFFD for face detection, implemented in Gluon. Original paper: 'LFFD: A Light and Fast Face Detector for Edge Devices,' https://arxiv.org/abs/1904.10633. """ __all__ = ['LFFD', 'lffd20x5s320v2_widerface', 'lffd25x8s560v1_widerface'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import conv3x3, conv1x1_block, conv3x3_block, MultiOutputSequential, ParallelConcurent from .resnet import ResUnit from .preresnet import PreResUnit class LffdDetectionBranch(HybridBlock): """ LFFD specific detection branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. use_bias : bool Whether the layer uses a bias vector. use_bn : bool Whether to use BatchNorm layer. """ def __init__(self, in_channels, out_channels, use_bias, use_bn, **kwargs): super(LffdDetectionBranch, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=in_channels, use_bias=use_bias, use_bn=use_bn) self.conv2 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, use_bias=use_bias, use_bn=use_bn, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class LffdDetectionBlock(HybridBlock): """ LFFD specific detection block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. use_bias : bool Whether the layer uses a bias vector. use_bn : bool Whether to use BatchNorm layer. """ def __init__(self, in_channels, mid_channels, use_bias, use_bn, **kwargs): super(LffdDetectionBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_bias=use_bias, use_bn=use_bn) self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(LffdDetectionBranch( in_channels=mid_channels, out_channels=4, use_bias=use_bias, use_bn=use_bn)) self.branches.add(LffdDetectionBranch( in_channels=mid_channels, out_channels=2, use_bias=use_bias, use_bn=use_bn)) def hybrid_forward(self, F, x): x = self.conv(x) x = self.branches(x) return x class LFFD(HybridBlock): """ LFFD model from 'LFFD: A Light and Fast Face Detector for Edge Devices,' https://arxiv.org/abs/1904.10633. Parameters: ---------- enc_channels : list of int Number of output channels for each encoder stage. dec_channels : int Number of output channels for each decoder stage. init_block_channels : int Number of output channels for the initial encoder unit. layers : list of int Number of units in each encoder stage. int_bends : list of int Number of internal bends for each encoder stage. use_preresnet : bool Whether to use PreResnet backbone instead of ResNet. receptive_field_center_starts : list of int The start location of the first receptive field of each scale. receptive_field_strides : list of int Receptive field stride for each scale. bbox_factors : list of float A half of bbox upper bound for each scale. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (640, 640) Spatial size of the expected input image. """ def __init__(self, enc_channels, dec_channels, init_block_channels, layers, int_bends, use_preresnet, receptive_field_center_starts, receptive_field_strides, bbox_factors, in_channels=3, in_size=(480, 640), **kwargs): super(LFFD, self).__init__(**kwargs) self.in_size = in_size self.receptive_field_center_starts = receptive_field_center_starts self.receptive_field_strides = receptive_field_strides self.bbox_factors = bbox_factors unit_class = PreResUnit if use_preresnet else ResUnit use_bias = True use_bn = False with self.name_scope(): self.encoder = MultiOutputSequential(return_last=False) self.encoder.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, strides=2, padding=0, use_bias=use_bias, use_bn=use_bn)) in_channels = init_block_channels for i, channels_per_stage in enumerate(enc_channels): layers_per_stage = layers[i] int_bends_per_stage = int_bends[i] stage = MultiOutputSequential(prefix="stage{}_".format(i + 1), multi_output=False, dual_output=True) stage.add(conv3x3( in_channels=in_channels, out_channels=channels_per_stage, strides=2, padding=0, use_bias=use_bias)) for j in range(layers_per_stage): unit = unit_class( in_channels=channels_per_stage, out_channels=channels_per_stage, strides=1, use_bias=use_bias, use_bn=use_bn, bottleneck=False) if layers_per_stage - j <= int_bends_per_stage: unit.do_output = True stage.add(unit) final_activ = nn.Activation("relu") final_activ.do_output = True stage.add(final_activ) stage.do_output2 = True in_channels = channels_per_stage self.encoder.add(stage) self.decoder = ParallelConcurent() k = 0 for i, channels_per_stage in enumerate(enc_channels): layers_per_stage = layers[i] int_bends_per_stage = int_bends[i] for j in range(layers_per_stage): if layers_per_stage - j <= int_bends_per_stage: self.decoder.add(LffdDetectionBlock( in_channels=channels_per_stage, mid_channels=dec_channels, use_bias=use_bias, use_bn=use_bn)) k += 1 self.decoder.add(LffdDetectionBlock( in_channels=channels_per_stage, mid_channels=dec_channels, use_bias=use_bias, use_bn=use_bn)) k += 1 def hybrid_forward(self, F, x): x = self.encoder(x) x = self.decoder(x) return x def get_lffd(blocks, use_preresnet, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create LFFD model with specific parameters. Parameters: ---------- blocks : int Number of blocks. use_preresnet : bool Whether to use PreResnet backbone instead of ResNet. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 20: layers = [3, 1, 1, 1, 1] enc_channels = [64, 64, 64, 128, 128] int_bends = [0, 0, 0, 0, 0] receptive_field_center_starts = [3, 7, 15, 31, 63] receptive_field_strides = [4, 8, 16, 32, 64] bbox_factors = [10.0, 20.0, 40.0, 80.0, 160.0] elif blocks == 25: layers = [4, 2, 1, 3] enc_channels = [64, 64, 128, 128] int_bends = [1, 1, 0, 2] receptive_field_center_starts = [3, 3, 7, 7, 15, 31, 31, 31] receptive_field_strides = [4, 4, 8, 8, 16, 32, 32, 32] bbox_factors = [7.5, 10.0, 20.0, 35.0, 55.0, 125.0, 200.0, 280.0] else: raise ValueError("Unsupported LFFD with number of blocks: {}".format(blocks)) dec_channels = 128 init_block_channels = 64 net = LFFD( enc_channels=enc_channels, dec_channels=dec_channels, init_block_channels=init_block_channels, layers=layers, int_bends=int_bends, use_preresnet=use_preresnet, receptive_field_center_starts=receptive_field_center_starts, receptive_field_strides=receptive_field_strides, bbox_factors=bbox_factors, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def lffd20x5s320v2_widerface(**kwargs): """ LFFD-320-20L-5S-V2 model for WIDER FACE from 'LFFD: A Light and Fast Face Detector for Edge Devices,' https://arxiv.org/abs/1904.10633. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_lffd(blocks=20, use_preresnet=True, model_name="lffd20x5s320v2_widerface", **kwargs) def lffd25x8s560v1_widerface(**kwargs): """ LFFD-560-25L-8S-V1 model for WIDER FACE from 'LFFD: A Light and Fast Face Detector for Edge Devices,' https://arxiv.org/abs/1904.10633. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_lffd(blocks=25, use_preresnet=False, model_name="lffd25x8s560v1_widerface", **kwargs) def _test(): import numpy as np import mxnet as mx in_size = (480, 640) pretrained = False models = [ (lffd20x5s320v2_widerface, 5), (lffd25x8s560v1_widerface, 8), ] for model, num_outs in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != lffd20x5s320v2_widerface or weight_count == 1520606) assert (model != lffd25x8s560v1_widerface or weight_count == 2290608) batch = 14 x = mx.nd.random.normal(shape=(batch, 3, in_size[0], in_size[1]), ctx=ctx) y = net(x) assert (len(y) == num_outs) if __name__ == "__main__": _test()
12,410
33.28453
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/sepreresnet.py
""" SE-PreResNet for ImageNet-1K, implemented in Gluon. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['SEPreResNet', 'sepreresnet10', 'sepreresnet12', 'sepreresnet14', 'sepreresnet16', 'sepreresnet18', 'sepreresnet26', 'sepreresnetbc26b', 'sepreresnet34', 'sepreresnetbc38b', 'sepreresnet50', 'sepreresnet50b', 'sepreresnet101', 'sepreresnet101b', 'sepreresnet152', 'sepreresnet152b', 'sepreresnet200', 'sepreresnet200b', 'SEPreResUnit'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, SEBlock from .preresnet import PreResBlock, PreResBottleneck, PreResInitBlock, PreResActivation class SEPreResUnit(HybridBlock): """ SE-PreResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer of the block. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, bottleneck, conv1_stride, **kwargs): super(SEPreResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = PreResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, conv1_stride=conv1_stride) else: self.body = PreResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) self.se = SEBlock(channels=out_channels) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, strides=strides) def hybrid_forward(self, F, x): identity = x x, x_pre_activ = self.body(x) x = self.se(x) if self.resize_identity: identity = self.identity_conv(x_pre_activ) x = x + identity return x class SEPreResNet(HybridBlock): """ SE-PreResNet model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(SEPreResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(SEPreResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_sepreresnet(blocks, bottleneck=None, conv1_stride=True, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create SE-PreResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif (blocks == 38) and bottleneck: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] elif blocks == 269: layers = [3, 30, 48, 8] else: raise ValueError("Unsupported SE-PreResNet with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = SEPreResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def sepreresnet10(**kwargs): """ SE-PreResNet-10 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=10, model_name="sepreresnet10", **kwargs) def sepreresnet12(**kwargs): """ SE-PreResNet-12 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=12, model_name="sepreresnet12", **kwargs) def sepreresnet14(**kwargs): """ SE-PreResNet-14 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=14, model_name="sepreresnet14", **kwargs) def sepreresnet16(**kwargs): """ SE-PreResNet-16 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=16, model_name="sepreresnet16", **kwargs) def sepreresnet18(**kwargs): """ SE-PreResNet-18 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=18, model_name="sepreresnet18", **kwargs) def sepreresnet26(**kwargs): """ SE-PreResNet-26 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=26, bottleneck=False, model_name="sepreresnet26", **kwargs) def sepreresnetbc26b(**kwargs): """ SE-PreResNet-BC-26b model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="sepreresnetbc26b", **kwargs) def sepreresnet34(**kwargs): """ SE-PreResNet-34 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=34, model_name="sepreresnet34", **kwargs) def sepreresnetbc38b(**kwargs): """ SE-PreResNet-BC-38b model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=38, bottleneck=True, conv1_stride=False, model_name="sepreresnetbc38b", **kwargs) def sepreresnet50(**kwargs): """ SE-PreResNet-50 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=50, model_name="sepreresnet50", **kwargs) def sepreresnet50b(**kwargs): """ SE-PreResNet-50 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=50, conv1_stride=False, model_name="sepreresnet50b", **kwargs) def sepreresnet101(**kwargs): """ SE-PreResNet-101 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=101, model_name="sepreresnet101", **kwargs) def sepreresnet101b(**kwargs): """ SE-PreResNet-101 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=101, conv1_stride=False, model_name="sepreresnet101b", **kwargs) def sepreresnet152(**kwargs): """ SE-PreResNet-152 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=152, model_name="sepreresnet152", **kwargs) def sepreresnet152b(**kwargs): """ SE-PreResNet-152 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=152, conv1_stride=False, model_name="sepreresnet152b", **kwargs) def sepreresnet200(**kwargs): """ SE-PreResNet-200 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=200, model_name="sepreresnet200", **kwargs) def sepreresnet200b(**kwargs): """ SE-PreResNet-200 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=200, conv1_stride=False, model_name="sepreresnet200b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ sepreresnet10, sepreresnet12, sepreresnet14, sepreresnet16, sepreresnet18, sepreresnet26, sepreresnetbc26b, sepreresnet34, sepreresnetbc38b, sepreresnet50, sepreresnet50b, sepreresnet101, sepreresnet101b, sepreresnet152, sepreresnet152b, sepreresnet200, sepreresnet200b, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sepreresnet10 or weight_count == 5461668) assert (model != sepreresnet12 or weight_count == 5536232) assert (model != sepreresnet14 or weight_count == 5833840) assert (model != sepreresnet16 or weight_count == 7022976) assert (model != sepreresnet18 or weight_count == 11776928) assert (model != sepreresnet26 or weight_count == 18092188) assert (model != sepreresnetbc26b or weight_count == 17388424) assert (model != sepreresnet34 or weight_count == 21957204) assert (model != sepreresnetbc38b or weight_count == 24019064) assert (model != sepreresnet50 or weight_count == 28080472) assert (model != sepreresnet50b or weight_count == 28080472) assert (model != sepreresnet101 or weight_count == 49319320) assert (model != sepreresnet101b or weight_count == 49319320) assert (model != sepreresnet152 or weight_count == 66814296) assert (model != sepreresnet152b or weight_count == 66814296) assert (model != sepreresnet200 or weight_count == 71828312) assert (model != sepreresnet200b or weight_count == 71828312) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
21,007
34.071786
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/resnext.py
""" ResNeXt for ImageNet-1K, implemented in Gluon. Original paper: 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. """ __all__ = ['ResNeXt', 'resnext14_16x4d', 'resnext14_32x2d', 'resnext14_32x4d', 'resnext26_16x4d', 'resnext26_32x2d', 'resnext26_32x4d', 'resnext38_32x4d', 'resnext50_32x4d', 'resnext101_32x4d', 'resnext101_64x4d', 'ResNeXtBottleneck', 'ResNeXtUnit'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block from .resnet import ResInitBlock class ResNeXtBottleneck(HybridBlock): """ ResNeXt bottleneck block for residual path in ResNeXt unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, bottleneck_factor=4, **kwargs): super(ResNeXtBottleneck, self).__init__(**kwargs) mid_channels = out_channels // bottleneck_factor D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=group_width, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=group_width, out_channels=group_width, strides=strides, groups=cardinality, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv1x1_block( in_channels=group_width, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class ResNeXtUnit(HybridBlock): """ ResNeXt unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, **kwargs): super(ResNeXtUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = ResNeXtBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class ResNeXt(HybridBlock): """ ResNeXt model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(ResNeXt, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(ResNeXtUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resnext(blocks, cardinality, bottleneck_width, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ResNeXt model with specific parameters. Parameters: ---------- blocks : int Number of blocks. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 14: layers = [1, 1, 1, 1] elif blocks == 26: layers = [2, 2, 2, 2] elif blocks == 38: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] else: raise ValueError("Unsupported ResNeXt with number of blocks: {}".format(blocks)) assert (sum(layers) * 3 + 2 == blocks) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = ResNeXt( channels=channels, init_block_channels=init_block_channels, cardinality=cardinality, bottleneck_width=bottleneck_width, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def resnext14_16x4d(**kwargs): """ ResNeXt-14 (16x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=14, cardinality=16, bottleneck_width=4, model_name="resnext14_16x4d", **kwargs) def resnext14_32x2d(**kwargs): """ ResNeXt-14 (32x2d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=14, cardinality=32, bottleneck_width=2, model_name="resnext14_32x2d", **kwargs) def resnext14_32x4d(**kwargs): """ ResNeXt-14 (32x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=14, cardinality=32, bottleneck_width=4, model_name="resnext14_32x4d", **kwargs) def resnext26_16x4d(**kwargs): """ ResNeXt-26 (16x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=26, cardinality=16, bottleneck_width=4, model_name="resnext26_16x4d", **kwargs) def resnext26_32x2d(**kwargs): """ ResNeXt-26 (32x2d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=26, cardinality=32, bottleneck_width=2, model_name="resnext26_32x2d", **kwargs) def resnext26_32x4d(**kwargs): """ ResNeXt-26 (32x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=26, cardinality=32, bottleneck_width=4, model_name="resnext26_32x4d", **kwargs) def resnext38_32x4d(**kwargs): """ ResNeXt-38 (32x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=38, cardinality=32, bottleneck_width=4, model_name="resnext38_32x4d", **kwargs) def resnext50_32x4d(**kwargs): """ ResNeXt-50 (32x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=50, cardinality=32, bottleneck_width=4, model_name="resnext50_32x4d", **kwargs) def resnext101_32x4d(**kwargs): """ ResNeXt-101 (32x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=101, cardinality=32, bottleneck_width=4, model_name="resnext101_32x4d", **kwargs) def resnext101_64x4d(**kwargs): """ ResNeXt-101 (64x4d) model from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext(blocks=101, cardinality=64, bottleneck_width=4, model_name="resnext101_64x4d", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ resnext14_16x4d, resnext14_32x2d, resnext14_32x4d, resnext26_16x4d, resnext26_32x2d, resnext26_32x4d, resnext38_32x4d, resnext50_32x4d, resnext101_32x4d, resnext101_64x4d, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resnext14_16x4d or weight_count == 7127336) assert (model != resnext14_32x2d or weight_count == 7029416) assert (model != resnext14_32x4d or weight_count == 9411880) assert (model != resnext26_16x4d or weight_count == 10119976) assert (model != resnext26_32x2d or weight_count == 9924136) assert (model != resnext26_32x4d or weight_count == 15389480) assert (model != resnext38_32x4d or weight_count == 21367080) assert (model != resnext50_32x4d or weight_count == 25028904) assert (model != resnext101_32x4d or weight_count == 44177704) assert (model != resnext101_64x4d or weight_count == 83455272) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
17,156
33.245509
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/jasper.py
""" Jasper/DR for ASR, implemented in Gluon. Original paper: 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. """ __all__ = ['Jasper', 'jasper5x3', 'jasper10x4', 'jasper10x5', 'get_jasper', 'MaskConv1d', 'NemoAudioReader', 'NemoMelSpecExtractor', 'CtcDecoder'] import os import numpy as np import mxnet as mx from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import BatchNormExtra, DualPathSequential, DualPathParallelConcurent def outmask_fill(F, x, x_len, value=0.0): """ Masked fill a tensor. Parameters: ---------- F : object MXNet tensor processing package. x : tensor Input tensor. x_len : tensor Tensor with lengths. value : float, default 0.0 Filled value. Returns: ------- tensor Resulted tensor. """ x = F.SequenceMask( data=x.swapaxes(1, 2), sequence_length=x_len, use_sequence_length=True, value=value, axis=1).swapaxes(1, 2) return x def masked_normalize2(F, x, x_len): """ Normalize a tensor with mask (scheme #2). Parameters: ---------- F : object MXNet tensor processing package. x : tensor Input tensor. x_len : tensor Tensor with lengths. Returns: ------- tensor Resulted tensor. """ x = outmask_fill(F, x, x_len) x_len_float = x_len.astype("float32").expand_dims(axis=1) x_mean = x.sum(axis=2) / x_len_float x2_mean = x.square().sum(axis=2) / x_len_float x_std = (x2_mean - x_mean.square()).sqrt() x = (x - x_mean.expand_dims(axis=2)) / x_std.expand_dims(axis=2) return x class NemoAudioReader(object): """ Audio Reader from NVIDIA NEMO toolkit. Parameters: ---------- desired_audio_sample_rate : int, default 16000 Desired audio sample rate. trunc_value : int or None, default None Value to truncate. """ def __init__(self, desired_audio_sample_rate=16000): super(NemoAudioReader, self).__init__() self.desired_audio_sample_rate = desired_audio_sample_rate def read_from_file(self, audio_file_path): """ Read audio from file. Parameters: ---------- audio_file_path : str Path to audio file. Returns: ------- np.array Audio data. """ from soundfile import SoundFile with SoundFile(audio_file_path, "r") as data: sample_rate = data.samplerate audio_data = data.read(dtype="float32") audio_data = audio_data.transpose() if sample_rate != self.desired_audio_sample_rate: from librosa.core import resample as lr_resample audio_data = lr_resample(y=audio_data, orig_sr=sample_rate, target_sr=self.desired_audio_sample_rate) if audio_data.ndim >= 2: audio_data = np.mean(audio_data, axis=1) return audio_data def read_from_files(self, audio_file_paths): """ Read audios from files. Parameters: ---------- audio_file_paths : list of str Paths to audio files. Returns: ------- list of np.array Audio data. """ assert (type(audio_file_paths) in (list, tuple)) audio_data_list = [] for audio_file_path in audio_file_paths: audio_data = self.read_from_file(audio_file_path) audio_data_list.append(audio_data) return audio_data_list class NemoMelSpecExtractor(HybridBlock): """ Mel-Spectrogram Extractor from NVIDIA NEMO toolkit. Parameters: ---------- sample_rate : int, default 16000 Sample rate of the input audio data. window_size_sec : float, default 0.02 Size of window for FFT in seconds. window_stride_sec : float, default 0.01 Stride of window for FFT in seconds. n_fft : int, default 512 Length of FT window. n_filters : int, default 64 Number of Mel spectrogram freq bins. preemph : float, default 0.97 Amount of pre emphasis to add to audio. dither : float, default 1.0e-05 Amount of white-noise dithering. """ def __init__(self, sample_rate=16000, window_size_sec=0.02, window_stride_sec=0.01, n_fft=512, n_filters=64, preemph=0.97, dither=1.0e-5, **kwargs): super(NemoMelSpecExtractor, self).__init__(**kwargs) self.log_zero_guard_value = 2 ** -24 win_length = int(window_size_sec * sample_rate) self.hop_length = int(window_stride_sec * sample_rate) self.n_filters = n_filters from scipy import signal as scipy_signal from librosa import stft as librosa_stft window_arr = scipy_signal.hann(win_length, sym=True) self.stft = lambda x: librosa_stft( x, n_fft=n_fft, hop_length=self.hop_length, win_length=win_length, window=window_arr, center=True) self.dither = dither self.preemph = preemph self.pad_align = 16 from librosa.filters import mel as librosa_mel fb_arr = librosa_mel( sr=sample_rate, n_fft=n_fft, n_mels=n_filters, fmin=0.0, fmax=(sample_rate / 2.0)) fb_arr = np.expand_dims(fb_arr, axis=0) with self.name_scope(): self.window = self.params.get( "window", grad_req="null", shape=window_arr.shape, init=mx.init.Constant(window_arr), allow_deferred_init=False, differentiable=False) self.fb = self.params.get( "fb", grad_req="null", shape=fb_arr.shape, init=mx.init.Constant(fb_arr), allow_deferred_init=False, differentiable=False) def hybrid_forward(self, F, x, x_len, window=None, fb=None): x_len = ((x_len.astype("float32") / self.hop_length).ceil()).astype("int64") if self.dither > 0: x += self.dither * F.random.normal_like(x) x = F.concat(x[:, :1], x[:, 1:] - self.preemph * x[:, :-1], dim=1) x = self.calc_real_stft(F, x) x = F.power(x, 2) x = F.batch_dot(fb.repeat(repeats=x.shape[0], axis=0), x) x = F.log(x + self.log_zero_guard_value) x = masked_normalize2(F, x, x_len) x = outmask_fill(F, x, x_len) x_len_max = x.shape[-1] pad_rem = x_len_max % self.pad_align if pad_rem != 0: x = x.expand_dims(1).pad(mode="constant", pad_width=(0, 0, 0, 0, 0, 0, 0, self.pad_align - pad_rem), constant_value=0).squeeze(1) return x, x_len def calc_real_stft(self, F, x): x_np = x.asnumpy() ys = [] for xi_np in x_np: ys.append(self.stft(xi_np)) x_np = np.array(ys) x_np = np.abs(x_np) x = F.array(x_np, ctx=x.context) return x def calc_flops(self, x): assert (x.shape[0] == 1) num_flops = x[0].size num_macs = 0 return num_flops, num_macs class CtcDecoder(object): """ CTC decoder (to decode a sequence of labels to words). Parameters: ---------- vocabulary : list of str Vocabulary of the dataset. """ def __init__(self, vocabulary): super().__init__() self.blank_id = len(vocabulary) self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) def __call__(self, predictions): """ Decode a sequence of labels to words. Parameters: ---------- predictions : np.array of int or list of list of int Tensor with predicted labels. Returns: ------- list of str Words. """ hypotheses = [] for prediction in predictions: decoded_prediction = [] previous = self.blank_id for p in prediction: if (p != previous or previous == self.blank_id) and p != self.blank_id: decoded_prediction.append(p) previous = p hypothesis = "".join([self.labels_map[c] for c in decoded_prediction]) hypotheses.append(hypothesis) return hypotheses def conv1d1(in_channels, out_channels, strides=1, groups=1, use_bias=False, **kwargs): """ 1-dim kernel version of the 1D convolution layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv1D( channels=out_channels, kernel_size=1, strides=strides, groups=groups, use_bias=use_bias, in_channels=in_channels, **kwargs) class MaskConv1d(nn.Conv1D): """ Masked 1D convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 1 int Convolution window size. strides : int or tuple/list of 1 int Strides of the convolution. padding : int or tuple/list of 1 int, default 0 Padding value for convolution layer. dilation : int or tuple/list of 1 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_mask : bool, default True Whether to use mask. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding=0, dilation=1, groups=1, use_bias=False, use_mask=True, **kwargs): super(MaskConv1d, self).__init__( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels, **kwargs) self.use_mask = use_mask if self.use_mask: self.kernel_size = kernel_size[0] if isinstance(kernel_size, (list, tuple)) else kernel_size self.strides = strides[0] if isinstance(strides, (list, tuple)) else strides self.padding = padding[0] if isinstance(padding, (list, tuple)) else padding self.dilation = dilation[0] if isinstance(dilation, (list, tuple)) else dilation def hybrid_forward(self, F, x, x_len, weight, bias=None): if self.use_mask: x = outmask_fill(F, x, x_len) x_len = ((x_len + 2 * self.padding - self.dilation * (self.kernel_size - 1) - 1) / self.strides + 1).floor() x = super(MaskConv1d, self).hybrid_forward(F, x, weight=weight, bias=bias) return x, x_len def mask_conv1d1(in_channels, out_channels, strides=1, groups=1, use_bias=False, **kwargs): """ Masked 1-dim kernel version of the 1D convolution layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. """ return MaskConv1d( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, groups=groups, use_bias=use_bias, **kwargs) class MaskConvBlock1d(HybridBlock): """ Masked 1D convolution block with batch normalization, activation, and dropout. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int Convolution window size. strides : int Strides of the convolution. padding : int Padding value for convolution layer. dilation : int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), dropout_rate=0.0, **kwargs): super(MaskConvBlock1d, self).__init__(**kwargs) self.activate = (activation is not None) self.use_bn = use_bn self.use_dropout = (dropout_rate != 0.0) with self.name_scope(): self.conv = MaskConv1d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias) if self.use_bn: self.bn = BatchNormExtra( in_channels=out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) if self.activate: self.activ = activation() if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x, x_len): x, x_len = self.conv(x, x_len) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) if self.use_dropout: x = self.dropout(x) return x, x_len def mask_conv1d1_block(in_channels, out_channels, strides=1, padding=0, **kwargs): """ 1-dim kernel version of the masked 1D convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int, default 1 Strides of the convolution. padding : int, default 0 Padding value for convolution layer. """ return MaskConvBlock1d( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=padding, **kwargs) class ChannelShuffle1d(HybridBlock): """ 1D version of the channel shuffle layer. Parameters: ---------- channels : int Number of channels. groups : int Number of groups. """ def __init__(self, channels, groups, **kwargs): super(ChannelShuffle1d, self).__init__(**kwargs) assert (channels % groups == 0) self.groups = groups def hybrid_forward(self, F, x): return x.reshape((0, -4, self.groups, -1, -2)).swapaxes(1, 2).reshape((0, -3, -2)) def __repr__(self): s = "{name}(groups={groups})" return s.format( name=self.__class__.__name__, groups=self.groups) class DwsConvBlock1d(HybridBlock): """ Depthwise version of the 1D standard convolution block with batch normalization, activation, dropout, and channel shuffle. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int Convolution window size. strides : int Strides of the convolution. padding : int Padding value for convolution layer. dilation : int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), dropout_rate=0.0, **kwargs): super(DwsConvBlock1d, self).__init__(**kwargs) self.activate = (activation is not None) self.use_bn = use_bn self.use_dropout = (dropout_rate != 0.0) self.use_channel_shuffle = (groups > 1) with self.name_scope(): self.dw_conv = MaskConv1d( in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=in_channels, use_bias=use_bias) self.pw_conv = mask_conv1d1( in_channels=in_channels, out_channels=out_channels, groups=groups, use_bias=use_bias) if self.use_channel_shuffle: self.shuffle = ChannelShuffle1d( channels=out_channels, groups=groups) if self.use_bn: self.bn = BatchNormExtra( in_channels=out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) if self.activate: self.activ = activation() if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x, x_len): x, x_len = self.dw_conv(x, x_len) x, x_len = self.pw_conv(x, x_len) if self.use_channel_shuffle: x = self.shuffle(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) if self.use_dropout: x = self.dropout(x) return x, x_len class JasperUnit(HybridBlock): """ Jasper unit with residual connection. Parameters: ---------- in_channels : int or list of int Number of input channels. out_channels : int Number of output channels. kernel_size : int Convolution window size. bn_epsilon : float Small float added to variance in Batch norm. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. repeat : int Count of body convolution blocks. use_dw : bool Whether to use depthwise block. use_dr : bool Whether to use dense residual scheme. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, kernel_size, bn_epsilon, dropout_rate, repeat, use_dw, use_dr, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(JasperUnit, self).__init__(**kwargs) self.use_dropout = (dropout_rate != 0.0) self.use_dr = use_dr block_class = DwsConvBlock1d if use_dw else MaskConvBlock1d with self.name_scope(): if self.use_dr: self.identity_block = DualPathParallelConcurent() for i, dense_in_channels_i in enumerate(in_channels): self.identity_block.add(mask_conv1d1_block( in_channels=dense_in_channels_i, out_channels=out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dropout_rate=0.0, activation=None)) in_channels = in_channels[-1] else: self.identity_block = mask_conv1d1_block( in_channels=in_channels, out_channels=out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dropout_rate=0.0, activation=None) self.body = DualPathSequential() for i in range(repeat): activation = (lambda: nn.Activation("relu")) if i < repeat - 1 else None dropout_rate_i = dropout_rate if i < repeat - 1 else 0.0 self.body.add(block_class( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=1, padding=(kernel_size // 2), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dropout_rate=dropout_rate_i, activation=activation)) in_channels = out_channels self.activ = nn.Activation("relu") if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x, x_len): if self.use_dr: x_len, y, y_len = x_len if type(x_len) is tuple else (x_len, None, None) y = [x] if y is None else y + [x] y_len = [x_len] if y_len is None else y_len + [x_len] identity, _ = self.identity_block(y, y_len) identity = F.stack(*identity, axis=1) identity = identity.sum(axis=1) else: identity, _ = self.identity_block(x, x_len) x, x_len = self.body(x, x_len) x = x + identity x = self.activ(x) if self.use_dropout: x = self.dropout(x) if self.use_dr: return x, (x_len, y, y_len) else: return x, x_len class JasperFinalBlock(HybridBlock): """ Jasper specific final block. Parameters: ---------- in_channels : int Number of input channels. channels : list of int Number of output channels for each block. kernel_sizes : list of int Kernel sizes for each block. bn_epsilon : float Small float added to variance in Batch norm. dropout_rates : list of int Dropout rates for each block. use_dw : bool Whether to use depthwise block. use_dr : bool Whether to use dense residual scheme. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, channels, kernel_sizes, bn_epsilon, dropout_rates, use_dw, use_dr, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(JasperFinalBlock, self).__init__(**kwargs) self.use_dr = use_dr conv1_class = DwsConvBlock1d if use_dw else MaskConvBlock1d with self.name_scope(): self.conv1 = conv1_class( in_channels=in_channels, out_channels=channels[-2], kernel_size=kernel_sizes[-2], strides=1, padding=(2 * kernel_sizes[-2] // 2 - 1), dilation=2, bn_epsilon=bn_epsilon, dropout_rate=dropout_rates[-2], bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv2 = MaskConvBlock1d( in_channels=channels[-2], out_channels=channels[-1], kernel_size=kernel_sizes[-1], strides=1, padding=(kernel_sizes[-1] // 2), bn_epsilon=bn_epsilon, dropout_rate=dropout_rates[-1], bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x, x_len): if self.use_dr: x_len = x_len[0] x, x_len = self.conv1(x, x_len) x, x_len = self.conv2(x, x_len) return x, x_len class Jasper(HybridBlock): """ Jasper/DR/QuartzNet model from 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. Parameters: ---------- channels : list of int Number of output channels for each unit and initial/final block. kernel_sizes : list of int Kernel sizes for each unit and initial/final block. bn_epsilon : float Small float added to variance in Batch norm. dropout_rates : list of int Dropout rates for each unit and initial/final block. repeat : int Count of body convolution blocks. use_dw : bool Whether to use depthwise block. use_dr : bool Whether to use dense residual scheme. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. from_audio : bool, default True Whether to treat input as audio instead of Mel-specs. dither : float, default 0.0 Amount of white-noise dithering. return_text : bool, default False Whether to return text instead of logits. vocabulary : list of str or None, default None Vocabulary of the dataset. in_channels : int, default 64 Number of input channels (audio features). classes : int, default 29 Number of classification classes (number of graphemes). """ def __init__(self, channels, kernel_sizes, bn_epsilon, dropout_rates, repeat, use_dw, use_dr, bn_use_global_stats=False, bn_cudnn_off=False, from_audio=True, dither=0.0, return_text=False, vocabulary=None, in_channels=64, classes=29, **kwargs): super(Jasper, self).__init__(**kwargs) self.in_size = in_channels self.classes = classes self.vocabulary = vocabulary self.from_audio = from_audio self.return_text = return_text with self.name_scope(): if self.from_audio: self.preprocessor = NemoMelSpecExtractor(dither=dither) self.features = DualPathSequential() init_block_class = DwsConvBlock1d if use_dw else MaskConvBlock1d self.features.add(init_block_class( in_channels=in_channels, out_channels=channels[0], kernel_size=kernel_sizes[0], strides=2, padding=(kernel_sizes[0] // 2), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dropout_rate=dropout_rates[0])) in_channels = channels[0] in_channels_list = [] for i, (out_channels, kernel_size, dropout_rate) in\ enumerate(zip(channels[1:-2], kernel_sizes[1:-2], dropout_rates[1:-2])): in_channels_list += [in_channels] self.features.add(JasperUnit( in_channels=(in_channels_list if use_dr else in_channels), out_channels=out_channels, kernel_size=kernel_size, bn_epsilon=bn_epsilon, dropout_rate=dropout_rate, repeat=repeat, use_dw=use_dw, use_dr=use_dr, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = out_channels self.features.add(JasperFinalBlock( in_channels=in_channels, channels=channels, kernel_sizes=kernel_sizes, bn_epsilon=bn_epsilon, dropout_rates=dropout_rates, use_dw=use_dw, use_dr=use_dr, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = channels[-1] self.output = conv1d1( in_channels=in_channels, out_channels=classes, use_bias=True) if self.return_text: self.ctc_decoder = CtcDecoder(vocabulary=vocabulary) def hybrid_forward(self, F, x, x_len=None): if x_len is None: assert (type(x) in (list, tuple)) x, x_len = x if self.from_audio: x, x_len = self.preprocessor(x, x_len) x, x_len = self.features(x, x_len) x = self.output(x) if self.return_text: greedy_predictions = x.swapaxes(1, 2).log_softmax(dim=-1).argmax(dim=-1, keepdim=False).asnumpy() return self.ctc_decoder(greedy_predictions) else: return x, x_len def get_jasper(version, use_dw=False, use_dr=False, bn_epsilon=1e-3, vocabulary=None, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Jasper/DR/QuartzNet model with specific parameters. Parameters: ---------- version : tuple of str Model type and configuration. use_dw : bool, default False Whether to use depthwise block. use_dr : bool, default False Whether to use dense residual scheme. bn_epsilon : float, default 1e-3 Small float added to variance in Batch norm. vocabulary : list of str or None, default None Vocabulary of the dataset. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ import numpy as np blocks, repeat = tuple(map(int, version[1].split("x"))) main_stage_repeat = blocks // 5 model_type = version[0] if model_type == "jasper": channels_per_stage = [256, 256, 384, 512, 640, 768, 896, 1024] kernel_sizes_per_stage = [11, 11, 13, 17, 21, 25, 29, 1] dropout_rates_per_stage = [0.2, 0.2, 0.2, 0.2, 0.3, 0.3, 0.4, 0.4] elif model_type == "quartznet": channels_per_stage = [256, 256, 256, 512, 512, 512, 512, 1024] kernel_sizes_per_stage = [33, 33, 39, 51, 63, 75, 87, 1] dropout_rates_per_stage = [0.0] * 8 else: raise ValueError("Unsupported Jasper family model type: {}".format(model_type)) stage_repeat = np.full((8,), 1) stage_repeat[1:-2] *= main_stage_repeat channels = sum([[a] * r for (a, r) in zip(channels_per_stage, stage_repeat)], []) kernel_sizes = sum([[a] * r for (a, r) in zip(kernel_sizes_per_stage, stage_repeat)], []) dropout_rates = sum([[a] * r for (a, r) in zip(dropout_rates_per_stage, stage_repeat)], []) net = Jasper( channels=channels, kernel_sizes=kernel_sizes, bn_epsilon=bn_epsilon, dropout_rates=dropout_rates, repeat=repeat, use_dw=use_dw, use_dr=use_dr, vocabulary=vocabulary, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def jasper5x3(**kwargs): """ Jasper 5x3 model from 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_jasper(version=("jasper", "5x3"), model_name="jasper5x3", **kwargs) def jasper10x4(**kwargs): """ Jasper 10x4 model from 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_jasper(version=("jasper", "10x4"), model_name="jasper10x4", **kwargs) def jasper10x5(**kwargs): """ Jasper 10x5 model from 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_jasper(version=("jasper", "10x5"), model_name="jasper10x5", **kwargs) def _calc_width(net): import numpy as np net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) return weight_count def _test(): import mxnet as mx pretrained = False from_audio = True # dither = 0.0 dither = 1.0e-5 audio_features = 64 classes = 29 models = [ jasper5x3, jasper10x4, jasper10x5, ] for model in models: net = model( in_channels=audio_features, classes=classes, from_audio=from_audio, dither=dither, pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != jasper5x3 or weight_count == 107681053) assert (model != jasper10x4 or weight_count == 261393693) assert (model != jasper10x5 or weight_count == 322286877) batch = 3 aud_scale = 640 if from_audio else 1 seq_len = np.random.randint(150, 250, batch) * aud_scale seq_len_max = seq_len.max() + 2 x_shape = (batch, seq_len_max) if from_audio else (batch, audio_features, seq_len_max) x = mx.nd.random.normal(shape=x_shape, ctx=ctx) x_len = mx.nd.array(seq_len, ctx=ctx, dtype=np.long) y, y_len = net(x, x_len) assert (y.shape[:2] == (batch, net.classes)) if from_audio: assert (y.shape[2] in range(seq_len_max // aud_scale * 2, seq_len_max // aud_scale * 2 + 9)) else: assert (y.shape[2] in [seq_len_max // 2, seq_len_max // 2 + 1]) if __name__ == "__main__": _test()
38,940
31.806234
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/resneta.py
""" ResNet(A) with average downsampling for ImageNet-1K, implemented in Gluon. Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. """ __all__ = ['ResNetA', 'resneta10', 'resnetabc14b', 'resneta18', 'resneta50b', 'resneta101b', 'resneta152b'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block from .resnet import ResBlock, ResBottleneck from .senet import SEInitBlock class ResADownBlock(HybridBlock): """ ResNet(A) downsample block for the identity branch of a residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. dilation : int or tuple/list of 2 int, default 1 Dilation value for the second convolution layer in bottleneck. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, strides, dilation=1, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(ResADownBlock, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=(strides if dilation == 1 else 1), strides=(strides if dilation == 1 else 1), ceil_mode=True, count_include_pad=False) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) def hybrid_forward(self, F, x): x = self.pool(x) x = self.conv(x) return x class ResAUnit(HybridBlock): """ ResNet(A) unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for the second convolution layer in bottleneck. dilation : int or tuple/list of 2 int, default 1 Dilation value for the second convolution layer in bottleneck. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. bottleneck : bool, default True Whether to use a bottleneck or simple block in units. conv1_stride : bool, default False Whether to use stride in the first or the second convolution layer of the block. """ def __init__(self, in_channels, out_channels, strides, padding=1, dilation=1, bn_use_global_stats=False, bn_cudnn_off=False, bottleneck=True, conv1_stride=False, **kwargs): super(ResAUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, padding=padding, dilation=dilation, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) if self.resize_identity: self.identity_block = ResADownBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, dilation=dilation, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_block(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class ResNetA(HybridBlock): """ ResNet(A) with average downsampling model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. dilated : bool, default False Whether to use dilation. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, bn_use_global_stats=False, bn_cudnn_off=False, dilated=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(ResNetA, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(SEInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): if dilated: strides = 2 if ((j == 0) and (i != 0) and (i < 2)) else 1 dilation = (2 ** max(0, i - 1 - int(j == 0))) else: strides = 2 if (j == 0) and (i != 0) else 1 dilation = 1 stage.add(ResAUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, padding=dilation, dilation=dilation, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add(stage) self.features.add(nn.GlobalAvgPool2D()) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resneta(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ResNet(A) with average downsampling model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. width_scale : float, default 1.0 Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif (blocks == 38) and bottleneck: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] else: raise ValueError("Unsupported ResNet(A) with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1.0: channels = [[int(cij * width_scale) if (i != len(channels) - 1) or (j != len(ci) - 1) else cij for j, cij in enumerate(ci)] for i, ci in enumerate(channels)] init_block_channels = int(init_block_channels * width_scale) net = ResNetA( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def resneta10(**kwargs): """ ResNet(A)-10 with average downsampling model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resneta(blocks=10, model_name="resneta10", **kwargs) def resnetabc14b(**kwargs): """ ResNet(A)-BC-14b with average downsampling model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resneta(blocks=14, bottleneck=True, conv1_stride=False, model_name="resnetabc14b", **kwargs) def resneta18(**kwargs): """ ResNet(A)-18 with average downsampling model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resneta(blocks=18, model_name="resneta18", **kwargs) def resneta50b(**kwargs): """ ResNet(A)-50 with average downsampling model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resneta(blocks=50, conv1_stride=False, model_name="resneta50b", **kwargs) def resneta101b(**kwargs): """ ResNet(A)-101 with average downsampling model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resneta(blocks=101, conv1_stride=False, model_name="resneta101b", **kwargs) def resneta152b(**kwargs): """ ResNet(A)-152 with average downsampling model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resneta(blocks=152, conv1_stride=False, model_name="resneta152b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ resneta10, resnetabc14b, resneta18, resneta50b, resneta101b, resneta152b, ] for model in models: net = model( pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resneta10 or weight_count == 5438024) assert (model != resnetabc14b or weight_count == 10084168) assert (model != resneta18 or weight_count == 11708744) assert (model != resneta50b or weight_count == 25576264) assert (model != resneta101b or weight_count == 44568392) assert (model != resneta152b or weight_count == 60212040) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
17,132
34.545643
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/resnesta.py
""" ResNeSt(A) with average downsampling for ImageNet-1K, implemented in Gluon. Original paper: 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. """ __all__ = ['ResNeStA', 'resnestabc14', 'resnesta18', 'resnestabc26', 'resnestabc38', 'resnesta50', 'resnesta101', 'resnesta152', 'resnesta200', 'resnesta269', 'ResNeStADownBlock'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, saconv3x3_block from .senet import SEInitBlock class ResNeStABlock(HybridBlock): """ Simple ResNeSt(A) block for residual path in ResNeSt(A) unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, strides, use_bias=False, use_bn=True, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(ResNeStABlock, self).__init__(**kwargs) self.resize = (strides > 1) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) if self.resize: self.pool = nn.AvgPool2D( pool_size=3, strides=strides, padding=1) self.conv2 = saconv3x3_block( in_channels=out_channels, out_channels=out_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) if self.resize: x = self.pool(x) x = self.conv2(x) return x class ResNeStABottleneck(HybridBlock): """ ResNeSt(A) bottleneck block for residual path in ResNeSt(A) unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, bn_cudnn_off=False, bottleneck_factor=4, **kwargs): super(ResNeStABottleneck, self).__init__(**kwargs) self.resize = (strides > 1) mid_channels = out_channels // bottleneck_factor with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv2 = saconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) if self.resize: self.pool = nn.AvgPool2D( pool_size=3, strides=strides, padding=1) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) if self.resize: x = self.pool(x) x = self.conv3(x) return x class ResNeStADownBlock(HybridBlock): """ ResNeSt(A) downsample block for the identity branch of a residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(ResNeStADownBlock, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=strides, strides=strides, ceil_mode=True, count_include_pad=False) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) def hybrid_forward(self, F, x): x = self.pool(x) x = self.conv(x) return x class ResNeStAUnit(HybridBlock): """ ResNeSt(A) unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. bottleneck : bool, default True Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, bn_cudnn_off=False, bottleneck=True, **kwargs): super(ResNeStAUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = ResNeStABottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) else: self.body = ResNeStABlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) if self.resize_identity: self.identity_block = ResNeStADownBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_block(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class ResNeStA(HybridBlock): """ ResNeSt(A) with average downsampling model from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. dropout_rate : float, default 0.0 Fraction of the input units to drop. Must be a number between 0 and 1. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, dropout_rate=0.0, bn_use_global_stats=False, bn_cudnn_off=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(ResNeStA, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(SEInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(ResNeStAUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, bottleneck=bottleneck)) in_channels = out_channels self.features.add(stage) self.features.add(nn.GlobalAvgPool2D()) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) if dropout_rate > 0.0: self.output.add(nn.Dropout(rate=dropout_rate)) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resnesta(blocks, bottleneck=None, width_scale=1.0, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ResNeSt(A) with average downsampling model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. width_scale : float, default 1.0 Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif (blocks == 38) and bottleneck: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] elif blocks == 269: layers = [3, 30, 48, 8] else: raise ValueError("Unsupported ResNeSt(A) with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if blocks >= 101: init_block_channels *= 2 if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1.0: channels = [[int(cij * width_scale) if (i != len(channels) - 1) or (j != len(ci) - 1) else cij for j, cij in enumerate(ci)] for i, ci in enumerate(channels)] init_block_channels = int(init_block_channels * width_scale) net = ResNeStA( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def resnestabc14(**kwargs): """ ResNeSt(A)-BC-14 with average downsampling model from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=14, bottleneck=True, model_name="resnestabc14", **kwargs) def resnesta18(**kwargs): """ ResNeSt(A)-18 with average downsampling model from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=18, model_name="resnesta18", **kwargs) def resnestabc26(**kwargs): """ ResNeSt(A)-BC-26 with average downsampling model from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=26, bottleneck=True, model_name="resnestabc26", **kwargs) def resnestabc38(**kwargs): """ ResNeSt(A)-BC-38 with average downsampling model from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=38, bottleneck=True, model_name="resnestabc38", **kwargs) def resnesta50(**kwargs): """ ResNeSt(A)-50 with average downsampling model with stride at the second convolution in bottleneck block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=50, model_name="resnesta50", **kwargs) def resnesta101(**kwargs): """ ResNeSt(A)-101 with average downsampling model with stride at the second convolution in bottleneck block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=101, model_name="resnesta101", **kwargs) def resnesta152(**kwargs): """ ResNeSt(A)-152 with average downsampling model with stride at the second convolution in bottleneck block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=152, model_name="resnesta152", **kwargs) def resnesta200(in_size=(256, 256), **kwargs): """ ResNeSt(A)-200 with average downsampling model with stride at the second convolution in bottleneck block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- in_size : tuple of two ints, default (256, 256) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=200, in_size=in_size, dropout_rate=0.2, model_name="resnesta200", **kwargs) def resnesta269(in_size=(320, 320), **kwargs): """ ResNeSt(A)-269 with average downsampling model with stride at the second convolution in bottleneck block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- in_size : tuple of two ints, default (320, 320) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnesta(blocks=269, in_size=in_size, dropout_rate=0.2, model_name="resnesta269", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (resnestabc14, 224), (resnesta18, 224), (resnestabc26, 224), (resnestabc38, 224), (resnesta50, 224), (resnesta101, 224), (resnesta152, 224), (resnesta200, 256), (resnesta269, 320), ] for model, size in models: net = model( pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resnestabc14 or weight_count == 10611688) assert (model != resnesta18 or weight_count == 12763784) assert (model != resnestabc26 or weight_count == 17069448) assert (model != resnestabc38 or weight_count == 23527208) assert (model != resnesta50 or weight_count == 27483240) assert (model != resnesta101 or weight_count == 48275016) assert (model != resnesta152 or weight_count == 65316040) assert (model != resnesta200 or weight_count == 70201544) assert (model != resnesta269 or weight_count == 110929480) batch = 14 x = mx.nd.zeros((batch, 3, size, size), ctx=ctx) y = net(x) assert (y.shape == (batch, 1000)) if __name__ == "__main__": _test()
22,512
33.849845
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/senet.py
""" SENet for ImageNet-1K, implemented in Gluon. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['SENet', 'senet16', 'senet28', 'senet40', 'senet52', 'senet103', 'senet154', 'SEInitBlock'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, SEBlock class SENetBottleneck(HybridBlock): """ SENet bottleneck block for residual path in SENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, **kwargs): super(SENetBottleneck, self).__init__(**kwargs) mid_channels = out_channels // 4 D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D group_width2 = group_width // 2 with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=group_width2, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=group_width2, out_channels=group_width, strides=strides, groups=cardinality, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv1x1_block( in_channels=group_width, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class SENetUnit(HybridBlock): """ SENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. identity_conv3x3 : bool, default False Whether to use 3x3 convolution in the identity link. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, identity_conv3x3, **kwargs): super(SENetUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = SENetBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats) self.se = SEBlock(channels=out_channels) if self.resize_identity: if identity_conv3x3: self.identity_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) else: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = self.se(x) x = x + identity x = self.activ(x) return x class SEInitBlock(HybridBlock): """ SENet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(SEInitBlock, self).__init__(**kwargs) mid_channels = out_channels // 2 with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv3 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.pool(x) return x class SENet(HybridBlock): """ SENet model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(SENet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(SEInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) identity_conv3x3 = (i != 0) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(SENetUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats, identity_conv3x3=identity_conv3x3)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dropout(rate=0.2)) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_senet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create SENet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 16: layers = [1, 1, 1, 1] cardinality = 32 elif blocks == 28: layers = [2, 2, 2, 2] cardinality = 32 elif blocks == 40: layers = [3, 3, 3, 3] cardinality = 32 elif blocks == 52: layers = [3, 4, 6, 3] cardinality = 32 elif blocks == 103: layers = [3, 4, 23, 3] cardinality = 32 elif blocks == 154: layers = [3, 8, 36, 3] cardinality = 64 else: raise ValueError("Unsupported SENet with number of blocks: {}".format(blocks)) bottleneck_width = 4 init_block_channels = 128 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = SENet( channels=channels, init_block_channels=init_block_channels, cardinality=cardinality, bottleneck_width=bottleneck_width, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def senet16(**kwargs): """ SENet-16 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_senet(blocks=16, model_name="senet16", **kwargs) def senet28(**kwargs): """ SENet-28 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_senet(blocks=28, model_name="senet28", **kwargs) def senet40(**kwargs): """ SENet-40 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_senet(blocks=40, model_name="senet40", **kwargs) def senet52(**kwargs): """ SENet-52 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_senet(blocks=52, model_name="senet52", **kwargs) def senet103(**kwargs): """ SENet-103 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_senet(blocks=103, model_name="senet103", **kwargs) def senet154(**kwargs): """ SENet-154 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_senet(blocks=154, model_name="senet154", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ senet16, senet28, senet40, senet52, senet103, senet154, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != senet16 or weight_count == 31366168) assert (model != senet28 or weight_count == 36453768) assert (model != senet40 or weight_count == 41541368) assert (model != senet52 or weight_count == 44659416) assert (model != senet103 or weight_count == 60963096) assert (model != senet154 or weight_count == 115088984) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
15,863
31.641975
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/diapreresnet_cifar.py
""" DIA-PreResNet for CIFAR/SVHN, implemented in Gluon. Original papers: 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. """ __all__ = ['CIFARDIAPreResNet', 'diapreresnet20_cifar10', 'diapreresnet20_cifar100', 'diapreresnet20_svhn', 'diapreresnet56_cifar10', 'diapreresnet56_cifar100', 'diapreresnet56_svhn', 'diapreresnet110_cifar10', 'diapreresnet110_cifar100', 'diapreresnet110_svhn', 'diapreresnet164bn_cifar10', 'diapreresnet164bn_cifar100', 'diapreresnet164bn_svhn', 'diapreresnet1001_cifar10', 'diapreresnet1001_cifar100', 'diapreresnet1001_svhn', 'diapreresnet1202_cifar10', 'diapreresnet1202_cifar100', 'diapreresnet1202_svhn'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3, DualPathSequential from .preresnet import PreResActivation from .diaresnet import DIAAttention from .diapreresnet import DIAPreResUnit class CIFARDIAPreResNet(HybridBlock): """ DIA-PreResNet model for CIFAR from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, bn_use_global_stats=False, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARDIAPreResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential( return_two=False, prefix="stage{}_".format(i + 1)) attention = DIAAttention( in_x_features=channels_per_stage[0], in_h_features=channels_per_stage[0]) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(DIAPreResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=False, attention=attention)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_diapreresnet_cifar(classes, blocks, bottleneck, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DIA-PreResNet model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. blocks : int Number of blocks. bottleneck : bool Whether to use a bottleneck or simple block in units. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (classes in [10, 100]) if bottleneck: assert ((blocks - 2) % 9 == 0) layers = [(blocks - 2) // 9] * 3 else: assert ((blocks - 2) % 6 == 0) layers = [(blocks - 2) // 6] * 3 channels_per_layers = [16, 32, 64] init_block_channels = 16 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if bottleneck: channels = [[cij * 4 for cij in ci] for ci in channels] net = CIFARDIAPreResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, classes=classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def diapreresnet20_cifar10(classes=10, **kwargs): """ DIA-PreResNet-20 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="diapreresnet20_cifar10", **kwargs) def diapreresnet20_cifar100(classes=100, **kwargs): """ DIA-PreResNet-20 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="diapreresnet20_cifar100", **kwargs) def diapreresnet20_svhn(classes=10, **kwargs): """ DIA-PreResNet-20 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="diapreresnet20_svhn", **kwargs) def diapreresnet56_cifar10(classes=10, **kwargs): """ DIA-PreResNet-56 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=56, bottleneck=False, model_name="diapreresnet56_cifar10", **kwargs) def diapreresnet56_cifar100(classes=100, **kwargs): """ DIA-PreResNet-56 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=56, bottleneck=False, model_name="diapreresnet56_cifar100", **kwargs) def diapreresnet56_svhn(classes=10, **kwargs): """ DIA-PreResNet-56 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=56, bottleneck=False, model_name="diapreresnet56_svhn", **kwargs) def diapreresnet110_cifar10(classes=10, **kwargs): """ DIA-PreResNet-110 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=110, bottleneck=False, model_name="diapreresnet110_cifar10", **kwargs) def diapreresnet110_cifar100(classes=100, **kwargs): """ DIA-PreResNet-110 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=110, bottleneck=False, model_name="diapreresnet110_cifar100", **kwargs) def diapreresnet110_svhn(classes=10, **kwargs): """ DIA-PreResNet-110 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=110, bottleneck=False, model_name="diapreresnet110_svhn", **kwargs) def diapreresnet164bn_cifar10(classes=10, **kwargs): """ DIA-PreResNet-164(BN) model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=164, bottleneck=True, model_name="diapreresnet164bn_cifar10", **kwargs) def diapreresnet164bn_cifar100(classes=100, **kwargs): """ DIA-PreResNet-164(BN) model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=164, bottleneck=True, model_name="diapreresnet164bn_cifar100", **kwargs) def diapreresnet164bn_svhn(classes=10, **kwargs): """ DIA-PreResNet-164(BN) model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=164, bottleneck=True, model_name="diapreresnet164bn_svhn", **kwargs) def diapreresnet1001_cifar10(classes=10, **kwargs): """ DIA-PreResNet-1001 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=1001, bottleneck=True, model_name="diapreresnet1001_cifar10", **kwargs) def diapreresnet1001_cifar100(classes=100, **kwargs): """ DIA-PreResNet-1001 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=1001, bottleneck=True, model_name="diapreresnet1001_cifar100", **kwargs) def diapreresnet1001_svhn(classes=10, **kwargs): """ DIA-PreResNet-1001 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=1001, bottleneck=True, model_name="diapreresnet1001_svhn", **kwargs) def diapreresnet1202_cifar10(classes=10, **kwargs): """ DIA-PreResNet-1202 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=1202, bottleneck=False, model_name="diapreresnet1202_cifar10", **kwargs) def diapreresnet1202_cifar100(classes=100, **kwargs): """ DIA-PreResNet-1202 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=1202, bottleneck=False, model_name="diapreresnet1202_cifar100", **kwargs) def diapreresnet1202_svhn(classes=10, **kwargs): """ DIA-PreResNet-1202 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(classes=classes, blocks=1202, bottleneck=False, model_name="diapreresnet1202_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (diapreresnet20_cifar10, 10), (diapreresnet20_cifar100, 100), (diapreresnet20_svhn, 10), (diapreresnet56_cifar10, 10), (diapreresnet56_cifar100, 100), (diapreresnet56_svhn, 10), (diapreresnet110_cifar10, 10), (diapreresnet110_cifar100, 100), (diapreresnet110_svhn, 10), (diapreresnet164bn_cifar10, 10), (diapreresnet164bn_cifar100, 100), (diapreresnet164bn_svhn, 10), (diapreresnet1001_cifar10, 10), (diapreresnet1001_cifar100, 100), (diapreresnet1001_svhn, 10), (diapreresnet1202_cifar10, 10), (diapreresnet1202_cifar100, 100), (diapreresnet1202_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != diapreresnet20_cifar10 or weight_count == 286674) assert (model != diapreresnet20_cifar100 or weight_count == 292524) assert (model != diapreresnet20_svhn or weight_count == 286674) assert (model != diapreresnet56_cifar10 or weight_count == 869970) assert (model != diapreresnet56_cifar100 or weight_count == 875820) assert (model != diapreresnet56_svhn or weight_count == 869970) assert (model != diapreresnet110_cifar10 or weight_count == 1744914) assert (model != diapreresnet110_cifar100 or weight_count == 1750764) assert (model != diapreresnet110_svhn or weight_count == 1744914) assert (model != diapreresnet164bn_cifar10 or weight_count == 1922106) assert (model != diapreresnet164bn_cifar100 or weight_count == 1945236) assert (model != diapreresnet164bn_svhn or weight_count == 1922106) assert (model != diapreresnet1001_cifar10 or weight_count == 10546554) assert (model != diapreresnet1001_cifar100 or weight_count == 10569684) assert (model != diapreresnet1001_svhn or weight_count == 10546554) assert (model != diapreresnet1202_cifar10 or weight_count == 19438226) assert (model != diapreresnet1202_cifar100 or weight_count == 19444076) assert (model != diapreresnet1202_svhn or weight_count == 19438226) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
22,523
36.855462
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/simplepose_coco.py
""" SimplePose for COCO Keypoint, implemented in Gluon. Original paper: 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. """ __all__ = ['SimplePose', 'simplepose_resnet18_coco', 'simplepose_resnet50b_coco', 'simplepose_resnet101b_coco', 'simplepose_resnet152b_coco', 'simplepose_resneta50b_coco', 'simplepose_resneta101b_coco', 'simplepose_resneta152b_coco', 'HeatmapMaxDetBlock'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import DeconvBlock, conv1x1, HeatmapMaxDetBlock from .resnet import resnet18, resnet50b, resnet101b, resnet152b from .resneta import resneta50b, resneta101b, resneta152b class SimplePose(HybridBlock): """ SimplePose model from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int Number of output channels for the backbone. channels : list of int Number of output channels for each decoder unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. return_heatmap : bool, default False Whether to return only heatmap. fixed_size : bool, default True Whether to expect fixed spatial size of input image. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (256, 192) Spatial size of the expected input image. keypoints : int, default 17 Number of keypoints. """ def __init__(self, backbone, backbone_out_channels, channels, bn_use_global_stats=False, bn_cudnn_off=True, return_heatmap=False, fixed_size=True, in_channels=3, in_size=(256, 192), keypoints=17, **kwargs): super(SimplePose, self).__init__(**kwargs) assert (in_channels == 3) self.in_size = in_size self.keypoints = keypoints self.return_heatmap = return_heatmap with self.name_scope(): self.backbone = backbone self.decoder = nn.HybridSequential(prefix="") in_channels = backbone_out_channels for out_channels in channels: self.decoder.add(DeconvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=4, strides=2, padding=1, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = out_channels self.decoder.add(conv1x1( in_channels=in_channels, out_channels=keypoints, use_bias=True)) self.heatmap_max_det = HeatmapMaxDetBlock( channels=keypoints, in_size=(in_size[0] // 4, in_size[1] // 4), fixed_size=fixed_size) def hybrid_forward(self, F, x): x = self.backbone(x) heatmap = self.decoder(x) if self.return_heatmap: return heatmap else: keypoints = self.heatmap_max_det(heatmap) return keypoints def get_simplepose(backbone, backbone_out_channels, keypoints, bn_cudnn_off, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create SimplePose model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int Number of output channels for the backbone. keypoints : int Number of keypoints. bn_cudnn_off : bool Whether to disable CUDNN batch normalization operator. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ channels = [256, 256, 256] net = SimplePose( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, bn_cudnn_off=bn_cudnn_off, keypoints=keypoints, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def simplepose_resnet18_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet-18 for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=512, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resnet18_coco", **kwargs) def simplepose_resnet50b_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet-50b for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resnet50b_coco", **kwargs) def simplepose_resnet101b_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet-101b for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet101b(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resnet101b_coco", **kwargs) def simplepose_resnet152b_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet-152b for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet152b(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resnet152b_coco", **kwargs) def simplepose_resneta50b_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet(A)-50b for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resneta50b(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resneta50b_coco", **kwargs) def simplepose_resneta101b_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet(A)-101b for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resneta101b(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resneta101b_coco", **kwargs) def simplepose_resneta152b_coco(pretrained_backbone=False, keypoints=17, bn_cudnn_off=True, **kwargs): """ SimplePose model on the base of ResNet(A)-152b for COCO Keypoint from 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. bn_cudnn_off : bool, default True Whether to disable CUDNN batch normalization operator. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resneta152b(pretrained=pretrained_backbone, bn_cudnn_off=bn_cudnn_off).features[:-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, bn_cudnn_off=bn_cudnn_off, model_name="simplepose_resneta152b_coco", **kwargs) def _test(): import numpy as np import mxnet as mx in_size = (256, 192) keypoints = 17 return_heatmap = True pretrained = False models = [ simplepose_resnet18_coco, simplepose_resnet50b_coco, simplepose_resnet101b_coco, simplepose_resnet152b_coco, simplepose_resneta50b_coco, simplepose_resneta101b_coco, simplepose_resneta152b_coco, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, return_heatmap=return_heatmap) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != simplepose_resnet18_coco or weight_count == 15376721) assert (model != simplepose_resnet50b_coco or weight_count == 33999697) assert (model != simplepose_resnet101b_coco or weight_count == 52991825) assert (model != simplepose_resnet152b_coco or weight_count == 68635473) assert (model != simplepose_resneta50b_coco or weight_count == 34018929) assert (model != simplepose_resneta101b_coco or weight_count == 53011057) assert (model != simplepose_resneta152b_coco or weight_count == 68654705) batch = 14 x = mx.nd.random.normal(shape=(batch, 3, in_size[0], in_size[1]), ctx=ctx) y = net(x) assert ((y.shape[0] == batch) and (y.shape[1] == keypoints)) if return_heatmap: assert ((y.shape[2] == x.shape[2] // 4) and (y.shape[3] == x.shape[3] // 4)) else: assert (y.shape[2] == 3) if __name__ == "__main__": _test()
15,538
39.571802
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/vovnet.py
""" VoVNet for ImageNet-1K, implemented in Gluon. Original paper: 'An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection,' https://arxiv.org/abs/1904.09730. """ __all__ = ['VoVNet', 'vovnet27s', 'vovnet39', 'vovnet57'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, SequentialConcurrent class VoVUnit(HybridBlock): """ VoVNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. branch_channels : int Number of output channels for each branch. num_branches : int Number of branches. resize : bool Whether to use resize block. use_residual : bool Whether to use residual block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, branch_channels, num_branches, resize, use_residual, bn_use_global_stats=False, **kwargs): super(VoVUnit, self).__init__(**kwargs) self.resize = resize self.use_residual = use_residual with self.name_scope(): if self.resize: self.pool = nn.MaxPool2D( pool_size=3, strides=2, ceil_mode=True) self.branches = SequentialConcurrent(prefix="") with self.branches.name_scope(): branch_in_channels = in_channels for i in range(num_branches): self.branches.add(conv3x3_block( in_channels=branch_in_channels, out_channels=branch_channels, bn_use_global_stats=bn_use_global_stats)) branch_in_channels = branch_channels self.concat_conv = conv1x1_block( in_channels=(in_channels + num_branches * branch_channels), out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): if self.resize: x = self.pool(x) if self.use_residual: identity = x x = self.branches(x) x = self.concat_conv(x) if self.use_residual: x = x + identity return x class VoVInitBlock(HybridBlock): """ VoVNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, **kwargs): super(VoVInitBlock, self).__init__(**kwargs) mid_channels = out_channels // 2 with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, strides=2, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, strides=2, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class VoVNet(HybridBlock): """ VoVNet model from 'An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection,' https://arxiv.org/abs/1904.09730. Parameters: ---------- channels : list of list of int Number of output channels for each unit. branch_channels : list of list of int Number of branch output channels for each unit. num_branches : int Number of branches for the each unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, branch_channels, num_branches, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(VoVNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes init_block_channels = 128 with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(VoVInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): use_residual = (j != 0) resize = (j == 0) and (i != 0) stage.add(VoVUnit( in_channels=in_channels, out_channels=out_channels, branch_channels=branch_channels[i][j], num_branches=num_branches, resize=resize, use_residual=use_residual, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_vovnet(blocks, slim=False, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. slim : bool, default False Whether to use a slim model. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 27: layers = [1, 1, 1, 1] elif blocks == 39: layers = [1, 1, 2, 2] elif blocks == 57: layers = [1, 1, 4, 3] else: raise ValueError("Unsupported VoVNet with number of blocks: {}".format(blocks)) assert (sum(layers) * 6 + 3 == blocks) num_branches = 5 channels_per_layers = [256, 512, 768, 1024] branch_channels_per_layers = [128, 160, 192, 224] if slim: channels_per_layers = [ci // 2 for ci in channels_per_layers] branch_channels_per_layers = [ci // 2 for ci in branch_channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] branch_channels = [[ci] * li for (ci, li) in zip(branch_channels_per_layers, layers)] net = VoVNet( channels=channels, branch_channels=branch_channels, num_branches=num_branches, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def vovnet27s(**kwargs): """ VoVNet-27-slim model from 'An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection,' https://arxiv.org/abs/1904.09730. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_vovnet(blocks=27, slim=True, model_name="vovnet27s", **kwargs) def vovnet39(**kwargs): """ VoVNet-39 model from 'An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection,' https://arxiv.org/abs/1904.09730. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_vovnet(blocks=39, model_name="vovnet39", **kwargs) def vovnet57(**kwargs): """ VoVNet-57 model from 'An Energy and GPU-Computation Efficient Backbone Network for Real-Time Object Detection,' https://arxiv.org/abs/1904.09730. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_vovnet(blocks=57, model_name="vovnet57", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ vovnet27s, vovnet39, vovnet57, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != vovnet27s or weight_count == 3525736) assert (model != vovnet39 or weight_count == 22600296) assert (model != vovnet57 or weight_count == 36640296) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
11,803
32.064426
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/crunetb.py
""" CRU-Net(b), implemented in Gluon. Original paper: 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. """ __all__ = ['CRUNetb', 'crunet56b', 'crunet116b'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3, pre_conv1x1_block, pre_conv3x3_block from .resnet import ResInitBlock from .preresnet import PreResActivation class CRUConvBlock(HybridBlock): """ CRU-Net specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. groups : int, default 1 Number of groups. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. It's used by PreResNet. shared_conv : HybridBlock, default None Shared convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, groups=1, bn_use_global_stats=False, return_preact=False, shared_conv=None, **kwargs): super(CRUConvBlock, self).__init__(**kwargs) self.return_preact = return_preact with self.name_scope(): self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") if shared_conv is None: self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, groups=groups, use_bias=False, in_channels=in_channels) else: self.conv = shared_conv def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) if self.return_preact: x_pre_activ = x x = self.conv(x) if self.return_preact: return x, x_pre_activ else: return x def cru_conv1x1_block(in_channels, out_channels, strides=1, bn_use_global_stats=False, return_preact=False, shared_conv=None): """ 1x1 version of the CRU-Net specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. shared_conv : HybridBlock, default None Shared convolution layer. """ return CRUConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, bn_use_global_stats=bn_use_global_stats, return_preact=return_preact, shared_conv=shared_conv) class ResBottleneck(HybridBlock): """ Pre-ResNeXt bottleneck block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, **kwargs): super(ResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // 4 D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D with self.name_scope(): self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=group_width, bn_use_global_stats=bn_use_global_stats) self.conv2 = pre_conv3x3_block( in_channels=group_width, out_channels=group_width, strides=strides, groups=cardinality, bn_use_global_stats=bn_use_global_stats) self.conv3 = pre_conv1x1_block( in_channels=group_width, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class CRUBottleneck(HybridBlock): """ CRU-Net bottleneck block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. group_width: int Group width parameter. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. shared_conv1 : HybridBlock, default None Shared convolution layer #1. shared_conv2 : HybridBlock, default None Shared convolution layer #2. """ def __init__(self, in_channels, out_channels, strides, group_width, bn_use_global_stats, shared_conv1=None, shared_conv2=None, **kwargs): super(CRUBottleneck, self).__init__(**kwargs) with self.name_scope(): self.conv1 = cru_conv1x1_block( in_channels=in_channels, out_channels=group_width, bn_use_global_stats=bn_use_global_stats, shared_conv=shared_conv1) if shared_conv2 is None: self.conv2 = conv3x3( in_channels=group_width, out_channels=group_width, strides=strides, groups=group_width) else: self.conv2 = shared_conv2 self.conv3 = pre_conv1x1_block( in_channels=group_width, out_channels=group_width, bn_use_global_stats=bn_use_global_stats) self.conv4 = pre_conv1x1_block( in_channels=group_width, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.conv4(x) return x class ResUnit(HybridBlock): """ CRU-Net residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, cardinality, bottleneck_width, bn_use_global_stats, **kwargs): super(ResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = pre_conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity return x class CRUUnit(HybridBlock): """ CRU-Net collective residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. group_width: int Group width parameter. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. shared_conv1 : HybridBlock, default None Shared convolution layer #1. shared_conv2 : HybridBlock, default None Shared convolution layer #2. """ def __init__(self, in_channels, out_channels, strides, group_width, bn_use_global_stats, shared_conv1=None, shared_conv2=None, **kwargs): super(CRUUnit, self).__init__(**kwargs) assert (strides == 1) or ((shared_conv1 is None) and (shared_conv2 is None)) self.resize_input = (in_channels != out_channels) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if self.resize_input: self.input_conv = pre_conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.body = CRUBottleneck( in_channels=out_channels, out_channels=out_channels, strides=strides, group_width=group_width, bn_use_global_stats=bn_use_global_stats, shared_conv1=shared_conv1, shared_conv2=shared_conv2) if self.resize_identity: # assert (self.input_conv.conv._kwargs["stride"][0] == strides) self.identity_conv = cru_conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, shared_conv=self.input_conv.conv if strides == 1 else None) def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x if self.resize_input: x = self.input_conv(x) x = self.body(x) x = x + identity return x class CRUNetb(HybridBlock): """ CRU-Net(b) model from 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. group_widths: list of int List of group width parameters. refresh_steps: list of int List of refresh step parameters. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, group_widths, refresh_steps, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(CRUNetb, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) group_width = group_widths[i] refresh_step = refresh_steps[i] with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 if group_width != 0: if ((refresh_step == 0) and (j == 0)) or ((refresh_step != 0) and (j % refresh_step == 0)): shared_conv1 = None shared_conv2 = None if (shared_conv2 is not None) and (shared_conv2._kwargs["stride"][0] != 1) and\ (strides == 1): shared_conv2 = None unit = CRUUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, group_width=group_width, bn_use_global_stats=bn_use_global_stats, shared_conv1=shared_conv1, shared_conv2=shared_conv2) if shared_conv1 is None: shared_conv1 = unit.body.conv1.conv if shared_conv2 is None: shared_conv2 = unit.body.conv2 stage.add(unit) else: stage.add(ResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_crunetb(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create CRU-Net(b) model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ cardinality = 32 bottleneck_width = 4 if blocks == 56: layers = [3, 4, 6, 3] group_widths = [0, 0, 640, 0] refresh_steps = [0, 0, 0, 0] elif blocks == 116: layers = [3, 6, 18, 3] group_widths = [0, 352, 704, 0] refresh_steps = [0, 0, 6, 0] else: raise ValueError("Unsupported CRU-Net(b) with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = CRUNetb( channels=channels, init_block_channels=init_block_channels, cardinality=cardinality, bottleneck_width=bottleneck_width, group_widths=group_widths, refresh_steps=refresh_steps, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def crunet56b(**kwargs): """ CRU-Net-56(b) model from 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_crunetb(blocks=56, model_name="crunet56b", **kwargs) def crunet116b(**kwargs): """ CRU-Net-116(b) model from 'Sharing Residual Units Through Collective Tensor Factorization To Improve Deep Neural Networks,' https://www.ijcai.org/proceedings/2018/88. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_crunetb(blocks=116, model_name="crunet116b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ crunet56b, crunet116b, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != crunet56b or weight_count == 26139432) assert (model != crunet116b or weight_count == 44321000) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
20,687
33.138614
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/espnetv2.py
""" ESPNetv2 for ImageNet-1K, implemented in Gluon. Original paper: 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. """ __all__ = ['ESPNetv2', 'espnetv2_wd2', 'espnetv2_w1', 'espnetv2_w5d4', 'espnetv2_w3d2', 'espnetv2_w2'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import PReLU2, conv3x3, conv1x1_block, conv3x3_block, DualPathSequential class PreActivation(HybridBlock): """ PreResNet like pure pre-activation block without convolution layer. Parameters: ---------- in_channels : int Number of input channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, bn_use_global_stats=False, **kwargs): super(PreActivation, self).__init__(**kwargs) with self.name_scope(): self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) self.activ = PReLU2(in_channels=in_channels) def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) return x class ShortcutBlock(HybridBlock): """ ESPNetv2 shortcut block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(ShortcutBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats, activation=(lambda: PReLU2(in_channels))) self.conv2 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class HierarchicalConcurrent(nn.HybridSequential): """ A container for hierarchical concatenation of blocks with parameters. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. """ def __init__(self, axis=1, **kwargs): super(HierarchicalConcurrent, self).__init__(**kwargs) self.axis = axis def hybrid_forward(self, F, x): out = [] y_prev = None for block in self._children.values(): y = block(x) if y_prev is not None: y = y + y_prev out.append(y) y_prev = y out = F.concat(*out, dim=self.axis) return out class ESPBlock(HybridBlock): """ ESPNetv2 block (so-called EESP block). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the branch convolution layers. dilations : list of int Dilation values for branches. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, dilations, bn_use_global_stats, **kwargs): super(ESPBlock, self).__init__(**kwargs) num_branches = len(dilations) assert (out_channels % num_branches == 0) self.downsample = (strides != 1) mid_channels = out_channels // num_branches with self.name_scope(): self.reduce_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, groups=num_branches, bn_use_global_stats=bn_use_global_stats, activation=(lambda: PReLU2(mid_channels))) self.branches = HierarchicalConcurrent(axis=1, prefix="") for i in range(num_branches): self.branches.add(conv3x3( in_channels=mid_channels, out_channels=mid_channels, strides=strides, padding=dilations[i], dilation=dilations[i], groups=mid_channels)) self.merge_conv = conv1x1_block( in_channels=out_channels, out_channels=out_channels, groups=num_branches, bn_use_global_stats=bn_use_global_stats, activation=None) self.preactiv = PreActivation(in_channels=out_channels) if not self.downsample: self.activ = PReLU2(out_channels) def hybrid_forward(self, F, x, x0): y = self.reduce_conv(x) y = self.branches(y) y = self.preactiv(y) y = self.merge_conv(y) if not self.downsample: y = y + x y = self.activ(y) return y, x0 class DownsampleBlock(HybridBlock): """ ESPNetv2 downsample block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. x0_channels : int Number of input channels for shortcut. dilations : list of int Dilation values for branches in EESP block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, x0_channels, dilations, bn_use_global_stats, **kwargs): super(DownsampleBlock, self).__init__(**kwargs) inc_channels = out_channels - in_channels with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=3, strides=2, padding=1) self.eesp = ESPBlock( in_channels=in_channels, out_channels=inc_channels, strides=2, dilations=dilations, bn_use_global_stats=bn_use_global_stats) self.shortcut_block = ShortcutBlock( in_channels=x0_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.activ = PReLU2(out_channels) def hybrid_forward(self, F, x, x0): y1 = self.pool(x) y2, _ = self.eesp(x, None) x = F.concat(y1, y2, dim=1) x0 = self.pool(x0) y3 = self.shortcut_block(x0) x = x + y3 x = self.activ(x) return x, x0 class ESPInitBlock(HybridBlock): """ ESPNetv2 initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(ESPInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=2, bn_use_global_stats=bn_use_global_stats, activation=(lambda: PReLU2(out_channels))) self.pool = nn.AvgPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x, x0): x = self.conv(x) x0 = self.pool(x0) return x, x0 class ESPFinalBlock(HybridBlock): """ ESPNetv2 final block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. final_groups : int Number of groups in the last convolution layer. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, final_groups, bn_use_global_stats, **kwargs): super(ESPFinalBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, groups=in_channels, bn_use_global_stats=bn_use_global_stats, activation=(lambda: PReLU2(in_channels))) self.conv2 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, groups=final_groups, bn_use_global_stats=bn_use_global_stats, activation=(lambda: PReLU2(out_channels))) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class ESPNetv2(HybridBlock): """ ESPNetv2 model from 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. final_block_channels : int Number of output channels for the final unit. final_block_groups : int Number of groups for the final unit. dilations : list of list of list of int Dilation values for branches in each unit. dropout_rate : float, default 0.2 Parameter of Dropout layer. Faction of the input units to drop. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, final_block_groups, dilations, dropout_rate=0.2, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000): super(ESPNetv2, self).__init__() self.in_size = in_size self.classes = classes x0_channels = in_channels with self.name_scope(): self.features = DualPathSequential( return_two=False, first_ordinals=0, last_ordinals=2, prefix="") self.features.add(ESPInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential(prefix="stage{}_".format(i + 1)) for j, out_channels in enumerate(channels_per_stage): if j == 0: unit = DownsampleBlock( in_channels=in_channels, out_channels=out_channels, x0_channels=x0_channels, dilations=dilations[i][j], bn_use_global_stats=bn_use_global_stats) else: unit = ESPBlock( in_channels=in_channels, out_channels=out_channels, strides=1, dilations=dilations[i][j], bn_use_global_stats=bn_use_global_stats) stage.add(unit) in_channels = out_channels self.features.add(stage) self.features.add(ESPFinalBlock( in_channels=in_channels, out_channels=final_block_channels, final_groups=final_block_groups, bn_use_global_stats=bn_use_global_stats)) in_channels = final_block_channels self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dropout(rate=dropout_rate)) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x, x) x = self.output(x) return x def get_espnetv2(width_scale, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ESPNetv2 model with specific parameters. Parameters: ---------- width_scale : float Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (width_scale <= 2.0) branches = 4 layers = [1, 4, 8, 4] max_dilation_list = [6, 5, 4, 3, 2] max_dilations = [[max_dilation_list[i]] + [max_dilation_list[i + 1]] * (li - 1) for (i, li) in enumerate(layers)] dilations = [[sorted([k + 1 if k < dij else 1 for k in range(branches)]) for dij in di] for di in max_dilations] base_channels = 32 weighed_base_channels = math.ceil(float(math.floor(base_channels * width_scale)) / branches) * branches channels_per_layers = [weighed_base_channels * pow(2, i + 1) for i in range(len(layers))] init_block_channels = base_channels if weighed_base_channels > base_channels else weighed_base_channels final_block_channels = 1024 if width_scale <= 1.5 else 1280 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = ESPNetv2( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, final_block_groups=branches, dilations=dilations, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ignore_extra=True, ctx=ctx) return net def espnetv2_wd2(**kwargs): """ ESPNetv2 x0.5 model from 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_espnetv2(width_scale=0.5, model_name="espnetv2_wd2", **kwargs) def espnetv2_w1(**kwargs): """ ESPNetv2 x1.0 model from 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_espnetv2(width_scale=1.0, model_name="espnetv2_w1", **kwargs) def espnetv2_w5d4(**kwargs): """ ESPNetv2 x1.25 model from 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_espnetv2(width_scale=1.25, model_name="espnetv2_w5d4", **kwargs) def espnetv2_w3d2(**kwargs): """ ESPNetv2 x1.5 model from 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_espnetv2(width_scale=1.5, model_name="espnetv2_w3d2", **kwargs) def espnetv2_w2(**kwargs): """ ESPNetv2 x2.0 model from 'ESPNetv2: A Light-weight, Power Efficient, and General Purpose Convolutional Neural Network,' https://arxiv.org/abs/1811.11431. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_espnetv2(width_scale=2.0, model_name="espnetv2_w2", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ espnetv2_wd2, espnetv2_w1, espnetv2_w5d4, espnetv2_w3d2, espnetv2_w2, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != espnetv2_wd2 or weight_count == 1241092) assert (model != espnetv2_w1 or weight_count == 1669592) assert (model != espnetv2_w5d4 or weight_count == 1964832) assert (model != espnetv2_w3d2 or weight_count == 2314120) assert (model != espnetv2_w2 or weight_count == 3497144) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
20,193
32.600666
118
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/shufflenet.py
""" ShuffleNet for ImageNet-1K, implemented in Gluon. Original paper: 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. """ __all__ = ['ShuffleNet', 'shufflenet_g1_w1', 'shufflenet_g2_w1', 'shufflenet_g3_w1', 'shufflenet_g4_w1', 'shufflenet_g8_w1', 'shufflenet_g1_w3d4', 'shufflenet_g3_w3d4', 'shufflenet_g1_wd2', 'shufflenet_g3_wd2', 'shufflenet_g1_wd4', 'shufflenet_g3_wd4'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv3x3, depthwise_conv3x3, ChannelShuffle class ShuffleUnit(HybridBlock): """ ShuffleNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. groups : int Number of groups in convolution layers. downsample : bool Whether do downsample. ignore_group : bool Whether ignore group value in the first convolution layer. """ def __init__(self, in_channels, out_channels, groups, downsample, ignore_group, **kwargs): super(ShuffleUnit, self).__init__(**kwargs) self.downsample = downsample mid_channels = out_channels // 4 if downsample: out_channels -= in_channels with self.name_scope(): self.compress_conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels, groups=(1 if ignore_group else groups)) self.compress_bn1 = nn.BatchNorm(in_channels=mid_channels) self.c_shuffle = ChannelShuffle( channels=mid_channels, groups=groups) self.dw_conv2 = depthwise_conv3x3( channels=mid_channels, strides=(2 if self.downsample else 1)) self.dw_bn2 = nn.BatchNorm(in_channels=mid_channels) self.expand_conv3 = conv1x1( in_channels=mid_channels, out_channels=out_channels, groups=groups) self.expand_bn3 = nn.BatchNorm(in_channels=out_channels) if downsample: self.avgpool = nn.AvgPool2D( pool_size=3, strides=2, padding=1) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): identity = x x = self.compress_conv1(x) x = self.compress_bn1(x) x = self.activ(x) x = self.c_shuffle(x) x = self.dw_conv2(x) x = self.dw_bn2(x) x = self.expand_conv3(x) x = self.expand_bn3(x) if self.downsample: identity = self.avgpool(identity) x = F.concat(x, identity, dim=1) else: x = x + identity x = self.activ(x) return x class ShuffleInitBlock(HybridBlock): """ ShuffleNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(ShuffleInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv3x3( in_channels=in_channels, out_channels=out_channels, strides=2) self.bn = nn.BatchNorm(in_channels=out_channels) self.activ = nn.Activation("relu") self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) x = self.pool(x) return x class ShuffleNet(HybridBlock): """ ShuffleNet model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. groups : int Number of groups in convolution layers. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, groups, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(ShuffleNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ShuffleInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): downsample = (j == 0) ignore_group = (i == 0) and (j == 0) stage.add(ShuffleUnit( in_channels=in_channels, out_channels=out_channels, groups=groups, downsample=downsample, ignore_group=ignore_group)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_shufflenet(groups, width_scale, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ShuffleNet model with specific parameters. Parameters: ---------- groups : int Number of groups in convolution layers. width_scale : float Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ init_block_channels = 24 layers = [4, 8, 4] if groups == 1: channels_per_layers = [144, 288, 576] elif groups == 2: channels_per_layers = [200, 400, 800] elif groups == 3: channels_per_layers = [240, 480, 960] elif groups == 4: channels_per_layers = [272, 544, 1088] elif groups == 8: channels_per_layers = [384, 768, 1536] else: raise ValueError("The {} of groups is not supported".format(groups)) channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1.0: channels = [[int(cij * width_scale) for cij in ci] for ci in channels] init_block_channels = int(init_block_channels * width_scale) net = ShuffleNet( channels=channels, init_block_channels=init_block_channels, groups=groups, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def shufflenet_g1_w1(**kwargs): """ ShuffleNet 1x (g=1) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=1, width_scale=1.0, model_name="shufflenet_g1_w1", **kwargs) def shufflenet_g2_w1(**kwargs): """ ShuffleNet 1x (g=2) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=2, width_scale=1.0, model_name="shufflenet_g2_w1", **kwargs) def shufflenet_g3_w1(**kwargs): """ ShuffleNet 1x (g=3) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=3, width_scale=1.0, model_name="shufflenet_g3_w1", **kwargs) def shufflenet_g4_w1(**kwargs): """ ShuffleNet 1x (g=4) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=4, width_scale=1.0, model_name="shufflenet_g4_w1", **kwargs) def shufflenet_g8_w1(**kwargs): """ ShuffleNet 1x (g=8) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=8, width_scale=1.0, model_name="shufflenet_g8_w1", **kwargs) def shufflenet_g1_w3d4(**kwargs): """ ShuffleNet 0.75x (g=1) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=1, width_scale=0.75, model_name="shufflenet_g1_w3d4", **kwargs) def shufflenet_g3_w3d4(**kwargs): """ ShuffleNet 0.75x (g=3) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=3, width_scale=0.75, model_name="shufflenet_g3_w3d4", **kwargs) def shufflenet_g1_wd2(**kwargs): """ ShuffleNet 0.5x (g=1) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=1, width_scale=0.5, model_name="shufflenet_g1_wd2", **kwargs) def shufflenet_g3_wd2(**kwargs): """ ShuffleNet 0.5x (g=3) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=3, width_scale=0.5, model_name="shufflenet_g3_wd2", **kwargs) def shufflenet_g1_wd4(**kwargs): """ ShuffleNet 0.25x (g=1) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=1, width_scale=0.25, model_name="shufflenet_g1_wd4", **kwargs) def shufflenet_g3_wd4(**kwargs): """ ShuffleNet 0.25x (g=3) model from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_shufflenet(groups=3, width_scale=0.25, model_name="shufflenet_g3_wd4", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ shufflenet_g1_w1, shufflenet_g2_w1, shufflenet_g3_w1, shufflenet_g4_w1, shufflenet_g8_w1, shufflenet_g1_w3d4, shufflenet_g3_w3d4, shufflenet_g1_wd2, shufflenet_g3_wd2, shufflenet_g1_wd4, shufflenet_g3_wd4, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != shufflenet_g1_w1 or weight_count == 1531936) assert (model != shufflenet_g2_w1 or weight_count == 1733848) assert (model != shufflenet_g3_w1 or weight_count == 1865728) assert (model != shufflenet_g4_w1 or weight_count == 1968344) assert (model != shufflenet_g8_w1 or weight_count == 2434768) assert (model != shufflenet_g1_w3d4 or weight_count == 975214) assert (model != shufflenet_g3_w3d4 or weight_count == 1238266) assert (model != shufflenet_g1_wd2 or weight_count == 534484) assert (model != shufflenet_g3_wd2 or weight_count == 718324) assert (model != shufflenet_g1_wd4 or weight_count == 209746) assert (model != shufflenet_g3_wd4 or weight_count == 305902) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
17,247
33.222222
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/bamresnet.py
""" BAM-ResNet for ImageNet-1K, implemented in Gluon. Original paper: 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. """ __all__ = ['BamResNet', 'bam_resnet18', 'bam_resnet34', 'bam_resnet50', 'bam_resnet101', 'bam_resnet152'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv1x1_block, conv3x3_block from .resnet import ResInitBlock, ResUnit class DenseBlock(HybridBlock): """ Standard dense block with Batch normalization and ReLU activation. Parameters: ---------- in_channels : int Number of input features. out_channels : int Number of output features. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(DenseBlock, self).__init__(**kwargs) with self.name_scope(): self.fc = nn.Dense( units=out_channels, in_units=in_channels) self.bn = nn.BatchNorm( in_channels=out_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.fc(x) x = self.bn(x) x = self.activ(x) return x class ChannelGate(HybridBlock): """ BAM channel gate block. Parameters: ---------- channels : int Number of input/output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. reduction_ratio : int, default 16 Channel reduction ratio. num_layers : int, default 1 Number of dense blocks. """ def __init__(self, channels, bn_use_global_stats, reduction_ratio=16, num_layers=1, **kwargs): super(ChannelGate, self).__init__(**kwargs) mid_channels = channels // reduction_ratio with self.name_scope(): self.pool = nn.GlobalAvgPool2D() self.flatten = nn.Flatten() self.init_fc = DenseBlock( in_channels=channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.main_fcs = nn.HybridSequential(prefix="") for i in range(num_layers - 1): self.main_fcs.add(DenseBlock( in_channels=mid_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats)) self.final_fc = nn.Dense( units=channels, in_units=mid_channels) def hybrid_forward(self, F, x): input = x x = self.pool(x) x = self.flatten(x) x = self.init_fc(x) x = self.main_fcs(x) x = self.final_fc(x) x = x.expand_dims(2).expand_dims(3).broadcast_like(input) return x class SpatialGate(HybridBlock): """ BAM spatial gate block. Parameters: ---------- channels : int Number of input/output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. reduction_ratio : int, default 16 Channel reduction ratio. num_dil_convs : int, default 2 Number of dilated convolutions. dilation : int, default 4 Dilation/padding value for corresponding convolutions. """ def __init__(self, channels, bn_use_global_stats, reduction_ratio=16, num_dil_convs=2, dilation=4, **kwargs): super(SpatialGate, self).__init__(**kwargs) mid_channels = channels // reduction_ratio with self.name_scope(): self.init_conv = conv1x1_block( in_channels=channels, out_channels=mid_channels, strides=1, use_bias=True, bn_use_global_stats=bn_use_global_stats) self.dil_convs = nn.HybridSequential(prefix="") for i in range(num_dil_convs): self.dil_convs.add(conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=1, padding=dilation, dilation=dilation, use_bias=True, bn_use_global_stats=bn_use_global_stats)) self.final_conv = conv1x1( in_channels=mid_channels, out_channels=1, strides=1, use_bias=True) def hybrid_forward(self, F, x): input = x x = self.init_conv(x) x = self.dil_convs(x) x = self.final_conv(x) x = x.broadcast_like(input) return x class BamBlock(HybridBlock): """ BAM attention block for BAM-ResNet. Parameters: ---------- channels : int Number of input/output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, channels, bn_use_global_stats, **kwargs): super(BamBlock, self).__init__(**kwargs) with self.name_scope(): self.ch_att = ChannelGate( channels=channels, bn_use_global_stats=bn_use_global_stats) self.sp_att = SpatialGate( channels=channels, bn_use_global_stats=bn_use_global_stats) self.sigmoid = nn.Activation("sigmoid") def hybrid_forward(self, F, x): att = 1 + self.sigmoid(self.ch_att(x) * self.sp_att(x)) x = x * att return x class BamResUnit(HybridBlock): """ BAM-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, bottleneck, **kwargs): super(BamResUnit, self).__init__(**kwargs) self.use_bam = (strides != 1) with self.name_scope(): if self.use_bam: self.bam = BamBlock( channels=in_channels, bn_use_global_stats=bn_use_global_stats) self.res_unit = ResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=False) def hybrid_forward(self, F, x): if self.use_bam: x = self.bam(x) x = self.res_unit(x) return x class BamResNet(HybridBlock): """ BAM-ResNet model from 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(BamResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(BamResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resnet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create BAM-ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. use_se : bool Whether to use SE block. width_scale : float Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 18: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] else: raise ValueError("Unsupported BAM-ResNet with number of blocks: {}".format(blocks)) init_block_channels = 64 if blocks < 50: channels_per_layers = [64, 128, 256, 512] bottleneck = False else: channels_per_layers = [256, 512, 1024, 2048] bottleneck = True channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = BamResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def bam_resnet18(**kwargs): """ BAM-ResNet-18 model from 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet(blocks=18, model_name="bam_resnet18", **kwargs) def bam_resnet34(**kwargs): """ BAM-ResNet-34 model from 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet(blocks=34, model_name="bam_resnet34", **kwargs) def bam_resnet50(**kwargs): """ BAM-ResNet-50 model from 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet(blocks=50, model_name="bam_resnet50", **kwargs) def bam_resnet101(**kwargs): """ BAM-ResNet-101 model from 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet(blocks=101, model_name="bam_resnet101", **kwargs) def bam_resnet152(**kwargs): """ BAM-ResNet-152 model from 'BAM: Bottleneck Attention Module,' https://arxiv.org/abs/1807.06514. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet(blocks=152, model_name="bam_resnet152", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ bam_resnet18, bam_resnet34, bam_resnet50, bam_resnet101, bam_resnet152, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != bam_resnet18 or weight_count == 11712503) assert (model != bam_resnet34 or weight_count == 21820663) assert (model != bam_resnet50 or weight_count == 25915099) assert (model != bam_resnet101 or weight_count == 44907227) assert (model != bam_resnet152 or weight_count == 60550875) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
16,255
31.253968
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/resattnet.py
""" ResAttNet for ImageNet-1K, implemented in Gluon. Original paper: 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. """ __all__ = ['ResAttNet', 'resattnet56', 'resattnet92', 'resattnet128', 'resattnet164', 'resattnet200', 'resattnet236', 'resattnet452'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv7x7_block, pre_conv1x1_block, pre_conv3x3_block, Hourglass class PreResBottleneck(HybridBlock): """ PreResNet bottleneck block for residual path in PreResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, **kwargs): super(PreResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // 4 with self.name_scope(): self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, return_preact=True) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, strides=strides) self.conv3 = pre_conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x, x_pre_activ = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x, x_pre_activ class ResBlock(HybridBlock): """ Residual block with pre-activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides=1, bn_use_global_stats=False, **kwargs): super(ResBlock, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = PreResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, strides=strides) def hybrid_forward(self, F, x): identity = x x, x_pre_activ = self.body(x) if self.resize_identity: identity = self.identity_conv(x_pre_activ) x = x + identity return x class InterpolationBlock(HybridBlock): """ Interpolation block. Parameters: ---------- size : tuple of 2 int Spatial size of the output tensor for the bilinear upsampling operation. """ def __init__(self, size, **kwargs): super(InterpolationBlock, self).__init__(**kwargs) self.size = size def hybrid_forward(self, F, x): return F.contrib.BilinearResize2D(x, height=self.size[0], width=self.size[1]) class DoubleSkipBlock(HybridBlock): """ Double skip connection block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(DoubleSkipBlock, self).__init__(**kwargs) with self.name_scope(): self.skip1 = ResBlock( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = x + self.skip1(x) return x class ResBlockSequence(HybridBlock): """ Sequence of residual blocks with pre-activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. length : int Length of sequence. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, length, bn_use_global_stats, **kwargs): super(ResBlockSequence, self).__init__(**kwargs) with self.name_scope(): self.blocks = nn.HybridSequential(prefix="") for i in range(length): self.blocks.add(ResBlock( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.blocks(x) return x class DownAttBlock(HybridBlock): """ Down sub-block for hourglass of attention block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. length : int Length of residual blocks list. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, length, bn_use_global_stats, **kwargs): super(DownAttBlock, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) self.res_blocks = ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=length, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.pool(x) x = self.res_blocks(x) return x class UpAttBlock(HybridBlock): """ Up sub-block for hourglass of attention block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. length : int Length of residual blocks list. size : tuple of 2 int Spatial size of the output tensor for the bilinear upsampling operation. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, length, size, bn_use_global_stats, **kwargs): super(UpAttBlock, self).__init__(**kwargs) with self.name_scope(): self.res_blocks = ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=length, bn_use_global_stats=bn_use_global_stats) self.upsample = InterpolationBlock(size) def hybrid_forward(self, F, x): x = self.res_blocks(x) x = self.upsample(x) return x class MiddleAttBlock(HybridBlock): """ Middle sub-block for attention block. Parameters: ---------- channels : int Number of input/output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, channels, bn_use_global_stats, **kwargs): super(MiddleAttBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = pre_conv1x1_block( in_channels=channels, out_channels=channels, bn_use_global_stats=bn_use_global_stats) self.conv2 = pre_conv1x1_block( in_channels=channels, out_channels=channels, bn_use_global_stats=bn_use_global_stats) self.sigmoid = nn.Activation("sigmoid") def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.sigmoid(x) return x class AttBlock(HybridBlock): """ Attention block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. hourglass_depth : int Depth of hourglass block. att_scales : list of int Attention block specific scales. in_size : tuple of 2 int Spatial size of the input tensor for the bilinear upsampling operation. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, hourglass_depth, att_scales, in_size, bn_use_global_stats, **kwargs): super(AttBlock, self).__init__(**kwargs) assert (len(att_scales) == 3) scale_factor = 2 scale_p, scale_t, scale_r = att_scales with self.name_scope(): self.init_blocks = ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=scale_p, bn_use_global_stats=bn_use_global_stats) down_seq = nn.HybridSequential(prefix="") up_seq = nn.HybridSequential(prefix="") skip_seq = nn.HybridSequential(prefix="") for i in range(hourglass_depth): down_seq.add(DownAttBlock( in_channels=in_channels, out_channels=out_channels, length=scale_r, bn_use_global_stats=bn_use_global_stats)) up_seq.add(UpAttBlock( in_channels=in_channels, out_channels=out_channels, length=scale_r, size=in_size, bn_use_global_stats=bn_use_global_stats)) in_size = tuple([x // scale_factor for x in in_size]) if i == 0: skip_seq.add(ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=scale_t, bn_use_global_stats=bn_use_global_stats)) else: skip_seq.add(DoubleSkipBlock( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats)) self.hg = Hourglass( down_seq=down_seq, up_seq=up_seq, skip_seq=skip_seq, return_first_skip=True) self.middle_block = MiddleAttBlock( channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.final_block = ResBlock( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.init_blocks(x) x, y = self.hg(x) x = self.middle_block(x) x = (1 + x) * y x = self.final_block(x) return x class ResAttInitBlock(HybridBlock): """ ResAttNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(ResAttInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv7x7_block( in_channels=in_channels, out_channels=out_channels, strides=2, bn_use_global_stats=bn_use_global_stats) self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv(x) x = self.pool(x) return x class PreActivation(HybridBlock): """ Pre-activation block without convolution layer. It's used by itself as the final block in PreResNet. Parameters: ---------- in_channels : int Number of input channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, bn_use_global_stats, **kwargs): super(PreActivation, self).__init__(**kwargs) with self.name_scope(): self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) return x class ResAttNet(HybridBlock): """ ResAttNet model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. attentions : list of list of int Whether to use a attention unit or residual one. att_scales : list of int Attention block specific scales. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, attentions, att_scales, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(ResAttNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResAttInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels in_size = tuple([x // 4 for x in in_size]) for i, channels_per_stage in enumerate(channels): hourglass_depth = len(channels) - 1 - i stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 if attentions[i][j]: stage.add(AttBlock( in_channels=in_channels, out_channels=out_channels, hourglass_depth=hourglass_depth, att_scales=att_scales, in_size=in_size, bn_use_global_stats=bn_use_global_stats)) else: stage.add(ResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels in_size = tuple([x // strides for x in in_size]) self.features.add(stage) self.features.add(PreActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resattnet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ResAttNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 56: att_layers = [1, 1, 1] att_scales = [1, 2, 1] elif blocks == 92: att_layers = [1, 2, 3] att_scales = [1, 2, 1] elif blocks == 128: att_layers = [2, 3, 4] att_scales = [1, 2, 1] elif blocks == 164: att_layers = [3, 4, 5] att_scales = [1, 2, 1] elif blocks == 200: att_layers = [4, 5, 6] att_scales = [1, 2, 1] elif blocks == 236: att_layers = [5, 6, 7] att_scales = [1, 2, 1] elif blocks == 452: att_layers = [5, 6, 7] att_scales = [2, 4, 3] else: raise ValueError("Unsupported ResAttNet with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] layers = att_layers + [2] channels = [[ci] * (li + 1) for (ci, li) in zip(channels_per_layers, layers)] attentions = [[0] + [1] * li for li in att_layers] + [[0] * 3] net = ResAttNet( channels=channels, init_block_channels=init_block_channels, attentions=attentions, att_scales=att_scales, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def resattnet56(**kwargs): """ ResAttNet-56 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=56, model_name="resattnet56", **kwargs) def resattnet92(**kwargs): """ ResAttNet-92 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=92, model_name="resattnet92", **kwargs) def resattnet128(**kwargs): """ ResAttNet-128 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=128, model_name="resattnet128", **kwargs) def resattnet164(**kwargs): """ ResAttNet-164 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=164, model_name="resattnet164", **kwargs) def resattnet200(**kwargs): """ ResAttNet-200 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=200, model_name="resattnet200", **kwargs) def resattnet236(**kwargs): """ ResAttNet-236 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=236, model_name="resattnet236", **kwargs) def resattnet452(**kwargs): """ ResAttNet-452 model from 'Residual Attention Network for Image Classification,' https://arxiv.org/abs/1704.06904. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resattnet(blocks=452, model_name="resattnet452", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ resattnet56, resattnet92, resattnet128, resattnet164, resattnet200, resattnet236, resattnet452, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resattnet56 or weight_count == 31810728) assert (model != resattnet92 or weight_count == 52466344) assert (model != resattnet128 or weight_count == 65294504) assert (model != resattnet164 or weight_count == 78122664) assert (model != resattnet200 or weight_count == 90950824) assert (model != resattnet236 or weight_count == 103778984) assert (model != resattnet452 or weight_count == 182285224) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
25,649
32.096774
117
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/centernet.py
""" CenterNet for ImageNet-1K, implemented in Gluon. Original paper: 'Objects as Points,' https://arxiv.org/abs/1904.07850. """ __all__ = ['CenterNet', 'centernet_resnet18_voc', 'centernet_resnet18_coco', 'centernet_resnet50b_voc', 'centernet_resnet50b_coco', 'centernet_resnet101b_voc', 'centernet_resnet101b_coco', 'CenterNetHeatmapMaxDet'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import conv1x1, conv3x3_block, DeconvBlock from .resnet import resnet18, resnet50b, resnet101b class CenterNetDecoderUnit(HybridBlock): """ CenterNet decoder unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, **kwargs): super(CenterNetDecoderUnit, self).__init__(**kwargs) with self.name_scope(): self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, use_bias=True, bn_use_global_stats=bn_use_global_stats) self.deconv = DeconvBlock( in_channels=out_channels, out_channels=out_channels, kernel_size=4, strides=2, padding=1, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv(x) x = self.deconv(x) return x class CenterNetHeadBlock(HybridBlock): """ CenterNet simple head block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(CenterNetHeadBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, use_bias=True, use_bn=False) self.conv2 = conv1x1( in_channels=in_channels, out_channels=out_channels, use_bias=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class CenterNetHeatmapBlock(HybridBlock): """ CenterNet heatmap block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. do_nms : bool Whether do NMS (or simply clip for training otherwise). """ def __init__(self, in_channels, out_channels, do_nms, **kwargs): super(CenterNetHeatmapBlock, self).__init__(**kwargs) self.do_nms = do_nms with self.name_scope(): self.head = CenterNetHeadBlock( in_channels=in_channels, out_channels=out_channels) self.sigmoid = nn.Activation("sigmoid") if self.do_nms: self.pool = nn.MaxPool2D( pool_size=3, strides=1, padding=1) def hybrid_forward(self, F, x): x = self.head(x) x = self.sigmoid(x) if self.do_nms: y = self.pool(x) x = x * F.broadcast_equal(y, x) else: eps = 1e-4 x = x.clip(a_min=eps, a_max=(1.0 - eps)) return x class CenterNetHeatmapMaxDet(HybridBlock): """ CenterNet decoder for heads (heatmap, wh, reg). Parameters: ---------- topk : int, default 40 Keep only `topk` detections. scale : int, default is 4 Downsampling scale factor. max_batch : int, default is 256 Maximal batch size. """ def __init__(self, topk=40, scale=4, max_batch=256, **kwargs): super(CenterNetHeatmapMaxDet, self).__init__(**kwargs) self.topk = topk self.scale = scale self.max_batch = max_batch def hybrid_forward(self, F, x): heatmap = x.slice_axis(axis=1, begin=0, end=-4) wh = x.slice_axis(axis=1, begin=-4, end=-2) reg = x.slice_axis(axis=1, begin=-2, end=None) _, _, out_h, out_w = heatmap.shape_array().split(num_outputs=4, axis=0) scores, indices = heatmap.reshape((0, -1)).topk(k=self.topk, ret_typ="both") indices = indices.astype(dtype="int64") topk_classes = F.broadcast_div(indices, (out_h * out_w)).astype(dtype="float32") topk_indices = F.broadcast_mod(indices, (out_h * out_w)) topk_ys = F.broadcast_div(topk_indices, out_w).astype(dtype="float32") topk_xs = F.broadcast_mod(topk_indices, out_w).astype(dtype="float32") center = reg.transpose((0, 2, 3, 1)).reshape((0, -1, 2)) wh = wh.transpose((0, 2, 3, 1)).reshape((0, -1, 2)) batch_indices = F.arange(self.max_batch).slice_like(center, axes=0).expand_dims(-1).repeat(self.topk, 1).\ astype(dtype="int64") reg_xs_indices = F.zeros_like(batch_indices, dtype="int64") reg_ys_indices = F.ones_like(batch_indices, dtype="int64") reg_xs = F.concat(batch_indices, topk_indices, reg_xs_indices, dim=0).reshape((3, -1)) reg_ys = F.concat(batch_indices, topk_indices, reg_ys_indices, dim=0).reshape((3, -1)) xs = F.gather_nd(center, reg_xs).reshape((-1, self.topk)) ys = F.gather_nd(center, reg_ys).reshape((-1, self.topk)) topk_xs = topk_xs + xs topk_ys = topk_ys + ys w = F.gather_nd(wh, reg_xs).reshape((-1, self.topk)) h = F.gather_nd(wh, reg_ys).reshape((-1, self.topk)) half_w = 0.5 * w half_h = 0.5 * h bboxes = F.stack(topk_xs - half_w, topk_ys - half_h, topk_xs + half_w, topk_ys + half_h, axis=-1) bboxes = bboxes * self.scale topk_classes = topk_classes.expand_dims(axis=-1) scores = scores.expand_dims(axis=-1) result = F.concat(bboxes, topk_classes, scores, dim=-1) return result def __repr__(self): s = "{name}(topk={topk}, scale={scale})" return s.format( name=self.__class__.__name__, topk=self.topk, scale=self.scale) def calc_flops(self, x): assert (x.shape[0] == 1) num_flops = 10 * x.size num_macs = 0 return num_flops, num_macs class CenterNet(HybridBlock): """ CenterNet model from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int Number of output channels for the backbone. channels : list of int Number of output channels for each decoder unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. return_heatmap : bool, default False Whether to return only heatmap. topk : int, default 40 Keep only `topk` detections. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (512, 512) Spatial size of the expected input image. classes : int, default 80 Number of classification classes. """ def __init__(self, backbone, backbone_out_channels, channels, bn_use_global_stats=False, return_heatmap=False, topk=40, in_channels=3, in_size=(512, 512), classes=80, **kwargs): super(CenterNet, self).__init__(**kwargs) self.in_size = in_size self.in_channels = in_channels self.return_heatmap = return_heatmap with self.name_scope(): self.backbone = backbone self.decoder = nn.HybridSequential(prefix="") in_channels = backbone_out_channels for out_channels in channels: self.decoder.add(CenterNetDecoderUnit( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels heads = HybridConcurrent(axis=1, prefix="") heads.add(CenterNetHeatmapBlock( in_channels=in_channels, out_channels=classes, do_nms=(not self.return_heatmap))) heads.add(CenterNetHeadBlock( in_channels=in_channels, out_channels=2)) heads.add(CenterNetHeadBlock( in_channels=in_channels, out_channels=2)) self.decoder.add(heads) if not self.return_heatmap: self.heatmap_max_det = CenterNetHeatmapMaxDet( topk=topk, scale=4) def hybrid_forward(self, F, x): x = self.backbone(x) x = self.decoder(x) if not self.return_heatmap: x = self.heatmap_max_det(x) return x def get_centernet(backbone, backbone_out_channels, classes, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create CenterNet model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int Number of output channels for the backbone. classes : int Number of classes. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. Returns: ------- HybridBlock A network. """ channels = [256, 128, 64] net = CenterNet( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, classes=classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def centernet_resnet18_voc(pretrained_backbone=False, classes=20, **kwargs): """ CenterNet model on the base of ResNet-101b for VOC Detection from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 20 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features[:-1] return get_centernet(backbone=backbone, backbone_out_channels=512, classes=classes, model_name="centernet_resnet18_voc", **kwargs) def centernet_resnet18_coco(pretrained_backbone=False, classes=80, **kwargs): """ CenterNet model on the base of ResNet-101b for COCO Detection from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 80 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features[:-1] return get_centernet(backbone=backbone, backbone_out_channels=512, classes=classes, model_name="centernet_resnet18_coco", **kwargs) def centernet_resnet50b_voc(pretrained_backbone=False, classes=20, **kwargs): """ CenterNet model on the base of ResNet-101b for VOC Detection from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 20 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features[:-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, classes=classes, model_name="centernet_resnet50b_voc", **kwargs) def centernet_resnet50b_coco(pretrained_backbone=False, classes=80, **kwargs): """ CenterNet model on the base of ResNet-101b for COCO Detection from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 80 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features[:-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, classes=classes, model_name="centernet_resnet50b_coco", **kwargs) def centernet_resnet101b_voc(pretrained_backbone=False, classes=20, **kwargs): """ CenterNet model on the base of ResNet-101b for VOC Detection from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 20 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet101b(pretrained=pretrained_backbone).features[:-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, classes=classes, model_name="centernet_resnet101b_voc", **kwargs) def centernet_resnet101b_coco(pretrained_backbone=False, classes=80, **kwargs): """ CenterNet model on the base of ResNet-101b for COCO Detection from 'Objects as Points,' https://arxiv.org/abs/1904.07850. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 80 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet101b(pretrained=pretrained_backbone).features[:-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, classes=classes, model_name="centernet_resnet101b_coco", **kwargs) def _test(): import numpy as np import mxnet as mx in_size = (512, 512) topk = 40 return_heatmap = False pretrained = False models = [ (centernet_resnet18_voc, 20), (centernet_resnet18_coco, 80), (centernet_resnet50b_voc, 20), (centernet_resnet50b_coco, 80), (centernet_resnet101b_voc, 20), (centernet_resnet101b_coco, 80), ] for model, classes in models: net = model(pretrained=pretrained, topk=topk, in_size=in_size, return_heatmap=return_heatmap) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != centernet_resnet18_voc or weight_count == 14215640) assert (model != centernet_resnet18_coco or weight_count == 14219540) assert (model != centernet_resnet50b_voc or weight_count == 30086104) assert (model != centernet_resnet50b_coco or weight_count == 30090004) assert (model != centernet_resnet101b_voc or weight_count == 49078232) assert (model != centernet_resnet101b_coco or weight_count == 49082132) batch = 14 x = mx.nd.random.normal(shape=(batch, 3, in_size[0], in_size[1]), ctx=ctx) y = net(x) assert (y.shape[0] == batch) if return_heatmap: assert (y.shape[1] == classes + 4) and (y.shape[2] == x.shape[2] // 4) and (y.shape[3] == x.shape[3] // 4) else: assert (y.shape[1] == topk) and (y.shape[2] == 6) if __name__ == "__main__": _test()
18,867
34.466165
118
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/xdensenet_cifar.py
""" X-DenseNet for CIFAR/SVHN, implemented in Gluon. Original paper: 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. """ __all__ = ['CIFARXDenseNet', 'xdensenet40_2_k24_bc_cifar10', 'xdensenet40_2_k24_bc_cifar100', 'xdensenet40_2_k24_bc_svhn', 'xdensenet40_2_k36_bc_cifar10', 'xdensenet40_2_k36_bc_cifar100', 'xdensenet40_2_k36_bc_svhn'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3 from .preresnet import PreResActivation from .densenet import TransitionBlock from .xdensenet import pre_xconv3x3_block, XDenseUnit class XDenseSimpleUnit(HybridBlock): """ X-DenseNet simple unit for CIFAR. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. expand_ratio : int Ratio of expansion. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, dropout_rate, expand_ratio, **kwargs): super(XDenseSimpleUnit, self).__init__(**kwargs) self.use_dropout = (dropout_rate != 0.0) inc_channels = out_channels - in_channels with self.name_scope(): self.conv = pre_xconv3x3_block( in_channels=in_channels, out_channels=inc_channels, bn_use_global_stats=bn_use_global_stats, expand_ratio=expand_ratio) if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x): identity = x x = self.conv(x) if self.use_dropout: x = self.dropout(x) x = F.concat(identity, x, dim=1) return x class CIFARXDenseNet(HybridBlock): """ X-DenseNet model for CIFAR from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. expand_ratio : int, default 2 Ratio of expansion. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, bn_use_global_stats=False, dropout_rate=0.0, expand_ratio=2, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARXDenseNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes unit_class = XDenseUnit if bottleneck else XDenseSimpleUnit with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): if i != 0: stage.add(TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2), bn_use_global_stats=bn_use_global_stats)) in_channels = in_channels // 2 for j, out_channels in enumerate(channels_per_stage): stage.add(unit_class( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, dropout_rate=dropout_rate, expand_ratio=expand_ratio)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_xdensenet_cifar(classes, blocks, growth_rate, bottleneck, expand_ratio=2, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create X-DenseNet model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. blocks : int Number of blocks. growth_rate : int Growth rate. bottleneck : bool Whether to use a bottleneck or simple block in units. expand_ratio : int, default 2 Ratio of expansion. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (classes in [10, 100]) if bottleneck: assert ((blocks - 4) % 6 == 0) layers = [(blocks - 4) // 6] * 3 else: assert ((blocks - 4) % 3 == 0) layers = [(blocks - 4) // 3] * 3 init_block_channels = 2 * growth_rate from functools import reduce channels = reduce( lambda xi, yi: xi + [reduce( lambda xj, yj: xj + [xj[-1] + yj], [growth_rate] * yi, [xi[-1][-1] // 2])[1:]], layers, [[init_block_channels * 2]])[1:] net = CIFARXDenseNet( channels=channels, init_block_channels=init_block_channels, classes=classes, bottleneck=bottleneck, expand_ratio=expand_ratio, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def xdensenet40_2_k24_bc_cifar10(classes=10, **kwargs): """ X-DenseNet-BC-40-2 (k=24) model for CIFAR-10 from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True, model_name="xdensenet40_2_k24_bc_cifar10", **kwargs) def xdensenet40_2_k24_bc_cifar100(classes=100, **kwargs): """ X-DenseNet-BC-40-2 (k=24) model for CIFAR-100 from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True, model_name="xdensenet40_2_k24_bc_cifar100", **kwargs) def xdensenet40_2_k24_bc_svhn(classes=10, **kwargs): """ X-DenseNet-BC-40-2 (k=24) model for SVHN from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True, model_name="xdensenet40_2_k24_bc_svhn", **kwargs) def xdensenet40_2_k36_bc_cifar10(classes=10, **kwargs): """ X-DenseNet-BC-40-2 (k=36) model for CIFAR-10 from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True, model_name="xdensenet40_2_k36_bc_cifar10", **kwargs) def xdensenet40_2_k36_bc_cifar100(classes=100, **kwargs): """ X-DenseNet-BC-40-2 (k=36) model for CIFAR-100 from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True, model_name="xdensenet40_2_k36_bc_cifar100", **kwargs) def xdensenet40_2_k36_bc_svhn(classes=10, **kwargs): """ X-DenseNet-BC-40-2 (k=36) model for SVHN from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True, model_name="xdensenet40_2_k36_bc_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (xdensenet40_2_k24_bc_cifar10, 10), (xdensenet40_2_k24_bc_cifar100, 100), (xdensenet40_2_k24_bc_svhn, 10), (xdensenet40_2_k36_bc_cifar10, 10), (xdensenet40_2_k36_bc_cifar100, 100), (xdensenet40_2_k36_bc_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != xdensenet40_2_k24_bc_cifar10 or weight_count == 690346) assert (model != xdensenet40_2_k24_bc_cifar100 or weight_count == 714196) assert (model != xdensenet40_2_k24_bc_svhn or weight_count == 690346) assert (model != xdensenet40_2_k36_bc_cifar10 or weight_count == 1542682) assert (model != xdensenet40_2_k36_bc_cifar100 or weight_count == 1578412) assert (model != xdensenet40_2_k36_bc_svhn or weight_count == 1542682) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
14,171
35.245524
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/ntsnet_cub.py
""" NTS-Net for CUB-200-2011, implemented in Gluon. Original paper: 'Learning to Navigate for Fine-grained Classification,' https://arxiv.org/abs/1809.00287. """ __all__ = ['NTSNet', 'ntsnet_cub'] import os import numpy as np import mxnet as mx from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv3x3 from .resnet import resnet50b def hard_nms(cdds, top_n=10, iou_thresh=0.25): """ Hard Non-Maximum Suppression. Parameters: ---------- cdds : np.array Borders. top_n : int, default 10 Number of top-K informative regions. iou_thresh : float, default 0.25 IoU threshold. Returns: ------- np.array Filtered borders. """ assert (type(cdds) == np.ndarray) assert (len(cdds.shape) == 2) assert (cdds.shape[1] >= 5) cdds = cdds.copy() indices = np.argsort(cdds[:, 0]) cdds = cdds[indices] cdd_results = [] res = cdds while res.any(): cdd = res[-1] cdd_results.append(cdd) if len(cdd_results) == top_n: return np.array(cdd_results) res = res[:-1] start_max = np.maximum(res[:, 1:3], cdd[1:3]) end_min = np.minimum(res[:, 3:5], cdd[3:5]) lengths = end_min - start_max intersec_map = lengths[:, 0] * lengths[:, 1] intersec_map[np.logical_or(lengths[:, 0] < 0, lengths[:, 1] < 0)] = 0 iou_map_cur = intersec_map / ((res[:, 3] - res[:, 1]) * (res[:, 4] - res[:, 2]) + (cdd[3] - cdd[1]) * ( cdd[4] - cdd[2]) - intersec_map) res = res[iou_map_cur < iou_thresh] return np.array(cdd_results) class NavigatorBranch(HybridBlock): """ Navigator branch block for Navigator unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, strides, **kwargs): super(NavigatorBranch, self).__init__(**kwargs) mid_channels = 128 with self.name_scope(): self.down_conv = conv3x3( in_channels=in_channels, out_channels=mid_channels, strides=strides, use_bias=True) self.activ = nn.Activation("relu") self.tidy_conv = conv1x1( in_channels=mid_channels, out_channels=out_channels, use_bias=True) self.flatten = nn.Flatten() def hybrid_forward(self, F, x): y = self.down_conv(x) y = self.activ(y) z = self.tidy_conv(y) z = self.flatten(z) return z, y class NavigatorUnit(HybridBlock): """ Navigator init. """ def __init__(self, **kwargs): super(NavigatorUnit, self).__init__(**kwargs) with self.name_scope(): self.branch1 = NavigatorBranch( in_channels=2048, out_channels=6, strides=1) self.branch2 = NavigatorBranch( in_channels=128, out_channels=6, strides=2) self.branch3 = NavigatorBranch( in_channels=128, out_channels=9, strides=2) def hybrid_forward(self, F, x): t1, x = self.branch1(x) t2, x = self.branch2(x) t3, _ = self.branch3(x) return F.concat(t1, t2, t3, dim=1) class NTSNet(HybridBlock): """ NTS-Net model from 'Learning to Navigate for Fine-grained Classification,' https://arxiv.org/abs/1809.00287. Parameters: ---------- backbone : nn.Sequential Feature extractor. aux : bool, default False Whether to output auxiliary results. top_n : int, default 4 Number of extra top-K informative regions. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, backbone, aux=False, top_n=4, in_channels=3, in_size=(448, 448), classes=200, **kwargs): super(NTSNet, self).__init__(**kwargs) assert (in_channels > 0) self.in_size = in_size self.classes = classes self.top_n = top_n self.aux = aux self.num_cat = 4 pad_side = 224 self.pad_width = (0, 0, 0, 0, pad_side, pad_side, pad_side, pad_side) _, edge_anchors, _ = self._generate_default_anchor_maps() self.edge_anchors = (edge_anchors + 224).astype(np.int) self.edge_anchors = np.concatenate( (self.edge_anchors.copy(), np.arange(0, len(self.edge_anchors)).reshape(-1, 1)), axis=1) with self.name_scope(): self.backbone = backbone self.backbone_tail = nn.HybridSequential(prefix="") self.backbone_tail.add(nn.GlobalAvgPool2D()) self.backbone_tail.add(nn.Flatten()) self.backbone_tail.add(nn.Dropout(rate=0.5)) self.backbone_classifier = nn.Dense( units=classes, in_units=(512 * 4)) self.navigator_unit = NavigatorUnit() self.concat_net = nn.Dense( units=classes, in_units=(2048 * (self.num_cat + 1))) if self.aux: self.partcls_net = nn.Dense( units=classes, in_units=(512 * 4)) def hybrid_forward(self, F, x): raw_pre_features = self.backbone(x) rpn_score = self.navigator_unit(raw_pre_features) all_cdds = [np.concatenate((y.reshape(-1, 1), self.edge_anchors.copy()), axis=1) for y in rpn_score.asnumpy()] top_n_cdds = [hard_nms(y, top_n=self.top_n, iou_thresh=0.25) for y in all_cdds] top_n_cdds = np.array(top_n_cdds) top_n_index = top_n_cdds[:, :, -1].astype(np.int64) top_n_index2 = mx.nd.array(np.array([np.repeat(np.arange(top_n_cdds.shape[0]), top_n_cdds.shape[1]), top_n_index.flatten()]), dtype=np.int64) top_n_prob = F.gather_nd(rpn_score, top_n_index2).reshape(x.shape[0], -1) batch = x.shape[0] part_imgs = mx.nd.zeros(shape=(batch, self.top_n, 3, 224, 224), ctx=x.context, dtype=x.dtype) x_pad = F.pad(x, mode="constant", pad_width=self.pad_width, constant_value=0) for i in range(batch): for j in range(self.top_n): y0, x0, y1, x1 = tuple(top_n_cdds[i][j, 1:5].astype(np.int64)) part_imgs[i:i + 1, j] = F.contrib.BilinearResize2D( x_pad[i:i + 1, :, y0:y1, x0:x1], height=224, width=224) part_imgs = part_imgs.reshape((batch * self.top_n, 3, 224, 224)) part_features = self.backbone_tail(self.backbone(part_imgs.detach())) part_feature = part_features.reshape((batch, self.top_n, -1)) part_feature = part_feature[:, :self.num_cat, :] part_feature = part_feature.reshape((batch, -1)) raw_features = self.backbone_tail(raw_pre_features.detach()) concat_out = F.concat(part_feature, raw_features, dim=1) concat_logits = self.concat_net(concat_out) if self.aux: raw_logits = self.backbone_classifier(raw_features) part_logits = self.partcls_net(part_features).reshape((batch, self.top_n, -1)) return concat_logits, raw_logits, part_logits, top_n_prob else: return concat_logits @staticmethod def _generate_default_anchor_maps(input_shape=(448, 448)): """ Generate default anchor maps. Parameters: ---------- input_shape : tuple of 2 int Input image size. Returns: ------- center_anchors : np.array anchors * 4 (oy, ox, h, w). edge_anchors : np.array anchors * 4 (y0, x0, y1, x1). anchor_area : np.array anchors * 1 (area). """ anchor_scale = [2 ** (1.0 / 3.0), 2 ** (2.0 / 3.0)] anchor_aspect_ratio = [0.667, 1, 1.5] anchors_setting = ( dict(layer="p3", stride=32, size=48, scale=anchor_scale, aspect_ratio=anchor_aspect_ratio), dict(layer="p4", stride=64, size=96, scale=anchor_scale, aspect_ratio=anchor_aspect_ratio), dict(layer="p5", stride=128, size=192, scale=[1, anchor_scale[0], anchor_scale[1]], aspect_ratio=anchor_aspect_ratio), ) center_anchors = np.zeros((0, 4), dtype=np.float32) edge_anchors = np.zeros((0, 4), dtype=np.float32) anchor_areas = np.zeros((0,), dtype=np.float32) input_shape = np.array(input_shape, dtype=int) for anchor_info in anchors_setting: stride = anchor_info["stride"] size = anchor_info["size"] scales = anchor_info["scale"] aspect_ratios = anchor_info["aspect_ratio"] output_map_shape = np.ceil(input_shape.astype(np.float32) / stride) output_map_shape = output_map_shape.astype(np.int) output_shape = tuple(output_map_shape) + (4, ) ostart = stride / 2.0 oy = np.arange(ostart, ostart + stride * output_shape[0], stride) oy = oy.reshape(output_shape[0], 1) ox = np.arange(ostart, ostart + stride * output_shape[1], stride) ox = ox.reshape(1, output_shape[1]) center_anchor_map_template = np.zeros(output_shape, dtype=np.float32) center_anchor_map_template[:, :, 0] = oy center_anchor_map_template[:, :, 1] = ox for anchor_scale in scales: for anchor_aspect_ratio in aspect_ratios: center_anchor_map = center_anchor_map_template.copy() center_anchor_map[:, :, 2] = size * anchor_scale / float(anchor_aspect_ratio) ** 0.5 center_anchor_map[:, :, 3] = size * anchor_scale * float(anchor_aspect_ratio) ** 0.5 edge_anchor_map = np.concatenate( (center_anchor_map[:, :, :2] - center_anchor_map[:, :, 2:4] / 2.0, center_anchor_map[:, :, :2] + center_anchor_map[:, :, 2:4] / 2.0), axis=-1) anchor_area_map = center_anchor_map[:, :, 2] * center_anchor_map[:, :, 3] center_anchors = np.concatenate((center_anchors, center_anchor_map.reshape(-1, 4))) edge_anchors = np.concatenate((edge_anchors, edge_anchor_map.reshape(-1, 4))) anchor_areas = np.concatenate((anchor_areas, anchor_area_map.reshape(-1))) return center_anchors, edge_anchors, anchor_areas def get_ntsnet(backbone, aux=False, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create NTS-Net model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. aux : bool, default False Whether to output auxiliary results. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ net = NTSNet( backbone=backbone, aux=aux, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx, ignore_extra=True) return net def ntsnet_cub(pretrained_backbone=False, aux=True, **kwargs): """ NTS-Net model from 'Learning to Navigate for Fine-grained Classification,' https://arxiv.org/abs/1809.00287. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features[:-1] return get_ntsnet(backbone=backbone, aux=aux, model_name="ntsnet_cub", **kwargs) def _test(): import numpy as np import mxnet as mx aux = True pretrained = False models = [ ntsnet_cub, ] for model in models: net = model(pretrained=pretrained, aux=aux) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) if aux: assert (model != ntsnet_cub or weight_count == 29033133) else: assert (model != ntsnet_cub or weight_count == 28623333) x = mx.nd.zeros((5, 3, 448, 448), ctx=ctx) ys = net(x) y = ys[0] if aux else ys assert (y.shape[0] == x.shape[0]) and (y.shape[1] == 200) if __name__ == "__main__": _test()
14,467
33.695444
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/proxylessnas_cub.py
""" ProxylessNAS for CUB-200-2011, implemented in Gluon. Original paper: 'ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware,' https://arxiv.org/abs/1812.00332. """ __all__ = ['proxylessnas_cpu_cub', 'proxylessnas_gpu_cub', 'proxylessnas_mobile_cub', 'proxylessnas_mobile14_cub'] from .proxylessnas import get_proxylessnas def proxylessnas_cpu_cub(classes=200, **kwargs): """ ProxylessNAS (CPU) model for CUB-200-2011 from 'ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware,' https://arxiv.org/abs/1812.00332. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_proxylessnas(classes=classes, version="cpu", model_name="proxylessnas_cpu_cub", **kwargs) def proxylessnas_gpu_cub(classes=200, **kwargs): """ ProxylessNAS (GPU) model for CUB-200-2011 from 'ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware,' https://arxiv.org/abs/1812.00332. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_proxylessnas(classes=classes, version="gpu", model_name="proxylessnas_gpu_cub", **kwargs) def proxylessnas_mobile_cub(classes=200, **kwargs): """ ProxylessNAS (Mobile) model for CUB-200-2011 from 'ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware,' https://arxiv.org/abs/1812.00332. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_proxylessnas(classes=classes, version="mobile", model_name="proxylessnas_mobile_cub", **kwargs) def proxylessnas_mobile14_cub(classes=200, **kwargs): """ ProxylessNAS (Mobile-14) model for CUB-200-2011 from 'ProxylessNAS: Direct Neural Architecture Search on Target Task and Hardware,' https://arxiv.org/abs/1812.00332. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_proxylessnas(classes=classes, version="mobile14", model_name="proxylessnas_mobile14_cub", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ proxylessnas_cpu_cub, proxylessnas_gpu_cub, proxylessnas_mobile_cub, proxylessnas_mobile14_cub, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != proxylessnas_cpu_cub or weight_count == 3215248) assert (model != proxylessnas_gpu_cub or weight_count == 5736648) assert (model != proxylessnas_mobile_cub or weight_count == 3055712) assert (model != proxylessnas_mobile14_cub or weight_count == 5423168) x = mx.nd.zeros((14, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (14, 200)) if __name__ == "__main__": _test()
4,484
33.767442
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/ibnresnet.py
""" IBN-ResNet for ImageNet-1K, implemented in Gluon. Original paper: 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. """ __all__ = ['IBNResNet', 'ibn_resnet50', 'ibn_resnet101', 'ibn_resnet152'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, IBN from .resnet import ResInitBlock class IBNConvBlock(HybridBlock): """ IBN-Net specific convolution block with BN/IBN normalization and ReLU activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_ibn : bool, default False Whether use Instance-Batch Normalization. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_ibn=False, bn_use_global_stats=False, activate=True, **kwargs): super(IBNConvBlock, self).__init__(**kwargs) self.activate = activate self.use_ibn = use_ibn with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) if self.use_ibn: self.ibn = IBN(channels=out_channels) else: self.bn = nn.BatchNorm( in_channels=out_channels, use_global_stats=bn_use_global_stats) if self.activate: self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.conv(x) if self.use_ibn: x = self.ibn(x) else: x = self.bn(x) if self.activate: x = self.activ(x) return x def ibn_conv1x1_block(in_channels, out_channels, strides=1, groups=1, use_bias=False, use_ibn=False, bn_use_global_stats=False, activate=True): """ 1x1 version of the IBN-Net specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_ibn : bool, default False Whether use Instance-Batch Normalization. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activate : bool, default True Whether activate the convolution block. """ return IBNConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, groups=groups, use_bias=use_bias, use_ibn=use_ibn, bn_use_global_stats=bn_use_global_stats, activate=activate) class IBNResBottleneck(HybridBlock): """ IBN-ResNet bottleneck block for residual path in IBN-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. conv1_ibn : bool Whether to use IBN normalization in the first convolution layer of the block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, conv1_ibn, bn_use_global_stats, **kwargs): super(IBNResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // 4 with self.name_scope(): self.conv1 = ibn_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_ibn=conv1_ibn, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class IBNResUnit(HybridBlock): """ IBN-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. conv1_ibn : bool Whether to use IBN normalization in the first convolution layer of the block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, conv1_ibn, bn_use_global_stats, **kwargs): super(IBNResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = IBNResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, conv1_ibn=conv1_ibn, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class IBNResNet(HybridBlock): """ IBN-ResNet model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(IBNResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 conv1_ibn = (out_channels < 2048) stage.add(IBNResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, conv1_ibn=conv1_ibn, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_ibnresnet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create IBN-ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] else: raise ValueError("Unsupported IBN-ResNet with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = IBNResNet( channels=channels, init_block_channels=init_block_channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def ibn_resnet50(**kwargs): """ IBN-ResNet-50 model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibnresnet(blocks=50, model_name="ibn_resnet50", **kwargs) def ibn_resnet101(**kwargs): """ IBN-ResNet-101 model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibnresnet(blocks=101, model_name="ibn_resnet101", **kwargs) def ibn_resnet152(**kwargs): """ IBN-ResNet-152 model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibnresnet(blocks=152, model_name="ibn_resnet152", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ ibn_resnet50, ibn_resnet101, ibn_resnet152, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ibn_resnet50 or weight_count == 25557032) assert (model != ibn_resnet101 or weight_count == 44549160) assert (model != ibn_resnet152 or weight_count == 60192808) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
14,932
31.81978
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/common.py
""" Common routines for models in Gluon. """ __all__ = ['round_channels', 'BreakBlock', 'get_activation_layer', 'ReLU6', 'PReLU2', 'HSigmoid', 'HSwish', 'Softmax', 'SelectableDense', 'BatchNormExtra', 'DenseBlock', 'ConvBlock1d', 'conv1x1', 'conv3x3', 'depthwise_conv3x3', 'ConvBlock', 'conv1x1_block', 'conv3x3_block', 'conv5x5_block', 'conv7x7_block', 'dwconv_block', 'dwconv3x3_block', 'dwconv5x5_block', 'dwsconv3x3_block', 'PreConvBlock', 'pre_conv1x1_block', 'pre_conv3x3_block', 'DeconvBlock', 'NormActivation', 'InterpolationBlock', 'ChannelShuffle', 'ChannelShuffle2', 'SEBlock', 'SABlock', 'SAConvBlock', 'saconv3x3_block', 'DucBlock', 'split', 'IBN', 'DualPathSequential', 'ParametricSequential', 'Concurrent', 'SequentialConcurrent', 'ParametricConcurrent', 'Hourglass', 'SesquialteralHourglass', 'MultiOutputSequential', 'ParallelConcurent', 'DualPathParallelConcurent', 'HeatmapMaxDetBlock'] import math from inspect import isfunction import mxnet as mx from mxnet.gluon import nn, HybridBlock def round_channels(channels, divisor=8): """ Round weighted channel number (make divisible operation). Parameters: ---------- channels : int or float Original number of channels. divisor : int, default 8 Alignment value. Returns: ------- int Weighted number of channels. """ rounded_channels = max(int(channels + divisor / 2.0) // divisor * divisor, divisor) if float(rounded_channels) < 0.9 * channels: rounded_channels += divisor return rounded_channels class BreakBlock(HybridBlock): """ Break coonnection block for hourglass. """ def __init__(self, prefix=None, params=None): super(BreakBlock, self).__init__(prefix=prefix, params=params) def hybrid_forward(self, F, x): return None def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) class ReLU6(HybridBlock): """ ReLU6 activation layer. """ def __init__(self, **kwargs): super(ReLU6, self).__init__(**kwargs) def hybrid_forward(self, F, x): return F.clip(x, 0.0, 6.0, name="relu6") def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) class PReLU2(HybridBlock): """ Parametric leaky version of a Rectified Linear Unit (with wide alpha). Parameters: ---------- in_channels : int Number of input channels. alpha_initializer : Initializer Initializer for the `embeddings` matrix. """ def __init__(self, in_channels=1, alpha_initializer=mx.init.Constant(0.25), **kwargs): super(PReLU2, self).__init__(**kwargs) with self.name_scope(): self.alpha = self.params.get("alpha", shape=(in_channels,), init=alpha_initializer) def hybrid_forward(self, F, x, alpha): return F.LeakyReLU(x, gamma=alpha, act_type="prelu", name="fwd") def __repr__(self): s = '{name}(in_channels={in_channels})' return s.format( name=self.__class__.__name__, in_channels=self.alpha.shape[0]) class HSigmoid(HybridBlock): """ Approximated sigmoid function, so-called hard-version of sigmoid from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. """ def __init__(self, **kwargs): super(HSigmoid, self).__init__(**kwargs) def hybrid_forward(self, F, x): return F.clip(x + 3.0, 0.0, 6.0, name="relu6") / 6.0 def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) class HSwish(HybridBlock): """ H-Swish activation function from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. """ def __init__(self, **kwargs): super(HSwish, self).__init__(**kwargs) def hybrid_forward(self, F, x): return x * F.clip(x + 3.0, 0.0, 6.0, name="relu6") / 6.0 def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) class Softmax(HybridBlock): """ Softmax activation function. Parameters: ---------- axis : int, default 1 Axis along which to do softmax. """ def __init__(self, axis=1, **kwargs): super(Softmax, self).__init__(**kwargs) self.axis = axis def hybrid_forward(self, F, x): return x.softmax(axis=self.axis) def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) def get_activation_layer(activation): """ Create activation layer from string/function. Parameters: ---------- activation : function, or str, or HybridBlock Activation function or name of activation function. Returns: ------- HybridBlock Activation layer. """ assert (activation is not None) if isfunction(activation): return activation() elif isinstance(activation, str): if activation == "relu6": return ReLU6() elif activation == "swish": return nn.Swish() elif activation == "hswish": return HSwish() elif activation == "hsigmoid": return HSigmoid() else: return nn.Activation(activation) else: assert (isinstance(activation, HybridBlock)) return activation class SelectableDense(HybridBlock): """ Selectable dense layer. Parameters: ---------- in_channels : int Number of input features. out_channels : int Number of output features. use_bias : bool, default False Whether the layer uses a bias vector. dtype : str or np.dtype, default 'float32' Data type of output embeddings. weight_initializer : str or `Initializer` Initializer for the `kernel` weights matrix. bias_initializer: str or `Initializer` Initializer for the bias vector. num_options : int, default 1 Number of selectable options. """ def __init__(self, in_channels, out_channels, use_bias=False, dtype="float32", weight_initializer=None, bias_initializer="zeros", num_options=1, **kwargs): super(SelectableDense, self).__init__(**kwargs) self.in_channels = in_channels self.out_channels = out_channels self.use_bias = use_bias self.num_options = num_options with self.name_scope(): self.weight = self.params.get( "weight", shape=(num_options, out_channels, in_channels), init=weight_initializer, dtype=dtype, allow_deferred_init=True) if use_bias: self.bias = self.params.get( "bias", shape=(num_options, out_channels), init=bias_initializer, dtype=dtype, allow_deferred_init=True) else: self.bias = None def hybrid_forward(self, F, x, indices, weight, bias=None): weight = F.take(weight, indices=indices, axis=0) x = x.expand_dims(axis=-1) x = F.batch_dot(weight, x) x = x.squeeze(axis=-1) if self.use_bias: bias = F.take(bias, indices=indices, axis=0) x += bias return x def __repr__(self): s = "{name}({layout}, {num_options})" shape = self.weight.shape return s.format(name=self.__class__.__name__, layout="{0} -> {1}".format(shape[1] if shape[1] else None, shape[0]), num_options=self.num_options) class BatchNormExtra(nn.BatchNorm): """ Batch normalization layer with extra parameters. """ def __init__(self, **kwargs): has_cudnn_off = ("cudnn_off" in kwargs) if has_cudnn_off: cudnn_off = kwargs["cudnn_off"] del kwargs["cudnn_off"] super(BatchNormExtra, self).__init__(**kwargs) if has_cudnn_off: self._kwargs["cudnn_off"] = cudnn_off class DenseBlock(HybridBlock): """ Standard dense block with Batch normalization and activation. Parameters: ---------- in_channels : int Number of input features. out_channels : int Number of output features. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): super(DenseBlock, self).__init__(**kwargs) self.activate = (activation is not None) self.use_bn = use_bn with self.name_scope(): self.fc = nn.Dense( units=out_channels, use_bias=use_bias, in_units=in_channels) if self.use_bn: self.bn = BatchNormExtra( in_channels=out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) if self.activate: self.activ = get_activation_layer(activation) def hybrid_forward(self, F, x): x = self.fc(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x class ConvBlock1d(HybridBlock): """ Standard 1D convolution block with Batch normalization and activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int Convolution window size. strides : int Strides of the convolution. padding : int Padding value for convolution layer. dilation : int Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): super(ConvBlock1d, self).__init__(**kwargs) self.activate = (activation is not None) self.use_bn = use_bn with self.name_scope(): self.conv = nn.Conv1D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) if self.use_bn: self.bn = BatchNormExtra( in_channels=out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) if self.activate: self.activ = get_activation_layer(activation) def hybrid_forward(self, F, x): x = self.conv(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x def conv1x1(in_channels, out_channels, strides=1, groups=1, use_bias=False): """ Convolution 1x1 layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2D( channels=out_channels, kernel_size=1, strides=strides, groups=groups, use_bias=use_bias, in_channels=in_channels) def conv3x3(in_channels, out_channels, strides=1, padding=1, dilation=1, groups=1, use_bias=False): """ Convolution 3x3 layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2D( channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) def depthwise_conv3x3(channels, strides=1, padding=1, dilation=1, use_bias=False): """ Depthwise convolution 3x3 layer. Parameters: ---------- channels : int Number of input/output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2D( channels=channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, groups=channels, use_bias=use_bias, in_channels=channels) class ConvBlock(HybridBlock): """ Standard convolution block with batch normalization and activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): super(ConvBlock, self).__init__(**kwargs) self.activate = (activation is not None) self.use_bn = use_bn with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) if self.use_bn: self.bn = BatchNormExtra( in_channels=out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) if self.activate: self.activ = get_activation_layer(activation) def hybrid_forward(self, F, x): x = self.conv(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x def conv1x1_block(in_channels, out_channels, strides=1, padding=0, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ 1x1 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 0 Padding value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=padding, groups=groups, use_bias=use_bias, use_bn=use_bn, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) def conv3x3_block(in_channels, out_channels, strides=1, padding=1, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ 3x3 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, use_bn=use_bn, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) def conv5x5_block(in_channels, out_channels, strides=1, padding=2, dilation=1, groups=1, use_bias=False, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ 5x5 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 2 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=5, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) def conv7x7_block(in_channels, out_channels, strides=1, padding=3, use_bias=False, use_bn=True, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ 7x7 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 3 Padding value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=7, strides=strides, padding=padding, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) def dwconv_block(in_channels, out_channels, kernel_size, strides, padding, dilation=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ Depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=out_channels, use_bias=use_bias, use_bn=use_bn, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) def dwconv3x3_block(in_channels, out_channels, strides=1, padding=1, dilation=1, use_bias=False, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ 3x3 depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, use_bias=use_bias, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) def dwconv5x5_block(in_channels, out_channels, strides=1, padding=2, dilation=1, use_bias=False, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): """ 5x5 depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 2 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ return dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=5, strides=strides, padding=padding, dilation=dilation, use_bias=use_bias, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation, **kwargs) class DwsConvBlock(HybridBlock): """ Depthwise separable convolution block with BatchNorms and activations at each convolution layers. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. dw_use_bn : bool, default True Whether to use BatchNorm layer (depthwise convolution block). pw_use_bn : bool, default True Whether to use BatchNorm layer (pointwise convolution block). bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. dw_activation : function or str or None, default nn.Activation('relu') Activation function after the depthwise convolution block. pw_activation : function or str or None, default nn.Activation('relu') Activation function after the pointwise convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, use_bias=False, dw_use_bn=True, pw_use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, dw_activation=(lambda: nn.Activation("relu")), pw_activation=(lambda: nn.Activation("relu")), **kwargs): super(DwsConvBlock, self).__init__(**kwargs) with self.name_scope(): self.dw_conv = dwconv_block( in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, use_bias=use_bias, use_bn=dw_use_bn, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=dw_activation) self.pw_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, use_bias=use_bias, use_bn=pw_use_bn, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=pw_activation) def hybrid_forward(self, F, x): x = self.dw_conv(x) x = self.pw_conv(x) return x def dwsconv3x3_block(in_channels, out_channels, strides=1, padding=1, dilation=1, use_bias=False, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, dw_activation=(lambda: nn.Activation("relu")), pw_activation=(lambda: nn.Activation("relu")), **kwargs): """ 3x3 depthwise separable version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. dw_activation : function or str or None, default nn.Activation('relu') Activation function after the depthwise convolution block. pw_activation : function or str or None, default nn.Activation('relu') Activation function after the pointwise convolution block. """ return DwsConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, use_bias=use_bias, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dw_activation=dw_activation, pw_activation=pw_activation, **kwargs) class PreConvBlock(HybridBlock): """ Convolution block with Batch normalization and ReLU pre-activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. It's used by PreResNet. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_bn=True, bn_use_global_stats=False, return_preact=False, activate=True, **kwargs): super(PreConvBlock, self).__init__(**kwargs) self.return_preact = return_preact self.activate = activate self.use_bn = use_bn with self.name_scope(): if self.use_bn: self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) if self.activate: self.activ = nn.Activation("relu") self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) def hybrid_forward(self, F, x): if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) if self.return_preact: x_pre_activ = x x = self.conv(x) if self.return_preact: return x, x_pre_activ else: return x def pre_conv1x1_block(in_channels, out_channels, strides=1, use_bias=False, use_bn=True, bn_use_global_stats=False, return_preact=False, activate=True): """ 1x1 version of the pre-activated convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. activate : bool, default True Whether activate the convolution block. """ return PreConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats, return_preact=return_preact, activate=activate) def pre_conv3x3_block(in_channels, out_channels, strides=1, padding=1, dilation=1, groups=1, use_bias=False, use_bn=True, bn_use_global_stats=False, return_preact=False, activate=True): """ 3x3 version of the pre-activated convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. return_preact : bool, default False Whether return pre-activation. activate : bool, default True Whether activate the convolution block. """ return PreConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats, return_preact=return_preact, activate=activate) class DeconvBlock(HybridBlock): """ Deconvolution block with batch normalization and activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the deconvolution. padding : int or tuple/list of 2 int Padding value for deconvolution layer. out_padding : int or tuple/list of 2 int Output padding value for deconvolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for deconvolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, out_padding=0, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): super(DeconvBlock, self).__init__(**kwargs) self.activate = (activation is not None) self.use_bn = use_bn with self.name_scope(): self.conv = nn.Conv2DTranspose( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, output_padding=out_padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) if self.use_bn: self.bn = BatchNormExtra( in_channels=out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) if self.activate: self.activ = get_activation_layer(activation) def hybrid_forward(self, F, x): x = self.conv(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x class NormActivation(HybridBlock): """ Activation block with preliminary batch normalization. It's used by itself as the final block in PreResNet. Parameters: ---------- in_channels : int Number of input channels. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), **kwargs): super(NormActivation, self).__init__(**kwargs) with self.name_scope(): self.bn = BatchNormExtra( in_channels=in_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) self.activ = get_activation_layer(activation) def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) return x class InterpolationBlock(HybridBlock): """ Interpolation block. Parameters: ---------- scale_factor : int Multiplier for spatial size. out_size : tuple of 2 int, default None Spatial size of the output tensor for the bilinear interpolation operation. bilinear : bool, default True Whether to use bilinear interpolation. up : bool, default True Whether to upsample or downsample. """ def __init__(self, scale_factor, out_size=None, bilinear=True, up=True, **kwargs): super(InterpolationBlock, self).__init__(**kwargs) self.scale_factor = scale_factor self.out_size = out_size self.bilinear = bilinear self.up = up def hybrid_forward(self, F, x, size=None): if self.bilinear or (size is not None): out_size = self.calc_out_size(x) if size is None else size return F.contrib.BilinearResize2D(x, height=out_size[0], width=out_size[1]) else: return F.UpSampling(x, scale=self.scale_factor, sample_type="nearest") def calc_out_size(self, x): if self.out_size is not None: return self.out_size if self.up: return tuple(s * self.scale_factor for s in x.shape[2:]) else: return tuple(s // self.scale_factor for s in x.shape[2:]) def __repr__(self): s = '{name}(scale_factor={scale_factor}, out_size={out_size}, bilinear={bilinear}, up={up})' return s.format( name=self.__class__.__name__, scale_factor=self.scale_factor, out_size=self.out_size, bilinear=self.bilinear, up=self.up) def calc_flops(self, x): assert (x.shape[0] == 1) if self.bilinear: num_flops = 9 * x.size else: num_flops = 4 * x.size num_macs = 0 return num_flops, num_macs def channel_shuffle(x, groups): """ Channel shuffle operation from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. Parameters: ---------- x : Symbol or NDArray Input tensor. groups : int Number of groups. Returns: ------- Symbol or NDArray Resulted tensor. """ return x.reshape((0, -4, groups, -1, -2)).swapaxes(1, 2).reshape((0, -3, -2)) class ChannelShuffle(HybridBlock): """ Channel shuffle layer. This is a wrapper over the same operation. It is designed to save the number of groups. Parameters: ---------- channels : int Number of channels. groups : int Number of groups. """ def __init__(self, channels, groups, **kwargs): super(ChannelShuffle, self).__init__(**kwargs) assert (channels % groups == 0) self.groups = groups def hybrid_forward(self, F, x): return channel_shuffle(x, self.groups) def __repr__(self): s = "{name}(groups={groups})" return s.format( name=self.__class__.__name__, groups=self.groups) def channel_shuffle2(x, channels_per_group): """ Channel shuffle operation from 'ShuffleNet: An Extremely Efficient Convolutional Neural Network for Mobile Devices,' https://arxiv.org/abs/1707.01083. The alternative version. Parameters: ---------- x : Symbol or NDArray Input tensor. channels_per_group : int Number of channels per group. Returns: ------- Symbol or NDArray Resulted tensor. """ return x.reshape((0, -4, channels_per_group, -1, -2)).swapaxes(1, 2).reshape((0, -3, -2)) class ChannelShuffle2(HybridBlock): """ Channel shuffle layer. This is a wrapper over the same operation. It is designed to save the number of groups. The alternative version. Parameters: ---------- channels : int Number of channels. groups : int Number of groups. """ def __init__(self, channels, groups, **kwargs): super(ChannelShuffle2, self).__init__(**kwargs) assert (channels % groups == 0) self.channels_per_group = channels // groups def hybrid_forward(self, F, x): return channel_shuffle2(x, self.channels_per_group) class SEBlock(HybridBlock): """ Squeeze-and-Excitation block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- channels : int Number of channels. reduction : int, default 16 Squeeze reduction value. mid_channels : int or None, default None Number of middle channels. round_mid : bool, default False Whether to round middle channel number (make divisible by 8). use_conv : bool, default True Whether to convolutional layers instead of fully-connected ones. activation : function, or str, or HybridBlock, default 'relu' Activation function after the first convolution. out_activation : function, or str, or HybridBlock, default 'sigmoid' Activation function after the last convolution. """ def __init__(self, channels, reduction=16, mid_channels=None, round_mid=False, use_conv=True, mid_activation=(lambda: nn.Activation("relu")), out_activation=(lambda: nn.Activation("sigmoid")), **kwargs): super(SEBlock, self).__init__(**kwargs) self.use_conv = use_conv if mid_channels is None: mid_channels = channels // reduction if not round_mid else round_channels(float(channels) / reduction) with self.name_scope(): if use_conv: self.conv1 = conv1x1( in_channels=channels, out_channels=mid_channels, use_bias=True) else: self.fc1 = nn.Dense( in_units=channels, units=mid_channels) self.activ = get_activation_layer(mid_activation) if use_conv: self.conv2 = conv1x1( in_channels=mid_channels, out_channels=channels, use_bias=True) else: self.fc2 = nn.Dense( in_units=mid_channels, units=channels) self.sigmoid = get_activation_layer(out_activation) def hybrid_forward(self, F, x): w = F.contrib.AdaptiveAvgPooling2D(x, output_size=1) if not self.use_conv: w = F.Flatten(w) w = self.conv1(w) if self.use_conv else self.fc1(w) w = self.activ(w) w = self.conv2(w) if self.use_conv else self.fc2(w) w = self.sigmoid(w) if not self.use_conv: w = w.expand_dims(2).expand_dims(3) x = F.broadcast_mul(x, w) return x class SABlock(HybridBlock): """ Split-Attention block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- out_channels : int Number of output channels. groups : int Number of channel groups (cardinality, without radix). radix : int Number of splits within a cardinal group. reduction : int, default 4 Squeeze reduction value. min_channels : int, default 32 Minimal number of squeezed channels. use_conv : bool, default True Whether to convolutional layers instead of fully-connected ones. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, out_channels, groups, radix, reduction=4, min_channels=32, use_conv=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(SABlock, self).__init__(**kwargs) self.groups = groups self.radix = radix self.use_conv = use_conv in_channels = out_channels * radix mid_channels = max(in_channels // reduction, min_channels) with self.name_scope(): if use_conv: self.conv1 = conv1x1( in_channels=out_channels, out_channels=mid_channels, use_bias=True) else: self.fc1 = nn.Dense( in_units=out_channels, units=mid_channels) self.bn = BatchNormExtra( in_channels=mid_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats, cudnn_off=bn_cudnn_off) self.activ = nn.Activation("relu") if use_conv: self.conv2 = conv1x1( in_channels=mid_channels, out_channels=in_channels, use_bias=True) else: self.fc2 = nn.Dense( in_units=mid_channels, units=in_channels) def hybrid_forward(self, F, x): x = x.reshape((0, -4, self.radix, -1, -2)) w = x.sum(axis=1) w = F.contrib.AdaptiveAvgPooling2D(w, output_size=1) if not self.use_conv: w = F.Flatten(w) w = self.conv1(w) if self.use_conv else self.fc1(w) w = self.bn(w) w = self.activ(w) w = self.conv2(w) if self.use_conv else self.fc2(w) w = w.reshape((0, self.groups, self.radix, -1)) w = w.swapaxes(1, 2) w = F.softmax(w, axis=1) w = w.reshape((0, self.radix, -1, 1, 1)) x = F.broadcast_mul(x, w) x = x.sum(axis=1) return x class SAConvBlock(HybridBlock): """ Split-Attention convolution block from 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. radix : int, default 2 Number of splits within a cardinal group. reduction : int, default 4 Squeeze reduction value. min_channels : int, default 32 Minimal number of squeezed channels. use_conv : bool, default True Whether to convolutional layers instead of fully-connected ones. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, use_bn=True, bn_epsilon=1e-5, bn_use_global_stats=False, bn_cudnn_off=False, activation=(lambda: nn.Activation("relu")), radix=2, reduction=4, min_channels=32, use_conv=True, **kwargs): super(SAConvBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = ConvBlock( in_channels=in_channels, out_channels=(out_channels * radix), kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=(groups * radix), use_bias=use_bias, use_bn=use_bn, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=activation) self.att = SABlock( out_channels=out_channels, groups=groups, radix=radix, reduction=reduction, min_channels=min_channels, use_conv=use_conv, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): x = self.conv(x) x = self.att(x) return x def saconv3x3_block(in_channels, out_channels, strides=1, padding=1, **kwargs): """ 3x3 version of the Split-Attention convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. """ return SAConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, **kwargs) class PixelShuffle(HybridBlock): """ Pixel-shuffle operation from 'Real-Time Single Image and Video Super-Resolution Using an Efficient Sub-Pixel Convolutional Neural Network,' https://arxiv.org/abs/1609.05158. Parameters: ---------- scale_factor : int Multiplier for spatial size. in_size : tuple of 2 int Spatial size of the input heatmap tensor. fixed_size : bool Whether to expect fixed spatial size of input image. """ def __init__(self, channels, scale_factor, in_size, fixed_size, **kwargs): super(PixelShuffle, self).__init__(**kwargs) assert (channels % scale_factor % scale_factor == 0) self.channels = channels self.scale_factor = scale_factor self.in_size = in_size self.fixed_size = fixed_size def hybrid_forward(self, F, x): f1 = self.scale_factor f2 = self.scale_factor if not self.fixed_size: x = x.reshape((0, -4, -1, f1 * f2, 0, 0)) x = x.reshape((0, 0, -4, f1, f2, 0, 0)) x = x.transpose((0, 1, 4, 2, 5, 3)) x = x.reshape((0, 0, -3, -3)) else: new_channels = self.channels // f1 // f2 h, w = self.in_size x = x.reshape((0, new_channels, f1 * f2, h, w)) x = x.reshape((0, new_channels, f1, f2, h, w)) x = x.transpose((0, 1, 4, 2, 5, 3)) x = x.reshape((0, new_channels, h * f1, w * f2)) return x class DucBlock(HybridBlock): """ Dense Upsampling Convolution (DUC) block from 'Understanding Convolution for Semantic Segmentation,' https://arxiv.org/abs/1702.08502. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. scale_factor : int Multiplier for spatial size. in_size : tuple of 2 int Spatial size of the input heatmap tensor. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, scale_factor, in_size, fixed_size, bn_use_global_stats, bn_cudnn_off, **kwargs): super(DucBlock, self).__init__(**kwargs) mid_channels = (scale_factor * scale_factor) * out_channels with self.name_scope(): self.conv = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.pix_shuffle = PixelShuffle( channels=mid_channels, scale_factor=scale_factor, in_size=in_size, fixed_size=fixed_size) def hybrid_forward(self, F, x): x = self.conv(x) x = self.pix_shuffle(x) return x def split(x, sizes, axis=1): """ Splits an array along a particular axis into multiple sub-arrays. Parameters: ---------- x : Symbol or NDArray Input tensor. sizes : tuple/list of int Sizes of chunks. axis : int, default 1 Axis along which to split. Returns: ------- Tuple of Symbol or NDArray Resulted tensor. """ x_outs = [] begin = 0 for size in sizes: end = begin + size x_outs += [x.slice_axis(axis=axis, begin=begin, end=end)] begin = end return tuple(x_outs) class IBN(HybridBlock): """ Instance-Batch Normalization block from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- channels : int Number of channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. inst_fraction : float, default 0.5 The first fraction of channels for normalization. inst_first : bool, default True Whether instance normalization be on the first part of channels. """ def __init__(self, channels, bn_use_global_stats=False, first_fraction=0.5, inst_first=True, **kwargs): super(IBN, self).__init__(**kwargs) self.inst_first = inst_first h1_channels = int(math.floor(channels * first_fraction)) h2_channels = channels - h1_channels self.split_sections = [h1_channels, h2_channels] if self.inst_first: self.inst_norm = nn.InstanceNorm( in_channels=h1_channels, scale=True) self.batch_norm = nn.BatchNorm( in_channels=h2_channels, use_global_stats=bn_use_global_stats) else: self.batch_norm = nn.BatchNorm( in_channels=h1_channels, use_global_stats=bn_use_global_stats) self.inst_norm = nn.InstanceNorm( in_channels=h2_channels, scale=True) def hybrid_forward(self, F, x): x1, x2 = split(x, sizes=self.split_sections, axis=1) if self.inst_first: x1 = self.inst_norm(x1) x2 = self.batch_norm(x2) else: x1 = self.batch_norm(x1) x2 = self.inst_norm(x2) x = F.concat(x1, x2, dim=1) return x class DualPathSequential(nn.HybridSequential): """ A sequential container for hybrid blocks with dual inputs/outputs. Blocks will be executed in the order they are added. Parameters: ---------- return_two : bool, default True Whether to return two output after execution. first_ordinals : int, default 0 Number of the first blocks with single input/output. last_ordinals : int, default 0 Number of the final blocks with single input/output. dual_path_scheme : function Scheme of dual path response for a block. dual_path_scheme_ordinal : function Scheme of dual path response for an ordinal block. """ def __init__(self, return_two=True, first_ordinals=0, last_ordinals=0, dual_path_scheme=(lambda block, x1, x2: block(x1, x2)), dual_path_scheme_ordinal=(lambda block, x1, x2: (block(x1), x2)), **kwargs): super(DualPathSequential, self).__init__(**kwargs) self.return_two = return_two self.first_ordinals = first_ordinals self.last_ordinals = last_ordinals self.dual_path_scheme = dual_path_scheme self.dual_path_scheme_ordinal = dual_path_scheme_ordinal def hybrid_forward(self, F, x1, x2=None): length = len(self._children.values()) for i, block in enumerate(self._children.values()): if (i < self.first_ordinals) or (i >= length - self.last_ordinals): x1, x2 = self.dual_path_scheme_ordinal(block, x1, x2) else: x1, x2 = self.dual_path_scheme(block, x1, x2) if self.return_two: return x1, x2 else: return x1 class ParametricSequential(nn.HybridSequential): """ A sequential container for blocks with parameters. Blocks will be executed in the order they are added. """ def __init__(self, **kwargs): super(ParametricSequential, self).__init__(**kwargs) def hybrid_forward(self, F, x, *args, **kwargs): for block in self._children.values(): x = block(x, *args, **kwargs) return x class Concurrent(nn.HybridSequential): """ A container for concatenation of blocks on the base of the sequential container. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. stack : bool, default False Whether to concatenate tensors along a new dimension. merge_type : str, default None Type of branch merging. branches : list of HybridBlock, default None Whether to concatenate tensors along a new dimension. """ def __init__(self, axis=1, stack=False, merge_type=None, branches=None, **kwargs): super(Concurrent, self).__init__(**kwargs) assert (merge_type is None) or (merge_type in ["cat", "stack", "sum"]) self.axis = axis self.stack = stack if merge_type is not None: self.merge_type = merge_type else: self.merge_type = "stack" if stack else "cat" if branches is not None: with self.name_scope(): for branch in branches: self.add(branch) def hybrid_forward(self, F, x): out = [] for block in self._children.values(): out.append(block(x)) if self.merge_type == "stack": out = F.stack(*out, axis=self.axis) elif self.merge_type == "cat": out = F.concat(*out, dim=self.axis) elif self.merge_type == "sum": out = F.stack(*out, axis=self.axis).sum(axis=self.axis) else: raise NotImplementedError() return out class SequentialConcurrent(nn.HybridSequential): """ A sequential container with concatenated outputs. Blocks will be executed in the order they are added. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. stack : bool, default False Whether to concatenate tensors along a new dimension. cat_input : bool, default True Whether to concatenate input tensor. """ def __init__(self, axis=1, stack=False, cat_input=True, **kwargs): super(SequentialConcurrent, self).__init__(**kwargs) self.axis = axis self.stack = stack self.cat_input = cat_input def hybrid_forward(self, F, x): out = [x] if self.cat_input else [] for block in self._children.values(): x = block(x) out.append(x) if self.stack: out = F.stack(*out, axis=self.axis) else: out = F.concat(*out, dim=self.axis) return out class ParametricConcurrent(nn.HybridSequential): """ A container for concatenation of blocks with parameters. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. """ def __init__(self, axis=1, **kwargs): super(ParametricConcurrent, self).__init__(**kwargs) self.axis = axis def hybrid_forward(self, F, x, *args, **kwargs): out = [] for block in self._children.values(): out.append(block(x, *args, **kwargs)) out = F.concat(*out, dim=self.axis) return out class Hourglass(HybridBlock): """ A hourglass block. Parameters: ---------- down_seq : nn.HybridSequential Down modules as sequential. up_seq : nn.HybridSequential Up modules as sequential. skip_seq : nn.HybridSequential Skip connection modules as sequential. merge_type : str, default 'add' Type of concatenation of up and skip outputs. return_first_skip : bool, default False Whether return the first skip connection output. Used in ResAttNet. """ def __init__(self, down_seq, up_seq, skip_seq, merge_type="add", return_first_skip=False, **kwargs): super(Hourglass, self).__init__(**kwargs) self.depth = len(down_seq) assert (merge_type in ["cat", "add"]) assert (len(up_seq) == self.depth) assert (len(skip_seq) in (self.depth, self.depth + 1)) self.merge_type = merge_type self.return_first_skip = return_first_skip self.extra_skip = (len(skip_seq) == self.depth + 1) with self.name_scope(): self.down_seq = down_seq self.up_seq = up_seq self.skip_seq = skip_seq def _merge(self, F, x, y): if y is not None: if self.merge_type == "cat": x = F.concat(x, y, dim=1) elif self.merge_type == "add": x = x + y return x def hybrid_forward(self, F, x): y = None down_outs = [x] for down_module in self.down_seq._children.values(): x = down_module(x) down_outs.append(x) for i in range(len(down_outs)): if i != 0: y = down_outs[self.depth - i] skip_module = self.skip_seq[self.depth - i] y = skip_module(y) x = self._merge(F, x, y) if i != len(down_outs) - 1: if (i == 0) and self.extra_skip: skip_module = self.skip_seq[self.depth] x = skip_module(x) up_module = self.up_seq[self.depth - 1 - i] x = up_module(x) if self.return_first_skip: return x, y else: return x class SesquialteralHourglass(HybridBlock): """ A sesquialteral hourglass block. Parameters: ---------- down1_seq : nn.Sequential The first down modules as sequential. skip1_seq : nn.Sequential The first skip connection modules as sequential. up_seq : nn.Sequential Up modules as sequential. skip2_seq : nn.Sequential The second skip connection modules as sequential. down2_seq : nn.Sequential The second down modules as sequential. merge_type : str, default 'cat' Type of concatenation of up and skip outputs. """ def __init__(self, down1_seq, skip1_seq, up_seq, skip2_seq, down2_seq, merge_type="cat", **kwargs): super(SesquialteralHourglass, self).__init__(**kwargs) assert (len(down1_seq) == len(up_seq)) assert (len(down1_seq) == len(down2_seq)) assert (len(skip1_seq) == len(skip2_seq)) assert (len(down1_seq) == len(skip1_seq) - 1) assert (merge_type in ["cat", "add"]) self.merge_type = merge_type self.depth = len(down1_seq) with self.name_scope(): self.down1_seq = down1_seq self.skip1_seq = skip1_seq self.up_seq = up_seq self.skip2_seq = skip2_seq self.down2_seq = down2_seq def _merge(self, F, x, y): if y is not None: if self.merge_type == "cat": x = F.concat(x, y, dim=1) elif self.merge_type == "add": x = x + y return x def hybrid_forward(self, F, x): y = self.skip1_seq[0](x) skip1_outs = [y] for i in range(self.depth): x = self.down1_seq[i](x) y = self.skip1_seq[i + 1](x) skip1_outs.append(y) x = skip1_outs[self.depth] y = self.skip2_seq[0](x) skip2_outs = [y] for i in range(self.depth): x = self.up_seq[i](x) y = skip1_outs[self.depth - 1 - i] x = self._merge(F, x, y) y = self.skip2_seq[i + 1](x) skip2_outs.append(y) x = self.skip2_seq[self.depth](x) for i in range(self.depth): x = self.down2_seq[i](x) y = skip2_outs[self.depth - 1 - i] x = self._merge(F, x, y) return x class MultiOutputSequential(nn.HybridSequential): """ A sequential container with multiple outputs. Blocks will be executed in the order they are added. Parameters: ---------- multi_output : bool, default True Whether to return multiple output. dual_output : bool, default False Whether to return dual output. return_last : bool, default True Whether to forcibly return last value. """ def __init__(self, multi_output=True, dual_output=False, return_last=True, **kwargs): super(MultiOutputSequential, self).__init__(**kwargs) self.multi_output = multi_output self.dual_output = dual_output self.return_last = return_last def hybrid_forward(self, F, x): outs = [] for block in self._children.values(): x = block(x) if hasattr(block, "do_output") and block.do_output: outs.append(x) elif hasattr(block, "do_output2") and block.do_output2: assert (type(x) == tuple) outs.extend(x[1]) x = x[0] if self.multi_output: return [x] + outs if self.return_last else outs elif self.dual_output: return x, outs else: return x class ParallelConcurent(nn.HybridSequential): """ A sequential container with multiple inputs and multiple outputs. Modules will be executed in the order they are added. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. merge_type : str, default 'list' Type of branch merging. """ def __init__(self, axis=1, merge_type="list", **kwargs): super(ParallelConcurent, self).__init__(**kwargs) assert (merge_type is None) or (merge_type in ["list", "cat", "stack", "sum"]) self.axis = axis self.merge_type = merge_type def hybrid_forward(self, F, x): out = [] for block, xi in zip(self._children.values(), x): out.append(block(xi)) if self.merge_type == "list": pass elif self.merge_type == "stack": out = F.stack(*out, axis=self.axis) elif self.merge_type == "cat": out = F.concat(*out, dim=self.axis) elif self.merge_type == "sum": out = F.stack(*out, axis=self.axis).sum(axis=self.axis) else: raise NotImplementedError() return out class DualPathParallelConcurent(nn.HybridSequential): """ A sequential container with multiple dual-path inputs and single/multiple outputs. Blocks will be executed in the order they are added. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. merge_type : str, default 'list' Type of branch merging. """ def __init__(self, axis=1, merge_type="list", **kwargs): super(DualPathParallelConcurent, self).__init__(**kwargs) assert (merge_type is None) or (merge_type in ["list", "cat", "stack", "sum"]) self.axis = axis self.merge_type = merge_type def hybrid_forward(self, F, x1, x2): x1_out = [] x2_out = [] for block, x1i, x2i in zip(self._children.values(), x1, x2): y1i, y2i = block(x1i, x2i) x1_out.append(y1i) x2_out.append(y2i) if self.merge_type == "list": pass elif self.merge_type == "stack": x1_out = F.stack(*x1_out, axis=self.axis) x2_out = F.stack(*x2_out, axis=self.axis) elif self.merge_type == "cat": x1_out = F.concat(*x1_out, dim=self.axis) x2_out = F.concat(*x2_out, dim=self.axis) elif self.merge_type == "sum": x1_out = F.stack(*x1_out, axis=self.axis).sum(axis=self.axis) x2_out = F.stack(*x2_out, axis=self.axis).sum(axis=self.axis) else: raise NotImplementedError() return x1_out, x2_out class HeatmapMaxDetBlock(HybridBlock): """ Heatmap maximum detector block (for human pose estimation task). Parameters: ---------- channels : int Number of channels. in_size : tuple of 2 int Spatial size of the input heatmap tensor. fixed_size : bool Whether to expect fixed spatial size of input image. tune : bool, default True Whether to tune point positions. """ def __init__(self, channels, in_size, fixed_size, tune=True, **kwargs): super(HeatmapMaxDetBlock, self).__init__(**kwargs) self.channels = channels self.in_size = in_size self.fixed_size = fixed_size self.tune = tune def hybrid_forward(self, F, x): # assert (not self.fixed_size) or (self.in_size == x.shape[2:]) vector_dim = 2 in_size = self.in_size if self.fixed_size else x.shape[2:] heatmap_vector = x.reshape((0, 0, -3)) indices = heatmap_vector.argmax(axis=vector_dim, keepdims=True) scores = heatmap_vector.max(axis=vector_dim, keepdims=True) scores_mask = (scores > 0.0) pts_x = (indices % in_size[1]) * scores_mask pts_y = (indices / in_size[1]).floor() * scores_mask pts = F.concat(pts_x, pts_y, scores, dim=vector_dim) if self.tune: batch = x.shape[0] for b in range(batch): for k in range(self.channels): hm = x[b, k, :, :] px = int(pts[b, k, 0].asscalar()) py = int(pts[b, k, 1].asscalar()) if (0 < px < in_size[1] - 1) and (0 < py < in_size[0] - 1): pts[b, k, 0] += (hm[py, px + 1] - hm[py, px - 1]).sign() * 0.25 pts[b, k, 1] += (hm[py + 1, px] - hm[py - 1, px]).sign() * 0.25 return pts def __repr__(self): s = "{name}(channels={channels}, in_size={in_size}, fixed_size={fixed_size})" return s.format( name=self.__class__.__name__, channels=self.channels, in_size=self.in_size, fixed_size=self.fixed_size) def calc_flops(self, x): assert (x.shape[0] == 1) num_flops = x.size + 26 * self.channels num_macs = 0 return num_flops, num_macs
84,759
32.501976
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/lwopenpose_cmupan.py
""" Lightweight OpenPose 2D/3D for CMU Panoptic, implemented in Gluon. Original paper: 'Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose,' https://arxiv.org/abs/1811.12004. """ __all__ = ['LwOpenPose', 'lwopenpose2d_mobilenet_cmupan_coco', 'lwopenpose3d_mobilenet_cmupan_coco', 'LwopDecoderFinalBlock'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv1x1_block, conv3x3_block, dwsconv3x3_block class LwopResBottleneck(HybridBlock): """ Bottleneck block for residual path in the residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. use_bias : bool, default True Whether the layer uses a bias vector. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck_factor : int, default 2 Bottleneck factor. squeeze_out : bool, default False Whether to squeeze the output channels. """ def __init__(self, in_channels, out_channels, strides, use_bias=True, bn_use_global_stats=False, bottleneck_factor=2, squeeze_out=False, **kwargs): super(LwopResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // bottleneck_factor if squeeze_out else in_channels // bottleneck_factor with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_bias=use_bias, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, use_bias=use_bias, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, use_bias=use_bias, activation=None, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class LwopResUnit(HybridBlock): """ ResNet-like residual unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. use_bias : bool, default True Whether the layer uses a bias vector. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck_factor : int, default 2 Bottleneck factor. squeeze_out : bool, default False Whether to squeeze the output channels. activate : bool, default False Whether to activate the sum. """ def __init__(self, in_channels, out_channels, strides=1, use_bias=True, bn_use_global_stats=False, bottleneck_factor=2, squeeze_out=False, activate=False, **kwargs): super(LwopResUnit, self).__init__(**kwargs) self.activate = activate self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = LwopResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, use_bias=use_bias, bn_use_global_stats=bn_use_global_stats, bottleneck_factor=bottleneck_factor, squeeze_out=squeeze_out) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, use_bias=use_bias, bn_use_global_stats=bn_use_global_stats, activation=None) if self.activate: self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity if self.activate: x = self.activ(x) return x class LwopEncoderFinalBlock(HybridBlock): """ Lightweight OpenPose 2D/3D specific encoder final block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, **kwargs): super(LwopEncoderFinalBlock, self).__init__(**kwargs) with self.name_scope(): self.pre_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, use_bias=True, use_bn=False, bn_use_global_stats=bn_use_global_stats) self.body = nn.HybridSequential(prefix="") for i in range(3): self.body.add(dwsconv3x3_block( in_channels=out_channels, out_channels=out_channels, dw_use_bn=False, pw_use_bn=False, bn_use_global_stats=bn_use_global_stats, dw_activation=(lambda: nn.ELU()), pw_activation=(lambda: nn.ELU()))) self.post_conv = conv3x3_block( in_channels=out_channels, out_channels=out_channels, use_bias=True, use_bn=False, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.pre_conv(x) x = x + self.body(x) x = self.post_conv(x) return x class LwopRefinementBlock(HybridBlock): """ Lightweight OpenPose 2D/3D specific refinement block for decoder units. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, **kwargs): super(LwopRefinementBlock, self).__init__(**kwargs) with self.name_scope(): self.pre_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, use_bias=True, use_bn=False, bn_use_global_stats=bn_use_global_stats) self.body = nn.HybridSequential(prefix="") self.body.add(conv3x3_block( in_channels=out_channels, out_channels=out_channels, use_bias=True, bn_use_global_stats=bn_use_global_stats)) self.body.add(conv3x3_block( in_channels=out_channels, out_channels=out_channels, padding=2, dilation=2, use_bias=True, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.pre_conv(x) x = x + self.body(x) return x class LwopDecoderBend(HybridBlock): """ Lightweight OpenPose 2D/3D specific decoder bend block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, mid_channels, out_channels, bn_use_global_stats=False, **kwargs): super(LwopDecoderBend, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_bias=True, use_bn=False, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, use_bias=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class LwopDecoderInitBlock(HybridBlock): """ Lightweight OpenPose 2D/3D specific decoder init block. Parameters: ---------- in_channels : int Number of input channels. keypoints : int Number of keypoints. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, keypoints, bn_use_global_stats=False, **kwargs): super(LwopDecoderInitBlock, self).__init__(**kwargs) num_heatmap = keypoints num_paf = 2 * keypoints bend_mid_channels = 512 with self.name_scope(): self.body = nn.HybridSequential(prefix="") for i in range(3): self.body.add(conv3x3_block( in_channels=in_channels, out_channels=in_channels, use_bias=True, use_bn=False, bn_use_global_stats=bn_use_global_stats)) self.heatmap_bend = LwopDecoderBend( in_channels=in_channels, mid_channels=bend_mid_channels, out_channels=num_heatmap, bn_use_global_stats=bn_use_global_stats) self.paf_bend = LwopDecoderBend( in_channels=in_channels, mid_channels=bend_mid_channels, out_channels=num_paf, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): y = self.body(x) heatmap = self.heatmap_bend(y) paf = self.paf_bend(y) y = F.concat(x, heatmap, paf, dim=1) return y class LwopDecoderUnit(HybridBlock): """ Lightweight OpenPose 2D/3D specific decoder init. Parameters: ---------- in_channels : int Number of input channels. keypoints : int Number of keypoints. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, keypoints, bn_use_global_stats=False, **kwargs): super(LwopDecoderUnit, self).__init__(**kwargs) num_heatmap = keypoints num_paf = 2 * keypoints self.features_channels = in_channels - num_heatmap - num_paf with self.name_scope(): self.body = nn.HybridSequential(prefix="") for i in range(5): self.body.add(LwopRefinementBlock( in_channels=in_channels, out_channels=self.features_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = self.features_channels self.heatmap_bend = LwopDecoderBend( in_channels=self.features_channels, mid_channels=self.features_channels, out_channels=num_heatmap, bn_use_global_stats=bn_use_global_stats) self.paf_bend = LwopDecoderBend( in_channels=self.features_channels, mid_channels=self.features_channels, out_channels=num_paf, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): features = F.slice_axis(x, axis=1, begin=0, end=self.features_channels) y = self.body(x) heatmap = self.heatmap_bend(y) paf = self.paf_bend(y) y = F.concat(features, heatmap, paf, dim=1) return y class LwopDecoderFeaturesBend(HybridBlock): """ Lightweight OpenPose 2D/3D specific decoder 3D features bend. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, mid_channels, out_channels, bn_use_global_stats=False, **kwargs): super(LwopDecoderFeaturesBend, self).__init__(**kwargs) with self.name_scope(): self.body = nn.HybridSequential(prefix="") for i in range(2): self.body.add(LwopRefinementBlock( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = mid_channels self.features_bend = LwopDecoderBend( in_channels=mid_channels, mid_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.body(x) x = self.features_bend(x) return x class LwopDecoderFinalBlock(HybridBlock): """ Lightweight OpenPose 2D/3D specific decoder final block for calcualation 3D poses. Parameters: ---------- in_channels : int Number of input channels. keypoints : int Number of keypoints. bottleneck_factor : int Bottleneck factor. calc_3d_features : bool Whether to calculate 3D features. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, keypoints, bottleneck_factor, calc_3d_features, bn_use_global_stats=False, **kwargs): super(LwopDecoderFinalBlock, self).__init__(**kwargs) self.num_heatmap_paf = 3 * keypoints self.calc_3d_features = calc_3d_features features_out_channels = self.num_heatmap_paf features_in_channels = in_channels - features_out_channels if self.calc_3d_features: with self.name_scope(): self.body = nn.HybridSequential(prefix="") for i in range(5): self.body.add(LwopResUnit( in_channels=in_channels, out_channels=features_in_channels, bottleneck_factor=bottleneck_factor, bn_use_global_stats=bn_use_global_stats)) in_channels = features_in_channels self.features_bend = LwopDecoderFeaturesBend( in_channels=features_in_channels, mid_channels=features_in_channels, out_channels=features_out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): heatmap_paf_2d = F.slice_axis(x, axis=1, begin=-self.num_heatmap_paf, end=None) if not self.calc_3d_features: return heatmap_paf_2d x = self.body(x) x = self.features_bend(x) y = F.concat(heatmap_paf_2d, x, dim=1) return y class LwOpenPose(HybridBlock): """ Lightweight OpenPose 2D/3D model from 'Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose,' https://arxiv.org/abs/1811.12004. Parameters: ---------- encoder_channels : list of list of int Number of output channels for each encoder unit. encoder_paddings : list of list of int Padding/dilation value for each encoder unit. encoder_init_block_channels : int Number of output channels for the encoder initial unit. encoder_final_block_channels : int Number of output channels for the encoder final unit. refinement_units : int Number of refinement blocks in the decoder. calc_3d_features : bool Whether to calculate 3D features. return_heatmap : bool, default True Whether to return only heatmap. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (256, 192) Spatial size of the expected input image. keypoints : int, default 19 Number of keypoints. """ def __init__(self, encoder_channels, encoder_paddings, encoder_init_block_channels, encoder_final_block_channels, refinement_units, calc_3d_features, return_heatmap=True, bn_use_global_stats=False, in_channels=3, in_size=(368, 368), keypoints=19, **kwargs): super(LwOpenPose, self).__init__(**kwargs) assert (in_channels == 3) self.in_size = in_size self.keypoints = keypoints self.return_heatmap = return_heatmap self.calc_3d_features = calc_3d_features num_heatmap_paf = 3 * keypoints with self.name_scope(): self.encoder = nn.HybridSequential(prefix="") backbone = nn.HybridSequential(prefix="") backbone.add(conv3x3_block( in_channels=in_channels, out_channels=encoder_init_block_channels, strides=2, bn_use_global_stats=bn_use_global_stats)) in_channels = encoder_init_block_channels for i, channels_per_stage in enumerate(encoder_channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 padding = encoder_paddings[i][j] stage.add(dwsconv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=strides, padding=padding, dilation=padding, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels backbone.add(stage) self.encoder.add(backbone) self.encoder.add(LwopEncoderFinalBlock( in_channels=in_channels, out_channels=encoder_final_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = encoder_final_block_channels self.decoder = nn.HybridSequential(prefix="") self.decoder.add(LwopDecoderInitBlock( in_channels=in_channels, keypoints=keypoints, bn_use_global_stats=bn_use_global_stats)) in_channels = encoder_final_block_channels + num_heatmap_paf for i in range(refinement_units): self.decoder.add(LwopDecoderUnit( in_channels=in_channels, keypoints=keypoints, bn_use_global_stats=bn_use_global_stats)) self.decoder.add(LwopDecoderFinalBlock( in_channels=in_channels, keypoints=keypoints, bottleneck_factor=2, calc_3d_features=calc_3d_features, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.encoder(x) x = self.decoder(x) if self.return_heatmap: return x else: return x def get_lwopenpose(calc_3d_features, keypoints, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Lightweight OpenPose 2D/3D model with specific parameters. Parameters: ---------- calc_3d_features : bool, default False Whether to calculate 3D features. keypoints : int Number of keypoints. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ encoder_channels = [[64], [128, 128], [256, 256, 512, 512, 512, 512, 512, 512]] encoder_paddings = [[1], [1, 1], [1, 1, 1, 2, 1, 1, 1, 1]] encoder_init_block_channels = 32 encoder_final_block_channels = 128 refinement_units = 1 net = LwOpenPose( encoder_channels=encoder_channels, encoder_paddings=encoder_paddings, encoder_init_block_channels=encoder_init_block_channels, encoder_final_block_channels=encoder_final_block_channels, refinement_units=refinement_units, calc_3d_features=calc_3d_features, keypoints=keypoints, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def lwopenpose2d_mobilenet_cmupan_coco(keypoints=19, **kwargs): """ Lightweight OpenPose 2D model on the base of MobileNet for CMU Panoptic from 'Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose,' https://arxiv.org/abs/1811.12004. Parameters: ---------- keypoints : int, default 19 Number of keypoints. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_lwopenpose(calc_3d_features=False, keypoints=keypoints, model_name="lwopenpose2d_mobilenet_cmupan_coco", **kwargs) def lwopenpose3d_mobilenet_cmupan_coco(keypoints=19, **kwargs): """ Lightweight OpenPose 3D model on the base of MobileNet for CMU Panoptic from 'Real-time 2D Multi-Person Pose Estimation on CPU: Lightweight OpenPose,' https://arxiv.org/abs/1811.12004. Parameters: ---------- keypoints : int, default 19 Number of keypoints. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_lwopenpose(calc_3d_features=True, keypoints=keypoints, model_name="lwopenpose3d_mobilenet_cmupan_coco", **kwargs) def _test(): import numpy as np import mxnet as mx in_size = (368, 368) keypoints = 19 return_heatmap = True pretrained = False models = [ (lwopenpose2d_mobilenet_cmupan_coco, "2d"), (lwopenpose3d_mobilenet_cmupan_coco, "3d"), ] for model, model_dim in models: net = model(pretrained=pretrained, in_size=in_size, return_heatmap=return_heatmap) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != lwopenpose2d_mobilenet_cmupan_coco or weight_count == 4091698) assert (model != lwopenpose3d_mobilenet_cmupan_coco or weight_count == 5085983) batch = 14 x = mx.nd.random.normal(shape=(batch, 3, in_size[0], in_size[1]), ctx=ctx) y = net(x) if model_dim == "2d": assert (y.shape == (batch, 3 * keypoints, in_size[0] // 8, in_size[0] // 8)) else: assert (y.shape == (batch, 6 * keypoints, in_size[0] // 8, in_size[0] // 8)) if __name__ == "__main__": _test()
26,308
35.138736
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/rir_cifar.py
""" RiR for CIFAR/SVHN, implemented in Gluon. Original paper: 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. """ __all__ = ['CIFARRiR', 'rir_cifar10', 'rir_cifar100', 'rir_svhn', 'RiRFinalBlock'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv3x3, conv1x1_block, conv3x3_block, DualPathSequential class PostActivation(HybridBlock): """ Pure pre-activation block without convolution layer. Parameters: ---------- in_channels : int Number of input channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, bn_use_global_stats, **kwargs): super(PostActivation, self).__init__(**kwargs) with self.name_scope(): self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) return x class RiRUnit(HybridBlock): """ RiR unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, **kwargs): super(RiRUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.res_pass_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, strides=strides) self.trans_pass_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, strides=strides) self.res_cross_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, strides=strides) self.trans_cross_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, strides=strides) self.res_postactiv = PostActivation( in_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.trans_postactiv = PostActivation( in_channels=out_channels, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, strides=strides) def hybrid_forward(self, F, x_res, x_trans): if self.resize_identity: x_res_identity = self.identity_conv(x_res) else: x_res_identity = x_res y_res = self.res_cross_conv(x_res) y_trans = self.trans_cross_conv(x_trans) x_res = self.res_pass_conv(x_res) x_trans = self.trans_pass_conv(x_trans) x_res = x_res + x_res_identity + y_trans x_trans = x_trans + y_res x_res = self.res_postactiv(x_res) x_trans = self.trans_postactiv(x_trans) return x_res, x_trans class RiRInitBlock(HybridBlock): """ RiR initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(RiRInitBlock, self).__init__(**kwargs) with self.name_scope(): self.res_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.trans_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x, _): x_res = self.res_conv(x) x_trans = self.trans_conv(x) return x_res, x_trans class RiRFinalBlock(HybridBlock): """ RiR final block. """ def __init__(self): super(RiRFinalBlock, self).__init__() def hybrid_forward(self, F, x_res, x_trans): x = F.concat(x_res, x_trans, dim=1) return x, None class CIFARRiR(HybridBlock): """ RiR model for CIFAR from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. final_block_channels : int Number of output channels for the final unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, bn_use_global_stats=False, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARRiR, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = DualPathSequential( return_two=False, first_ordinals=0, last_ordinals=0, prefix="") self.features.add(RiRInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential(prefix="stage{}_".format(i + 1)) for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(RiRUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(RiRFinalBlock()) in_channels = final_block_channels self.output = nn.HybridSequential(prefix="") self.output.add(conv1x1_block( in_channels=in_channels, out_channels=classes, bn_use_global_stats=bn_use_global_stats, activation=None)) self.output.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output.add(nn.Flatten()) def hybrid_forward(self, F, x): x = self.features(x, x) x = self.output(x) return x def get_rir_cifar(classes, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create RiR model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ channels = [[48, 48, 48, 48], [96, 96, 96, 96, 96, 96], [192, 192, 192, 192, 192, 192]] init_block_channels = 48 final_block_channels = 384 net = CIFARRiR( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, classes=classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def rir_cifar10(classes=10, **kwargs): """ RiR model for CIFAR-10 from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_rir_cifar(classes=classes, model_name="rir_cifar10", **kwargs) def rir_cifar100(classes=100, **kwargs): """ RiR model for CIFAR-100 from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_rir_cifar(classes=classes, model_name="rir_cifar100", **kwargs) def rir_svhn(classes=10, **kwargs): """ RiR model for SVHN from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_rir_cifar(classes=classes, model_name="rir_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (rir_cifar10, 10), (rir_cifar100, 100), (rir_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != rir_cifar10 or weight_count == 9492980) assert (model != rir_cifar100 or weight_count == 9527720) assert (model != rir_svhn or weight_count == 9492980) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
12,481
31.761155
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/diapreresnet.py
""" DIA-PreResNet for ImageNet-1K, implemented in Gluon. Original papers: 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. """ __all__ = ['DIAPreResNet', 'diapreresnet10', 'diapreresnet12', 'diapreresnet14', 'diapreresnetbc14b', 'diapreresnet16', 'diapreresnet18', 'diapreresnet26', 'diapreresnetbc26b', 'diapreresnet34', 'diapreresnetbc38b', 'diapreresnet50', 'diapreresnet50b', 'diapreresnet101', 'diapreresnet101b', 'diapreresnet152', 'diapreresnet152b', 'diapreresnet200', 'diapreresnet200b', 'diapreresnet269b', 'DIAPreResUnit'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, DualPathSequential from .preresnet import PreResBlock, PreResBottleneck, PreResInitBlock, PreResActivation from .diaresnet import DIAAttention class DIAPreResUnit(HybridBlock): """ DIA-PreResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer of the block. attention : nn.Module, default None Attention module. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, bottleneck, conv1_stride, attention=None, **kwargs): super(DIAPreResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = PreResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, conv1_stride=conv1_stride) else: self.body = PreResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, strides=strides) self.attention = attention def hybrid_forward(self, F, x, hc=None): identity = x x, x_pre_activ = self.body(x) if self.resize_identity: identity = self.identity_conv(x_pre_activ) x, hc = self.attention(x, hc) x = x + identity return x, hc class DIAPreResNet(HybridBlock): """ DIA-PreResNet model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(DIAPreResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential( return_two=False, prefix="stage{}_".format(i + 1)) attention = DIAAttention( in_x_features=channels_per_stage[0], in_h_features=channels_per_stage[0]) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(DIAPreResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=conv1_stride, attention=attention)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_diapreresnet(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DIA-PreResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. width_scale : float, default 1.0 Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif (blocks == 38) and bottleneck: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] elif blocks == 269: layers = [3, 30, 48, 8] else: raise ValueError("Unsupported DIA-PreResNet with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1.0: channels = [[int(cij * width_scale) if (i != len(channels) - 1) or (j != len(ci) - 1) else cij for j, cij in enumerate(ci)] for i, ci in enumerate(channels)] init_block_channels = int(init_block_channels * width_scale) net = DIAPreResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def diapreresnet10(**kwargs): """ DIA-PreResNet-10 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=10, model_name="diapreresnet10", **kwargs) def diapreresnet12(**kwargs): """ DIA-PreResNet-12 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=12, model_name="diapreresnet12", **kwargs) def diapreresnet14(**kwargs): """ DIA-PreResNet-14 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=14, model_name="diapreresnet14", **kwargs) def diapreresnetbc14b(**kwargs): """ DIA-PreResNet-BC-14b model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=14, bottleneck=True, conv1_stride=False, model_name="diapreresnetbc14b", **kwargs) def diapreresnet16(**kwargs): """ DIA-PreResNet-16 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=16, model_name="diapreresnet16", **kwargs) def diapreresnet18(**kwargs): """ DIA-PreResNet-18 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=18, model_name="diapreresnet18", **kwargs) def diapreresnet26(**kwargs): """ DIA-PreResNet-26 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=26, bottleneck=False, model_name="diapreresnet26", **kwargs) def diapreresnetbc26b(**kwargs): """ DIA-PreResNet-BC-26b model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="diapreresnetbc26b", **kwargs) def diapreresnet34(**kwargs): """ DIA-PreResNet-34 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=34, model_name="diapreresnet34", **kwargs) def diapreresnetbc38b(**kwargs): """ DIA-PreResNet-BC-38b model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=38, bottleneck=True, conv1_stride=False, model_name="diapreresnetbc38b", **kwargs) def diapreresnet50(**kwargs): """ DIA-PreResNet-50 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=50, model_name="diapreresnet50", **kwargs) def diapreresnet50b(**kwargs): """ DIA-PreResNet-50 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=50, conv1_stride=False, model_name="diapreresnet50b", **kwargs) def diapreresnet101(**kwargs): """ DIA-PreResNet-101 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=101, model_name="diapreresnet101", **kwargs) def diapreresnet101b(**kwargs): """ DIA-PreResNet-101 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=101, conv1_stride=False, model_name="diapreresnet101b", **kwargs) def diapreresnet152(**kwargs): """ DIA-PreResNet-152 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=152, model_name="diapreresnet152", **kwargs) def diapreresnet152b(**kwargs): """ DIA-PreResNet-152 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=152, conv1_stride=False, model_name="diapreresnet152b", **kwargs) def diapreresnet200(**kwargs): """ DIA-PreResNet-200 model from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=200, model_name="diapreresnet200", **kwargs) def diapreresnet200b(**kwargs): """ DIA-PreResNet-200 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=200, conv1_stride=False, model_name="diapreresnet200b", **kwargs) def diapreresnet269b(**kwargs): """ DIA-PreResNet-269 model with stride at the second convolution in bottleneck block from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_diapreresnet(blocks=269, conv1_stride=False, model_name="diapreresnet269b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ diapreresnet10, diapreresnet12, diapreresnet14, diapreresnetbc14b, diapreresnet16, diapreresnet18, diapreresnet26, diapreresnetbc26b, diapreresnet34, diapreresnetbc38b, diapreresnet50, diapreresnet50b, diapreresnet101, diapreresnet101b, diapreresnet152, diapreresnet152b, diapreresnet200, diapreresnet200b, diapreresnet269b, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != diapreresnet10 or weight_count == 6295688) assert (model != diapreresnet12 or weight_count == 6369672) assert (model != diapreresnet14 or weight_count == 6665096) assert (model != diapreresnetbc14b or weight_count == 24016424) assert (model != diapreresnet16 or weight_count == 7845768) assert (model != diapreresnet18 or weight_count == 12566408) assert (model != diapreresnet26 or weight_count == 18837128) assert (model != diapreresnetbc26b or weight_count == 29946664) assert (model != diapreresnet34 or weight_count == 22674568) assert (model != diapreresnetbc38b or weight_count == 35876904) assert (model != diapreresnet50 or weight_count == 39508520) assert (model != diapreresnet50b or weight_count == 39508520) assert (model != diapreresnet101 or weight_count == 58500648) assert (model != diapreresnet101b or weight_count == 58500648) assert (model != diapreresnet152 or weight_count == 74144296) assert (model != diapreresnet152b or weight_count == 74144296) assert (model != diapreresnet200 or weight_count == 78625320) assert (model != diapreresnet200b or weight_count == 78625320) assert (model != diapreresnet269b or weight_count == 116024872) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
23,989
35.293495
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/jasperdr.py
""" Jasper DR (Dense Residual) for ASR, implemented in Gluon. Original paper: 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. """ __all__ = ['jasperdr10x5_en', 'jasperdr10x5_en_nr'] from .jasper import get_jasper def jasperdr10x5_en(classes=29, **kwargs): """ Jasper DR 10x5 model for English language from 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. Parameters: ---------- classes : int, default 29 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_jasper(classes=classes, version=("jasper", "10x5"), use_dr=True, model_name="jasperdr10x5_en", **kwargs) def jasperdr10x5_en_nr(classes=29, **kwargs): """ Jasper DR 10x5 model for English language (with presence of noise) from 'Jasper: An End-to-End Convolutional Neural Acoustic Model,' https://arxiv.org/abs/1904.03288. Parameters: ---------- classes : int, default 29 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_jasper(classes=classes, version=("jasper", "10x5"), use_dr=True, model_name="jasperdr10x5_en_nr", **kwargs) def _calc_width(net): import numpy as np net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) return weight_count def _test(): import numpy as np import mxnet as mx pretrained = False audio_features = 64 models = [ jasperdr10x5_en, jasperdr10x5_en_nr, ] for model in models: net = model( in_channels=audio_features, pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != jasperdr10x5_en or weight_count == 332632349) assert (model != jasperdr10x5_en_nr or weight_count == 332632349) batch = 3 seq_len = np.random.randint(60, 150, batch) seq_len_max = seq_len.max() + 2 x = mx.nd.random.normal(shape=(batch, audio_features, seq_len_max), ctx=ctx) x_len = mx.nd.array(seq_len, ctx=ctx, dtype=np.long) y, y_len = net(x, x_len) assert (y.shape[:2] == (batch, net.classes)) assert (y.shape[2] in [seq_len_max // 2, seq_len_max // 2 + 1]) if __name__ == "__main__": _test()
3,244
30.504854
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/deeplabv3.py
""" DeepLabv3 for image segmentation, implemented in Gluon. Original paper: 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. """ __all__ = ['DeepLabv3', 'deeplabv3_resnetd50b_voc', 'deeplabv3_resnetd101b_voc', 'deeplabv3_resnetd152b_voc', 'deeplabv3_resnetd50b_coco', 'deeplabv3_resnetd101b_coco', 'deeplabv3_resnetd152b_coco', 'deeplabv3_resnetd50b_ade20k', 'deeplabv3_resnetd101b_ade20k', 'deeplabv3_resnetd50b_cityscapes', 'deeplabv3_resnetd101b_cityscapes'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import conv1x1, conv1x1_block, conv3x3_block from .resnetd import resnetd50b, resnetd101b, resnetd152b class DeepLabv3FinalBlock(HybridBlock): """ DeepLabv3 final block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, bottleneck_factor=4, **kwargs): super(DeepLabv3FinalBlock, self).__init__(**kwargs) assert (in_channels % bottleneck_factor == 0) mid_channels = in_channels // bottleneck_factor with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels) self.dropout = nn.Dropout(rate=0.1) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, use_bias=True) def hybrid_forward(self, F, x, out_size): x = self.conv1(x) x = self.dropout(x) x = self.conv2(x) x = F.contrib.BilinearResize2D(x, height=out_size[0], width=out_size[1]) return x class ASPPAvgBranch(HybridBlock): """ ASPP branch with average pooling. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. upscale_out_size : tuple of 2 int or None Spatial size of output image for the bilinear upsampling operation. """ def __init__(self, in_channels, out_channels, upscale_out_size, **kwargs): super(ASPPAvgBranch, self).__init__(**kwargs) self.upscale_out_size = upscale_out_size with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels) def hybrid_forward(self, F, x): in_size = self.upscale_out_size if self.upscale_out_size is not None else x.shape[2:] x = F.contrib.AdaptiveAvgPooling2D(x, output_size=1) x = self.conv(x) x = F.contrib.BilinearResize2D(x, height=in_size[0], width=in_size[1]) return x class AtrousSpatialPyramidPooling(HybridBlock): """ Atrous Spatial Pyramid Pooling (ASPP) module. Parameters: ---------- in_channels : int Number of input channels. upscale_out_size : tuple of 2 int Spatial size of the input tensor for the bilinear upsampling operation. """ def __init__(self, in_channels, upscale_out_size, **kwargs): super(AtrousSpatialPyramidPooling, self).__init__(**kwargs) atrous_rates = [12, 24, 36] assert (in_channels % 8 == 0) mid_channels = in_channels // 8 project_in_channels = 5 * mid_channels with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(conv1x1_block( in_channels=in_channels, out_channels=mid_channels)) for atrous_rate in atrous_rates: self.branches.add(conv3x3_block( in_channels=in_channels, out_channels=mid_channels, padding=atrous_rate, dilation=atrous_rate)) self.branches.add(ASPPAvgBranch( in_channels=in_channels, out_channels=mid_channels, upscale_out_size=upscale_out_size)) self.conv = conv1x1_block( in_channels=project_in_channels, out_channels=mid_channels) self.dropout = nn.Dropout(rate=0.5) def hybrid_forward(self, F, x): x = self.branches(x) x = self.conv(x) x = self.dropout(x) return x class DeepLabv3(HybridBlock): """ DeepLabv3 model from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int, default 2048 Number of output channels form feature extractor. aux : bool, default False Whether to output an auxiliary result. fixed_size : bool, default True Whether to expect fixed spatial size of input image. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (480, 480) Spatial size of the expected input image. classes : int, default 21 Number of segmentation classes. """ def __init__(self, backbone, backbone_out_channels=2048, aux=False, fixed_size=True, in_channels=3, in_size=(480, 480), classes=21, **kwargs): super(DeepLabv3, self).__init__(**kwargs) assert (in_channels > 0) assert ((in_size[0] % 8 == 0) and (in_size[1] % 8 == 0)) self.in_size = in_size self.classes = classes self.aux = aux self.fixed_size = fixed_size with self.name_scope(): self.backbone = backbone pool_out_size = (self.in_size[0] // 8, self.in_size[1] // 8) if fixed_size else None self.pool = AtrousSpatialPyramidPooling( in_channels=backbone_out_channels, upscale_out_size=pool_out_size) pool_out_channels = backbone_out_channels // 8 self.final_block = DeepLabv3FinalBlock( in_channels=pool_out_channels, out_channels=classes, bottleneck_factor=1) if self.aux: aux_out_channels = backbone_out_channels // 2 self.aux_block = DeepLabv3FinalBlock( in_channels=aux_out_channels, out_channels=classes, bottleneck_factor=4) def hybrid_forward(self, F, x): in_size = self.in_size if self.fixed_size else x.shape[2:] x, y = self.backbone(x) x = self.pool(x) x = self.final_block(x, in_size) if self.aux: y = self.aux_block(y, in_size) return x, y else: return x def get_deeplabv3(backbone, classes, aux=False, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DeepLabv3 model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. classes : int Number of segmentation classes. aux : bool, default False Whether to output an auxiliary result. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ net = DeepLabv3( backbone=backbone, classes=classes, aux=aux, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx, ignore_extra=True) return net def deeplabv3_resnetd50b_voc(pretrained_backbone=False, classes=21, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-50b for Pascal VOC from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 21 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd50b_voc", **kwargs) def deeplabv3_resnetd101b_voc(pretrained_backbone=False, classes=21, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-101b for Pascal VOC from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 21 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd101b_voc", **kwargs) def deeplabv3_resnetd152b_voc(pretrained_backbone=False, classes=21, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-152b for Pascal VOC from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 21 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd152b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd152b_voc", **kwargs) def deeplabv3_resnetd50b_coco(pretrained_backbone=False, classes=21, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-50b for COCO from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 21 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd50b_coco", **kwargs) def deeplabv3_resnetd101b_coco(pretrained_backbone=False, classes=21, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-101b for COCO from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 21 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd101b_coco", **kwargs) def deeplabv3_resnetd152b_coco(pretrained_backbone=False, classes=21, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-152b for COCO from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 21 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd152b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd152b_coco", **kwargs) def deeplabv3_resnetd50b_ade20k(pretrained_backbone=False, classes=150, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-50b for ADE20K from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 150 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd50b_ade20k", **kwargs) def deeplabv3_resnetd101b_ade20k(pretrained_backbone=False, classes=150, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-101b for ADE20K from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 150 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd101b_ade20k", **kwargs) def deeplabv3_resnetd50b_cityscapes(pretrained_backbone=False, classes=19, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-50b for Cityscapes from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 19 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd50b_cityscapes", **kwargs) def deeplabv3_resnetd101b_cityscapes(pretrained_backbone=False, classes=19, aux=True, **kwargs): """ DeepLabv3 model on the base of ResNet(D)-101b for Cityscapes from 'Rethinking Atrous Convolution for Semantic Image Segmentation,' https://arxiv.org/abs/1706.05587. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. classes : int, default 19 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features[:-1] return get_deeplabv3(backbone=backbone, classes=classes, aux=aux, model_name="deeplabv3_resnetd101b_cityscapes", **kwargs) def _test(): import numpy as np import mxnet as mx in_size = (480, 480) aux = False pretrained = False models = [ (deeplabv3_resnetd50b_voc, 21), (deeplabv3_resnetd101b_voc, 21), (deeplabv3_resnetd152b_voc, 21), (deeplabv3_resnetd50b_coco, 21), (deeplabv3_resnetd101b_coco, 21), (deeplabv3_resnetd152b_coco, 21), (deeplabv3_resnetd50b_ade20k, 150), (deeplabv3_resnetd101b_ade20k, 150), (deeplabv3_resnetd50b_cityscapes, 19), (deeplabv3_resnetd101b_cityscapes, 19), ] for model, classes in models: net = model(pretrained=pretrained, in_size=in_size, aux=aux) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) if aux: assert (model != deeplabv3_resnetd50b_voc or weight_count == 42127850) assert (model != deeplabv3_resnetd101b_voc or weight_count == 61119978) assert (model != deeplabv3_resnetd152b_voc or weight_count == 76763626) assert (model != deeplabv3_resnetd50b_coco or weight_count == 42127850) assert (model != deeplabv3_resnetd101b_coco or weight_count == 61119978) assert (model != deeplabv3_resnetd152b_coco or weight_count == 76763626) assert (model != deeplabv3_resnetd50b_ade20k or weight_count == 42194156) assert (model != deeplabv3_resnetd101b_ade20k or weight_count == 61186284) assert (model != deeplabv3_resnetd50b_cityscapes or weight_count == 42126822) assert (model != deeplabv3_resnetd101b_cityscapes or weight_count == 61118950) else: assert (model != deeplabv3_resnetd50b_voc or weight_count == 39762645) assert (model != deeplabv3_resnetd101b_voc or weight_count == 58754773) assert (model != deeplabv3_resnetd152b_voc or weight_count == 74398421) assert (model != deeplabv3_resnetd50b_coco or weight_count == 39762645) assert (model != deeplabv3_resnetd101b_coco or weight_count == 58754773) assert (model != deeplabv3_resnetd152b_coco or weight_count == 74398421) assert (model != deeplabv3_resnetd50b_ade20k or weight_count == 39795798) assert (model != deeplabv3_resnetd101b_ade20k or weight_count == 58787926) assert (model != deeplabv3_resnetd50b_cityscapes or weight_count == 39762131) assert (model != deeplabv3_resnetd101b_cityscapes or weight_count == 58754259) x = mx.nd.zeros((1, 3, in_size[0], in_size[1]), ctx=ctx) ys = net(x) y = ys[0] if aux else ys assert ((y.shape[0] == x.shape[0]) and (y.shape[1] == classes) and (y.shape[2] == x.shape[2]) and (y.shape[3] == x.shape[3])) if __name__ == "__main__": _test()
22,813
38.884615
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/fpenet.py
""" FPENet for image segmentation, implemented in Gluon. Original paper: 'Feature Pyramid Encoding Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1909.08599. """ __all__ = ['FPENet', 'fpenet_cityscapes'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv1x1_block, conv3x3_block, SEBlock, InterpolationBlock, MultiOutputSequential class FPEBlock(HybridBlock): """ FPENet block. Parameters: ---------- channels : int Number of input/output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(FPEBlock, self).__init__(**kwargs) dilations = [1, 2, 4, 8] assert (channels % len(dilations) == 0) mid_channels = channels // len(dilations) with self.name_scope(): self.blocks = nn.HybridSequential(prefix="") for i, dilation in enumerate(dilations): self.blocks.add(conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, groups=mid_channels, dilation=dilation, padding=dilation, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) def hybrid_forward(self, F, x): xs = F.split(x, axis=1, num_outputs=len(self.blocks._children)) ys = [] for bi, xsi in zip(self.blocks._children.values(), xs): if len(ys) == 0: ys.append(bi(xsi)) else: ys.append(bi(xsi + ys[-1])) x = F.concat(*ys, dim=1) return x class FPEUnit(HybridBlock): """ FPENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bottleneck_factor : int Bottleneck factor. use_se : bool Whether to use SE-module. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, strides, bottleneck_factor, use_se, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(FPEUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) self.use_se = use_se mid1_channels = in_channels * bottleneck_factor with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid1_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.block = FPEBlock(channels=mid1_channels) self.conv2 = conv1x1_block( in_channels=mid1_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) if self.use_se: self.se = SEBlock(channels=out_channels) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.conv1(x) x = self.block(x) x = self.conv2(x) if self.use_se: x = self.se(x) x = x + identity x = self.activ(x) return x class FPEStage(HybridBlock): """ FPENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. layers : int Number of layers. use_se : bool Whether to use SE-module. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, layers, use_se, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(FPEStage, self).__init__(**kwargs) self.use_block = (layers > 1) with self.name_scope(): if self.use_block: self.down = FPEUnit( in_channels=in_channels, out_channels=out_channels, strides=2, bottleneck_factor=4, use_se=use_se, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.blocks = nn.HybridSequential(prefix="") for i in range(layers - 1): self.blocks.add(FPEUnit( in_channels=out_channels, out_channels=out_channels, strides=1, bottleneck_factor=1, use_se=use_se, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) else: self.down = FPEUnit( in_channels=in_channels, out_channels=out_channels, strides=1, bottleneck_factor=1, use_se=use_se, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): x = self.down(x) if self.use_block: y = self.blocks(x) x = x + y return x class MEUBlock(HybridBlock): """ FPENet specific mutual embedding upsample (MEU) block. Parameters: ---------- in_channels_high : int Number of input channels for x_high. in_channels_low : int Number of input channels for x_low. out_channels : int Number of output channels. out_size : tuple of 2 int Spatial size of output image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels_high, in_channels_low, out_channels, out_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(MEUBlock, self).__init__(**kwargs) with self.name_scope(): self.conv_high = conv1x1_block( in_channels=in_channels_high, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) self.conv_low = conv1x1_block( in_channels=in_channels_low, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) self.conv_w_high = conv1x1( in_channels=out_channels, out_channels=out_channels) self.conv_w_low = conv1x1( in_channels=1, out_channels=1) self.sigmoid = nn.Activation("sigmoid") self.relu = nn.Activation("relu") self.up = InterpolationBlock(scale_factor=2, out_size=out_size) def hybrid_forward(self, F, x_high, x_low): x_high = self.conv_high(x_high) x_low = self.conv_low(x_low) w_high = F.contrib.AdaptiveAvgPooling2D(x_high, output_size=1) w_high = self.conv_w_high(w_high) w_high = self.relu(w_high) w_high = self.sigmoid(w_high) w_low = x_low.mean(axis=1, keepdims=True) w_low = self.conv_w_low(w_low) w_low = self.sigmoid(w_low) x_high = self.up(x_high) x_high = F.broadcast_mul(x_high, w_low) x_low = F.broadcast_mul(x_low, w_high) out = x_high + x_low return out class FPENet(HybridBlock): """ FPENet model from 'Feature Pyramid Encoding Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1909.08599. Parameters: ---------- layers : list of int Number of layers for each unit. channels : list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. meu_channels : list of int Number of output channels for MEU blocks. use_se : bool Whether to use SE-module. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. aux : bool, default False Whether to output an auxiliary result. fixed_size : bool, default False Whether to expect fixed spatial size of input image. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (1024, 2048) Spatial size of the expected input image. classes : int, default 19 Number of segmentation classes. """ def __init__(self, layers, channels, init_block_channels, meu_channels, use_se, bn_use_global_stats=False, bn_cudnn_off=False, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), classes=19, **kwargs): super(FPENet, self).__init__(**kwargs) assert (aux is not None) assert (fixed_size is not None) assert ((in_size[0] % 8 == 0) and (in_size[1] % 8 == 0)) self.in_size = in_size self.classes = classes self.fixed_size = fixed_size with self.name_scope(): self.stem = conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) in_channels = init_block_channels self.encoder = MultiOutputSequential(return_last=False) for i, (layers_i, out_channels) in enumerate(zip(layers, channels)): stage = FPEStage( in_channels=in_channels, out_channels=out_channels, layers=layers_i, use_se=use_se, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) stage.do_output = True self.encoder.add(stage) in_channels = out_channels self.meu1 = MEUBlock( in_channels_high=channels[-1], in_channels_low=channels[-2], out_channels=meu_channels[0], out_size=((in_size[0] // 4, in_size[1] // 4) if fixed_size else None), bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.meu2 = MEUBlock( in_channels_high=meu_channels[0], in_channels_low=channels[-3], out_channels=meu_channels[1], out_size=((in_size[0] // 2, in_size[1] // 2) if fixed_size else None), bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) in_channels = meu_channels[1] self.classifier = conv1x1( in_channels=in_channels, out_channels=classes, use_bias=True) self.up = InterpolationBlock( scale_factor=2, out_size=(in_size if fixed_size else None)) def hybrid_forward(self, F, x): x = self.stem(x) y = self.encoder(x) x = self.meu1(y[2], y[1]) x = self.meu2(x, y[0]) x = self.classifier(x) x = self.up(x) return x def get_fpenet(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create FPENet model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ width = 16 channels = [int(width * (2 ** i)) for i in range(3)] init_block_channels = width layers = [1, 3, 9] meu_channels = [64, 32] use_se = False net = FPENet( layers=layers, channels=channels, init_block_channels=init_block_channels, meu_channels=meu_channels, use_se=use_se, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx, ignore_extra=True) return net def fpenet_cityscapes(classes=19, **kwargs): """ FPENet model for Cityscapes from 'Feature Pyramid Encoding Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1909.08599. Parameters: ---------- classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fpenet(classes=classes, model_name="fpenet_cityscapes", **kwargs) def _calc_width(net): import numpy as np net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) return weight_count def _test(): import mxnet as mx pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ fpenet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net.hybridize() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fpenet_cityscapes or weight_count == 115125) batch = 4 x = mx.nd.random.normal(shape=(batch, 3, in_size[0], in_size[1]), ctx=ctx) y = net(x) assert (y.shape == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
16,864
32.662675
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/irevnet.py
""" i-RevNet for ImageNet-1K, implemented in Gluon. Original paper: 'i-RevNet: Deep Invertible Networks,' https://arxiv.org/abs/1802.07088. """ __all__ = ['IRevNet', 'irevnet301', 'IRevDownscale', 'IRevSplitBlock', 'IRevMergeBlock'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3, pre_conv3x3_block, DualPathSequential class IRevDualPathSequential(DualPathSequential): """ An invertible sequential container for hybrid blocks with dual inputs/outputs. Blocks will be executed in the order they are added. Parameters: ---------- return_two : bool, default True Whether to return two output after execution. first_ordinals : int, default 0 Number of the first blocks with single input/output. last_ordinals : int, default 0 Number of the final blocks with single input/output. dual_path_scheme : function Scheme of dual path response for a block. dual_path_scheme_ordinal : function Scheme of dual path response for an ordinal block. last_noninvertible : int, default 0 Number of the final blocks skipped during inverse. """ def __init__(self, return_two=True, first_ordinals=0, last_ordinals=0, dual_path_scheme=(lambda block, x1, x2: block(x1, x2)), dual_path_scheme_ordinal=(lambda block, x1, x2: (block(x1), x2)), last_noninvertible=0, **kwargs): super(IRevDualPathSequential, self).__init__( return_two=return_two, first_ordinals=first_ordinals, last_ordinals=last_ordinals, dual_path_scheme=dual_path_scheme, dual_path_scheme_ordinal=dual_path_scheme_ordinal, **kwargs) self.last_noninvertible = last_noninvertible def inverse(self, x1, x2=None): length = len(self._children.values()) for i, block in enumerate(reversed(self._children.values())): if i < self.last_noninvertible: pass elif (i < self.last_ordinals) or (i >= length - self.first_ordinals): x1, x2 = self.dual_path_scheme_ordinal(block.inverse, x1, x2) else: x1, x2 = self.dual_path_scheme(block.inverse, x1, x2) if self.return_two: return x1, x2 else: return x1 class IRevDownscale(HybridBlock): """ i-RevNet specific downscale (so-called psi-block). Parameters: ---------- scale : int Scale (downscale) value. """ def __init__(self, scale, **kwargs): super(IRevDownscale, self).__init__(**kwargs) self.scale = scale def hybrid_forward(self, F, x): batch, x_channels, x_height, x_width = x.shape y_channels = x_channels * self.scale * self.scale assert (x_height % self.scale == 0) y_height = x_height // self.scale y = x.transpose(axes=(0, 2, 3, 1)) d2_split_seq = y.split(axis=2, num_outputs=(y.shape[2] // self.scale)) d2_split_seq = [t.reshape(batch, y_height, y_channels) for t in d2_split_seq] y = F.stack(*d2_split_seq, axis=1) y = y.transpose(axes=(0, 3, 2, 1)) return y def inverse(self, y): import mxnet.ndarray as F scale_sqr = self.scale * self.scale batch, y_channels, y_height, y_width = y.shape assert (y_channels % scale_sqr == 0) x_channels = y_channels // scale_sqr x_height = y_height * self.scale x_width = y_width * self.scale x = y.transpose(axes=(0, 2, 3, 1)) x = x.reshape(batch, y_height, y_width, scale_sqr, x_channels) d3_split_seq = x.split(axis=3, num_outputs=(x.shape[3] // self.scale)) d3_split_seq = [t.reshape(batch, y_height, x_width, x_channels) for t in d3_split_seq] x = F.stack(*d3_split_seq, axis=0) x = x.swapaxes(0, 1).transpose(axes=(0, 2, 1, 3, 4)).reshape(batch, x_height, x_width, x_channels) x = x.transpose(axes=(0, 3, 1, 2)) return x class IRevInjectivePad(HybridBlock): """ i-RevNet channel zero padding block. Parameters: ---------- padding : int Size of the padding. """ def __init__(self, padding, **kwargs): super(IRevInjectivePad, self).__init__(**kwargs) self.padding = padding def hybrid_forward(self, F, x): x = x.transpose(axes=(0, 2, 1, 3)) x = F.pad(x, mode="constant", pad_width=(0, 0, 0, 0, 0, self.padding, 0, 0), constant_value=0) x = x.transpose(axes=(0, 2, 1, 3)) return x def inverse(self, x): return x[:, :x.shape[1] - self.padding, :, :] class IRevSplitBlock(HybridBlock): """ iRevNet split block. """ def __init__(self, **kwargs): super(IRevSplitBlock, self).__init__(**kwargs) def hybrid_forward(self, F, x, _): x1, x2 = F.split(x, axis=1, num_outputs=2) return x1, x2 def inverse(self, x1, x2): import mxnet.ndarray as F x = F.concat(x1, x2, dim=1) return x, None class IRevMergeBlock(HybridBlock): """ iRevNet merge block. """ def __init__(self, **kwargs): super(IRevMergeBlock, self).__init__(**kwargs) def hybrid_forward(self, F, x1, x2): x = F.concat(x1, x2, dim=1) return x, x def inverse(self, x, _): import mxnet.ndarray as F x1, x2 = F.split(x, axis=1, num_outputs=2) return x1, x2 class IRevBottleneck(HybridBlock): """ iRevNet bottleneck block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the branch convolution layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. preactivate : bool Whether use pre-activation for the first convolution block. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, preactivate, **kwargs): super(IRevBottleneck, self).__init__(**kwargs) mid_channels = out_channels // 4 with self.name_scope(): if preactivate: self.conv1 = pre_conv3x3_block( in_channels=in_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) else: self.conv1 = conv3x3( in_channels=in_channels, out_channels=mid_channels, strides=strides) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.conv3 = pre_conv3x3_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class IRevUnit(HybridBlock): """ iRevNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the branch convolution layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. preactivate : bool Whether use pre-activation for the first convolution block. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, preactivate, **kwargs): super(IRevUnit, self).__init__(**kwargs) if not preactivate: in_channels = in_channels // 2 padding = 2 * (out_channels - in_channels) self.do_padding = (padding != 0) and (strides == 1) self.do_downscale = (strides != 1) with self.name_scope(): if self.do_padding: self.pad = IRevInjectivePad(padding) self.bottleneck = IRevBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, preactivate=preactivate) if self.do_downscale: self.psi = IRevDownscale(strides) def hybrid_forward(self, F, x1, x2): if self.do_padding: x = F.concat(x1, x2, dim=1) x = self.pad(x) x1, x2 = F.split(x, axis=1, num_outputs=2) fx2 = self.bottleneck(x2) if self.do_downscale: x1 = self.psi(x1) x2 = self.psi(x2) y1 = fx2 + x1 return x2, y1 def inverse(self, x2, y1): import mxnet.ndarray as F if self.do_downscale: x2 = self.psi.inverse(x2) fx2 = - self.bottleneck(x2) x1 = fx2 + y1 if self.do_downscale: x1 = self.psi.inverse(x1) if self.do_padding: x = F.concat(x1, x2, dim=1) x = self.pad.inverse(x) x1, x2 = F.split(x, axis=1, num_outputs=2) return x1, x2 class IRevPostActivation(HybridBlock): """ iRevNet specific post-activation block. Parameters: ---------- in_channels : int Number of input channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, bn_use_global_stats=False, **kwargs): super(IRevPostActivation, self).__init__(**kwargs) with self.name_scope(): self.bn = nn.BatchNorm( in_channels=in_channels, use_global_stats=bn_use_global_stats) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) return x class IRevNet(HybridBlock): """ i-RevNet model from 'i-RevNet: Deep Invertible Networks,' https://arxiv.org/abs/1802.07088. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. init_block_channels : int Number of output channels for the initial unit. final_block_channels : int Number of output channels for the final unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(IRevNet, self).__init__(**kwargs) assert (in_channels > 0) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = IRevDualPathSequential( first_ordinals=1, last_ordinals=2, last_noninvertible=2) self.features.add(IRevDownscale(scale=2)) in_channels = init_block_channels self.features.add(IRevSplitBlock()) for i, channels_per_stage in enumerate(channels): stage = IRevDualPathSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) else 1 preactivate = not ((i == 0) and (j == 0)) stage.add(IRevUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, preactivate=preactivate)) in_channels = out_channels self.features.add(stage) in_channels = final_block_channels self.features.add(IRevMergeBlock()) self.features.add(IRevPostActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x, return_out_bij=False): x, out_bij = self.features(x) x = self.output(x) if return_out_bij: return x, out_bij else: return x def inverse(self, out_bij): x, _ = self.features.inverse(out_bij) return x def get_irevnet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create i-RevNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 301: layers = [6, 16, 72, 6] else: raise ValueError("Unsupported i-RevNet with number of blocks: {}".format(blocks)) assert (sum(layers) * 3 + 1 == blocks) channels_per_layers = [24, 96, 384, 1536] init_block_channels = 12 final_block_channels = 3072 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = IRevNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ignore_extra=True, ctx=ctx) return net def irevnet301(**kwargs): """ i-RevNet-301 model from 'i-RevNet: Deep Invertible Networks,' https://arxiv.org/abs/1802.07088. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_irevnet(blocks=301, model_name="irevnet301", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ irevnet301, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != irevnet301 or weight_count == 125120356) x = mx.nd.random.randn(2, 3, 224, 224, ctx=ctx) y = net(x) assert (y.shape == (2, 1000)) y, out_bij = net(x, True) x_ = net.inverse(out_bij) assert (x_.shape == (2, 3, 224, 224)) assert ((np.max(np.abs(x.asnumpy() - x_.asnumpy())) < 1e-4) or (np.max(np.abs(y.asnumpy()) > 1e10))) if __name__ == "__main__": _test()
17,306
31.59322
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/model_store.py
""" Model store which provides pretrained models. """ __all__ = ['get_model_file'] import os import zipfile import logging from mxnet.gluon.utils import download, check_sha1 _model_sha1 = {name: (error, checksum, repo_release_tag) for name, error, checksum, repo_release_tag in [ ('alexnet', '1610', '4dd7cfb6275229c3b3889ec7ad29f8d48a499193', 'v0.0.481'), ('alexnetb', '1705', '0181007ac2ba5d5c051346af1eea4a00c36f34e9', 'v0.0.485'), ('zfnet', '1678', '3299fdce9712697f9e03e1a0cf741745553a28cf', 'v0.0.395'), ('zfnetb', '1459', '7a654810eced689323e414030165cf1392c9684f', 'v0.0.400'), ('vgg11', '1016', '3d78e0ec95d358577acf8b2e2f768a72ec319ee3', 'v0.0.381'), ('vgg13', '0950', 'd2bcaaf3704afb47b2818066255b3de65d0c03e1', 'v0.0.388'), ('vgg16', '0832', '22fe503aa438ebfd1f3128b676db21a06de3cd59', 'v0.0.401'), ('vgg19', '0767', 'e198aa1f6519b3331e5ebbe782ca35fd2aca6213', 'v0.0.420'), ('bn_vgg11', '0934', '3f79cab1e9e0bab0d57c0f80a6c9782e5cdc765d', 'v0.0.339'), ('bn_vgg13', '0887', '540243b0a3eb2ae3974c2d75cf82e655fa4f0dda', 'v0.0.353'), ('bn_vgg16', '0757', '90441925c7c19c9db6bf96824c2440774c0e54df', 'v0.0.359'), ('bn_vgg19', '0689', 'cd8f4229e5e757ba951afc6eb1a4ec7a471022f7', 'v0.0.360'), ('bn_vgg11b', '0975', '685ae89dcf3916b9c6e887d4c5bc441a34253744', 'v0.0.407'), ('bn_vgg13b', '0912', 'fc678318391d45465bda4aa17744dc0d148135e3', 'v0.0.488'), ('bn_vgg16b', '0775', '77dad99b4f72176870655ffbf59e96f3b9570a0b', 'v0.0.489'), ('bn_vgg19b', '0735', '8d9a132bfd9230304e588ec159170ee3173d64ba', 'v0.0.490'), ('bninception', '0754', '75225419bae387931f6c8ccc19dc48fb1cc8cdae', 'v0.0.405'), ('resnet10', '1220', '83ebb0937d85142c3f827be56f08a57a5996c90b', 'v0.0.569'), ('resnet12', '1203', 'a41948541130b9ebf8285412b38c668198e4ad4f', 'v0.0.485'), ('resnet14', '1086', '5d9f22a7aa6c821eb69fc043af2da040f36110eb', 'v0.0.491'), ('resnetbc14b', '1033', '4ff348bebc7696cd219be605668bb76546dd9641', 'v0.0.481'), ('resnet16', '0978', '2c28373c9ce1f50ff6a4b47a16a3d1bc2e557c75', 'v0.0.493'), ('resnet18_wd4', '1740', 'a74ea15d056a84133e02706a6c7f8ed8f50c8462', 'v0.0.262'), ('resnet18_wd2', '1284', '9a5154065311c8ffbbc57b20a443b205c7a910fa', 'v0.0.263'), ('resnet18_w3d4', '1066', '1a574a4198a5bbf01572c2b3f091eb824ff8196e', 'v0.0.266'), ('resnet18', '0867', '711ed8ab624c4c580e90321989aef6174ad2991d', 'v0.0.478'), ('resnet26', '0823', 'a2746eb21d73c3c85edacabe36d680988986f890', 'v0.0.489'), ('resnetbc26b', '0758', '2b5e8d0888936a340ea13c7e8ba30b237cd62f1c', 'v0.0.313'), ('resnet34', '0743', '5cdeeccda6f87fe13aed279213006061a8b42037', 'v0.0.291'), ('resnetbc38b', '0672', '820944641ba54f4aaa43d2a305ab52b9dcb740c7', 'v0.0.328'), ('resnet50', '0604', 'a71d1d2a8e8e4259742bbd67c386623233b57c6c', 'v0.0.329'), ('resnet50b', '0611', 'ca12f8d804000bf5202e2e3838dec7ef6b772149', 'v0.0.308'), ('resnet101', '0516', '75a3105c609aedb89d63dd3dc47b578f41fff6be', 'v0.0.499'), ('resnet101b', '0512', 'af5c4233b28c8b7acd3b3ebe02f9f2eda2c77824', 'v0.0.357'), ('resnet152', '0444', '15f2dd80f964225e685e7564cecd04911feb97ee', 'v0.0.518'), ('resnet152b', '0429', '45a3ad0f989852a4456c907fa3b40b00dd24937c', 'v0.0.517'), ('preresnet10', '1401', '2b96c0818dbabc422e98d8fbfc9b684c023922ed', 'v0.0.249'), ('preresnet12', '1321', 'b628efb5415784075e18b6734b1ba1e5c7280dee', 'v0.0.257'), ('preresnet14', '1218', 'd65fa6287414d9412e34ac0df6921eaa5646a2b6', 'v0.0.260'), ('preresnetbc14b', '1151', 'c712a235b75ad4956411bab265dfd924c748726e', 'v0.0.315'), ('preresnet16', '1081', '5b00b55f74adb9ee4a6ba5f946aafd48b4d8aa47', 'v0.0.261'), ('preresnet18_wd4', '1778', '3d949d1ae20b9188a423b56a1f7a89b4bcecc3d2', 'v0.0.272'), ('preresnet18_wd2', '1319', '63e55c24bc0ae93a8f8daefa4b35dc3e70147f65', 'v0.0.273'), ('preresnet18_w3d4', '1068', 'eb5698616757fd0947851f62c33fc4d7b4a5f23a', 'v0.0.274'), ('preresnet18', '0951', '71279a0b7339f1efd12bed737219a9ed76175a9d', 'v0.0.140'), ('preresnet26', '0834', 'c2ecba0948934c28d459b7f87fbc1489420fd4fb', 'v0.0.316'), ('preresnetbc26b', '0786', '265f591f320db0915c18c16f4ed0e2e53ee46567', 'v0.0.325'), ('preresnet34', '0751', 'ba9c829e72d54f8b02cf32ea202c195d36568467', 'v0.0.300'), ('preresnetbc38b', '0633', '809d2defea82276fdd7ff2faafc1f6ffe57c93b5', 'v0.0.348'), ('preresnet50', '0620', '50f13b2d3fd197c8aa721745adaf2d6615fd8c16', 'v0.0.330'), ('preresnet50b', '0632', '951de2dc558f94f489ce62fedf979ccc08361641', 'v0.0.307'), ('preresnet101', '0534', 'ea9a1724662744fe349c0aa8dd0ec42708fe88c7', 'v0.0.504'), ('preresnet101b', '0540', '3839a4733a8a614bb6b7b4f555759bb5b8013d42', 'v0.0.351'), ('preresnet152', '0446', '3b41bd93a7a2807e8cc88ae75ec0cde69af1e786', 'v0.0.510'), ('preresnet152b', '0438', 'ebf71d023e88d006c55946e3c6cf38c5d0cffeb2', 'v0.0.523'), ('preresnet200b', '0446', '70135e6618bd8e870574a2667d3ec592238b3a3c', 'v0.0.529'), ('preresnet269b', '0501', '6e56119dc52a38af690bb41d4ae1683007558233', 'v0.0.545'), ('resnext14_16x4d', '1223', '1f8072e8d01d1427941a06bbca896211e98e2b75', 'v0.0.370'), ('resnext14_32x2d', '1247', '2ca8cc2544045c21d0a7ca740483097491ba0855', 'v0.0.371'), ('resnext14_32x4d', '1110', '9be6190e328c15a06703be3ba922d707c2f4d8e7', 'v0.0.327'), ('resnext26_32x2d', '0850', 'a1fb4451be6336d9f648ccc2c2dedacc5704904a', 'v0.0.373'), ('resnext26_32x4d', '0721', '5264d7efd606e1c95a2480050e9f03a7a2f02b09', 'v0.0.332'), ('resnext50_32x4d', '0545', '748e53b31fe716b5fdd059107d22bdfec3a31839', 'v0.0.498'), ('resnext101_32x4d', '0418', 'c07faa66647bd25dd222b44d8782c59fcc2e3708', 'v0.0.530'), ('resnext101_64x4d', '0439', '178e79057f27afec04219aa1a1a96abbbd8763a9', 'v0.0.544'), ('seresnet10', '1169', '675d4b5bbaaab1b87988df96819cd19e5323512c', 'v0.0.486'), ('seresnet12', '1176', 'df10dd88759750bb042ce7b09dd8a78f32d78e12', 'v0.0.544'), ('seresnet14', '1095', '579e2fecfe37d1c9422f6521fef610bfde6fcaaa', 'v0.0.545'), ('seresnet16', '0972', 'ef50864856c2120082e3db7d4193f31ec77624c6', 'v0.0.545'), ('seresnet18', '0920', '85a6b1da19645419cc3075852588cc7e7da5715f', 'v0.0.355'), ('seresnet26', '0803', '9f9004192240ae0125399d2f6acbb5359027039d', 'v0.0.363'), ('seresnetbc26b', '0682', '15ae6e19b1626107029df18bf3f5140e6fcb2b02', 'v0.0.366'), ('seresnetbc38b', '0575', 'f80f0c3c2612e1334204df1575b8a7cd18f851ff', 'v0.0.374'), ('seresnet50', '0560', 'e75ef498abfc021f356378b2c806de8927287fe7', 'v0.0.441'), ('seresnet50b', '0533', '0d8f0d23bb980621095e41de69e3a68d7aaeba45', 'v0.0.387'), ('seresnet101', '0441', 'f9cb9df623683520522c5fe8577e8a93cf253f46', 'v0.0.533'), ('seresnet101b', '0462', '59fae71a5db6ad1f1b5a79f7339920aa1191ed70', 'v0.0.460'), ('seresnet152', '0430', 'c712f55410938861c32740fe93770b0d54e094a3', 'v0.0.538'), ('sepreresnet10', '1221', 'c3ab94ee5c938f53a8a61849be3a463458444a1f', 'v0.0.544'), ('sepreresnet12', '1180', 'd6b7260b4931e8806368c42b8bb605b9a8afaa23', 'v0.0.543'), ('sepreresnet16', '0956', 'b5e4e06f6f5394f7557986193700a8ecf23ffdc7', 'v0.0.543'), ('sepreresnet18', '0881', 'b018c6e38601b0719445e302ea64196924319814', 'v0.0.543'), ('sepreresnet26', '0804', 'a2e48457a44958ed7f20940997f419080b1d14e8', 'v0.0.543'), ('sepreresnetbc26b', '0636', '33c94c9dd2986d0643f7c432d5203294ecf2124e', 'v0.0.399'), ('sepreresnetbc38b', '0563', 'd8f0fbd35b840743e312b9ec944c7fc141824ace', 'v0.0.409'), ('sepreresnet50b', '0532', '5b620ff7175c674ef41647e2e79992dceefeec19', 'v0.0.461'), ('seresnext50_32x4d', '0433', 'd74d1d0a5bd1449562e9f1de92b1389c647fdd13', 'v0.0.505'), ('seresnext101_32x4d', '0444', '3c89aec5b98c1bfca13bc8c21e07dba82243c012', 'v0.0.529'), ('seresnext101_64x4d', '0408', '2e039a419e4eb84ce93d1515027099c4dc13f5a1', 'v0.0.561'), ('senet16', '0806', 'ba26802160725af926e5a3217f8848bd3a6599fd', 'v0.0.341'), ('senet28', '0591', 'd5297a35a2e56ecc892499c1bd4373125a40b783', 'v0.0.356'), ('senet154', '0440', 'c0e2d2b9e70894e28e4f957296229bdd89b70607', 'v0.0.522'), ('resnestabc14', '0634', '4b0cbe8c59e9d764027f51670096037e2f058970', 'v0.0.493'), ('resnesta18', '0689', '8f37b6927751b27a72c1b8c625de265da9b31570', 'v0.0.489'), ('resnestabc26', '0470', 'f88d49d7bd6cdf1f448afdf4bc5dadd9921798dd', 'v0.0.495'), ('resnesta50', '0438', '6ce54a97d6e6fe8aa66755669b8f47a0790bfeb6', 'v0.0.531'), ('resnesta101', '0399', 'ab6c6f89407d7b2cf89406f7bb62a48e86d7cabc', 'v0.0.465'), ('resnesta152', '0451', '532c448de0011e90924b3c1e829f9525c6a5420e', 'v0.0.540'), ('resnesta200', '0340', '3bd1f0c8d9a2862b89590c1c490a38d6d982522c', 'v0.0.465'), ('resnesta269', '0336', '8333862a0c016432f6cd5b45f7e9de7ee6e3f319', 'v0.0.465'), ('ibn_resnet50', '0559', '0f75710a144ea1483e6235f6909f8b8c3555ec01', 'v0.0.495'), ('ibn_resnet101', '0489', 'db938be9060c0ea48913fa0fa9794bf6bcacea7f', 'v0.0.552'), ('ibnb_resnet50', '0579', '146aa6e143d582cd33af23c942e3e0b2fe152817', 'v0.0.552'), ('ibn_resnext101_32x4d', '0487', '1398597e25f0be2d7b05f190c5c53f7b8d5265e6', 'v0.0.553'), ('ibn_densenet121', '0646', '82ee3ff44fd308107ec07d759c52a99e94af7041', 'v0.0.493'), ('ibn_densenet169', '0608', 'b509f33947b2f205f03e49ce2d5ee87855bbf708', 'v0.0.500'), ('airnet50_1x64d_r2', '0523', '1f982e0e8b3af40a5d67ca854f18ca837aa332f4', 'v0.0.522'), ('airnet50_1x64d_r16', '0544', '09a9f13b9d5f3e27eb26271070cc4bf8dd72cf9f', 'v0.0.519'), ('airnext50_32x4d_r2', '0504', '664dd077a978562a376e8ab70b5d740893873111', 'v0.0.521'), ('bam_resnet50', '0538', 'fa612c3da552cd26094634c359413db69d0ef97b', 'v0.0.499'), ('cbam_resnet50', '0488', 'ebd4af235d763f4ecce6dd4e09b508ba61b561f7', 'v0.0.537'), ('scnet50', '0511', '359d35d017e838d3c3c2bdf1cc83074610dc3c5b', 'v0.0.493'), ('scnet101', '0446', 'bc5cade042e78813db4ce5878d733daa0ce7f89d', 'v0.0.507'), ('scneta50', '0459', 'c1f6f0286ed942bbc468395b82b7be84312f3c73', 'v0.0.509'), ('regnetx002', '1038', '7800b310f45b4666ef6c862bc8a2573f65ddaa40', 'v0.0.475'), ('regnetx004', '0855', 'b933a72fe7304a31391a9e9fcdf7ffc47ea05353', 'v0.0.479'), ('regnetx006', '0756', 'd41aa087bd288266105142b88b03b67d74352a46', 'v0.0.482'), ('regnetx008', '0724', '79309908dee31eda64e824a9b3bd33c0afaaf5a8', 'v0.0.482'), ('regnetx016', '0613', '018dbe2d89d2bac26f41d98b09ba57b049ed4cfe', 'v0.0.486'), ('regnetx032', '0568', '6d4372fcf0b9d6e3edefa99e72dd411b6a0c676c', 'v0.0.492'), ('regnetx040', '0469', 'c22092c7ceeac713a9102c697f23bec0b7cfaec0', 'v0.0.495'), ('regnetx064', '0458', '473342937e02dce6244a62780f98aed1f5eb831d', 'v0.0.535'), ('regnetx080', '0466', '4086910ad9df1693db658c86f874f686a279175c', 'v0.0.515'), ('regnetx120', '0518', 'ea20b368214230ce65c14218aedba07b68235e38', 'v0.0.542'), ('regnetx160', '0456', '941840a38b20e5b5c6ca092bed3f2ea0ff602895', 'v0.0.532'), ('regnetx320', '0394', 'a120c8d2d181dacffde8a3029292ee59da12e9be', 'v0.0.548'), ('regnety002', '0953', 'b37fcac05e59ba5e751701da8e81195bbfbf3db8', 'v0.0.476'), ('regnety004', '0747', '5626bdf45eaf0f23ddc6e1b68f6ad2db7ca119cf', 'v0.0.481'), ('regnety006', '0697', '81372679b5f9601a5bc72caa13aff378d2ac4233', 'v0.0.483'), ('regnety008', '0645', 'd92881be768a267f1cd2b540d087884bbe93f644', 'v0.0.483'), ('regnety016', '0568', 'c4541a25e92ebf31d9bddb4975738c08fe929836', 'v0.0.486'), ('regnety032', '0413', '9066698526cbc930706fe00b21ca56c31ad7e2e4', 'v0.0.473'), ('regnety040', '0467', '6039a215ee79a7855156c5166df4c6b34f8d501d', 'v0.0.494'), ('regnety064', '0445', '8ae16f971eacad369df091e68c26e63b8c6da9a5', 'v0.0.513'), ('regnety080', '0436', '4cd7fc61d54ec63bfeff11246e8258188351e7ed', 'v0.0.516'), ('regnety120', '0431', 'effa35b601ffbdced20342d9bfe3917a8279880a', 'v0.0.526'), ('regnety160', '0430', '8273156944dea5a439ef55e7ed84a10524be090c', 'v0.0.527'), ('regnety320', '0373', 'd26a5f80c1d67e3527ddbdce75683370c5799e06', 'v0.0.550'), ('pyramidnet101_a360', '0520', '3a98a2bfc4e74b00b5fff20783613281121a6d37', 'v0.0.507'), ('diracnet18v2', '1117', '27601f6fa54e3b10d77981f30650d7a9d4bce91e', 'v0.0.111'), ('diracnet34v2', '0946', '1faa6f1245e152d1a3e12de4b5dc1ba554bc3bb8', 'v0.0.111'), ('crunet56', '0536', 'd7112ed431630f4162fb2ebe7a7c0666fa1718a8', 'v0.0.558'), ('densenet121', '0685', 'd3a1fae8b311343498f736e494d60d32e35debfb', 'v0.0.314'), ('densenet161', '0592', '29897d410ea5ae427278df060de578911df74667', 'v0.0.432'), ('densenet169', '0605', '9c045c864828e773f92f998199821fc0c21e0eb4', 'v0.0.406'), ('densenet201', '0590', '89aa8c295b21fdd682df0027d2232e6dabf2cace', 'v0.0.426'), ('condensenet74_c4_g4', '0864', 'cde68fa2fcc9197e336717a17753a15a6efd7596', 'v0.0.4'), ('condensenet74_c8_g8', '1049', '4cf4a08e7fb46f5821049dcae97ae442b0ceb546', 'v0.0.4'), ('peleenet', '0979', '758d3cf992a6c92731069be091be2e8ebd3209e2', 'v0.0.496'), ('wrn50_2', '0606', '5ff344cae22c8ce3f21e042be5a5d899681d89a0', 'v0.0.520'), ('drnc26', '0711', 'e69a4e7bc94aab58e05880c886e8e829584845c3', 'v0.0.508'), ('drnc42', '0614', '6ee5008a3e66476380f4fadf3cf2633c32a596c9', 'v0.0.556'), ('drnc58', '0515', '418d18b26f32347e695b85a60d2bf658409be28d', 'v0.0.559'), ('drnd22', '0744', 'aecf15fc2171d29d8ea765afa540e4ddb40a9b19', 'v0.0.498'), ('drnd38', '0624', '4a86d6ebd970e02792068c46c80ae00686d759cc', 'v0.0.552'), ('drnd54', '0497', '117ca29b47a6e640269d455dfcc12e9958f5f1e1', 'v0.0.554'), ('drnd105', '0487', '52286b345e48cc873bbc0239cdd91b03beae655d', 'v0.0.564'), ('dpn68', '0658', '07251919c08640c94375670cbc5f0fbc312ed59b', 'v0.0.310'), ('dpn98', '0422', '1ace57dc702d5b6ca7eccb657036f682605d7efc', 'v0.0.540'), ('dpn131', '0477', 'e814a53cc6cddc0b6f2a11393a88c44723a8ec0c', 'v0.0.534'), ('darknet_tiny', '1746', '16501793621fbcb137f2dfb901760c1f621fa5ec', 'v0.0.69'), ('darknet_ref', '1668', '3011b4e14b629f80da54ab57bef305d588f748ab', 'v0.0.64'), ('darknet53', '0550', '1d0b4f620071d544a3a1ccbe437267b8925ecc63', 'v0.0.501'), ('irevnet301', '0735', '4378cc1061da0377a3c7fc00c6be49c3875c997a', 'v0.0.564'), ('bagnet9', '2536', '6b5d69dac8fd780e9ac26b2b964370445662c610', 'v0.0.553'), ('bagnet17', '1523', 'cc7b9a74ec742972fc289686bdaf5b1f5731121f', 'v0.0.558'), ('bagnet33', '1041', '3b20fa3a31aa08f216d7091749f59154b6357397', 'v0.0.561'), ('dla34', '0705', '557c5f4f1db66481af6101c75d9ba52b486eda25', 'v0.0.486'), ('dla46c', '1286', '5b38b67fecf2d701b736eb23e1301b6dd7eb5fb9', 'v0.0.282'), ('dla46xc', '1225', 'e570f5f00a098b0de34e657f9d8caeda524d39f3', 'v0.0.293'), ('dla60', '0554', '88b141c4ca81598dbe4333bd4ddd5a1554772348', 'v0.0.494'), ('dla60x', '0553', '58924af84faf6ac0d980b047ccb505619925d97d', 'v0.0.493'), ('dla60xc', '1074', '1b4e4048847e1ba060eb76538ee09e760f40be11', 'v0.0.289'), ('dla102', '0517', '0e5d954cee3bb260a679425524deab61871bd0bf', 'v0.0.505'), ('dla102x', '0470', 'ea82787ce2b3c79cff32ccf54461507cc09c58ce', 'v0.0.528'), ('dla102x2', '0423', 'dde259b3525fb6dae7e445f6e534a13159b3b365', 'v0.0.542'), ('dla169', '0460', '71971da4e881084851c9b581e719ce4edfb8292d', 'v0.0.539'), ('fishnet150', '0466', 'ed21862d14bb0a4e9d1723b7c639e455a7b141c0', 'v0.0.502'), ('espnetv2_wd2', '1971', '2821c339eb68154129e2864f1b5d74d876a08d77', 'v0.0.567'), ('espnetv2_w1', '1383', '3e031b749057f0237b3ebd7954b175b4bb13200f', 'v0.0.569'), ('espnetv2_w5d4', '1226', '2a498c6d631c54dfacdd5cf9b4c6a07a6db82b2d', 'v0.0.564'), ('espnetv2_w3d2', '1086', 'e869dabdd427a4ab38016d375d5e1b2f29758cde', 'v0.0.566'), ('espnetv2_w2', '0941', 'ef6b1cc026609da7a8df81360e36ecb04985ad28', 'v0.0.566'), ('dicenet_wd5', '3051', '46077eac47e2a299dc1d4edc0128bd3cbd8e9ea9', 'v0.0.563'), ('dicenet_wd2', '2308', 'c03d6d00b8f9ecc8ad87c1ee4bd82a0291a0af06', 'v0.0.561'), ('dicenet_w3d4', '1618', 'eefcffc69e24e6ba0df06db03d0fef8be6afdfe1', 'v0.0.567'), ('dicenet_w1', '1411', '93e2e0df3b9cb62c8ba08d3f3d61237e20fce9e2', 'v0.0.513'), ('dicenet_w5d4', '1251', '710737608fb6e2282cdd3903d4a703eb1989b594', 'v0.0.515'), ('dicenet_w3d2', '1144', 'bb2a902d7adbe96472157e131153f48dfea9f37f', 'v0.0.522'), ('dicenet_w7d8', '1081', '37b9276b8b97c87554652d6963a566cbd189b054', 'v0.0.527'), ('dicenet_w2', '0915', 'c301319d2d2867f155227389ea7402106858c6d4', 'v0.0.569'), ('hrnet_w18_small_v1', '0873', '1060c1c562770adb94115de1cad797684dbb5703', 'v0.0.492'), ('hrnet_w18_small_v2', '0602', '1f319e1b1990818754b97104467cfecf3401ae4e', 'v0.0.499'), ('hrnetv2_w18', '0500', '0a70090294c726c1f303e9114d23c9373695eb44', 'v0.0.508'), ('hrnetv2_w30', '0508', 'eb6afc492537295615dc465236934a9a970c1bef', 'v0.0.525'), ('hrnetv2_w32', '0496', '9c6c47dcce100ff006509ed559158862a53436a6', 'v0.0.528'), ('hrnetv2_w40', '0481', '2382694cceff60b32b4702e64945706f183e3e1c', 'v0.0.534'), ('hrnetv2_w44', '0486', 'f5882024ac9513c2d6c2b41bc774862e392f69ad', 'v0.0.541'), ('hrnetv2_w48', '0484', '936bb78f8c187ca3cc96c7564fa3ff8ad9cb68a6', 'v0.0.541'), ('hrnetv2_w64', '0478', '4c18c51411db532c3b8168f0a1ae3e461e703d68', 'v0.0.543'), ('vovnet27s', '0980', '2a44d455a86e740bd1be61d204553af213ca3ac6', 'v0.0.551'), ('vovnet39', '0548', '20b60ee6bb8c59c684f8ffed7fbda3d76b1a7280', 'v0.0.493'), ('vovnet57', '0510', 'ed3cad779b94c4f4f531d6c37f90f1a5209d3b1c', 'v0.0.505'), ('selecsls42b', '0596', 'f5a35c74880fbe94fbe3770a96968fc81186b42b', 'v0.0.493'), ('selecsls60', '0511', '960edec5159b56bdba4ce606df203c8ce14cb8ba', 'v0.0.496'), ('selecsls60b', '0537', '7f83801b1c158502d93a4523bdecda43017448d5', 'v0.0.495'), ('hardnet39ds', '0864', '72e8423ee0b496c10b48ae687a417385d9667394', 'v0.0.485'), ('hardnet68ds', '0738', '012bf3ac31f38c78d5cdef1367cc3c27447236af', 'v0.0.487'), ('hardnet68', '0702', '6667306df893290d38510a76fd42ec82e05e2f68', 'v0.0.557'), ('hardnet85', '0572', '3baa0a7dd204196fd1afa631cf6de082a5cc0a36', 'v0.0.495'), ('squeezenet_v1_0', '1734', 'e6f8b0e8253cef1c5c071dfaf2df5fdfc6a64f8c', 'v0.0.128'), ('squeezenet_v1_1', '1739', 'd7a1483aaa1053c7cd0cf08529b2b87ed2781b35', 'v0.0.88'), ('squeezeresnet_v1_0', '1767', '66474b9b6a771055b28c37b70621c026a1ef6ef4', 'v0.0.178'), ('squeezeresnet_v1_1', '1784', '26064b82773e7a7175d6038976a73abfcd5ed2be', 'v0.0.70'), ('sqnxt23_w1', '1866', '73b700c40de5f7be9d2cf4ed30cc8935c670a3c3', 'v0.0.171'), ('sqnxt23v5_w1', '1743', '7a83722e7d362cef950d8534020f837caf9e6314', 'v0.0.172'), ('sqnxt23_w3d2', '1321', '4d733bcd19f1e502ebc46b52f0b69d959636902e', 'v0.0.210'), ('sqnxt23v5_w3d2', '1268', '4f98bbd3841d8d09a067100841f64ce3eccf184a', 'v0.0.212'), ('sqnxt23_w2', '1063', '95d9b55a5e857298bdb7974db6e3dbd9ecc94401', 'v0.0.240'), ('sqnxt23v5_w2', '1024', '707246f323bc95d0ea2d5608e9e85ae9fe59773a', 'v0.0.216'), ('shufflenet_g1_wd4', '3677', 'ee58f36811d023e1b2e651469c470e588c93f9d3', 'v0.0.134'), ('shufflenet_g3_wd4', '3617', 'bd08e3ed6aff4993cf5363fe8acaf0b22394bea0', 'v0.0.135'), ('shufflenet_g1_wd2', '2238', 'f77dcd18d3b759a3046bd4a2443c40e4ff455313', 'v0.0.174'), ('shufflenet_g3_wd2', '2060', 'ea6737a54bce651a0e8c0b533b982799842cb1c8', 'v0.0.167'), ('shufflenet_g1_w3d4', '1675', '2f1530aa72ee04e3599c5296b590a835d9d50e7f', 'v0.0.218'), ('shufflenet_g3_w3d4', '1609', 'e008e926f370af28e587f349384238d240a0fc02', 'v0.0.219'), ('shufflenet_g1_w1', '1350', '01934ee8f4bf7eaf4e36dd6442debb84ca2a2849', 'v0.0.223'), ('shufflenet_g2_w1', '1332', 'f5a1479fd8523032ee17a4de00fefd33ff4d31e6', 'v0.0.241'), ('shufflenet_g3_w1', '1329', 'ac58d62c5f277c0e9e5a119cc1f48cb1fcfc8306', 'v0.0.244'), ('shufflenet_g4_w1', '1310', '73c039ebf56f9561dd6eecc4cbad1ab1db168ed1', 'v0.0.245'), ('shufflenet_g8_w1', '1319', '9a50ddd9ce67ec697e3ed085d6c39e3d265f5719', 'v0.0.250'), ('shufflenetv2_wd2', '1830', '156953de22d0e749c987da4a58e0e53a5fb18291', 'v0.0.90'), ('shufflenetv2_w1', '1123', '27435039ab7794c86ceab11bd93a19a5ecab78d2', 'v0.0.133'), ('shufflenetv2_w3d2', '0913', 'f132506c9fa5f0eb27398f9936b53423d0cd5b66', 'v0.0.288'), ('shufflenetv2_w2', '0823', '2d67ac62057103fd2ed4790ea0058e0922abdd0f', 'v0.0.301'), ('shufflenetv2b_wd2', '1782', '845a9c43cf4a9873f89c6116634e74329b977e64', 'v0.0.157'), ('shufflenetv2b_w1', '1101', 'f679702f7c626161413320160c6c9c199de9b667', 'v0.0.161'), ('shufflenetv2b_w3d2', '0879', '4022da3a5922127b1acf5327bd9f1d4d55726e05', 'v0.0.203'), ('shufflenetv2b_w2', '0810', '7429df751916bf24bd7fb86bc137ae36275b9d19', 'v0.0.242'), ('menet108_8x1_g3', '2030', 'aa07f925180834389cfd3bf50cb22d2501225118', 'v0.0.89'), ('menet128_8x1_g4', '1913', '0c890a76fb23c0af50fdec076cb16d0f0ee70355', 'v0.0.103'), ('menet160_8x1_g8', '2028', '4f28279a94e631f6a51735de5ea29703cca69845', 'v0.0.154'), ('menet228_12x1_g3', '1289', '2dc2eec7c9ebb41c459450e1843503b5ac7ecb3a', 'v0.0.131'), ('menet256_12x1_g4', '1216', '7caf63d15190648e266a4e7520c3ad677716f388', 'v0.0.152'), ('menet348_12x1_g3', '0936', '62c72b0b56460f062d4da7155bd64a524f42fb88', 'v0.0.173'), ('menet352_12x1_g8', '1167', '5892fea4e44eb27814a9b092a1a06eb81cea7844', 'v0.0.198'), ('menet456_24x1_g3', '0780', '7a89b32c89f878ac63fc96ddc71cb1a5e91c84d6', 'v0.0.237'), ('mobilenet_wd4', '2218', '3185cdd29b3b964ad51fdd7820bd65f091cf281f', 'v0.0.62'), ('mobilenet_wd2', '1330', '94f13ae1375b48892d8ecbb4a253bb583fe27277', 'v0.0.156'), ('mobilenet_w3d4', '1051', '6361d4b4192b5fc68f3409100d825e8edb28876b', 'v0.0.130'), ('mobilenet_w1', '0865', 'eafd91e9369abb09726f2168aba24453b17fc22e', 'v0.0.155'), ('mobilenetb_wd4', '2165', '2070764e0b3be74922eb5fa0a4342c693821ba90', 'v0.0.481'), ('mobilenetb_wd2', '1271', '799ef980b2726a77d4b68d99f520d9d6bc7d86dc', 'v0.0.480'), ('mobilenetb_w3d4', '1020', 'b01c8bacda6f8e26b34a0313a4dc3883511760f7', 'v0.0.481'), ('mobilenetb_w1', '0788', '82664eb4c1f2ddd0ac163f50263237f7667223f3', 'v0.0.489'), ('fdmobilenet_wd4', '3053', 'd4f18e5b4ed63e5426eafbf5db7f8e2a97c28581', 'v0.0.177'), ('fdmobilenet_wd2', '1969', '242b9fa82d54f54f08b4bdbb194b7c89030e7bc4', 'v0.0.83'), ('fdmobilenet_w3d4', '1601', 'cb10c3e129706d3023d752e7402965af08f91ca7', 'v0.0.159'), ('fdmobilenet_w1', '1312', '95fa0092aac013c88243771faf66ef1134b7574d', 'v0.0.162'), ('mobilenetv2_wd4', '2412', 'd92b5b2dbb52e27354ddd673e6fd240a0cf27175', 'v0.0.137'), ('mobilenetv2_wd2', '1442', 'd7c586c716e3ea85e793f7c5aaf9cae2a907117b', 'v0.0.170'), ('mobilenetv2_w3d4', '1044', '768454f4bdaae337c180bb81248b8c5b8d31040b', 'v0.0.230'), ('mobilenetv2_w1', '0864', '6e58b1cb96852e4c6de6fc9cd11241384af21df9', 'v0.0.213'), ('mobilenetv2b_wd4', '2338', '77ba7e8d41542d311e240dab75e4d29fa0677fb9', 'v0.0.483'), ('mobilenetv2b_wd2', '1373', '3bfc8a592a0881c2cb025f52a09fb5057a7896be', 'v0.0.486'), ('mobilenetv2b_w3d4', '1064', '5d4dc4e5622043697382272183a3d0bd43dbc218', 'v0.0.483'), ('mobilenetv2b_w1', '0884', 'ab0ea3993e7c533f0aea5793331dc6302d715e9c', 'v0.0.483'), ('mobilenetv3_large_w1', '0729', 'db741a9938acc8a5fd9544aaf53b41aebb98e021', 'v0.0.491'), ('igcv3_wd4', '2830', '71abf6e0b6bff1d3a3938bfea7c752b59ac05e9d', 'v0.0.142'), ('igcv3_wd2', '1703', '145b7089e1d0e0ce88f17393a357d5bb4ae37734', 'v0.0.132'), ('igcv3_w3d4', '1096', '3c7c86fc43df2e5cf95a451ebe07fccf2d9dc076', 'v0.0.207'), ('igcv3_w1', '0900', 'e2c3da1cffd8e42da7a052b80db2f86758c8d35b', 'v0.0.243'), ('mnasnet_b1', '0723', 'a6f74cf912fa5b1ee4bb9825f3143a0c1ced03be', 'v0.0.493'), ('mnasnet_a1', '0705', '3efe98a3bd6ea0a8cb0902a4dba424c134145d5f', 'v0.0.486'), ('darts', '0756', 'c2c7c33ba60d1052f95bcae72128fc47b1214cff', 'v0.0.485'), ('proxylessnas_cpu', '0750', '256da7c8a05cd87a59e30e314b22dc1d4565946e', 'v0.0.324'), ('proxylessnas_gpu', '0724', 'd9ce80964e37fb30bddcc552f1d68361b1a94873', 'v0.0.333'), ('proxylessnas_mobile', '0780', 'b8bb5a64f333562475dcfc09eeb7e603d6e66afb', 'v0.0.326'), ('proxylessnas_mobile14', '0651', 'f08baec85343104994b821581cde3ee965a2c593', 'v0.0.331'), ('fbnet_cb', '0761', '3db688f2fa465bc93bc546b4432389fe33aec5b3', 'v0.0.486'), ('xception', '0511', '9755eb77589d627d816fe9806a2b2aa931b6c2f6', 'v0.0.544'), ('inceptionv3', '0533', '4a05a7ffcb5963d44678933596ef85a0f724ebab', 'v0.0.552'), ('inceptionv4', '0488', 'a828ae6ccccb3178257d20e5f17cc8c6ac4348cc', 'v0.0.543'), ('inceptionresnetv1', '0481', '3cc3b4959349e56d6b29619b07ae020e19039f67', 'v0.0.552'), ('inceptionresnetv2', '0470', '4ea29355ae39ececb733e949dfdcbdff72272559', 'v0.0.547'), ('polynet', '0453', '742803144e5a2a6148212570726350da09adf3f6', 'v0.0.96'), ('nasnet_4a1056', '0790', 'f89dd74f47e42c35c9a1182f248df1d319524db7', 'v0.0.495'), ('nasnet_6a4032', '0424', '73cca5fee009db77412c5fca7c826b3563752757', 'v0.0.101'), ('pnasnet5large', '0428', '998a548f44ac1b1ac6c4959a721f2675ab5c48b9', 'v0.0.114'), ('spnasnet', '0776', '09cc881e024d69ab3bd88bef332becb6147a0651', 'v0.0.490'), ('efficientnet_b0', '0722', '041a8346ad6a13dbb66384d9a41ed20d959d3e77', 'v0.0.364'), ('efficientnet_b1', '0626', '455dcb2a05a1295c8f6b728d4f6c72c507e1369a', 'v0.0.376'), ('efficientnet_b0b', '0670', '8892ba581b0d81dbc8fe56c49d81d3dd007b1db8', 'v0.0.403'), ('efficientnet_b1b', '0565', 'c29a1b67804b70856ed3cc329256b80e3afc04ad', 'v0.0.403'), ('efficientnet_b2b', '0516', '7532826e5c14f7ade9e8ea9ac92044d817575b06', 'v0.0.403'), ('efficientnet_b3b', '0431', '1e342ec2160e6de813f9bd4d0ab9ce3b5749780e', 'v0.0.403'), ('efficientnet_b4b', '0376', 'b60e177974539fc6dcc9fef3b90591bfc9949514', 'v0.0.403'), ('efficientnet_b5b', '0334', 'cd70ae717ddca72430efe91cf7a3c4e28bcd61ac', 'v0.0.403'), ('efficientnet_b6b', '0312', 'f581d9f046032e28e532082fa49bfd373952db4f', 'v0.0.403'), ('efficientnet_b7b', '0311', '2b8a6040588aea44b57df89e2d9239d906737508', 'v0.0.403'), ('efficientnet_b0c', '0646', '81eabd2992ba7bb80c1c1a7e20373e7c65aa1286', 'v0.0.433'), ('efficientnet_b1c', '0555', '10b5589de6ee9af3c67f0ae35a424b4adc0a9e35', 'v0.0.433'), ('efficientnet_b2c', '0489', '6f649ece72d0334e5191da78767f2ac9149e85f0', 'v0.0.433'), ('efficientnet_b3c', '0434', 'e1e2a1b7f3457bdd8f46a5858472ad4acd3d2362', 'v0.0.433'), ('efficientnet_b4c', '0359', 'cdb2012d6688b0208527d36817589095a1db1031', 'v0.0.433'), ('efficientnet_b5c', '0302', '3240f368eb5e7ed78b9ac0d68800fbea9b220e9b', 'v0.0.433'), ('efficientnet_b6c', '0285', 'e71a1ccc3e876a7d64818c83518dd0e64c630d3e', 'v0.0.433'), ('efficientnet_b7c', '0277', 'feea7daf3478645131c94319c71141df559dce19', 'v0.0.433'), ('efficientnet_b8c', '0270', '050ec6358583d0a48c74b02563474ecc1d9dacba', 'v0.0.433'), ('efficientnet_edge_small_b', '0629', '5b398abc73c4c4870d88c62452ff919bec2440c9', 'v0.0.434'), ('efficientnet_edge_medium_b', '0553', '0b3c86d49b3684d19e6589030aeceb918faa648c', 'v0.0.434'), ('efficientnet_edge_large_b', '0477', '055436da0fa440933b906c528cb34b200e28f73c', 'v0.0.434'), ('mixnet_s', '0703', '135aa0426712a9f60afe2cfff2df9607f3fc2d68', 'v0.0.493'), ('mixnet_m', '0631', '0881aba9281e3cae3e7cddd75d385c3cccdf7e25', 'v0.0.493'), ('mixnet_l', '0557', '94cbf10cb34c3e3940117d86c28bcc6dbfe9e005', 'v0.0.500'), ('resneta10', '1159', 'a66e01d9f567ce747e7c255adac848a33c54a3a5', 'v0.0.484'), ('resnetabc14b', '0956', '6f8c36067feb2d27b5e0813bf18ef29038db0d70', 'v0.0.477'), ('resneta18', '0802', '225dd3ae0eac3ce3815732ac4ab0e25df709a75c', 'v0.0.486'), ('resneta50b', '0534', '28eff48a72d892802dde424db3fd0e1a9c12be16', 'v0.0.492'), ('resneta101b', '0442', 'cdf5f7ac2ad291cdc35b351e9220a490c8cc996c', 'v0.0.532'), ('resneta152b', '0424', 'deaaaabf413821eb543621826c368493b9041438', 'v0.0.524'), ('resnetd50b', '0549', '17d6004b5c6c1b97cfb47377ae5076810c5d88be', 'v0.0.296'), ('resnetd101b', '0461', 'fead1bcb86bba2be4ed7f0033fa972dc613e3280', 'v0.0.296'), ('resnetd152b', '0467', 'd0fe2fe09c6462de17aca4a72bbcb08b76a66e02', 'v0.0.296'), ('nin_cifar10', '0743', '9696dc1a8f67e7aa233836bcbdb99625769b1e86', 'v0.0.175'), ('nin_cifar100', '2839', 'eed0e9af2cd8e5aa77bb063204525812dbd9190f', 'v0.0.183'), ('nin_svhn', '0376', '7cb750180b0a981007194461bf57cfd90eb59c88', 'v0.0.270'), ('resnet20_cifar10', '0597', '13c5ab19145591d75873da3497be1dd1bd2afd46', 'v0.0.163'), ('resnet20_cifar100', '2964', '4e1443526ee96648bfe4d4954871a97a9c9622f4', 'v0.0.180'), ('resnet20_svhn', '0343', '7ac0d94a4563c9611092ce08f2124a3828103139', 'v0.0.265'), ('resnet56_cifar10', '0452', 'a73e63e9d0f3f7adde59b4142323c0dd05930de7', 'v0.0.163'), ('resnet56_cifar100', '2488', '590977100774a289b91088245dd2bd0cbe6567e6', 'v0.0.181'), ('resnet56_svhn', '0275', 'e676e4216a771b7d0339e87284c7ebb03af8ed25', 'v0.0.265'), ('resnet110_cifar10', '0369', 'f89f1c4d9fdd9e5cd00949a872211376979ff703', 'v0.0.163'), ('resnet110_cifar100', '2280', '6c5fa14bb4ced2dffe6ee1536306687aae57f9cb', 'v0.0.190'), ('resnet110_svhn', '0245', '0570b5942680cf88c66ae9a76c0e7ff0a41e71a6', 'v0.0.265'), ('resnet164bn_cifar10', '0368', 'e7941eeeddef9336664522eaa3af92d77128cac0', 'v0.0.179'), ('resnet164bn_cifar100', '2044', 'c7db7b5e6fbe6dc0f9501d25784f1a107c6e0315', 'v0.0.182'), ('resnet164bn_svhn', '0242', '8cdce67452d2780c7c69f4d0b979e80189d4bff8', 'v0.0.267'), ('resnet272bn_cifar10', '0333', '99dc36ca2abc91f3f82db181a14c5364cd5526be', 'v0.0.368'), ('resnet272bn_cifar100', '2007', '088af5c23634fe75206081d946fc82fdc9e999ad', 'v0.0.368'), ('resnet272bn_svhn', '0243', '39d741c8d081ebd2266a114e82363839ffdf8ebb', 'v0.0.368'), ('resnet542bn_cifar10', '0343', 'e687b254e1eace223ceef39ad17106e61b8649ba', 'v0.0.369'), ('resnet542bn_cifar100', '1932', 'df8bd5264c1db11dd545f62e9c750c7976edccc9', 'v0.0.369'), ('resnet542bn_svhn', '0234', '4f78075cbcba196fc8f5297b71730906c1bf7d8a', 'v0.0.369'), ('resnet1001_cifar10', '0328', 'bb979d53089138b5060b418cad6c8ad9a940bf81', 'v0.0.201'), ('resnet1001_cifar100', '1979', '692d9516620bc8b7a4da30a98ebcb7432243f5e9', 'v0.0.254'), ('resnet1001_svhn', '0241', '031fb0ce5e5ddbebca2fd7d856d63ddd147fe933', 'v0.0.408'), ('resnet1202_cifar10', '0353', '377510a63595e544333f6f57523222cd845744a8', 'v0.0.214'), ('resnet1202_cifar100', '2156', '1d94f9ccdd81e1785ea6ec02a861a4a05f39e5c9', 'v0.0.410'), ('preresnet20_cifar10', '0651', 'daa895737a34edda75c40f2d8566660590c84a3f', 'v0.0.164'), ('preresnet20_cifar100', '3022', '37f15365d48768f792f4551bd6ccf5259bc70530', 'v0.0.187'), ('preresnet20_svhn', '0322', '608cee12c0bc3cb59feea96386f6c12c6da91ba5', 'v0.0.269'), ('preresnet56_cifar10', '0449', 'cb37cb9d4524d4e0f5724aeed9face455f527efc', 'v0.0.164'), ('preresnet56_cifar100', '2505', '4c39e83f567f15d6ee0d69bf2dcaccd62067dfe5', 'v0.0.188'), ('preresnet56_svhn', '0280', 'b974c2c96a18ff2278f1d33df58c8537f9139ed9', 'v0.0.269'), ('preresnet110_cifar10', '0386', 'd6d4b7bd9f154eca242482a7559413d5c7b6d465', 'v0.0.164'), ('preresnet110_cifar100', '2267', '18cf4161c67c03e50cff7eb30988a559f3f97260', 'v0.0.191'), ('preresnet110_svhn', '0279', '6804450b744fa922d9ec22aa4c792d3a5da812f6', 'v0.0.269'), ('preresnet164bn_cifar10', '0364', '7ecf30cb818f80908ef4a77af4660c1080d0df81', 'v0.0.196'), ('preresnet164bn_cifar100', '2018', 'a20557c8968c04d8d07e40fdc5b0d1ec1fb3339d', 'v0.0.192'), ('preresnet164bn_svhn', '0258', '4aeee06affea89767c058fe1650b7476f05d8563', 'v0.0.269'), ('preresnet272bn_cifar10', '0325', '944ba29df55afcf2789399de552e91578edd4295', 'v0.0.389'), ('preresnet272bn_cifar100', '1963', '38e296beff0cd92697235c717a801ec422cdafe3', 'v0.0.389'), ('preresnet272bn_svhn', '0234', '7ff97873447fbfb1d823fd43439e34351f149c13', 'v0.0.389'), ('preresnet542bn_cifar10', '0314', 'ac40a67bb3b7f02179ff3c4fa0d6533ff3e2dd9f', 'v0.0.391'), ('preresnet542bn_cifar100', '1871', 'd536ad01fc40fe19605a3409efb995bf8593aa29', 'v0.0.391'), ('preresnet542bn_svhn', '0236', '3a4633f14e96cce30086ef4149b52c0b7cbccce6', 'v0.0.391'), ('preresnet1001_cifar10', '0265', '50507ff74b6047abe6d04af6471d9bacafa05e24', 'v0.0.209'), ('preresnet1001_cifar100', '1841', '185e033d77e61cec588196e3fe8bf8dcb43acfab', 'v0.0.283'), ('preresnet1202_cifar10', '0339', '942cf6f22d80b5428256825234a252b8d6ebbe9d', 'v0.0.246'), ('resnext20_1x64d_cifar10', '0433', '0661d12e534a87bd5f3305d541afce5730d45492', 'v0.0.365'), ('resnext20_1x64d_cifar100', '2197', 'e7073542469be79876e3b3aeccb767638a136b93', 'v0.0.365'), ('resnext20_1x64d_svhn', '0298', '3c7febc8eee0887ebc54d5e1a4411a46066f0624', 'v0.0.365'), ('resnext20_2x32d_cifar10', '0453', 'afb48ca4764efaff92e4cc50ec11566993180ac1', 'v0.0.365'), ('resnext20_2x32d_cifar100', '2255', '995281ee5b0d021bf6e9e984802e30a0042e0290', 'v0.0.365'), ('resnext20_2x32d_svhn', '0296', '54189677f599b4f42d2469ad75921165f84d82dc', 'v0.0.365'), ('resnext20_2x64d_cifar10', '0403', '6f0c138fe13a73c9f149053065f1f430761e75be', 'v0.0.365'), ('resnext20_2x64d_cifar100', '2060', '5f6dfa3ff5f1b0e3e1441f554a80c504c1991a03', 'v0.0.365'), ('resnext20_2x64d_svhn', '0283', '9c77f074dcd6333a859da6ddd833a662031406ea', 'v0.0.365'), ('resnext20_4x16d_cifar10', '0470', 'ae1ba8697ec0a62a58edb9c87ef0d9f290d3c857', 'v0.0.365'), ('resnext20_4x16d_cifar100', '2304', '2c9d578ab797ef887b3e8c51d897a8e720880bfd', 'v0.0.365'), ('resnext20_4x16d_svhn', '0317', '6691c8f56f7a63b5638408356a7f9fe4931d7ac9', 'v0.0.365'), ('resnext20_4x32d_cifar10', '0373', 'cf6960607fb51269cb80d6900eadefd16c6bb511', 'v0.0.365'), ('resnext20_4x32d_cifar100', '2131', '2c558efca7c66c412843603cff137e9565d7c016', 'v0.0.365'), ('resnext20_4x32d_svhn', '0298', '1da9a7bf6bed55f89b1c19a9b6ff97f7fa674c81', 'v0.0.365'), ('resnext20_8x8d_cifar10', '0466', '280e5f89e24ca765a771e733f53377ede244207f', 'v0.0.365'), ('resnext20_8x8d_cifar100', '2282', '363f03e8040da3a2f1cfe88df99481b23382e28c', 'v0.0.365'), ('resnext20_8x8d_svhn', '0318', 'c1536efbb2d7c78ee5b9323b43e3d0d56e9d2aa4', 'v0.0.365'), ('resnext20_8x16d_cifar10', '0404', '4d7f72818137674c3c17257427b0c5788b250320', 'v0.0.365'), ('resnext20_8x16d_cifar100', '2172', '3fc47c7072039fd5437ed276e92a0840c94fe82e', 'v0.0.365'), ('resnext20_8x16d_svhn', '0301', '41b28fd308e8195ee74050a2c56e2085552a55e0', 'v0.0.365'), ('resnext20_16x4d_cifar10', '0404', '426b5b2f994bc8154cb17957cf7d487b156d4fe2', 'v0.0.365'), ('resnext20_16x4d_cifar100', '2282', '508d32271ea1d34d40305a67d64cf1547ca58ae6', 'v0.0.365'), ('resnext20_16x4d_svhn', '0321', '854df3b71743fe871b5cd9b72ff6197b34c57b5e', 'v0.0.365'), ('resnext20_16x8d_cifar10', '0394', 'f81d05668bf293e5d84150ec54f94c4a50b79d9c', 'v0.0.365'), ('resnext20_16x8d_cifar100', '2173', 'a246aea51d1fd165c9f6db2aa528a2afa6a03dcd', 'v0.0.365'), ('resnext20_16x8d_svhn', '0293', '31f4b14e3113085a6b9a8e384a3131af764afe82', 'v0.0.365'), ('resnext20_32x2d_cifar10', '0461', '2d6ee8362c497b7782b91f85d5c9fd0f4ac0470e', 'v0.0.365'), ('resnext20_32x2d_cifar100', '2322', 'ce65201429a93cf9175c0e2b8601ceaf538a2438', 'v0.0.365'), ('resnext20_32x2d_svhn', '0327', '2499ff6de736f1241bc58422ae9d1d39d8ac1014', 'v0.0.365'), ('resnext20_32x4d_cifar10', '0420', 'a365893948daea8ef5b20b61bc3519fe199381af', 'v0.0.365'), ('resnext20_32x4d_cifar100', '2213', '5b2ffba877ea3dd0aa86d7c38ab5467d4ed856e6', 'v0.0.365'), ('resnext20_32x4d_svhn', '0309', 'ddbef9ac7da748e162ac886621833a5aef84cb86', 'v0.0.365'), ('resnext20_64x1d_cifar10', '0493', '6618e9ac6df3bc2614c23ba0e39ea0d0d5d21eec', 'v0.0.365'), ('resnext20_64x1d_cifar100', '2353', '9c789af45dd1455b4b7fe94a21709bb9dc8c411b', 'v0.0.365'), ('resnext20_64x1d_svhn', '0342', '2591ea440078f6c089513ace548ed79fb0f9bb76', 'v0.0.365'), ('resnext20_64x2d_cifar10', '0438', '32fe188b44b9e7f80ddec8024799be871dd63c4e', 'v0.0.365'), ('resnext20_64x2d_cifar100', '2235', '62fcc38a8e37550ace0f67927d2f5a74990de920', 'v0.0.365'), ('resnext20_64x2d_svhn', '0314', '4c01490b9566eeb2aaf27446ca8f5bac621894f0', 'v0.0.365'), ('resnext29_32x4d_cifar10', '0315', 'c8a1beda8ba616dc9af682d3ac172bfdd7a2472d', 'v0.0.169'), ('resnext29_32x4d_cifar100', '1950', '5f2eedcdd5cea6fdec1508f261f556a953ae28c2', 'v0.0.200'), ('resnext29_32x4d_svhn', '0280', 'dcb6aef96fbd76aa249e8f834093e2384b898404', 'v0.0.275'), ('resnext29_16x64d_cifar10', '0241', '76b97a4dd6185602a8ca8bdd77a70f8ddfcd4e83', 'v0.0.176'), ('resnext29_16x64d_cifar100', '1693', '1fcec90d6425e0405c61a1e90a80701ea556beca', 'v0.0.322'), ('resnext29_16x64d_svhn', '0268', 'c57307f3bf70d0f39d6cfb1dc2c82d8ef9e89603', 'v0.0.358'), ('resnext56_1x64d_cifar10', '0287', '8edd977c69cfec45604f1b0dbbe36fb2b65f122d', 'v0.0.367'), ('resnext56_1x64d_cifar100', '1825', 'b78642c10e21afc477c27076e2ab9ec6854cca13', 'v0.0.367'), ('resnext56_1x64d_svhn', '0242', '860c610caa23e742093d7c7677253ee5b917af44', 'v0.0.367'), ('resnext56_2x32d_cifar10', '0301', 'd0284dff7382d969c59e602f66b25a72b8d9e232', 'v0.0.367'), ('resnext56_2x32d_cifar100', '1786', '32205070280db6bd6431a6b7ad0537524af5a6f0', 'v0.0.367'), ('resnext56_2x32d_svhn', '0246', 'ffb8df9ba6a5053c79dbdecf7d0b6f1d248fcfdf', 'v0.0.367'), ('resnext56_4x16d_cifar10', '0311', 'add022e7dd21245720e789a0a16a7475d15b5702', 'v0.0.367'), ('resnext56_4x16d_cifar100', '1809', '366de7b5abba9b09d1ffde87e3e9f379db73d691', 'v0.0.367'), ('resnext56_4x16d_svhn', '0244', 'f7b697f924eeb64ed14d21f576a4a82007937bbe', 'v0.0.367'), ('resnext56_8x8d_cifar10', '0307', '4f0b72469f9890079471db355e9bfb405d443317', 'v0.0.367'), ('resnext56_8x8d_cifar100', '1806', '827a485e2d5f3154251d6453563a73a511028445', 'v0.0.367'), ('resnext56_8x8d_svhn', '0247', 'f0550cd045d9f4a4f8648f0c9c517421a34f4a11', 'v0.0.367'), ('resnext56_16x4d_cifar10', '0312', '93d71b610a356f8e285dde08ad4fe96d79efcadd', 'v0.0.367'), ('resnext56_16x4d_cifar100', '1824', '9cb7a1326870878315dccb50b0be51968d6ddfa7', 'v0.0.367'), ('resnext56_16x4d_svhn', '0256', '943386bd2e0fabba153955d80171016a95482609', 'v0.0.367'), ('resnext56_32x2d_cifar10', '0314', 'ea8b43351206a567f865a9a53ecae65790da8a60', 'v0.0.367'), ('resnext56_32x2d_cifar100', '1860', '3f65de935e532fe3c18324f645260983c61c0c76', 'v0.0.367'), ('resnext56_32x2d_svhn', '0253', 'ba8c809dcaa9f6f6e3b2e12dfd9df1b123d11be7', 'v0.0.367'), ('resnext56_64x1d_cifar10', '0341', '12a684ad58d1d3407d61b7de4b2cedc042225d20', 'v0.0.367'), ('resnext56_64x1d_cifar100', '1816', 'b80f4315d53963dd3d1ac856d3c2dbf660b836eb', 'v0.0.367'), ('resnext56_64x1d_svhn', '0255', '144bab62a5ed27bcb2465d9ab8b9b4ce0bf31dc5', 'v0.0.367'), ('resnext272_1x64d_cifar10', '0255', 'c1a3fddc4de9f6ebee2a588bd585ede68974da4f', 'v0.0.372'), ('resnext272_1x64d_cifar100', '1911', 'e0b3656a204c40e0b4ee6ade1e8d262c521e842d', 'v0.0.372'), ('resnext272_1x64d_svhn', '0235', '025ee7b915fa25d617e9774b0303f35488445d78', 'v0.0.372'), ('resnext272_2x32d_cifar10', '0274', '23b391ce58d9694ea35240ec3b76ce8b0ebf66b8', 'v0.0.375'), ('resnext272_2x32d_cifar100', '1834', '4802083b5fda38a4a30b8a2b1f24eb4b8fdf55ad', 'v0.0.375'), ('resnext272_2x32d_svhn', '0244', 'b65ddfe317dabceb4d4d7e910ca07c0c575ad9b8', 'v0.0.375'), ('seresnet20_cifar10', '0601', '3411e5ad1060975c45fe6d0d836755a92e3bb27c', 'v0.0.362'), ('seresnet20_cifar100', '2854', '184ad148171fd9244bf8570eee6647a996678ab4', 'v0.0.362'), ('seresnet20_svhn', '0323', 'a3a3c67731eb8bea0cd3af3b8b2f88c1cc70987e', 'v0.0.362'), ('seresnet56_cifar10', '0413', '21bac136e4cac21abb3e08b60254c73b16f0190f', 'v0.0.362'), ('seresnet56_cifar100', '2294', '989d4d9227c4fc33440c267f1e1ac324fd246ad4', 'v0.0.362'), ('seresnet56_svhn', '0264', '63a155acd6407b5e96516b86f6a7cd9e6855c372', 'v0.0.362'), ('seresnet110_cifar10', '0363', 'fa3f09a88d24282e938488c4588968e273770605', 'v0.0.362'), ('seresnet110_cifar100', '2086', '5345be4166268ce2cd44d88eb9edba9f86ccc864', 'v0.0.362'), ('seresnet110_svhn', '0235', 'd129498ad625983d92048f32b80de5d16987779a', 'v0.0.362'), ('seresnet164bn_cifar10', '0339', '11c923152587746a5539a9e4f140db847b9b61c1', 'v0.0.362'), ('seresnet164bn_cifar100', '1995', '6c9dc66b86de6be67df1e59d4aff4a592d7d98b8', 'v0.0.362'), ('seresnet164bn_svhn', '0245', 'd97ea6c83b0fd1da3a976d518977b19bd466d015', 'v0.0.362'), ('seresnet272bn_cifar10', '0339', 'da4073add21614f22b231d8663867007f3f2312d', 'v0.0.390'), ('seresnet272bn_cifar100', '1907', '754af9375f060f55f8c3393a70659df460b8d47d', 'v0.0.390'), ('seresnet272bn_svhn', '0238', '9ffe8acad2a03cd98f6746d3f1528fe5f294aea4', 'v0.0.390'), ('seresnet542bn_cifar10', '0347', 'e64d9ca4b98349973b802572af6625879bc3c4a4', 'v0.0.385'), ('seresnet542bn_cifar100', '1887', 'cd76c769c06b886d2a94268c95691dfc905bac64', 'v0.0.385'), ('seresnet542bn_svhn', '0226', '05ce3771c46aa8b7af7c2056398f83b2b9f116db', 'v0.0.385'), ('sepreresnet20_cifar10', '0618', 'e55551e6e35d04fe8e35d24a5e3d608a08e8dfa2', 'v0.0.379'), ('sepreresnet20_cifar100', '2831', 'ee5d3bd66ce643950e54d54825036b98b13b31cd', 'v0.0.379'), ('sepreresnet20_svhn', '0324', 'd5bb6768cc3134137371a832c5ebc289d982a8db', 'v0.0.379'), ('sepreresnet56_cifar10', '0451', '56c299345242bcfca52c83f77cfff80b7058b1fe', 'v0.0.379'), ('sepreresnet56_cifar100', '2305', '313a7a30a129c174ecaf3de6a94a5fd34dc8d711', 'v0.0.379'), ('sepreresnet56_svhn', '0271', 'f556af3db771e6e4e5b847bd4eddef71b879b8d0', 'v0.0.379'), ('sepreresnet110_cifar10', '0454', '67eea1cc03f76ee0054d39f004ab10f7a70978bb', 'v0.0.379'), ('sepreresnet110_cifar100', '2261', '3291a56be67afd1154d9d2d05e2e1411c12dcb4a', 'v0.0.379'), ('sepreresnet110_svhn', '0259', '5c09cacbcf786e18509947c40de695883d6b3328', 'v0.0.379'), ('sepreresnet164bn_cifar10', '0373', 'ac72ac7fa9d78e66a719717f922f738bbe7f9699', 'v0.0.379'), ('sepreresnet164bn_cifar100', '2005', 'd93993672a414e8f19cb12d2489be930e8605b8f', 'v0.0.379'), ('sepreresnet164bn_svhn', '0256', 'a45d1a65092900bb768969c52f26c61292838caa', 'v0.0.379'), ('sepreresnet272bn_cifar10', '0339', '3e47d575280a70c1726d6b2eb8d0d7c069a3e472', 'v0.0.379'), ('sepreresnet272bn_cifar100', '1913', 'd243b0580717b197bf9eb32a8959c83d21a3124f', 'v0.0.379'), ('sepreresnet272bn_svhn', '0249', '34b910cdfa34bf318f58c3312bc074059c7c669f', 'v0.0.379'), ('sepreresnet542bn_cifar10', '0308', '05f7d4a6bfb1af1825b734eac99b3b46dd8c4b91', 'v0.0.382'), ('sepreresnet542bn_cifar100', '1945', '4dd0e21d02fef2ae1fe3fb3a8cd8c72db11bb685', 'v0.0.382'), ('sepreresnet542bn_svhn', '0247', '456035daf4daecd909e957da090078deba6cb449', 'v0.0.382'), ('pyramidnet110_a48_cifar10', '0372', '35b94d0575c2081a142e71955c8ceea8c51ec5e5', 'v0.0.184'), ('pyramidnet110_a48_cifar100', '2095', '00fd42a00492b2bbb28cacfb7b1a6c63072c37a3', 'v0.0.186'), ('pyramidnet110_a48_svhn', '0247', 'd8a5c6e20b6cc01989a52f9e307caf640169ed0a', 'v0.0.281'), ('pyramidnet110_a84_cifar10', '0298', '81710d7ab90838a8a299bf5f50aed2a3fa41f0e3', 'v0.0.185'), ('pyramidnet110_a84_cifar100', '1887', '6712d5dc69452f2fde1fcc3ee32c3164dcaffc4e', 'v0.0.199'), ('pyramidnet110_a84_svhn', '0243', '473cc640c4ad0a1642500c84f2ef848498d12a37', 'v0.0.392'), ('pyramidnet110_a270_cifar10', '0251', '1e769ce50ef915a807ee99907912c87766fff60f', 'v0.0.194'), ('pyramidnet110_a270_cifar100', '1710', '2732fc6430085192189fd7ccfd287881cc5a6c0d', 'v0.0.319'), ('pyramidnet110_a270_svhn', '0238', '034be5421b598e84f395f421182f664495ca62ca', 'v0.0.393'), ('pyramidnet164_a270_bn_cifar10', '0242', 'c4a79ea3d84344b9d352074122e37f593ee98fd2', 'v0.0.264'), ('pyramidnet164_a270_bn_cifar100', '1670', '08f46c7ff99e9c3fd7b5262e34dc8a00b316646f', 'v0.0.312'), ('pyramidnet164_a270_bn_svhn', '0233', '27b67f1494ed508e0192e3e0f09ac86e32e1e734', 'v0.0.396'), ('pyramidnet200_a240_bn_cifar10', '0244', '52f4d43ec4d952f847c3a8e0503d5a4e6286679c', 'v0.0.268'), ('pyramidnet200_a240_bn_cifar100', '1609', 'e61e7e7eb6675aaf7a18461fea9bb3a53538d43b', 'v0.0.317'), ('pyramidnet200_a240_bn_svhn', '0232', '02bf262e70d9b3ce8038255d0abdec2bc5161f6d', 'v0.0.397'), ('pyramidnet236_a220_bn_cifar10', '0247', '1bd295a7fb834f639b238ffee818b3bde4126c81', 'v0.0.285'), ('pyramidnet236_a220_bn_cifar100', '1634', 'f066b3c6a4d217c42f5e8872fe23d343afe378ec', 'v0.0.312'), ('pyramidnet236_a220_bn_svhn', '0235', '1a0c0711f013035c0e05145501c93fa2519603ea', 'v0.0.398'), ('pyramidnet272_a200_bn_cifar10', '0239', 'd7b23c5460f059ac82ebc7b2cd992a203e098476', 'v0.0.284'), ('pyramidnet272_a200_bn_cifar100', '1619', '486e942734d91cd62d6bcbc283e1d7b56b734507', 'v0.0.312'), ('pyramidnet272_a200_bn_svhn', '0240', 'dcd9af34f57708f598bba723824bf3525f6e42c7', 'v0.0.404'), ('densenet40_k12_cifar10', '0561', '28dc0035549e51dcb53d1360707bd6f1558a5dcd', 'v0.0.193'), ('densenet40_k12_cifar100', '2490', '908f02ba7dbd7b8138f264193189e762a5590b1c', 'v0.0.195'), ('densenet40_k12_svhn', '0305', '645564c186a4e807293a68fb388803e36916e7b2', 'v0.0.278'), ('densenet40_k12_bc_cifar10', '0643', '7fdeda31c5accbddf47ab0f0b9a32cff723bf70d', 'v0.0.231'), ('densenet40_k12_bc_cifar100', '2841', '35cd8e6a2ae0896a8af2b689e076057fa19efa9b', 'v0.0.232'), ('densenet40_k12_bc_svhn', '0320', '6f2f98243fac9da22be26681bcd0a4d08e0f4baf', 'v0.0.279'), ('densenet40_k24_bc_cifar10', '0452', '13fa807e095b44ecaf3882e488b33a890d9d1e29', 'v0.0.220'), ('densenet40_k24_bc_cifar100', '2267', '2c4ef7c4bbe7f64784ad18b3845f4bf533f2ce57', 'v0.0.221'), ('densenet40_k24_bc_svhn', '0290', '03e136dd71bc85966fd2a4cb15692cfff3886df2', 'v0.0.280'), ('densenet40_k36_bc_cifar10', '0404', '4c154567e25619994a2f86371afbf1ad1e7475e9', 'v0.0.224'), ('densenet40_k36_bc_cifar100', '2050', 'd7275d39bcf439151c3bbeb707efa54943714b03', 'v0.0.225'), ('densenet40_k36_bc_svhn', '0260', 'b81ec8d662937851beecc62f36209fd8db464265', 'v0.0.311'), ('densenet100_k12_cifar10', '0366', '4e371ccb315d0fcd727a76255ca62ae9e92059cc', 'v0.0.205'), ('densenet100_k12_cifar100', '1964', '2ed5ec27a4d4a63876a4cacf52be53c91fbecb5f', 'v0.0.206'), ('densenet100_k12_svhn', '0260', '3e2b34b2087fe507a3672bfce1520747fca58046', 'v0.0.311'), ('densenet100_k24_cifar10', '0313', '9f795bac946d1390cf59f686b730fe512c406bd2', 'v0.0.252'), ('densenet100_k24_cifar100', '1808', '9bfa3e9c736a80906d163380cb361b940c2188bf', 'v0.0.318'), ('densenet100_k12_bc_cifar10', '0416', '6685d1f4844b092471f7d03dfc3fa64a302008e6', 'v0.0.189'), ('densenet100_k12_bc_cifar100', '2119', 'fbd8a54c1c9e4614f950b8473f8524d25caba4a7', 'v0.0.208'), ('densenet190_k40_bc_cifar10', '0252', '87b15be0620c0adff249d33540c20314188b16d7', 'v0.0.286'), ('densenet250_k24_bc_cifar10', '0267', 'dad68693d83a276d14a87dce6cebc5aceebca775', 'v0.0.290'), ('densenet250_k24_bc_cifar100', '1739', '598e91b7906f427296ab72cf40032f0846a52d91', 'v0.0.303'), ('xdensenet40_2_k24_bc_cifar10', '0531', '66c9d384d3ef4ec4095c9759bb8b7986f2f58e26', 'v0.0.226'), ('xdensenet40_2_k24_bc_cifar100', '2396', '73d5ba88a39b971457b9cea2cd72d1e05ab4d165', 'v0.0.227'), ('xdensenet40_2_k24_bc_svhn', '0287', '745f374b398bce378903af8c71cb3c67f6891d7f', 'v0.0.306'), ('xdensenet40_2_k36_bc_cifar10', '0437', 'e9bf419295f833b56fa3da27218107ed42310307', 'v0.0.233'), ('xdensenet40_2_k36_bc_cifar100', '2165', '78b6e754d90774d7b6ec3d811e6e57192148cfbf', 'v0.0.234'), ('xdensenet40_2_k36_bc_svhn', '0274', '4377e8918c1e008201aafc448f642642474eab14', 'v0.0.306'), ('wrn16_10_cifar10', '0293', 'ecf1c17c0814763095df562cb27d15a5aeb51836', 'v0.0.166'), ('wrn16_10_cifar100', '1895', 'bcb5c89ca71ffc99bc09b861b339724047724659', 'v0.0.204'), ('wrn16_10_svhn', '0278', '76f4e1361f9eca82fa4c2764b530f57280a34cfe', 'v0.0.271'), ('wrn28_10_cifar10', '0239', '16f3c8a249993f23b0f81d9ce3650faef5e455d8', 'v0.0.166'), ('wrn28_10_cifar100', '1788', '67ec43c6e913d43c8936809f04b0780035a24835', 'v0.0.320'), ('wrn28_10_svhn', '0271', 'fcd7a6b03a552b22ec25ee9a3833dc260976a757', 'v0.0.276'), ('wrn40_8_cifar10', '0237', '3b81d261706b751f5b731149b05fa92f500218e8', 'v0.0.166'), ('wrn40_8_cifar100', '1803', '114f6be2d5f8d561a5e3b4106fac30028defe300', 'v0.0.321'), ('wrn40_8_svhn', '0254', 'be7a21da6bc958c79725d7a29502c6a781cc67d9', 'v0.0.277'), ('wrn20_10_1bit_cifar10', '0326', 'c1a8ba4f1e1336a289c4b2eec75e25445b511ca6', 'v0.0.302'), ('wrn20_10_1bit_cifar100', '1904', 'adae01d6bec92d4fe388cddbb7f7eb598b1655d1', 'v0.0.302'), ('wrn20_10_1bit_svhn', '0273', 'ce9f819cf117fa66af112d9cbb0b65568623118d', 'v0.0.302'), ('wrn20_10_32bit_cifar10', '0314', '355496184493a55323c99bad9f79b0803548d373', 'v0.0.302'), ('wrn20_10_32bit_cifar100', '1812', 'd064f38aeaa14e9a2f4e9893ef6cca65615c53f9', 'v0.0.302'), ('wrn20_10_32bit_svhn', '0259', 'd9e8b46e180a34c0a765e22d24741f3849fca13a', 'v0.0.302'), ('ror3_56_cifar10', '0543', 'ee31a69a0503b41878c49d8925ac8e7ee813293b', 'v0.0.228'), ('ror3_56_cifar100', '2549', '4334559313cd9291af3d6ec0df144b21e695228b', 'v0.0.229'), ('ror3_56_svhn', '0269', '56617cf90e0902e88686af14939605c45d1170cf', 'v0.0.287'), ('ror3_110_cifar10', '0435', '0359916596cba01dfa481f105094c1047f592980', 'v0.0.235'), ('ror3_110_cifar100', '2364', 'b8c4d317241f54990180443d7fd9702d79c57ccc', 'v0.0.236'), ('ror3_110_svhn', '0257', '0677b7dfee32659a92719a5a16a7f387a5635f0b', 'v0.0.287'), ('ror3_164_cifar10', '0393', 'cc11aa06d928d0805279baccbf2b82371c31f503', 'v0.0.294'), ('ror3_164_cifar100', '2234', 'eb6a7fb8128240d84843a8e39adb00f606b6e2cf', 'v0.0.294'), ('ror3_164_svhn', '0273', 'b008c1b01386aca1803a1286607c5e1f843fc919', 'v0.0.294'), ('rir_cifar10', '0328', '5bed6f3506055b3ab5c4780a540cfebe014490ec', 'v0.0.292'), ('rir_cifar100', '1923', 'c42563834a971e18eacfc2287585aa2efa8af3eb', 'v0.0.292'), ('rir_svhn', '0268', '1c0718deaef5836efca4d5ded6140f0cd51424ab', 'v0.0.292'), ('shakeshakeresnet20_2x16d_cifar10', '0515', 'a7b8a2f77457e151da5d5ad3b9a2473594fecfc0', 'v0.0.215'), ('shakeshakeresnet20_2x16d_cifar100', '2922', 'e46e31a7d8308b57d9c0687000c40f15623998c2', 'v0.0.247'), ('shakeshakeresnet20_2x16d_svhn', '0317', '7a48fde5e1ccd5ff695892adf7094c15368ec778', 'v0.0.295'), ('shakeshakeresnet26_2x32d_cifar10', '0317', '21e60e626765001aaaf4eb26f7cb8f4a69ea3dc1', 'v0.0.217'), ('shakeshakeresnet26_2x32d_cifar100', '1880', 'bd46a7418374e3b3c844b33e12b09b6a98eb4e6e', 'v0.0.222'), ('shakeshakeresnet26_2x32d_svhn', '0262', 'f1dbb8ef162d9ec56478e2579272f85ed78ad896', 'v0.0.295'), ('diaresnet20_cifar10', '0622', '3e47641d76c1992652d8f973294f4763ecef1987', 'v0.0.340'), ('diaresnet20_cifar100', '2771', '3a58490ea95538ad5809c05739b4362088ea6961', 'v0.0.342'), ('diaresnet20_svhn', '0323', '579535ddc8b7c9becfe9bf97393ab33d9d5e7d0b', 'v0.0.342'), ('diaresnet56_cifar10', '0505', '45df69745c9692168697a7b980ade080ef7af07d', 'v0.0.340'), ('diaresnet56_cifar100', '2435', 'e45b7f281bb63c90104ff79d1519b4785a975a92', 'v0.0.342'), ('diaresnet56_svhn', '0268', '8f2c0574380bf14b0e9711d6370b2898f337cab0', 'v0.0.342'), ('diaresnet110_cifar10', '0410', '56f547ec833f419ea216f51439de50287dfef3c3', 'v0.0.340'), ('diaresnet110_cifar100', '2211', 'e99fad4ef0b2e7f09376beb314d672db7c3b6a55', 'v0.0.342'), ('diaresnet110_svhn', '0247', 'c587ac09f45fd7a29adfc1da62ad50174fd248ec', 'v0.0.342'), ('diaresnet164bn_cifar10', '0350', '533e7c6a30fce31c4f65686782cf761e7913750c', 'v0.0.340'), ('diaresnet164bn_cifar100', '1953', '43fa3821ab72e94187c12f7f950a2343649b3657', 'v0.0.342'), ('diaresnet164bn_svhn', '0244', 'eba062dce4033fd85ff78c2530b363b3768c036e', 'v0.0.342'), ('diapreresnet20_cifar10', '0642', 'ec36098cfbbb889fdd124083e785d0e21ba34792', 'v0.0.343'), ('diapreresnet20_cifar100', '2837', '32f0f1be9aa1da73f8fdb74f27ebaa49e7f9ace6', 'v0.0.343'), ('diapreresnet20_svhn', '0303', 'e33be387b0e71a4b0597558157ffbdb79c6db30c', 'v0.0.343'), ('diapreresnet56_cifar10', '0483', 'cba6950f21643a70b8e61b7197ca9cee9b2d0545', 'v0.0.343'), ('diapreresnet56_cifar100', '2505', 'c9f8bd4380d35e3806e1697c1c8d80bf7341c04e', 'v0.0.343'), ('diapreresnet56_svhn', '0280', '98a2a0bab42ff2605bd2cd4e63280b3631b042cb', 'v0.0.343'), ('diapreresnet110_cifar10', '0425', 'f4eae5abe2edebb1e224f7cf092ba02a873eb781', 'v0.0.343'), ('diapreresnet110_cifar100', '2269', '78d79bab215a5dc7221859d0b2688d040a55afb2', 'v0.0.343'), ('diapreresnet110_svhn', '0242', 'decb3765e92f5620580fe6b440ee2a82811d412e', 'v0.0.343'), ('diapreresnet164bn_cifar10', '0356', '9cf07392dc9714324e470fd50efb92ef286296ac', 'v0.0.343'), ('diapreresnet164bn_cifar100', '1999', '1625154f3cce7e131f25d8ee0b315b3fcc6fb760', 'v0.0.343'), ('diapreresnet164bn_svhn', '0256', '8476c5c9176abf21ea380dd00074b0ec30bbc530', 'v0.0.343'), ('resnet10_cub', '2765', '9dab9a498c380e6b7447827e00996d7cc61cc414', 'v0.0.335'), ('resnet12_cub', '2658', 'a46b8ec2d8dcd66a628dcfcb617acb15ef786b95', 'v0.0.336'), ('resnet14_cub', '2435', '0b9801b2e3aa3908bbc98f50d3ae3e986652742b', 'v0.0.337'), ('resnet16_cub', '2321', '031374ada9830869372a63e132c2477a04425444', 'v0.0.338'), ('resnet18_cub', '2330', 'e72712003928ed70ccf44b953e9cec4f78a75eea', 'v0.0.344'), ('resnet26_cub', '2252', '61cce1ea575f650a7e12a08b1a09335afa6cb605', 'v0.0.345'), ('seresnet10_cub', '2739', '7060c03f78bc60df09288b433eb6117c0e167210', 'v0.0.361'), ('seresnet12_cub', '2604', 'ee095118bde6e05ea102f5b945401a7221b7b7fb', 'v0.0.361'), ('seresnet14_cub', '2363', '5d2049d53c0445d7c66c849a5cd805ce39a37ddb', 'v0.0.361'), ('seresnet16_cub', '2321', '576e58eff57730a516094b8ba79452092187b693', 'v0.0.361'), ('seresnet18_cub', '2308', '3d2496d66efd6a00ca516c1a7a5a091f90043237', 'v0.0.361'), ('seresnet26_cub', '2251', '8d54edb2800b2ff071ee5beedde285eb9553bc22', 'v0.0.361'), ('mobilenet_w1_cub', '2346', 'efcad3dcf1975552f15028255a15f86a16b60987', 'v0.0.346'), ('proxylessnas_mobile_cub', '2188', '36d33231029b466638b3b1f8b2d1392e22d1afa7', 'v0.0.347'), ('ntsnet_cub', '1326', '75ae8cdcf4beb1ab60c1a983c9f143baaebbdea0', 'v0.0.334'), ('pspnet_resnetd101b_voc', '8144', 'e15319bf5428637e7fc00dcd426dd458ac937b08', 'v0.0.297'), ('pspnet_resnetd50b_ade20k', '3687', 'f0dcdf734f8f32a879dec3c4e7fe61d629244030', 'v0.0.297'), ('pspnet_resnetd101b_ade20k', '3797', 'c1280aeab8daa31c0893f7551d70130c2b68214a', 'v0.0.297'), ('pspnet_resnetd101b_cityscapes', '7172', 'd5ad2fa4c4208f439ab0b98267babe0c4d9e6e94', 'v0.0.297'), ('pspnet_resnetd101b_coco', '6741', '87582b79c48c4e995de808ff0cbc162c55b52031', 'v0.0.297'), ('deeplabv3_resnetd101b_voc', '8024', '8ee3099c5c983ef1cc0ce23b23d91db40b2986b8', 'v0.0.298'), ('deeplabv3_resnetd152b_voc', '8120', '88fb315dc3c58a84f325e63105fbfe322932073f', 'v0.0.298'), ('deeplabv3_resnetd50b_ade20k', '3713', '5d5e2f74008ab3637a05b6b1357c9c339296188c', 'v0.0.298'), ('deeplabv3_resnetd101b_ade20k', '3784', '6224836f8f31a00be1718a530a20670136bb3958', 'v0.0.298'), ('deeplabv3_resnetd101b_coco', '6773', '74dc9914078e47feb3ff64fba717d1d4040d8235', 'v0.0.298'), ('deeplabv3_resnetd152b_coco', '6899', 'edd79b4ca095f1674e7a68ee0dc8ed8bcd0b6a26', 'v0.0.298'), ('fcn8sd_resnetd101b_voc', '8040', 'f6c67c75bce4f9a3e17bf555369c0c9332ab5c1f', 'v0.0.299'), ('fcn8sd_resnetd50b_ade20k', '3339', '9856c5ee8186d1ac4b0eb5177c73e76c4cd63bb0', 'v0.0.299'), ('fcn8sd_resnetd101b_ade20k', '3588', '081774b2fb373d7b759cda2160fa0d2599b1c5f1', 'v0.0.299'), ('fcn8sd_resnetd101b_coco', '6011', '05e97cc5f5fcdf1c5ec5c617062d43adfe150d88', 'v0.0.299'), ('icnet_resnetd50b_cityscapes', '6402', '6c8f86a53526ae107e58d5f645bc4de0da9c1bb1', 'v0.0.457'), ('fastscnn_cityscapes', '6576', '9e0d75e56bde8d1643d3ff0053e55114c0a77ee9', 'v0.0.474'), ('sinet_cityscapes', '6031', '47d8ae7824bd297bbf25c2f33e3d4a86a503aefa', 'v0.0.437'), ('bisenet_resnet18_celebamaskhq', '0000', 'd72f0cf3101625bb4265e4cf5ae557b994f84d67', 'v0.0.462'), ('danet_resnetd50b_cityscapes', '6799', '9880a0eb9523ba2c1f98025f7d7115a2a4c1f376', 'v0.0.468'), ('danet_resnetd101b_cityscapes', '6810', 'ea69dcea31f5b250254a226f70df130875ff18b2', 'v0.0.468'), ('alphapose_fastseresnet101b_coco', '7415', '70082a53d3cdeb7ebe4c7d8c16a6a39830f1ed23', 'v0.0.454'), ('simplepose_resnet18_coco', '6631', '5a6198e5103a28faab4e49c687121634c7f7d196', 'v0.0.455'), ('simplepose_resnet50b_coco', '7102', '6315ffa72993eea3f573ae1cc23f84d0275f0fbe', 'v0.0.455'), ('simplepose_resnet101b_coco', '7244', '0491ab951827782492ffbd8fa57aa6dd599c7e9f', 'v0.0.455'), ('simplepose_resnet152b_coco', '7253', '4590c1c555eb54c4e9afdc83fa8a199132afd212', 'v0.0.455'), ('simplepose_resneta50b_coco', '7170', 'fa09a84ee2e085ad6b641c4fe0cc483651861789', 'v0.0.455'), ('simplepose_resneta101b_coco', '7297', '7ddd6cb20bcd626e05e8a627601bf19b704ba6a9', 'v0.0.455'), ('simplepose_resneta152b_coco', '7344', '9ec1a3dc2a23a19cb7f5ac5466cebba8069a0f93', 'v0.0.455'), ('simplepose_mobile_resnet18_coco', '6625', '8ff93eed70ac73503c8c1e346ddd1ade5d9e3edf', 'v0.0.456'), ('simplepose_mobile_resnet50b_coco', '7110', 'e0f2e587ffdf5e074a29f877890b13b99a58c6c2', 'v0.0.456'), ('simplepose_mobile_mobilenet_w1_coco', '6410', '0867e5aa76d5ec37cde08c71de8324a9c2913922', 'v0.0.456'), ('simplepose_mobile_mobilenetv2b_w1_coco', '6374', '07e9c6295a8aa2b7bb9a1d0b7c716141e7ee71dc', 'v0.0.456'), ('simplepose_mobile_mobilenetv3_small_w1_coco', '5434', 'cb837c0e32edec68dc8598d71599ea7404936f96', 'v0.0.456'), ('simplepose_mobile_mobilenetv3_large_w1_coco', '6367', '7ba036a5ade736042531a0fe500ecec368dbf157', 'v0.0.456'), ('lwopenpose2d_mobilenet_cmupan_coco', '3999', 'b4a22e7c2a05e53fe22185002c48e81f76c2d918', 'v0.0.458'), ('lwopenpose3d_mobilenet_cmupan_coco', '3999', '4658738ec27d46ee01f2cad4aa975914e9f7108c', 'v0.0.458'), ('ibppose_coco', '6486', '024d1fafb7471572129ccbb07d662f6d8ccdc758', 'v0.0.459'), ('jasperdr10x5_en', '2190', '1ce0ab1cd891294988d7f34d2153fabd968aa876', 'v0.0.555'), ('jasperdr10x5_en_nr', '1789', '49b0b7718e69265c288c163fa6bbb68eb4db79e4', 'v0.0.555'), ('quartznet5x5_en_ls', '4468', 'b37c8cb2e84b2573ed4364d3786e5d06f36cba92', 'v0.0.555'), ('quartznet15x5_en', '1677', 'bf63d3ffbd117d7e6c36bcd7725c998b8ab06754', 'v0.0.555'), ('quartznet15x5_en_nr', '1774', '5a70b9ec078945392c9e5a4161a36f431b1966ab', 'v0.0.555'), ('quartznet15x5_de', '1166', '5e254c91fe57bb2ed5bae62eb5067fc4922f4184', 'v0.0.555'), ('quartznet15x5_fr', '1388', 'f2236953caa996832118c7a412c839395237b5c2', 'v0.0.555'), ('quartznet15x5_it', '1502', '2df788c3c06415237e254ea91ab2994b35c32e71', 'v0.0.555'), ('quartznet15x5_es', '1295', '117352a802814689275b03a161c537ae91e68a9f', 'v0.0.555'), ('quartznet15x5_ca', '0842', 'da40489e324b6a790638d8ed164be3728f4caf3c', 'v0.0.555'), ('quartznet15x5_pl', '1359', '0df08d125314b5a728cd22b77bb8c6dc2b9c2cd1', 'v0.0.555'), ('quartznet15x5_ru', '1648', 'aecf49e137f8041cebc5fcfe8d1834420f03d98b', 'v0.0.555'), ('quartznet15x5_ru34', '0968', 'ff446c0cca42687330ade49dfb6bac05d7251803', 'v0.0.555'), ]} imgclsmob_repo_url = 'https://github.com/osmr/imgclsmob' def get_model_name_suffix_data(model_name): if model_name not in _model_sha1: raise ValueError("Pretrained model for {name} is not available.".format(name=model_name)) error, sha1_hash, repo_release_tag = _model_sha1[model_name] return error, sha1_hash, repo_release_tag def get_model_file(model_name, local_model_store_dir_path=os.path.join("~", ".mxnet", "models")): """ Return location for the pretrained on local file system. This function will download from online model zoo when model cannot be found or has mismatch. The root directory will be created if it doesn't exist. Parameters: ---------- model_name : str Name of the model. local_model_store_dir_path : str, default $MXNET_HOME/models Location for keeping the model parameters. Returns: ------- file_path Path to the requested pretrained model file. """ error, sha1_hash, repo_release_tag = get_model_name_suffix_data(model_name) short_sha1 = sha1_hash[:8] file_name = "{name}-{error}-{short_sha1}.params".format( name=model_name, error=error, short_sha1=short_sha1) local_model_store_dir_path = os.path.expanduser(local_model_store_dir_path) file_path = os.path.join(local_model_store_dir_path, file_name) if os.path.exists(file_path): if check_sha1(file_path, sha1_hash): return file_path else: logging.warning("Mismatch in the content of model file detected. Downloading again.") else: logging.info("Model file not found. Downloading to {}.".format(file_path)) if not os.path.exists(local_model_store_dir_path): os.makedirs(local_model_store_dir_path) zip_file_path = file_path + ".zip" download( url="{repo_url}/releases/download/{repo_release_tag}/{file_name}.zip".format( repo_url=imgclsmob_repo_url, repo_release_tag=repo_release_tag, file_name=file_name), path=zip_file_path, overwrite=True) with zipfile.ZipFile(zip_file_path) as zf: zf.extractall(local_model_store_dir_path) os.remove(zip_file_path) if check_sha1(file_path, sha1_hash): return file_path else: raise ValueError("Downloaded file has different hash. Please try again.")
61,759
83.718793
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/fastseresnet.py
""" Fast-SE-ResNet for ImageNet-1K, implemented in Gluon. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['FastSEResNet', 'fastseresnet101b'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, SEBlock from .resnet import ResBlock, ResBottleneck, ResInitBlock class FastSEResUnit(HybridBlock): """ Fast-SE-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer of the block. use_se : bool Whether to use SE-module. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, bottleneck, conv1_stride, use_se, **kwargs): super(FastSEResUnit, self).__init__(**kwargs) self.use_se = use_se self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) if self.use_se: self.se = SEBlock( channels=out_channels, reduction=1, use_conv=False) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) if self.use_se: x = self.se(x) x = x + identity x = self.activ(x) return x class FastSEResNet(HybridBlock): """ Fast-SE-ResNet model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(FastSEResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 use_se = (j == 0) stage.add(FastSEResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=conv1_stride, use_se=use_se)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_fastseresnet(blocks, bottleneck=None, conv1_stride=True, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Fast-SE-ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif (blocks == 38) and bottleneck: layers = [3, 3, 3, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] else: raise ValueError("Unsupported Fast-SE-ResNet with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = FastSEResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def fastseresnet101b(**kwargs): """ Fast-SE-ResNet-101 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fastseresnet(blocks=101, conv1_stride=False, model_name="fastseresnet101b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ fastseresnet101b, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fastseresnet101b or weight_count == 55697960) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
10,500
32.336508
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/ibnbresnet.py
""" IBN(b)-ResNet for ImageNet-1K, implemented in Gluon. Original paper: 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. """ __all__ = ['IBNbResNet', 'ibnb_resnet50', 'ibnb_resnet101', 'ibnb_resnet152'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block from .resnet import ResBottleneck class IBNbConvBlock(HybridBlock): """ IBN(b)-ResNet specific convolution block with Instance normalization and ReLU activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, activate=True, **kwargs): super(IBNbConvBlock, self).__init__(**kwargs) self.activate = activate with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) self.inst_norm = nn.InstanceNorm( in_channels=out_channels, scale=True) if self.activate: self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.conv(x) x = self.inst_norm(x) if self.activate: x = self.activ(x) return x def ibnb_conv7x7_block(in_channels, out_channels, strides=1, padding=3, use_bias=False, activate=True): """ 7x7 version of the IBN(b)-ResNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 3 Padding value for convolution layer. use_bias : bool, default False Whether the layer uses a bias vector. activate : bool, default True Whether activate the convolution block. """ return IBNbConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=7, strides=strides, padding=padding, use_bias=use_bias, activate=activate) class IBNbResUnit(HybridBlock): """ IBN(b)-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. use_inst_norm : bool Whether to use instance normalization. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, use_inst_norm, bn_use_global_stats, **kwargs): super(IBNbResUnit, self).__init__(**kwargs) self.use_inst_norm = use_inst_norm self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, conv1_stride=False) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) if self.use_inst_norm: self.inst_norm = nn.InstanceNorm( in_channels=out_channels, scale=True) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity if self.use_inst_norm: x = self.inst_norm(x) x = self.activ(x) return x class IBNbResInitBlock(HybridBlock): """ IBN(b)-ResNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(IBNbResInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = ibnb_conv7x7_block( in_channels=in_channels, out_channels=out_channels, strides=2) self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv(x) x = self.pool(x) return x class IBNbResNet(HybridBlock): """ IBN(b)-ResNet model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(IBNbResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(IBNbResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 use_inst_norm = (i < 2) and (j == len(channels_per_stage) - 1) stage.add(IBNbResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, use_inst_norm=use_inst_norm, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_ibnbresnet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create IBN(b)-ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] else: raise ValueError("Unsupported IBN(b)-ResNet with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = IBNbResNet( channels=channels, init_block_channels=init_block_channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def ibnb_resnet50(**kwargs): """ IBN(b)-ResNet-50 model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibnbresnet(blocks=50, model_name="ibnb_resnet50", **kwargs) def ibnb_resnet101(**kwargs): """ IBN(b)-ResNet-101 model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibnbresnet(blocks=101, model_name="ibnb_resnet101", **kwargs) def ibnb_resnet152(**kwargs): """ IBN(b)-ResNet-152 model from 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibnbresnet(blocks=152, model_name="ibnb_resnet152", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ ibnb_resnet50, ibnb_resnet101, ibnb_resnet152, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ibnb_resnet50 or weight_count == 25558568) assert (model != ibnb_resnet101 or weight_count == 44550696) assert (model != ibnb_resnet152 or weight_count == 60194344) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
13,442
31.007143
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/polynet.py
""" PolyNet for ImageNet-1K, implemented in Gluon. Original paper: 'PolyNet: A Pursuit of Structural Diversity in Very Deep Networks,' https://arxiv.org/abs/1611.05725. """ __all__ = ['PolyNet', 'polynet'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import ConvBlock, conv1x1_block, conv3x3_block, ParametricSequential, ParametricConcurrent class PolyConv(HybridBlock): """ PolyNet specific convolution block. A block that is used inside poly-N (poly-2, poly-3, and so on) modules. The Convolution layer is shared between all Inception blocks inside a poly-N module. BatchNorm layers are not shared between Inception blocks and therefore the number of BatchNorm layers is equal to the number of Inception blocks inside a poly-N module. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. num_blocks : int Number of blocks (BatchNorm layers). """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, bn_use_global_stats, num_blocks, **kwargs): super(PolyConv, self).__init__(**kwargs) with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, use_bias=False, in_channels=in_channels) for i in range(num_blocks): setattr(self, "bn{}".format(i + 1), nn.BatchNorm( in_channels=out_channels, use_global_stats=bn_use_global_stats)) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x, index): x = self.conv(x) bn = getattr(self, "bn{}".format(index + 1)) x = bn(x) x = self.activ(x) return x def poly_conv1x1(in_channels, out_channels, bn_use_global_stats, num_blocks): """ 1x1 version of the PolyNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. num_blocks : int Number of blocks (BatchNorm layers). """ return PolyConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=1, padding=0, bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks) class MaxPoolBranch(HybridBlock): """ PolyNet specific max pooling branch block. """ def __init__(self, **kwargs): super(MaxPoolBranch, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=0) def hybrid_forward(self, F, x): x = self.pool(x) return x class Conv1x1Branch(HybridBlock): """ PolyNet specific convolutional 1x1 branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(Conv1x1Branch, self).__init__(**kwargs) with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv(x) return x class Conv3x3Branch(HybridBlock): """ PolyNet specific convolutional 3x3 branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(Conv3x3Branch, self).__init__(**kwargs) with self.name_scope(): self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=2, padding=0, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv(x) return x class ConvSeqBranch(HybridBlock): """ PolyNet specific convolutional sequence branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of tuple of int List of numbers of output channels. kernel_size_list : list of tuple of int or tuple of tuple/list of 2 int List of convolution window sizes. strides_list : list of tuple of int or tuple of tuple/list of 2 int List of strides of the convolution. padding_list : list of tuple of int or tuple of tuple/list of 2 int List of padding values for convolution layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels_list, kernel_size_list, strides_list, padding_list, bn_use_global_stats, **kwargs): super(ConvSeqBranch, self).__init__(**kwargs) assert (len(out_channels_list) == len(kernel_size_list)) assert (len(out_channels_list) == len(strides_list)) assert (len(out_channels_list) == len(padding_list)) with self.name_scope(): self.conv_list = nn.HybridSequential(prefix="") for i, (out_channels, kernel_size, strides, padding) in enumerate(zip( out_channels_list, kernel_size_list, strides_list, padding_list)): self.conv_list.add(ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels def hybrid_forward(self, F, x): x = self.conv_list(x) return x class PolyConvSeqBranch(HybridBlock): """ PolyNet specific convolutional sequence branch block with internal PolyNet specific convolution blocks. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of tuple of int List of numbers of output channels. kernel_size_list : list of tuple of int or tuple of tuple/list of 2 int List of convolution window sizes. strides_list : list of tuple of int or tuple of tuple/list of 2 int List of strides of the convolution. padding_list : list of tuple of int or tuple of tuple/list of 2 int List of padding values for convolution layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. num_blocks : int Number of blocks for PolyConv. """ def __init__(self, in_channels, out_channels_list, kernel_size_list, strides_list, padding_list, bn_use_global_stats, num_blocks, **kwargs): super(PolyConvSeqBranch, self).__init__(**kwargs) assert (len(out_channels_list) == len(kernel_size_list)) assert (len(out_channels_list) == len(strides_list)) assert (len(out_channels_list) == len(padding_list)) with self.name_scope(): self.conv_list = ParametricSequential(prefix="") for i, (out_channels, kernel_size, strides, padding) in enumerate(zip( out_channels_list, kernel_size_list, strides_list, padding_list)): self.conv_list.add(PolyConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks)) in_channels = out_channels def hybrid_forward(self, F, x, index): x = self.conv_list(x, index) return x class TwoWayABlock(HybridBlock): """ PolyNet type Inception-A block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(TwoWayABlock, self).__init__(**kwargs) in_channels = 384 with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(32, 48, 64), kernel_size_list=(1, 3, 3), strides_list=(1, 1, 1), padding_list=(0, 1, 1), bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(32, 32), kernel_size_list=(1, 3), strides_list=(1, 1), padding_list=(0, 1), bn_use_global_stats=bn_use_global_stats)) self.branches.add(Conv1x1Branch( in_channels=in_channels, out_channels=32, bn_use_global_stats=bn_use_global_stats)) self.conv = conv1x1_block( in_channels=128, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.branches(x) x = self.conv(x) return x class TwoWayBBlock(HybridBlock): """ PolyNet type Inception-B block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(TwoWayBBlock, self).__init__(**kwargs) in_channels = 1152 with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(128, 160, 192), kernel_size_list=(1, (1, 7), (7, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 3), (3, 0)), bn_use_global_stats=bn_use_global_stats)) self.branches.add(Conv1x1Branch( in_channels=in_channels, out_channels=192, bn_use_global_stats=bn_use_global_stats)) self.conv = conv1x1_block( in_channels=384, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.branches(x) x = self.conv(x) return x class TwoWayCBlock(HybridBlock): """ PolyNet type Inception-C block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(TwoWayCBlock, self).__init__(**kwargs) in_channels = 2048 with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 224, 256), kernel_size_list=(1, (1, 3), (3, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 1), (1, 0)), bn_use_global_stats=bn_use_global_stats)) self.branches.add(Conv1x1Branch( in_channels=in_channels, out_channels=192, bn_use_global_stats=bn_use_global_stats)) self.conv = conv1x1_block( in_channels=448, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.branches(x) x = self.conv(x) return x class PolyPreBBlock(HybridBlock): """ PolyNet type PolyResidual-Pre-B block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. num_blocks : int Number of blocks (BatchNorm layers). """ def __init__(self, bn_use_global_stats, num_blocks, **kwargs): super(PolyPreBBlock, self).__init__(**kwargs) in_channels = 1152 with self.name_scope(): self.branches = ParametricConcurrent(axis=1, prefix="") self.branches.add(PolyConvSeqBranch( in_channels=in_channels, out_channels_list=(128, 160, 192), kernel_size_list=(1, (1, 7), (7, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 3), (3, 0)), bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks)) self.branches.add(poly_conv1x1( in_channels=in_channels, out_channels=192, bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks)) def hybrid_forward(self, F, x, index): x = self.branches(x, index) return x class PolyPreCBlock(HybridBlock): """ PolyNet type PolyResidual-Pre-C block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. num_blocks : int Number of blocks (BatchNorm layers). """ def __init__(self, bn_use_global_stats, num_blocks, **kwargs): super(PolyPreCBlock, self).__init__(**kwargs) in_channels = 2048 with self.name_scope(): self.branches = ParametricConcurrent(axis=1, prefix="") self.branches.add(PolyConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 224, 256), kernel_size_list=(1, (1, 3), (3, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 1), (1, 0)), bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks)) self.branches.add(poly_conv1x1( in_channels=in_channels, out_channels=192, bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks)) def hybrid_forward(self, F, x, index): x = self.branches(x, index) return x def poly_res_b_block(bn_use_global_stats): """ PolyNet type PolyResidual-Res-B block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ return conv1x1_block( in_channels=384, out_channels=1152, strides=1, bn_use_global_stats=bn_use_global_stats, activation=None) def poly_res_c_block(bn_use_global_stats): """ PolyNet type PolyResidual-Res-C block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ return conv1x1_block( in_channels=448, out_channels=2048, strides=1, bn_use_global_stats=bn_use_global_stats, activation=None) class MultiResidual(HybridBlock): """ Base class for constructing N-way modules (2-way, 3-way, and so on). Actually it is for 2-way modules. Parameters: ---------- scale : float, default 1.0 Scale value for each residual branch. res_block : HybridBlock class Residual branch block. num_blocks : int Number of residual branches. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, scale, res_block, num_blocks, bn_use_global_stats, **kwargs): super(MultiResidual, self).__init__(**kwargs) assert (num_blocks >= 1) self.scale = scale self.num_blocks = num_blocks with self.name_scope(): for i in range(num_blocks): setattr(self, "res_block{}".format(i + 1), res_block(bn_use_global_stats=bn_use_global_stats)) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): out = x for i in range(self.num_blocks): res_block = getattr(self, "res_block{}".format(i + 1)) out = out + self.scale * res_block(x) out = self.activ(out) return out class PolyResidual(HybridBlock): """ The other base class for constructing N-way poly-modules. Actually it is for 3-way poly-modules. Parameters: ---------- scale : float, default 1.0 Scale value for each residual branch. res_block : HybridBlock class Residual branch block. num_blocks : int Number of residual branches. pre_block : HybridBlock class Preliminary block. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, scale, res_block, num_blocks, pre_block, bn_use_global_stats, **kwargs): super(PolyResidual, self).__init__(**kwargs) assert (num_blocks >= 1) self.scale = scale self.num_blocks = num_blocks with self.name_scope(): self.pre_block = pre_block( bn_use_global_stats=bn_use_global_stats, num_blocks=num_blocks) for i in range(num_blocks): setattr(self, "res_block{}".format(i + 1), res_block(bn_use_global_stats=bn_use_global_stats)) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): out = x for index in range(self.num_blocks): x = self.pre_block(x, index) res_block = getattr(self, "res_block{}".format(index + 1)) x = res_block(x) out = out + self.scale * x x = self.activ(x) out = self.activ(out) return out class PolyBaseUnit(HybridBlock): """ PolyNet unit base class. Parameters: ---------- two_way_scale : float Scale value for 2-way stage. two_way_block : HybridBlock class Residual branch block for 2-way-stage. poly_scale : float, default 0.0 Scale value for 2-way stage. poly_res_block : HybridBlock class, default None Residual branch block for poly-stage. poly_pre_block : HybridBlock class, default None Preliminary branch block for poly-stage. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, two_way_scale, two_way_block, poly_scale=0.0, poly_res_block=None, poly_pre_block=None, bn_use_global_stats=False, **kwargs): super(PolyBaseUnit, self).__init__(**kwargs) with self.name_scope(): if poly_res_block is not None: assert (poly_scale != 0.0) assert (poly_pre_block is not None) self.poly = PolyResidual( scale=poly_scale, res_block=poly_res_block, num_blocks=3, pre_block=poly_pre_block, bn_use_global_stats=bn_use_global_stats) else: assert (poly_scale == 0.0) assert (poly_pre_block is None) self.poly = None self.twoway = MultiResidual( scale=two_way_scale, res_block=two_way_block, num_blocks=2, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): if self.poly is not None: x = self.poly(x) x = self.twoway(x) return x class PolyAUnit(PolyBaseUnit): """ PolyNet type A unit. Parameters: ---------- two_way_scale : float Scale value for 2-way stage. poly_scale : float Scale value for 2-way stage. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, two_way_scale, poly_scale=0.0, bn_use_global_stats=False, **kwargs): super(PolyAUnit, self).__init__( two_way_scale=two_way_scale, two_way_block=TwoWayABlock, bn_use_global_stats=bn_use_global_stats, **kwargs) assert (poly_scale == 0.0) class PolyBUnit(PolyBaseUnit): """ PolyNet type B unit. Parameters: ---------- two_way_scale : float Scale value for 2-way stage. poly_scale : float Scale value for 2-way stage. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, two_way_scale, poly_scale, bn_use_global_stats, **kwargs): super(PolyBUnit, self).__init__( two_way_scale=two_way_scale, two_way_block=TwoWayBBlock, poly_scale=poly_scale, poly_res_block=poly_res_b_block, poly_pre_block=PolyPreBBlock, bn_use_global_stats=bn_use_global_stats, **kwargs) class PolyCUnit(PolyBaseUnit): """ PolyNet type C unit. Parameters: ---------- two_way_scale : float Scale value for 2-way stage. poly_scale : float Scale value for 2-way stage. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, two_way_scale, poly_scale, bn_use_global_stats, **kwargs): super(PolyCUnit, self).__init__( two_way_scale=two_way_scale, two_way_block=TwoWayCBlock, poly_scale=poly_scale, poly_res_block=poly_res_c_block, poly_pre_block=PolyPreCBlock, bn_use_global_stats=bn_use_global_stats, **kwargs) class ReductionAUnit(HybridBlock): """ PolyNet type Reduction-A unit. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(ReductionAUnit, self).__init__(**kwargs) in_channels = 384 with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(256, 256, 384), kernel_size_list=(1, 3, 3), strides_list=(1, 1, 2), padding_list=(0, 1, 0), bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(384,), kernel_size_list=(3,), strides_list=(2,), padding_list=(0,), bn_use_global_stats=bn_use_global_stats)) self.branches.add(MaxPoolBranch()) def hybrid_forward(self, F, x): x = self.branches(x) return x class ReductionBUnit(HybridBlock): """ PolyNet type Reduction-B unit. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(ReductionBUnit, self).__init__(**kwargs) in_channels = 1152 with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(256, 256, 256), kernel_size_list=(1, 3, 3), strides_list=(1, 1, 2), padding_list=(0, 1, 0), bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(256, 256), kernel_size_list=(1, 3), strides_list=(1, 2), padding_list=(0, 0), bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(256, 384), kernel_size_list=(1, 3), strides_list=(1, 2), padding_list=(0, 0), bn_use_global_stats=bn_use_global_stats)) self.branches.add(MaxPoolBranch()) def hybrid_forward(self, F, x): x = self.branches(x) return x class PolyBlock3a(HybridBlock): """ PolyNet type Mixed-3a block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(PolyBlock3a, self).__init__(**kwargs) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(MaxPoolBranch()) self.branches.add(Conv3x3Branch( in_channels=64, out_channels=96, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class PolyBlock4a(HybridBlock): """ PolyNet type Mixed-4a block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(PolyBlock4a, self).__init__(**kwargs) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=160, out_channels_list=(64, 96), kernel_size_list=(1, 3), strides_list=(1, 1), padding_list=(0, 0), bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=160, out_channels_list=(64, 64, 64, 96), kernel_size_list=(1, (7, 1), (1, 7), 3), strides_list=(1, 1, 1, 1), padding_list=(0, (3, 0), (0, 3), 0), bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class PolyBlock5a(HybridBlock): """ PolyNet type Mixed-5a block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(PolyBlock5a, self).__init__(**kwargs) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(MaxPoolBranch()) self.branches.add(Conv3x3Branch( in_channels=192, out_channels=192, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class PolyInitBlock(HybridBlock): """ PolyNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, bn_use_global_stats, **kwargs): super(PolyInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=32, strides=2, padding=0, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=32, out_channels=32, padding=0, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv3x3_block( in_channels=32, out_channels=64, bn_use_global_stats=bn_use_global_stats) self.block1 = PolyBlock3a(bn_use_global_stats=bn_use_global_stats) self.block2 = PolyBlock4a(bn_use_global_stats=bn_use_global_stats) self.block3 = PolyBlock5a(bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.block1(x) x = self.block2(x) x = self.block3(x) return x class PolyNet(HybridBlock): """ PolyNet model from 'PolyNet: A Pursuit of Structural Diversity in Very Deep Networks,' https://arxiv.org/abs/1611.05725. Parameters: ---------- two_way_scales : list of list of floats Two way scale values for each normal unit. poly_scales : list of list of floats Three way scale values for each normal unit. dropout_rate : float, default 0.2 Fraction of the input units to drop. Must be a number between 0 and 1. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (331, 331) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, two_way_scales, poly_scales, dropout_rate=0.2, bn_use_global_stats=False, in_channels=3, in_size=(331, 331), classes=1000, **kwargs): super(PolyNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes normal_units = [PolyAUnit, PolyBUnit, PolyCUnit] reduction_units = [ReductionAUnit, ReductionBUnit] with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(PolyInitBlock( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) for i, (two_way_scales_per_stage, poly_scales_per_stage) in enumerate(zip(two_way_scales, poly_scales)): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, (two_way_scale, poly_scale) in enumerate(zip(two_way_scales_per_stage, poly_scales_per_stage)): if (j == 0) and (i != 0): unit = reduction_units[i - 1] stage.add(unit(bn_use_global_stats=bn_use_global_stats)) else: unit = normal_units[i] stage.add(unit( two_way_scale=two_way_scale, poly_scale=poly_scale, bn_use_global_stats=bn_use_global_stats)) self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=9, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dropout(rate=dropout_rate)) self.output.add(nn.Dense( units=classes, in_units=2048)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_polynet(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create PolyNet model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ two_way_scales = [ [1.000000, 0.992308, 0.984615, 0.976923, 0.969231, 0.961538, 0.953846, 0.946154, 0.938462, 0.930769], [0.000000, 0.915385, 0.900000, 0.884615, 0.869231, 0.853846, 0.838462, 0.823077, 0.807692, 0.792308, 0.776923], [0.000000, 0.761538, 0.746154, 0.730769, 0.715385, 0.700000]] poly_scales = [ [0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000, 0.000000], [0.000000, 0.923077, 0.907692, 0.892308, 0.876923, 0.861538, 0.846154, 0.830769, 0.815385, 0.800000, 0.784615], [0.000000, 0.769231, 0.753846, 0.738462, 0.723077, 0.707692]] net = PolyNet( two_way_scales=two_way_scales, poly_scales=poly_scales, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def polynet(**kwargs): """ PolyNet model from 'PolyNet: A Pursuit of Structural Diversity in Very Deep Networks,' https://arxiv.org/abs/1611.05725. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_polynet(model_name="polynet", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ polynet, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != polynet or weight_count == 95366600) x = mx.nd.zeros((1, 3, 331, 331), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
38,413
32.52007
122
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/resnet_cifar.py
""" ResNet for CIFAR/SVHN, implemented in Gluon. Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. """ __all__ = ['CIFARResNet', 'resnet20_cifar10', 'resnet20_cifar100', 'resnet20_svhn', 'resnet56_cifar10', 'resnet56_cifar100', 'resnet56_svhn', 'resnet110_cifar10', 'resnet110_cifar100', 'resnet110_svhn', 'resnet164bn_cifar10', 'resnet164bn_cifar100', 'resnet164bn_svhn', 'resnet272bn_cifar10', 'resnet272bn_cifar100', 'resnet272bn_svhn', 'resnet542bn_cifar10', 'resnet542bn_cifar100', 'resnet542bn_svhn', 'resnet1001_cifar10', 'resnet1001_cifar100', 'resnet1001_svhn', 'resnet1202_cifar10', 'resnet1202_cifar100', 'resnet1202_svhn'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3_block from .resnet import ResUnit class CIFARResNet(HybridBlock): """ ResNet model for CIFAR from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, bn_use_global_stats=False, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(ResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=False)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resnet_cifar(classes, blocks, bottleneck, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ResNet model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. blocks : int Number of blocks. bottleneck : bool Whether to use a bottleneck or simple block in units. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (classes in [10, 100]) if bottleneck: assert ((blocks - 2) % 9 == 0) layers = [(blocks - 2) // 9] * 3 else: assert ((blocks - 2) % 6 == 0) layers = [(blocks - 2) // 6] * 3 channels_per_layers = [16, 32, 64] init_block_channels = 16 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if bottleneck: channels = [[cij * 4 for cij in ci] for ci in channels] net = CIFARResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, classes=classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def resnet20_cifar10(classes=10, **kwargs): """ ResNet-20 model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="resnet20_cifar10", **kwargs) def resnet20_cifar100(classes=100, **kwargs): """ ResNet-20 model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="resnet20_cifar100", **kwargs) def resnet20_svhn(classes=10, **kwargs): """ ResNet-20 model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="resnet20_svhn", **kwargs) def resnet56_cifar10(classes=10, **kwargs): """ ResNet-56 model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=56, bottleneck=False, model_name="resnet56_cifar10", **kwargs) def resnet56_cifar100(classes=100, **kwargs): """ ResNet-56 model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=56, bottleneck=False, model_name="resnet56_cifar100", **kwargs) def resnet56_svhn(classes=10, **kwargs): """ ResNet-56 model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=56, bottleneck=False, model_name="resnet56_svhn", **kwargs) def resnet110_cifar10(classes=10, **kwargs): """ ResNet-110 model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=110, bottleneck=False, model_name="resnet110_cifar10", **kwargs) def resnet110_cifar100(classes=100, **kwargs): """ ResNet-110 model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=110, bottleneck=False, model_name="resnet110_cifar100", **kwargs) def resnet110_svhn(classes=10, **kwargs): """ ResNet-110 model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=110, bottleneck=False, model_name="resnet110_svhn", **kwargs) def resnet164bn_cifar10(classes=10, **kwargs): """ ResNet-164(BN) model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=164, bottleneck=True, model_name="resnet164bn_cifar10", **kwargs) def resnet164bn_cifar100(classes=100, **kwargs): """ ResNet-164(BN) model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=164, bottleneck=True, model_name="resnet164bn_cifar100", **kwargs) def resnet164bn_svhn(classes=10, **kwargs): """ ResNet-164(BN) model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=164, bottleneck=True, model_name="resnet164bn_svhn", **kwargs) def resnet272bn_cifar10(classes=10, **kwargs): """ ResNet-272(BN) model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=272, bottleneck=True, model_name="resnet272bn_cifar10", **kwargs) def resnet272bn_cifar100(classes=100, **kwargs): """ ResNet-272(BN) model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=272, bottleneck=True, model_name="resnet272bn_cifar100", **kwargs) def resnet272bn_svhn(classes=10, **kwargs): """ ResNet-272(BN) model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=272, bottleneck=True, model_name="resnet272bn_svhn", **kwargs) def resnet542bn_cifar10(classes=10, **kwargs): """ ResNet-542(BN) model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=542, bottleneck=True, model_name="resnet542bn_cifar10", **kwargs) def resnet542bn_cifar100(classes=100, **kwargs): """ ResNet-542(BN) model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=542, bottleneck=True, model_name="resnet542bn_cifar100", **kwargs) def resnet542bn_svhn(classes=10, **kwargs): """ ResNet-272(BN) model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=542, bottleneck=True, model_name="resnet542bn_svhn", **kwargs) def resnet1001_cifar10(classes=10, **kwargs): """ ResNet-1001 model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=1001, bottleneck=True, model_name="resnet1001_cifar10", **kwargs) def resnet1001_cifar100(classes=100, **kwargs): """ ResNet-1001 model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=1001, bottleneck=True, model_name="resnet1001_cifar100", **kwargs) def resnet1001_svhn(classes=10, **kwargs): """ ResNet-1001 model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=1001, bottleneck=True, model_name="resnet1001_svhn", **kwargs) def resnet1202_cifar10(classes=10, **kwargs): """ ResNet-1202 model for CIFAR-10 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=1202, bottleneck=False, model_name="resnet1202_cifar10", **kwargs) def resnet1202_cifar100(classes=100, **kwargs): """ ResNet-1202 model for CIFAR-100 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=1202, bottleneck=False, model_name="resnet1202_cifar100", **kwargs) def resnet1202_svhn(classes=10, **kwargs): """ ResNet-1202 model for SVHN from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnet_cifar(classes=classes, blocks=1202, bottleneck=False, model_name="resnet1202_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (resnet20_cifar10, 10), (resnet20_cifar100, 100), (resnet20_svhn, 10), (resnet56_cifar10, 10), (resnet56_cifar100, 100), (resnet56_svhn, 10), (resnet110_cifar10, 10), (resnet110_cifar100, 100), (resnet110_svhn, 10), (resnet164bn_cifar10, 10), (resnet164bn_cifar100, 100), (resnet164bn_svhn, 10), (resnet272bn_cifar10, 10), (resnet272bn_cifar100, 100), (resnet272bn_svhn, 10), (resnet542bn_cifar10, 10), (resnet542bn_cifar100, 100), (resnet542bn_svhn, 10), (resnet1001_cifar10, 10), (resnet1001_cifar100, 100), (resnet1001_svhn, 10), (resnet1202_cifar10, 10), (resnet1202_cifar100, 100), (resnet1202_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resnet20_cifar10 or weight_count == 272474) assert (model != resnet20_cifar100 or weight_count == 278324) assert (model != resnet20_svhn or weight_count == 272474) assert (model != resnet56_cifar10 or weight_count == 855770) assert (model != resnet56_cifar100 or weight_count == 861620) assert (model != resnet56_svhn or weight_count == 855770) assert (model != resnet110_cifar10 or weight_count == 1730714) assert (model != resnet110_cifar100 or weight_count == 1736564) assert (model != resnet110_svhn or weight_count == 1730714) assert (model != resnet164bn_cifar10 or weight_count == 1704154) assert (model != resnet164bn_cifar100 or weight_count == 1727284) assert (model != resnet164bn_svhn or weight_count == 1704154) assert (model != resnet272bn_cifar10 or weight_count == 2816986) assert (model != resnet272bn_cifar100 or weight_count == 2840116) assert (model != resnet272bn_svhn or weight_count == 2816986) assert (model != resnet542bn_cifar10 or weight_count == 5599066) assert (model != resnet542bn_cifar100 or weight_count == 5622196) assert (model != resnet542bn_svhn or weight_count == 5599066) assert (model != resnet1001_cifar10 or weight_count == 10328602) assert (model != resnet1001_cifar100 or weight_count == 10351732) assert (model != resnet1001_svhn or weight_count == 10328602) assert (model != resnet1202_cifar10 or weight_count == 19424026) assert (model != resnet1202_cifar100 or weight_count == 19429876) assert (model != resnet1202_svhn or weight_count == 19424026) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
25,521
36.09593
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/nasnet.py
""" NASNet-A for ImageNet-1K, implemented in Gluon. Original paper: 'Learning Transferable Architectures for Scalable Image Recognition,' https://arxiv.org/abs/1707.07012. """ __all__ = ['NASNet', 'nasnet_4a1056', 'nasnet_6a4032', 'nasnet_dual_path_sequential'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, DualPathSequential class NasDualPathScheme(object): """ NASNet specific scheme of dual path response for a block in a DualPathSequential module. Parameters: ---------- can_skip_input : bool Whether can skip input for some blocks. """ def __init__(self, can_skip_input): super(NasDualPathScheme, self).__init__() self.can_skip_input = can_skip_input """ Scheme function. Parameters: ---------- block : nn.HybridBlock A block. x : Tensor Current processed tensor. x_prev : Tensor Previous processed tensor. Returns: ------- x_next : Tensor Next processed tensor. x : Tensor Current processed tensor. """ def __call__(self, block, x, x_prev): x_next = block(x, x_prev) if type(x_next) == tuple: x_next, x = x_next if self.can_skip_input and hasattr(block, 'skip_input') and block.skip_input: x = x_prev return x_next, x def nasnet_dual_path_scheme_ordinal(block, x, _): """ NASNet specific scheme of dual path response for an ordinal block with dual inputs/outputs in a DualPathSequential block. Parameters: ---------- block : nn.HybridBlock A block. x : Tensor Current processed tensor. Returns: ------- x_next : Tensor Next processed tensor. x : Tensor Current processed tensor. """ return block(x), x def nasnet_dual_path_sequential(return_two=True, first_ordinals=0, last_ordinals=0, can_skip_input=False, **kwargs): """ NASNet specific dual path sequential container. Parameters: ---------- return_two : bool, default True Whether to return two output after execution. first_ordinals : int, default 0 Number of the first blocks with single input/output. last_ordinals : int, default 0 Number of the final blocks with single input/output. dual_path_scheme : function Scheme of dual path response for a block. dual_path_scheme_ordinal : function Scheme of dual path response for an ordinal block. can_skip_input : bool, default False Whether can skip input for some blocks. """ return DualPathSequential( return_two=return_two, first_ordinals=first_ordinals, last_ordinals=last_ordinals, dual_path_scheme=NasDualPathScheme(can_skip_input=can_skip_input), dual_path_scheme_ordinal=nasnet_dual_path_scheme_ordinal, **kwargs) def nasnet_batch_norm(channels): """ NASNet specific Batch normalization layer. Parameters: ---------- channels : int Number of channels in input data. """ return nn.BatchNorm( momentum=0.1, epsilon=0.001, in_channels=channels) def nasnet_avgpool1x1_s2(): """ NASNet specific 1x1 Average pooling layer with stride 2. """ return nn.AvgPool2D( pool_size=1, strides=2, count_include_pad=False) def nasnet_avgpool3x3_s1(): """ NASNet specific 3x3 Average pooling layer with stride 1. """ return nn.AvgPool2D( pool_size=3, strides=1, padding=1, count_include_pad=False) def nasnet_avgpool3x3_s2(): """ NASNet specific 3x3 Average pooling layer with stride 2. """ return nn.AvgPool2D( pool_size=3, strides=2, padding=1, count_include_pad=False) def process_with_padding(x, F, process=(lambda x: x), pad_width=(0, 0, 0, 0, 1, 0, 1, 0)): """ Auxiliary decorator for layer with NASNet specific extra padding. Parameters: ---------- x : NDArray Input tensor. F : module Gluon API module. process : function, default (lambda x: x) a decorated layer pad_width : tuple of int, default (0, 0, 0, 0, 1, 0, 1, 0) Whether the layer uses a bias vector. Returns: ------- NDArray Resulted tensor. """ x = F.pad(x, mode="constant", pad_width=pad_width, constant_value=0) x = process(x) x = F.slice(x, begin=(None, None, 1, 1), end=(None, None, None, None)) return x class NasMaxPoolBlock(HybridBlock): """ NASNet specific Max pooling layer with extra padding. Parameters: ---------- extra_padding : bool, default False Whether to use extra padding. """ def __init__(self, extra_padding=False, **kwargs): super(NasMaxPoolBlock, self).__init__(**kwargs) self.extra_padding = extra_padding with self.name_scope(): self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): if self.extra_padding: x = process_with_padding(x, F, self.pool) else: x = self.pool(x) return x class NasAvgPoolBlock(HybridBlock): """ NASNet specific 3x3 Average pooling layer with extra padding. Parameters: ---------- extra_padding : bool, default False Whether to use extra padding. """ def __init__(self, extra_padding=False, **kwargs): super(NasAvgPoolBlock, self).__init__(**kwargs) self.extra_padding = extra_padding with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=3, strides=2, padding=1, count_include_pad=False) def hybrid_forward(self, F, x): if self.extra_padding: x = process_with_padding(x, F, self.pool) else: x = self.pool(x) return x class NasConv(HybridBlock): """ NASNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. groups : int Number of groups. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, groups, **kwargs): super(NasConv, self).__init__(**kwargs) with self.name_scope(): self.activ = nn.Activation("relu") self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, groups=groups, use_bias=False, in_channels=in_channels) self.bn = nasnet_batch_norm(channels=out_channels) def hybrid_forward(self, F, x): x = self.activ(x) x = self.conv(x) x = self.bn(x) return x def nas_conv1x1(in_channels, out_channels): """ 1x1 version of the NASNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ return NasConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=1, padding=0, groups=1) class DwsConv(HybridBlock): """ Standard depthwise separable convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. use_bias : bool, default False Whether the layers use a bias vector. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, use_bias=False, **kwargs): super(DwsConv, self).__init__(**kwargs) with self.name_scope(): self.dw_conv = nn.Conv2D( channels=in_channels, kernel_size=kernel_size, strides=strides, padding=padding, groups=in_channels, use_bias=use_bias, in_channels=in_channels) self.pw_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, use_bias=use_bias) def hybrid_forward(self, F, x): x = self.dw_conv(x) x = self.pw_conv(x) return x class NasDwsConv(HybridBlock): """ NASNet specific depthwise separable convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. extra_padding : bool, default False Whether to use extra padding. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, extra_padding=False, **kwargs): super(NasDwsConv, self).__init__(**kwargs) self.extra_padding = extra_padding with self.name_scope(): self.activ = nn.Activation(activation="relu") self.conv = DwsConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, use_bias=False) self.bn = nasnet_batch_norm(channels=out_channels) def hybrid_forward(self, F, x): x = self.activ(x) if self.extra_padding: x = process_with_padding(x, F, self.conv) else: x = self.conv(x) x = self.bn(x) return x class DwsBranch(HybridBlock): """ NASNet specific block with depthwise separable convolution layers. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. extra_padding : bool, default False Whether to use extra padding. stem : bool, default False Whether to use squeeze reduction if False. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, extra_padding=False, stem=False, **kwargs): super(DwsBranch, self).__init__(**kwargs) assert (not stem) or (not extra_padding) mid_channels = out_channels if stem else in_channels with self.name_scope(): self.conv1 = NasDwsConv( in_channels=in_channels, out_channels=mid_channels, kernel_size=kernel_size, strides=strides, padding=padding, extra_padding=extra_padding) self.conv2 = NasDwsConv( in_channels=mid_channels, out_channels=out_channels, kernel_size=kernel_size, strides=1, padding=padding) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x def dws_branch_k3_s1_p1(in_channels, out_channels, extra_padding=False): """ 3x3/1/1 version of the NASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. extra_padding : bool, default False Whether to use extra padding. """ return DwsBranch( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=1, padding=1, extra_padding=extra_padding) def dws_branch_k5_s1_p2(in_channels, out_channels, extra_padding=False): """ 5x5/1/2 version of the NASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. extra_padding : bool, default False Whether to use extra padding. """ return DwsBranch( in_channels=in_channels, out_channels=out_channels, kernel_size=5, strides=1, padding=2, extra_padding=extra_padding) def dws_branch_k5_s2_p2(in_channels, out_channels, extra_padding=False, stem=False): """ 5x5/2/2 version of the NASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. extra_padding : bool, default False Whether to use extra padding. stem : bool, default False Whether to use squeeze reduction if False. """ return DwsBranch( in_channels=in_channels, out_channels=out_channels, kernel_size=5, strides=2, padding=2, extra_padding=extra_padding, stem=stem) def dws_branch_k7_s2_p3(in_channels, out_channels, extra_padding=False, stem=False): """ 7x7/2/3 version of the NASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. extra_padding : bool, default False Whether to use extra padding. stem : bool, default False Whether to use squeeze reduction if False. """ return DwsBranch( in_channels=in_channels, out_channels=out_channels, kernel_size=7, strides=2, padding=3, extra_padding=extra_padding, stem=stem) class NasPathBranch(HybridBlock): """ NASNet specific `path` branch (auxiliary block). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. extra_padding : bool, default False Whether to use extra padding. """ def __init__(self, in_channels, out_channels, extra_padding=False, **kwargs): super(NasPathBranch, self).__init__(**kwargs) self.extra_padding = extra_padding with self.name_scope(): self.avgpool = nasnet_avgpool1x1_s2() self.conv = conv1x1( in_channels=in_channels, out_channels=out_channels) def hybrid_forward(self, F, x): if self.extra_padding: x = process_with_padding(x, F, pad_width=(0, 0, 0, 0, 0, 1, 0, 1)) x = self.avgpool(x) x = self.conv(x) return x class NasPathBlock(HybridBlock): """ NASNet specific `path` block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(NasPathBlock, self).__init__(**kwargs) mid_channels = out_channels // 2 with self.name_scope(): self.activ = nn.Activation("relu") self.path1 = NasPathBranch( in_channels=in_channels, out_channels=mid_channels) self.path2 = NasPathBranch( in_channels=in_channels, out_channels=mid_channels, extra_padding=True) self.bn = nasnet_batch_norm(channels=out_channels) def hybrid_forward(self, F, x): x = self.activ(x) x1 = self.path1(x) x2 = self.path2(x) x = F.concat(x1, x2, dim=1) x = self.bn(x) return x class Stem1Unit(HybridBlock): """ NASNet Stem1 unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(Stem1Unit, self).__init__(**kwargs) mid_channels = out_channels // 4 with self.name_scope(): self.conv1x1 = nas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5_s2_p2( in_channels=mid_channels, out_channels=mid_channels) self.comb0_right = dws_branch_k7_s2_p3( in_channels=in_channels, out_channels=mid_channels, stem=True) self.comb1_left = NasMaxPoolBlock(extra_padding=False) self.comb1_right = dws_branch_k7_s2_p3( in_channels=in_channels, out_channels=mid_channels, stem=True) self.comb2_left = nasnet_avgpool3x3_s2() self.comb2_right = dws_branch_k5_s2_p2( in_channels=in_channels, out_channels=mid_channels, stem=True) self.comb3_right = nasnet_avgpool3x3_s1() self.comb4_left = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) self.comb4_right = NasMaxPoolBlock(extra_padding=False) def hybrid_forward(self, F, x, _=None): x_left = self.conv1x1(x) x_right = x x0 = self.comb0_left(x_left) + self.comb0_right(x_right) x1 = self.comb1_left(x_left) + self.comb1_right(x_right) x2 = self.comb2_left(x_left) + self.comb2_right(x_right) x3 = x1 + self.comb3_right(x0) x4 = self.comb4_left(x0) + self.comb4_right(x_left) x_out = F.concat(x1, x2, x3, x4, dim=1) return x_out class Stem2Unit(HybridBlock): """ NASNet Stem2 unit. Parameters: ---------- in_channels : int Number of input channels. prev_in_channels : int Number of input channels in previous input. out_channels : int Number of output channels. extra_padding : bool Whether to use extra padding. """ def __init__(self, in_channels, prev_in_channels, out_channels, extra_padding, **kwargs): super(Stem2Unit, self).__init__(**kwargs) mid_channels = out_channels // 4 with self.name_scope(): self.conv1x1 = nas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.path = NasPathBlock( in_channels=prev_in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5_s2_p2( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb0_right = dws_branch_k7_s2_p3( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb1_left = NasMaxPoolBlock(extra_padding=extra_padding) self.comb1_right = dws_branch_k7_s2_p3( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb2_left = NasAvgPoolBlock(extra_padding=extra_padding) self.comb2_right = dws_branch_k5_s2_p2( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb3_right = nasnet_avgpool3x3_s1() self.comb4_left = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb4_right = NasMaxPoolBlock(extra_padding=extra_padding) def hybrid_forward(self, F, x, x_prev): x_left = self.conv1x1(x) x_right = self.path(x_prev) x0 = self.comb0_left(x_left) + self.comb0_right(x_right) x1 = self.comb1_left(x_left) + self.comb1_right(x_right) x2 = self.comb2_left(x_left) + self.comb2_right(x_right) x3 = x1 + self.comb3_right(x0) x4 = self.comb4_left(x0) + self.comb4_right(x_left) x_out = F.concat(x1, x2, x3, x4, dim=1) return x_out class FirstUnit(HybridBlock): """ NASNet First unit. Parameters: ---------- in_channels : int Number of input channels. prev_in_channels : int Number of input channels in previous input. out_channels : int Number of output channels. """ def __init__(self, in_channels, prev_in_channels, out_channels, **kwargs): super(FirstUnit, self).__init__(**kwargs) mid_channels = out_channels // 6 with self.name_scope(): self.conv1x1 = nas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.path = NasPathBlock( in_channels=prev_in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5_s1_p2( in_channels=mid_channels, out_channels=mid_channels) self.comb0_right = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) self.comb1_left = dws_branch_k5_s1_p2( in_channels=mid_channels, out_channels=mid_channels) self.comb1_right = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) self.comb2_left = nasnet_avgpool3x3_s1() self.comb3_left = nasnet_avgpool3x3_s1() self.comb3_right = nasnet_avgpool3x3_s1() self.comb4_left = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) def hybrid_forward(self, F, x, x_prev): x_left = self.conv1x1(x) x_right = self.path(x_prev) x0 = self.comb0_left(x_left) + self.comb0_right(x_right) x1 = self.comb1_left(x_right) + self.comb1_right(x_right) x2 = self.comb2_left(x_left) + x_right x3 = self.comb3_left(x_right) + self.comb3_right(x_right) x4 = self.comb4_left(x_left) + x_left x_out = F.concat(x_right, x0, x1, x2, x3, x4, dim=1) return x_out class NormalUnit(HybridBlock): """ NASNet Normal unit. Parameters: ---------- in_channels : int Number of input channels. prev_in_channels : int Number of input channels in previous input. out_channels : int Number of output channels. """ def __init__(self, in_channels, prev_in_channels, out_channels, **kwargs): super(NormalUnit, self).__init__(**kwargs) mid_channels = out_channels // 6 with self.name_scope(): self.conv1x1_prev = nas_conv1x1( in_channels=prev_in_channels, out_channels=mid_channels) self.conv1x1 = nas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5_s1_p2( in_channels=mid_channels, out_channels=mid_channels) self.comb0_right = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) self.comb1_left = dws_branch_k5_s1_p2( in_channels=mid_channels, out_channels=mid_channels) self.comb1_right = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) self.comb2_left = nasnet_avgpool3x3_s1() self.comb3_left = nasnet_avgpool3x3_s1() self.comb3_right = nasnet_avgpool3x3_s1() self.comb4_left = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels) def hybrid_forward(self, F, x, x_prev): x_left = self.conv1x1(x) x_right = self.conv1x1_prev(x_prev) x0 = self.comb0_left(x_left) + self.comb0_right(x_right) x1 = self.comb1_left(x_right) + self.comb1_right(x_right) x2 = self.comb2_left(x_left) + x_right x3 = self.comb3_left(x_right) + self.comb3_right(x_right) x4 = self.comb4_left(x_left) + x_left x_out = F.concat(x_right, x0, x1, x2, x3, x4, dim=1) return x_out class ReductionBaseUnit(HybridBlock): """ NASNet Reduction base unit. Parameters: ---------- in_channels : int Number of input channels. prev_in_channels : int Number of input channels in previous input. out_channels : int Number of output channels. extra_padding : bool, default True Whether to use extra padding. """ def __init__(self, in_channels, prev_in_channels, out_channels, extra_padding=True, **kwargs): super(ReductionBaseUnit, self).__init__(**kwargs) self.skip_input = True mid_channels = out_channels // 4 with self.name_scope(): self.conv1x1_prev = nas_conv1x1( in_channels=prev_in_channels, out_channels=mid_channels) self.conv1x1 = nas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5_s2_p2( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb0_right = dws_branch_k7_s2_p3( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb1_left = NasMaxPoolBlock(extra_padding=extra_padding) self.comb1_right = dws_branch_k7_s2_p3( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb2_left = NasAvgPoolBlock(extra_padding=extra_padding) self.comb2_right = dws_branch_k5_s2_p2( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb3_right = nasnet_avgpool3x3_s1() self.comb4_left = dws_branch_k3_s1_p1( in_channels=mid_channels, out_channels=mid_channels, extra_padding=extra_padding) self.comb4_right = NasMaxPoolBlock(extra_padding=extra_padding) def hybrid_forward(self, F, x, x_prev): x_left = self.conv1x1(x) x_right = self.conv1x1_prev(x_prev) x0 = self.comb0_left(x_left) + self.comb0_right(x_right) x1 = self.comb1_left(x_left) + self.comb1_right(x_right) x2 = self.comb2_left(x_left) + self.comb2_right(x_right) x3 = x1 + self.comb3_right(x0) x4 = self.comb4_left(x0) + self.comb4_right(x_left) x_out = F.concat(x1, x2, x3, x4, dim=1) return x_out class Reduction1Unit(ReductionBaseUnit): """ NASNet Reduction1 unit. Parameters: ---------- in_channels : int Number of input channels. prev_in_channels : int Number of input channels in previous input. out_channels : int Number of output channels. """ def __init__(self, in_channels, prev_in_channels, out_channels): super(Reduction1Unit, self).__init__( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels, extra_padding=True) class Reduction2Unit(ReductionBaseUnit): """ NASNet Reduction2 unit. Parameters: ---------- in_channels : int Number of input channels. prev_in_channels : int Number of input channels in previous input. out_channels : int Number of output channels. extra_padding : bool Whether to use extra padding. """ def __init__(self, in_channels, prev_in_channels, out_channels, extra_padding): super(Reduction2Unit, self).__init__( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels, extra_padding=extra_padding) class NASNetInitBlock(HybridBlock): """ NASNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(NASNetInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=3, strides=2, padding=0, use_bias=False, in_channels=in_channels) self.bn = nasnet_batch_norm(channels=out_channels) def hybrid_forward(self, F, x): x = self.conv(x) x = self.bn(x) return x class NASNet(HybridBlock): """ NASNet-A model from 'Learning Transferable Architectures for Scalable Image Recognition,' https://arxiv.org/abs/1707.07012. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. stem_blocks_channels : list of 2 int Number of output channels for the Stem units. final_pool_size : int Size of the pooling windows for final pool. extra_padding : bool Whether to use extra padding. skip_reduction_layer_input : bool Whether to skip the reduction layers when calculating the previous layer to connect to. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, stem_blocks_channels, final_pool_size, extra_padding, skip_reduction_layer_input, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(NASNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes reduction_units = [Reduction1Unit, Reduction2Unit] with self.name_scope(): self.features = nasnet_dual_path_sequential( return_two=False, first_ordinals=1, last_ordinals=2, prefix="") self.features.add(NASNetInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels out_channels = stem_blocks_channels[0] self.features.add(Stem1Unit( in_channels=in_channels, out_channels=out_channels)) prev_in_channels = in_channels in_channels = out_channels out_channels = stem_blocks_channels[1] self.features.add(Stem2Unit( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels, extra_padding=extra_padding)) prev_in_channels = in_channels in_channels = out_channels for i, channels_per_stage in enumerate(channels): stage = nasnet_dual_path_sequential( can_skip_input=skip_reduction_layer_input, prefix='stage{}_'.format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): if (j == 0) and (i != 0): unit = reduction_units[i - 1] elif ((i == 0) and (j == 0)) or ((i != 0) and (j == 1)): unit = FirstUnit else: unit = NormalUnit if unit == Reduction2Unit: stage.add(Reduction2Unit( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels, extra_padding=extra_padding)) else: stage.add(unit( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels)) prev_in_channels = in_channels in_channels = out_channels self.features.add(stage) self.features.add(nn.Activation("relu")) self.features.add(nn.AvgPool2D( pool_size=final_pool_size, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dropout(rate=0.5)) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_nasnet(repeat, penultimate_filters, init_block_channels, final_pool_size, extra_padding, skip_reduction_layer_input, in_size, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create NASNet-A model with specific parameters. Parameters: ---------- repeat : int NNumber of cell repeats. penultimate_filters : int Number of filters in the penultimate layer of the network. init_block_channels : int Number of output channels for the initial unit. final_pool_size : int Size of the pooling windows for final pool. extra_padding : bool Whether to use extra padding. skip_reduction_layer_input : bool Whether to skip the reduction layers when calculating the previous layer to connect to. in_size : tuple of two ints Spatial size of the expected input image. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ stem_blocks_channels = [1, 2] reduct_channels = [[], [8], [16]] norm_channels = [6, 12, 24] channels = [rci + [nci] * repeat for rci, nci in zip(reduct_channels, norm_channels)] base_channel_chunk = penultimate_filters // channels[-1][-1] stem_blocks_channels = [(ci * base_channel_chunk) for ci in stem_blocks_channels] channels = [[(cij * base_channel_chunk) for cij in ci] for ci in channels] net = NASNet( channels=channels, init_block_channels=init_block_channels, stem_blocks_channels=stem_blocks_channels, final_pool_size=final_pool_size, extra_padding=extra_padding, skip_reduction_layer_input=skip_reduction_layer_input, in_size=in_size, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def nasnet_4a1056(**kwargs): """ NASNet-A 4@1056 (NASNet-A-Mobile) model from 'Learning Transferable Architectures for Scalable Image Recognition,' https://arxiv.org/abs/1707.07012. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_nasnet( repeat=4, penultimate_filters=1056, init_block_channels=32, final_pool_size=7, extra_padding=True, skip_reduction_layer_input=False, in_size=(224, 224), model_name="nasnet_4a1056", **kwargs) def nasnet_6a4032(**kwargs): """ NASNet-A 6@4032 (NASNet-A-Large) model from 'Learning Transferable Architectures for Scalable Image Recognition,' https://arxiv.org/abs/1707.07012. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_nasnet( repeat=6, penultimate_filters=4032, init_block_channels=96, final_pool_size=11, extra_padding=False, skip_reduction_layer_input=True, in_size=(331, 331), model_name="nasnet_6a4032", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ nasnet_4a1056, nasnet_6a4032, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != nasnet_4a1056 or weight_count == 5289978) assert (model != nasnet_6a4032 or weight_count == 88753150) x = mx.nd.zeros((1, 3, net.in_size[0], net.in_size[1]), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
41,341
29.443299
118
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/resnext_cifar.py
""" ResNeXt for CIFAR/SVHN, implemented in Gluon. Original paper: 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. """ __all__ = ['CIFARResNeXt', 'resnext20_1x64d_cifar10', 'resnext20_1x64d_cifar100', 'resnext20_1x64d_svhn', 'resnext20_2x32d_cifar10', 'resnext20_2x32d_cifar100', 'resnext20_2x32d_svhn', 'resnext20_2x64d_cifar10', 'resnext20_2x64d_cifar100', 'resnext20_2x64d_svhn', 'resnext20_4x16d_cifar10', 'resnext20_4x16d_cifar100', 'resnext20_4x16d_svhn', 'resnext20_4x32d_cifar10', 'resnext20_4x32d_cifar100', 'resnext20_4x32d_svhn', 'resnext20_8x8d_cifar10', 'resnext20_8x8d_cifar100', 'resnext20_8x8d_svhn', 'resnext20_8x16d_cifar10', 'resnext20_8x16d_cifar100', 'resnext20_8x16d_svhn', 'resnext20_16x4d_cifar10', 'resnext20_16x4d_cifar100', 'resnext20_16x4d_svhn', 'resnext20_16x8d_cifar10', 'resnext20_16x8d_cifar100', 'resnext20_16x8d_svhn', 'resnext20_32x2d_cifar10', 'resnext20_32x2d_cifar100', 'resnext20_32x2d_svhn', 'resnext20_32x4d_cifar10', 'resnext20_32x4d_cifar100', 'resnext20_32x4d_svhn', 'resnext20_64x1d_cifar10', 'resnext20_64x1d_cifar100', 'resnext20_64x1d_svhn', 'resnext20_64x2d_cifar10', 'resnext20_64x2d_cifar100', 'resnext20_64x2d_svhn', 'resnext29_32x4d_cifar10', 'resnext29_32x4d_cifar100', 'resnext29_32x4d_svhn', 'resnext29_16x64d_cifar10', 'resnext29_16x64d_cifar100', 'resnext29_16x64d_svhn', 'resnext56_1x64d_cifar10', 'resnext56_1x64d_cifar100', 'resnext56_1x64d_svhn', 'resnext56_2x32d_cifar10', 'resnext56_2x32d_cifar100', 'resnext56_2x32d_svhn', 'resnext56_4x16d_cifar10', 'resnext56_4x16d_cifar100', 'resnext56_4x16d_svhn', 'resnext56_8x8d_cifar10', 'resnext56_8x8d_cifar100', 'resnext56_8x8d_svhn', 'resnext56_16x4d_cifar10', 'resnext56_16x4d_cifar100', 'resnext56_16x4d_svhn', 'resnext56_32x2d_cifar10', 'resnext56_32x2d_cifar100', 'resnext56_32x2d_svhn', 'resnext56_64x1d_cifar10', 'resnext56_64x1d_cifar100', 'resnext56_64x1d_svhn', 'resnext272_1x64d_cifar10', 'resnext272_1x64d_cifar100', 'resnext272_1x64d_svhn', 'resnext272_2x32d_cifar10', 'resnext272_2x32d_cifar100', 'resnext272_2x32d_svhn'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3_block from .resnext import ResNeXtUnit class CIFARResNeXt(HybridBlock): """ ResNeXt model for CIFAR from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, bn_use_global_stats=False, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARResNeXt, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(ResNeXtUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, cardinality=cardinality, bottleneck_width=bottleneck_width, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_resnext_cifar(classes, blocks, cardinality, bottleneck_width, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ ResNeXt model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. blocks : int Number of blocks. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (blocks - 2) % 9 == 0 layers = [(blocks - 2) // 9] * 3 channels_per_layers = [256, 512, 1024] init_block_channels = 64 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = CIFARResNeXt( channels=channels, init_block_channels=init_block_channels, cardinality=cardinality, bottleneck_width=bottleneck_width, classes=classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def resnext20_1x64d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (1x64d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=1, bottleneck_width=64, model_name="resnext20_1x64d_cifar10", **kwargs) def resnext20_1x64d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (1x64d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=1, bottleneck_width=64, model_name="resnext20_1x64d_cifar100", **kwargs) def resnext20_1x64d_svhn(classes=10, **kwargs): """ ResNeXt-20 (1x64d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=1, bottleneck_width=64, model_name="resnext20_1x64d_svhn", **kwargs) def resnext20_2x32d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (2x32d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=2, bottleneck_width=32, model_name="resnext20_2x32d_cifar10", **kwargs) def resnext20_2x32d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (2x32d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=2, bottleneck_width=32, model_name="resnext20_2x32d_cifar100", **kwargs) def resnext20_2x32d_svhn(classes=10, **kwargs): """ ResNeXt-20 (2x32d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=2, bottleneck_width=32, model_name="resnext20_2x32d_svhn", **kwargs) def resnext20_2x64d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (2x64d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=2, bottleneck_width=64, model_name="resnext20_2x64d_cifar10", **kwargs) def resnext20_2x64d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (2x64d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=2, bottleneck_width=64, model_name="resnext20_2x64d_cifar100", **kwargs) def resnext20_2x64d_svhn(classes=10, **kwargs): """ ResNeXt-20 (2x64d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=2, bottleneck_width=64, model_name="resnext20_2x64d_svhn", **kwargs) def resnext20_4x16d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (4x16d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=4, bottleneck_width=16, model_name="resnext20_4x16d_cifar10", **kwargs) def resnext20_4x16d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (4x16d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=4, bottleneck_width=16, model_name="resnext20_4x16d_cifar100", **kwargs) def resnext20_4x16d_svhn(classes=10, **kwargs): """ ResNeXt-20 (4x16d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=4, bottleneck_width=16, model_name="resnext20_4x16d_svhn", **kwargs) def resnext20_4x32d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (4x32d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=4, bottleneck_width=32, model_name="resnext20_4x32d_cifar10", **kwargs) def resnext20_4x32d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (4x32d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=4, bottleneck_width=32, model_name="resnext20_4x32d_cifar100", **kwargs) def resnext20_4x32d_svhn(classes=10, **kwargs): """ ResNeXt-20 (4x32d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=4, bottleneck_width=32, model_name="resnext20_4x32d_svhn", **kwargs) def resnext20_8x8d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (8x8d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=8, bottleneck_width=8, model_name="resnext20_8x8d_cifar10", **kwargs) def resnext20_8x8d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (8x8d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=8, bottleneck_width=8, model_name="resnext20_8x8d_cifar100", **kwargs) def resnext20_8x8d_svhn(classes=10, **kwargs): """ ResNeXt-20 (8x8d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=8, bottleneck_width=8, model_name="resnext20_8x8d_svhn", **kwargs) def resnext20_8x16d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (8x16d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=8, bottleneck_width=16, model_name="resnext20_8x16d_cifar10", **kwargs) def resnext20_8x16d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (8x16d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=8, bottleneck_width=16, model_name="resnext20_8x16d_cifar100", **kwargs) def resnext20_8x16d_svhn(classes=10, **kwargs): """ ResNeXt-20 (8x16d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=8, bottleneck_width=16, model_name="resnext20_8x16d_svhn", **kwargs) def resnext20_16x4d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (16x4d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=16, bottleneck_width=4, model_name="resnext20_16x4d_cifar10", **kwargs) def resnext20_16x4d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (16x4d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=16, bottleneck_width=4, model_name="resnext20_16x4d_cifar100", **kwargs) def resnext20_16x4d_svhn(classes=10, **kwargs): """ ResNeXt-20 (16x4d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=16, bottleneck_width=4, model_name="resnext20_16x4d_svhn", **kwargs) def resnext20_16x8d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (16x8d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=16, bottleneck_width=8, model_name="resnext20_16x8d_cifar10", **kwargs) def resnext20_16x8d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (16x8d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=16, bottleneck_width=8, model_name="resnext20_16x8d_cifar100", **kwargs) def resnext20_16x8d_svhn(classes=10, **kwargs): """ ResNeXt-20 (16x8d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=16, bottleneck_width=8, model_name="resnext20_16x8d_svhn", **kwargs) def resnext20_32x2d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (32x2d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=32, bottleneck_width=2, model_name="resnext20_32x2d_cifar10", **kwargs) def resnext20_32x2d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (32x2d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=32, bottleneck_width=2, model_name="resnext20_32x2d_cifar100", **kwargs) def resnext20_32x2d_svhn(classes=10, **kwargs): """ ResNeXt-20 (32x2d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=32, bottleneck_width=2, model_name="resnext20_32x2d_svhn", **kwargs) def resnext20_32x4d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (32x4d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=32, bottleneck_width=4, model_name="resnext20_32x4d_cifar10", **kwargs) def resnext20_32x4d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (32x4d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=32, bottleneck_width=4, model_name="resnext20_32x4d_cifar100", **kwargs) def resnext20_32x4d_svhn(classes=10, **kwargs): """ ResNeXt-20 (32x4d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=32, bottleneck_width=4, model_name="resnext20_32x4d_svhn", **kwargs) def resnext20_64x1d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (64x1d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=64, bottleneck_width=1, model_name="resnext20_64x1d_cifar10", **kwargs) def resnext20_64x1d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (64x1d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=64, bottleneck_width=1, model_name="resnext20_64x1d_cifar100", **kwargs) def resnext20_64x1d_svhn(classes=10, **kwargs): """ ResNeXt-20 (64x1d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=64, bottleneck_width=1, model_name="resnext20_64x1d_svhn", **kwargs) def resnext20_64x2d_cifar10(classes=10, **kwargs): """ ResNeXt-20 (64x2d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=64, bottleneck_width=2, model_name="resnext20_64x2d_cifar10", **kwargs) def resnext20_64x2d_cifar100(classes=100, **kwargs): """ ResNeXt-20 (64x2d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=64, bottleneck_width=2, model_name="resnext20_64x2d_cifar100", **kwargs) def resnext20_64x2d_svhn(classes=10, **kwargs): """ ResNeXt-20 (64x1d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=20, cardinality=64, bottleneck_width=2, model_name="resnext20_64x2d_svhn", **kwargs) def resnext29_32x4d_cifar10(classes=10, **kwargs): """ ResNeXt-29 (32x4d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=29, cardinality=32, bottleneck_width=4, model_name="resnext29_32x4d_cifar10", **kwargs) def resnext29_32x4d_cifar100(classes=100, **kwargs): """ ResNeXt-29 (32x4d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=29, cardinality=32, bottleneck_width=4, model_name="resnext29_32x4d_cifar100", **kwargs) def resnext29_32x4d_svhn(classes=10, **kwargs): """ ResNeXt-29 (32x4d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=29, cardinality=32, bottleneck_width=4, model_name="resnext29_32x4d_svhn", **kwargs) def resnext29_16x64d_cifar10(classes=10, **kwargs): """ ResNeXt-29 (16x64d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=29, cardinality=16, bottleneck_width=64, model_name="resnext29_16x64d_cifar10", **kwargs) def resnext29_16x64d_cifar100(classes=100, **kwargs): """ ResNeXt-29 (16x64d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=29, cardinality=16, bottleneck_width=64, model_name="resnext29_16x64d_cifar100", **kwargs) def resnext29_16x64d_svhn(classes=10, **kwargs): """ ResNeXt-29 (16x64d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=29, cardinality=16, bottleneck_width=64, model_name="resnext29_16x64d_svhn", **kwargs) def resnext56_1x64d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (1x64d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=1, bottleneck_width=64, model_name="resnext56_1x64d_cifar10", **kwargs) def resnext56_1x64d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (1x64d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=1, bottleneck_width=64, model_name="resnext56_1x64d_cifar100", **kwargs) def resnext56_1x64d_svhn(classes=10, **kwargs): """ ResNeXt-56 (1x64d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=1, bottleneck_width=64, model_name="resnext56_1x64d_svhn", **kwargs) def resnext56_2x32d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (2x32d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=2, bottleneck_width=32, model_name="resnext56_2x32d_cifar10", **kwargs) def resnext56_2x32d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (2x32d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=2, bottleneck_width=32, model_name="resnext56_2x32d_cifar100", **kwargs) def resnext56_2x32d_svhn(classes=10, **kwargs): """ ResNeXt-56 (2x32d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=2, bottleneck_width=32, model_name="resnext56_2x32d_svhn", **kwargs) def resnext56_4x16d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (4x16d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=4, bottleneck_width=16, model_name="resnext56_4x16d_cifar10", **kwargs) def resnext56_4x16d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (4x16d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=4, bottleneck_width=16, model_name="resnext56_4x16d_cifar100", **kwargs) def resnext56_4x16d_svhn(classes=10, **kwargs): """ ResNeXt-56 (4x16d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=4, bottleneck_width=16, model_name="resnext56_4x16d_svhn", **kwargs) def resnext56_8x8d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (8x8d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=8, bottleneck_width=8, model_name="resnext56_8x8d_cifar10", **kwargs) def resnext56_8x8d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (8x8d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=8, bottleneck_width=8, model_name="resnext56_8x8d_cifar100", **kwargs) def resnext56_8x8d_svhn(classes=10, **kwargs): """ ResNeXt-56 (8x8d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=8, bottleneck_width=8, model_name="resnext56_8x8d_svhn", **kwargs) def resnext56_16x4d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (16x4d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=16, bottleneck_width=4, model_name="resnext56_16x4d_cifar10", **kwargs) def resnext56_16x4d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (16x4d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=16, bottleneck_width=4, model_name="resnext56_16x4d_cifar100", **kwargs) def resnext56_16x4d_svhn(classes=10, **kwargs): """ ResNeXt-56 (16x4d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=16, bottleneck_width=4, model_name="resnext56_16x4d_svhn", **kwargs) def resnext56_32x2d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (32x2d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=32, bottleneck_width=2, model_name="resnext56_32x2d_cifar10", **kwargs) def resnext56_32x2d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (32x2d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=32, bottleneck_width=2, model_name="resnext56_32x2d_cifar100", **kwargs) def resnext56_32x2d_svhn(classes=10, **kwargs): """ ResNeXt-56 (32x2d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=32, bottleneck_width=2, model_name="resnext56_32x2d_svhn", **kwargs) def resnext56_64x1d_cifar10(classes=10, **kwargs): """ ResNeXt-56 (64x1d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=64, bottleneck_width=1, model_name="resnext56_64x1d_cifar10", **kwargs) def resnext56_64x1d_cifar100(classes=100, **kwargs): """ ResNeXt-56 (64x1d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=64, bottleneck_width=1, model_name="resnext56_64x1d_cifar100", **kwargs) def resnext56_64x1d_svhn(classes=10, **kwargs): """ ResNeXt-56 (64x1d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=56, cardinality=64, bottleneck_width=1, model_name="resnext56_64x1d_svhn", **kwargs) def resnext272_1x64d_cifar10(classes=10, **kwargs): """ ResNeXt-272 (1x64d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=272, cardinality=1, bottleneck_width=64, model_name="resnext272_1x64d_cifar10", **kwargs) def resnext272_1x64d_cifar100(classes=100, **kwargs): """ ResNeXt-272 (1x64d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=272, cardinality=1, bottleneck_width=64, model_name="resnext272_1x64d_cifar100", **kwargs) def resnext272_1x64d_svhn(classes=10, **kwargs): """ ResNeXt-272 (1x64d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=272, cardinality=1, bottleneck_width=64, model_name="resnext272_1x64d_svhn", **kwargs) def resnext272_2x32d_cifar10(classes=10, **kwargs): """ ResNeXt-272 (2x32d) model for CIFAR-10 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=272, cardinality=2, bottleneck_width=32, model_name="resnext272_2x32d_cifar10", **kwargs) def resnext272_2x32d_cifar100(classes=100, **kwargs): """ ResNeXt-272 (2x32d) model for CIFAR-100 from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=272, cardinality=2, bottleneck_width=32, model_name="resnext272_2x32d_cifar100", **kwargs) def resnext272_2x32d_svhn(classes=10, **kwargs): """ ResNeXt-272 (2x32d) model for SVHN from 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_resnext_cifar(classes=classes, blocks=272, cardinality=2, bottleneck_width=32, model_name="resnext272_2x32d_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (resnext20_1x64d_cifar10, 10), (resnext20_1x64d_cifar100, 100), (resnext20_1x64d_svhn, 10), (resnext20_2x32d_cifar10, 10), (resnext20_2x32d_cifar100, 100), (resnext20_2x32d_svhn, 10), (resnext20_2x64d_cifar10, 10), (resnext20_2x64d_cifar100, 100), (resnext20_2x64d_svhn, 10), (resnext20_4x16d_cifar10, 10), (resnext20_4x16d_cifar100, 100), (resnext20_4x16d_svhn, 10), (resnext20_4x32d_cifar10, 10), (resnext20_4x32d_cifar100, 100), (resnext20_4x32d_svhn, 10), (resnext20_8x8d_cifar10, 10), (resnext20_8x8d_cifar100, 100), (resnext20_8x8d_svhn, 10), (resnext20_8x16d_cifar10, 10), (resnext20_8x16d_cifar100, 100), (resnext20_8x16d_svhn, 10), (resnext20_16x4d_cifar10, 10), (resnext20_16x4d_cifar100, 100), (resnext20_16x4d_svhn, 10), (resnext20_16x8d_cifar10, 10), (resnext20_16x8d_cifar100, 100), (resnext20_16x8d_svhn, 10), (resnext20_32x2d_cifar10, 10), (resnext20_32x2d_cifar100, 100), (resnext20_32x2d_svhn, 10), (resnext20_32x4d_cifar10, 10), (resnext20_32x4d_cifar100, 100), (resnext20_32x4d_svhn, 10), (resnext20_64x1d_cifar10, 10), (resnext20_64x1d_cifar100, 100), (resnext20_64x1d_svhn, 10), (resnext20_64x2d_cifar10, 10), (resnext20_64x2d_cifar100, 100), (resnext20_64x2d_svhn, 10), (resnext29_32x4d_cifar10, 10), (resnext29_32x4d_cifar100, 100), (resnext29_32x4d_svhn, 10), (resnext29_16x64d_cifar10, 10), (resnext29_16x64d_cifar100, 100), (resnext29_16x64d_svhn, 10), (resnext56_1x64d_cifar10, 10), (resnext56_1x64d_cifar100, 100), (resnext56_1x64d_svhn, 10), (resnext56_2x32d_cifar10, 10), (resnext56_2x32d_cifar100, 100), (resnext56_2x32d_svhn, 10), (resnext56_4x16d_cifar10, 10), (resnext56_4x16d_cifar100, 100), (resnext56_4x16d_svhn, 10), (resnext56_8x8d_cifar10, 10), (resnext56_8x8d_cifar100, 100), (resnext56_8x8d_svhn, 10), (resnext56_16x4d_cifar10, 10), (resnext56_16x4d_cifar100, 100), (resnext56_16x4d_svhn, 10), (resnext56_32x2d_cifar10, 10), (resnext56_32x2d_cifar100, 100), (resnext56_32x2d_svhn, 10), (resnext56_64x1d_cifar10, 10), (resnext56_64x1d_cifar100, 100), (resnext56_64x1d_svhn, 10), (resnext272_1x64d_cifar10, 10), (resnext272_1x64d_cifar100, 100), (resnext272_1x64d_svhn, 10), (resnext272_2x32d_cifar10, 10), (resnext272_2x32d_cifar100, 100), (resnext272_2x32d_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resnext20_1x64d_cifar10 or weight_count == 3446602) assert (model != resnext20_1x64d_cifar100 or weight_count == 3538852) assert (model != resnext20_1x64d_svhn or weight_count == 3446602) assert (model != resnext20_2x32d_cifar10 or weight_count == 2672458) assert (model != resnext20_2x32d_cifar100 or weight_count == 2764708) assert (model != resnext20_2x32d_svhn or weight_count == 2672458) assert (model != resnext20_2x64d_cifar10 or weight_count == 6198602) assert (model != resnext20_2x64d_cifar100 or weight_count == 6290852) assert (model != resnext20_2x64d_svhn or weight_count == 6198602) assert (model != resnext20_4x16d_cifar10 or weight_count == 2285386) assert (model != resnext20_4x16d_cifar100 or weight_count == 2377636) assert (model != resnext20_4x16d_svhn or weight_count == 2285386) assert (model != resnext20_4x32d_cifar10 or weight_count == 4650314) assert (model != resnext20_4x32d_cifar100 or weight_count == 4742564) assert (model != resnext20_4x32d_svhn or weight_count == 4650314) assert (model != resnext20_8x8d_cifar10 or weight_count == 2091850) assert (model != resnext20_8x8d_cifar100 or weight_count == 2184100) assert (model != resnext20_8x8d_svhn or weight_count == 2091850) assert (model != resnext20_8x16d_cifar10 or weight_count == 3876170) assert (model != resnext20_8x16d_cifar100 or weight_count == 3968420) assert (model != resnext20_8x16d_svhn or weight_count == 3876170) assert (model != resnext20_16x4d_cifar10 or weight_count == 1995082) assert (model != resnext20_16x4d_cifar100 or weight_count == 2087332) assert (model != resnext20_16x4d_svhn or weight_count == 1995082) assert (model != resnext20_16x8d_cifar10 or weight_count == 3489098) assert (model != resnext20_16x8d_cifar100 or weight_count == 3581348) assert (model != resnext20_16x8d_svhn or weight_count == 3489098) assert (model != resnext20_32x2d_cifar10 or weight_count == 1946698) assert (model != resnext20_32x2d_cifar100 or weight_count == 2038948) assert (model != resnext20_32x2d_svhn or weight_count == 1946698) assert (model != resnext20_32x4d_cifar10 or weight_count == 3295562) assert (model != resnext20_32x4d_cifar100 or weight_count == 3387812) assert (model != resnext20_32x4d_svhn or weight_count == 3295562) assert (model != resnext20_64x1d_cifar10 or weight_count == 1922506) assert (model != resnext20_64x1d_cifar100 or weight_count == 2014756) assert (model != resnext20_64x1d_svhn or weight_count == 1922506) assert (model != resnext20_64x2d_cifar10 or weight_count == 3198794) assert (model != resnext20_64x2d_cifar100 or weight_count == 3291044) assert (model != resnext20_64x2d_svhn or weight_count == 3198794) assert (model != resnext29_32x4d_cifar10 or weight_count == 4775754) assert (model != resnext29_32x4d_cifar100 or weight_count == 4868004) assert (model != resnext29_32x4d_svhn or weight_count == 4775754) assert (model != resnext29_16x64d_cifar10 or weight_count == 68155210) assert (model != resnext29_16x64d_cifar100 or weight_count == 68247460) assert (model != resnext29_16x64d_svhn or weight_count == 68155210) assert (model != resnext56_1x64d_cifar10 or weight_count == 9317194) assert (model != resnext56_1x64d_cifar100 or weight_count == 9409444) assert (model != resnext56_1x64d_svhn or weight_count == 9317194) assert (model != resnext56_2x32d_cifar10 or weight_count == 6994762) assert (model != resnext56_2x32d_cifar100 or weight_count == 7087012) assert (model != resnext56_2x32d_svhn or weight_count == 6994762) assert (model != resnext56_4x16d_cifar10 or weight_count == 5833546) assert (model != resnext56_4x16d_cifar100 or weight_count == 5925796) assert (model != resnext56_4x16d_svhn or weight_count == 5833546) assert (model != resnext56_8x8d_cifar10 or weight_count == 5252938) assert (model != resnext56_8x8d_cifar100 or weight_count == 5345188) assert (model != resnext56_8x8d_svhn or weight_count == 5252938) assert (model != resnext56_16x4d_cifar10 or weight_count == 4962634) assert (model != resnext56_16x4d_cifar100 or weight_count == 5054884) assert (model != resnext56_16x4d_svhn or weight_count == 4962634) assert (model != resnext56_32x2d_cifar10 or weight_count == 4817482) assert (model != resnext56_32x2d_cifar100 or weight_count == 4909732) assert (model != resnext56_32x2d_svhn or weight_count == 4817482) assert (model != resnext56_64x1d_cifar10 or weight_count == 4744906) assert (model != resnext56_64x1d_cifar100 or weight_count == 4837156) assert (model != resnext56_64x1d_svhn or weight_count == 4744906) assert (model != resnext272_1x64d_cifar10 or weight_count == 44540746) assert (model != resnext272_1x64d_cifar100 or weight_count == 44632996) assert (model != resnext272_1x64d_svhn or weight_count == 44540746) assert (model != resnext272_2x32d_cifar10 or weight_count == 32928586) assert (model != resnext272_2x32d_cifar100 or weight_count == 33020836) assert (model != resnext272_2x32d_svhn or weight_count == 32928586) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
71,735
39.075978
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/densenet_cifar.py
""" DenseNet for CIFAR/SVHN, implemented in Gluon. Original paper: 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. """ __all__ = ['CIFARDenseNet', 'densenet40_k12_cifar10', 'densenet40_k12_cifar100', 'densenet40_k12_svhn', 'densenet40_k12_bc_cifar10', 'densenet40_k12_bc_cifar100', 'densenet40_k12_bc_svhn', 'densenet40_k24_bc_cifar10', 'densenet40_k24_bc_cifar100', 'densenet40_k24_bc_svhn', 'densenet40_k36_bc_cifar10', 'densenet40_k36_bc_cifar100', 'densenet40_k36_bc_svhn', 'densenet100_k12_cifar10', 'densenet100_k12_cifar100', 'densenet100_k12_svhn', 'densenet100_k24_cifar10', 'densenet100_k24_cifar100', 'densenet100_k24_svhn', 'densenet100_k12_bc_cifar10', 'densenet100_k12_bc_cifar100', 'densenet100_k12_bc_svhn', 'densenet190_k40_bc_cifar10', 'densenet190_k40_bc_cifar100', 'densenet190_k40_bc_svhn', 'densenet250_k24_bc_cifar10', 'densenet250_k24_bc_cifar100', 'densenet250_k24_bc_svhn'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3, pre_conv3x3_block from .preresnet import PreResActivation from .densenet import DenseUnit, TransitionBlock class DenseSimpleUnit(HybridBlock): """ DenseNet simple unit for CIFAR. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, dropout_rate, **kwargs): super(DenseSimpleUnit, self).__init__(**kwargs) self.use_dropout = (dropout_rate != 0.0) inc_channels = out_channels - in_channels with self.name_scope(): self.conv = pre_conv3x3_block( in_channels=in_channels, out_channels=inc_channels, bn_use_global_stats=bn_use_global_stats) if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x): identity = x x = self.conv(x) if self.use_dropout: x = self.dropout(x) x = F.concat(identity, x, dim=1) return x class CIFARDenseNet(HybridBlock): """ DenseNet model for CIFAR from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, bn_use_global_stats=False, dropout_rate=0.0, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARDenseNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes unit_class = DenseUnit if bottleneck else DenseSimpleUnit with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): if i != 0: stage.add(TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2), bn_use_global_stats=bn_use_global_stats)) in_channels = in_channels // 2 for j, out_channels in enumerate(channels_per_stage): stage.add(unit_class( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, dropout_rate=dropout_rate)) in_channels = out_channels self.features.add(stage) self.features.add(PreResActivation( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_densenet_cifar(classes, blocks, growth_rate, bottleneck, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DenseNet model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. blocks : int Number of blocks. growth_rate : int Growth rate. bottleneck : bool Whether to use a bottleneck or simple block in units. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (classes in [10, 100]) if bottleneck: assert ((blocks - 4) % 6 == 0) layers = [(blocks - 4) // 6] * 3 else: assert ((blocks - 4) % 3 == 0) layers = [(blocks - 4) // 3] * 3 init_block_channels = 2 * growth_rate from functools import reduce channels = reduce( lambda xi, yi: xi + [reduce( lambda xj, yj: xj + [xj[-1] + yj], [growth_rate] * yi, [xi[-1][-1] // 2])[1:]], layers, [[init_block_channels * 2]])[1:] net = CIFARDenseNet( channels=channels, init_block_channels=init_block_channels, classes=classes, bottleneck=bottleneck, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def densenet40_k12_cifar10(classes=10, **kwargs): """ DenseNet-40 (k=12) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=12, bottleneck=False, model_name="densenet40_k12_cifar10", **kwargs) def densenet40_k12_cifar100(classes=100, **kwargs): """ DenseNet-40 (k=12) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=12, bottleneck=False, model_name="densenet40_k12_cifar100", **kwargs) def densenet40_k12_svhn(classes=10, **kwargs): """ DenseNet-40 (k=12) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=12, bottleneck=False, model_name="densenet40_k12_svhn", **kwargs) def densenet40_k12_bc_cifar10(classes=10, **kwargs): """ DenseNet-BC-40 (k=12) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=12, bottleneck=True, model_name="densenet40_k12_bc_cifar10", **kwargs) def densenet40_k12_bc_cifar100(classes=100, **kwargs): """ DenseNet-BC-40 (k=12) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=12, bottleneck=True, model_name="densenet40_k12_bc_cifar100", **kwargs) def densenet40_k12_bc_svhn(classes=10, **kwargs): """ DenseNet-BC-40 (k=12) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=12, bottleneck=True, model_name="densenet40_k12_bc_svhn", **kwargs) def densenet40_k24_bc_cifar10(classes=10, **kwargs): """ DenseNet-BC-40 (k=24) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True, model_name="densenet40_k24_bc_cifar10", **kwargs) def densenet40_k24_bc_cifar100(classes=100, **kwargs): """ DenseNet-BC-40 (k=24) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True, model_name="densenet40_k24_bc_cifar100", **kwargs) def densenet40_k24_bc_svhn(classes=10, **kwargs): """ DenseNet-BC-40 (k=24) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=24, bottleneck=True, model_name="densenet40_k24_bc_svhn", **kwargs) def densenet40_k36_bc_cifar10(classes=10, **kwargs): """ DenseNet-BC-40 (k=36) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True, model_name="densenet40_k36_bc_cifar10", **kwargs) def densenet40_k36_bc_cifar100(classes=100, **kwargs): """ DenseNet-BC-40 (k=36) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True, model_name="densenet40_k36_bc_cifar100", **kwargs) def densenet40_k36_bc_svhn(classes=10, **kwargs): """ DenseNet-BC-40 (k=36) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=40, growth_rate=36, bottleneck=True, model_name="densenet40_k36_bc_svhn", **kwargs) def densenet100_k12_cifar10(classes=10, **kwargs): """ DenseNet-100 (k=12) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=12, bottleneck=False, model_name="densenet100_k12_cifar10", **kwargs) def densenet100_k12_cifar100(classes=100, **kwargs): """ DenseNet-100 (k=12) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=12, bottleneck=False, model_name="densenet100_k12_cifar100", **kwargs) def densenet100_k12_svhn(classes=10, **kwargs): """ DenseNet-100 (k=12) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=12, bottleneck=False, model_name="densenet100_k12_svhn", **kwargs) def densenet100_k24_cifar10(classes=10, **kwargs): """ DenseNet-100 (k=24) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=24, bottleneck=False, model_name="densenet100_k24_cifar10", **kwargs) def densenet100_k24_cifar100(classes=100, **kwargs): """ DenseNet-100 (k=24) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=24, bottleneck=False, model_name="densenet100_k24_cifar100", **kwargs) def densenet100_k24_svhn(classes=10, **kwargs): """ DenseNet-100 (k=24) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=24, bottleneck=False, model_name="densenet100_k24_svhn", **kwargs) def densenet100_k12_bc_cifar10(classes=10, **kwargs): """ DenseNet-BC-100 (k=12) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=12, bottleneck=True, model_name="densenet100_k12_bc_cifar10", **kwargs) def densenet100_k12_bc_cifar100(classes=100, **kwargs): """ DenseNet-BC-100 (k=12) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=12, bottleneck=True, model_name="densenet100_k12_bc_cifar100", **kwargs) def densenet100_k12_bc_svhn(classes=10, **kwargs): """ DenseNet-BC-100 (k=12) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=100, growth_rate=12, bottleneck=True, model_name="densenet100_k12_bc_svhn", **kwargs) def densenet190_k40_bc_cifar10(classes=10, **kwargs): """ DenseNet-BC-190 (k=40) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=190, growth_rate=40, bottleneck=True, model_name="densenet190_k40_bc_cifar10", **kwargs) def densenet190_k40_bc_cifar100(classes=100, **kwargs): """ DenseNet-BC-190 (k=40) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=190, growth_rate=40, bottleneck=True, model_name="densenet190_k40_bc_cifar100", **kwargs) def densenet190_k40_bc_svhn(classes=10, **kwargs): """ DenseNet-BC-190 (k=40) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=190, growth_rate=40, bottleneck=True, model_name="densenet190_k40_bc_svhn", **kwargs) def densenet250_k24_bc_cifar10(classes=10, **kwargs): """ DenseNet-BC-250 (k=24) model for CIFAR-10 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=250, growth_rate=24, bottleneck=True, model_name="densenet250_k24_bc_cifar10", **kwargs) def densenet250_k24_bc_cifar100(classes=100, **kwargs): """ DenseNet-BC-250 (k=24) model for CIFAR-100 from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=250, growth_rate=24, bottleneck=True, model_name="densenet250_k24_bc_cifar100", **kwargs) def densenet250_k24_bc_svhn(classes=10, **kwargs): """ DenseNet-BC-250 (k=24) model for SVHN from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_densenet_cifar(classes=classes, blocks=250, growth_rate=24, bottleneck=True, model_name="densenet250_k24_bc_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (densenet40_k12_cifar10, 10), (densenet40_k12_cifar100, 100), (densenet40_k12_svhn, 10), (densenet40_k12_bc_cifar10, 10), (densenet40_k12_bc_cifar100, 100), (densenet40_k12_bc_svhn, 10), (densenet40_k24_bc_cifar10, 10), (densenet40_k24_bc_cifar100, 100), (densenet40_k24_bc_svhn, 10), (densenet40_k36_bc_cifar10, 10), (densenet40_k36_bc_cifar100, 100), (densenet40_k36_bc_svhn, 10), (densenet100_k12_cifar10, 10), (densenet100_k12_cifar100, 100), (densenet100_k12_svhn, 10), (densenet100_k24_cifar10, 10), (densenet100_k24_cifar100, 100), (densenet100_k24_svhn, 10), (densenet100_k12_bc_cifar10, 10), (densenet100_k12_bc_cifar100, 100), (densenet100_k12_bc_svhn, 10), (densenet190_k40_bc_cifar10, 10), (densenet190_k40_bc_cifar100, 100), (densenet190_k40_bc_svhn, 10), (densenet250_k24_bc_cifar10, 10), (densenet250_k24_bc_cifar100, 100), (densenet250_k24_bc_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != densenet40_k12_cifar10 or weight_count == 599050) assert (model != densenet40_k12_cifar100 or weight_count == 622360) assert (model != densenet40_k12_svhn or weight_count == 599050) assert (model != densenet40_k12_bc_cifar10 or weight_count == 176122) assert (model != densenet40_k12_bc_cifar100 or weight_count == 188092) assert (model != densenet40_k12_bc_svhn or weight_count == 176122) assert (model != densenet40_k24_bc_cifar10 or weight_count == 690346) assert (model != densenet40_k24_bc_cifar100 or weight_count == 714196) assert (model != densenet40_k24_bc_svhn or weight_count == 690346) assert (model != densenet40_k36_bc_cifar10 or weight_count == 1542682) assert (model != densenet40_k36_bc_cifar100 or weight_count == 1578412) assert (model != densenet40_k36_bc_svhn or weight_count == 1542682) assert (model != densenet100_k12_cifar10 or weight_count == 4068490) assert (model != densenet100_k12_cifar100 or weight_count == 4129600) assert (model != densenet100_k12_svhn or weight_count == 4068490) assert (model != densenet100_k24_cifar10 or weight_count == 16114138) assert (model != densenet100_k24_cifar100 or weight_count == 16236268) assert (model != densenet100_k24_svhn or weight_count == 16114138) assert (model != densenet100_k12_bc_cifar10 or weight_count == 769162) assert (model != densenet100_k12_bc_cifar100 or weight_count == 800032) assert (model != densenet100_k12_bc_svhn or weight_count == 769162) assert (model != densenet190_k40_bc_cifar10 or weight_count == 25624430) assert (model != densenet190_k40_bc_cifar100 or weight_count == 25821620) assert (model != densenet190_k40_bc_svhn or weight_count == 25624430) assert (model != densenet250_k24_bc_cifar10 or weight_count == 15324406) assert (model != densenet250_k24_bc_cifar100 or weight_count == 15480556) assert (model != densenet250_k24_bc_svhn or weight_count == 15324406) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
32,370
37.354265
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/bninception.py
""" BN-Inception for ImageNet-1K, implemented in Gluon. Original paper: 'Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,' https://arxiv.org/abs/1502.03167. """ __all__ = ['BNInception', 'bninception'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import conv1x1_block, conv3x3_block, conv7x7_block class Inception3x3Branch(HybridBlock): """ BN-Inception 3x3 branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of intermediate channels. strides : int or tuple/list of 2 int, default 1 Strides of the second convolution. use_bias : bool, default True Whether the convolution layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layers. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, mid_channels, strides=1, use_bias=True, use_bn=True, bn_use_global_stats=False, **kwargs): super(Inception3x3Branch, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, strides=strides, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class InceptionDouble3x3Branch(HybridBlock): """ BN-Inception double 3x3 branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of intermediate channels. strides : int or tuple/list of 2 int, default 1 Strides of the second convolution. use_bias : bool, default True Whether the convolution layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layers. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, mid_channels, strides=1, use_bias=True, use_bn=True, bn_use_global_stats=False, **kwargs): super(InceptionDouble3x3Branch, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, strides=strides, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class InceptionPoolBranch(HybridBlock): """ BN-Inception avg-pool branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. avg_pool : bool Whether use average pooling or max pooling. use_bias : bool Whether the convolution layer uses a bias vector. use_bn : bool Whether to use BatchNorm layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, avg_pool, use_bias, use_bn, bn_use_global_stats, **kwargs): super(InceptionPoolBranch, self).__init__(**kwargs) with self.name_scope(): if avg_pool: self.pool = nn.AvgPool2D( pool_size=3, strides=1, padding=1, ceil_mode=True, count_include_pad=True) else: self.pool = nn.MaxPool2D( pool_size=3, strides=1, padding=1, ceil_mode=True) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.pool(x) x = self.conv(x) return x class StemBlock(HybridBlock): """ BN-Inception stem block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of intermediate channels. use_bias : bool Whether the convolution layer uses a bias vector. use_bn : bool Whether to use BatchNorm layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, mid_channels, use_bias, use_bn, bn_use_global_stats, **kwargs): super(StemBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv7x7_block( in_channels=in_channels, out_channels=mid_channels, strides=2, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) self.pool1 = nn.MaxPool2D( pool_size=3, strides=2, padding=0, ceil_mode=True) self.conv2 = Inception3x3Branch( in_channels=mid_channels, out_channels=out_channels, mid_channels=mid_channels, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats) self.pool2 = nn.MaxPool2D( pool_size=3, strides=2, padding=0, ceil_mode=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.pool1(x) x = self.conv2(x) x = self.pool2(x) return x class InceptionBlock(HybridBlock): """ BN-Inception unit. Parameters: ---------- in_channels : int Number of input channels. mid1_channels_list : list of int Number of pre-middle channels for branches. mid2_channels_list : list of int Number of middle channels for branches. avg_pool : bool Whether use average pooling or max pooling. use_bias : bool Whether the convolution layer uses a bias vector. use_bn : bool Whether to use BatchNorm layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, mid1_channels_list, mid2_channels_list, avg_pool, use_bias, use_bn, bn_use_global_stats, **kwargs): super(InceptionBlock, self).__init__(**kwargs) assert (len(mid1_channels_list) == 2) assert (len(mid2_channels_list) == 4) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(conv1x1_block( in_channels=in_channels, out_channels=mid2_channels_list[0], use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) self.branches.add(Inception3x3Branch( in_channels=in_channels, out_channels=mid2_channels_list[1], mid_channels=mid1_channels_list[0], use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) self.branches.add(InceptionDouble3x3Branch( in_channels=in_channels, out_channels=mid2_channels_list[2], mid_channels=mid1_channels_list[1], use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) self.branches.add(InceptionPoolBranch( in_channels=in_channels, out_channels=mid2_channels_list[3], avg_pool=avg_pool, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class ReductionBlock(HybridBlock): """ BN-Inception reduction block. Parameters: ---------- in_channels : int Number of input channels. mid1_channels_list : list of int Number of pre-middle channels for branches. mid2_channels_list : list of int Number of middle channels for branches. use_bias : bool Whether the convolution layer uses a bias vector. use_bn : bool Whether to use BatchNorm layers. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, mid1_channels_list, mid2_channels_list, use_bias, use_bn, bn_use_global_stats, **kwargs): super(ReductionBlock, self).__init__(**kwargs) assert (len(mid1_channels_list) == 2) assert (len(mid2_channels_list) == 4) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(Inception3x3Branch( in_channels=in_channels, out_channels=mid2_channels_list[1], mid_channels=mid1_channels_list[0], strides=2, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) self.branches.add(InceptionDouble3x3Branch( in_channels=in_channels, out_channels=mid2_channels_list[2], mid_channels=mid1_channels_list[1], strides=2, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) self.branches.add(nn.MaxPool2D( pool_size=3, strides=2, padding=0, ceil_mode=True)) def hybrid_forward(self, F, x): x = self.branches(x) return x class BNInception(HybridBlock): """ BN-Inception model from 'Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,' https://arxiv.org/abs/1502.03167. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels_list : list of int Number of output channels for the initial unit. mid1_channels_list : list of list of list of int Number of pre-middle channels for each unit. mid2_channels_list : list of list of list of int Number of middle channels for each unit. use_bias : bool, default True Whether the convolution layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layers. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels_list, mid1_channels_list, mid2_channels_list, use_bias=True, use_bn=True, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(BNInception, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(StemBlock( in_channels=in_channels, out_channels=init_block_channels_list[1], mid_channels=init_block_channels_list[0], use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels_list[-1] for i, channels_per_stage in enumerate(channels): mid1_channels_list_i = mid1_channels_list[i] mid2_channels_list_i = mid2_channels_list[i] stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): if (j == 0) and (i != 0): stage.add(ReductionBlock( in_channels=in_channels, mid1_channels_list=mid1_channels_list_i[j], mid2_channels_list=mid2_channels_list_i[j], use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) else: avg_pool = (i != len(channels) - 1) or (j != len(channels_per_stage) - 1) stage.add(InceptionBlock( in_channels=in_channels, mid1_channels_list=mid1_channels_list_i[j], mid2_channels_list=mid2_channels_list_i[j], avg_pool=avg_pool, use_bias=use_bias, use_bn=use_bn, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_bninception(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create BN-Inception model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ init_block_channels_list = [64, 192] channels = [[256, 320], [576, 576, 576, 608, 608], [1056, 1024, 1024]] mid1_channels_list = [ [[64, 64], [64, 64]], [[128, 64], # 3c [64, 96], # 4a [96, 96], # 4a [128, 128], # 4c [128, 160]], # 4d [[128, 192], # 4e [192, 160], # 5a [192, 192]], ] mid2_channels_list = [ [[64, 64, 96, 32], [64, 96, 96, 64]], [[0, 160, 96, 0], # 3c [224, 96, 128, 128], # 4a [192, 128, 128, 128], # 4b [160, 160, 160, 128], # 4c [96, 192, 192, 128]], # 4d [[0, 192, 256, 0], # 4e [352, 320, 224, 128], # 5a [352, 320, 224, 128]], ] net = BNInception( channels=channels, init_block_channels_list=init_block_channels_list, mid1_channels_list=mid1_channels_list, mid2_channels_list=mid2_channels_list, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def bninception(**kwargs): """ BN-Inception model from 'Batch Normalization: Accelerating Deep Network Training by Reducing Internal Covariate Shift,' https://arxiv.org/abs/1502.03167. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_bninception(model_name="bninception", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ bninception, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != bninception or weight_count == 11295240) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
19,996
33.008503
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/msdnet.py
""" MSDNet for ImageNet-1K, implemented in Gluon. Original paper: 'Multi-Scale Dense Networks for Resource Efficient Image Classification,' https://arxiv.org/abs/1703.09844. """ __all__ = ['MSDNet', 'msdnet22', 'MultiOutputSequential', 'MSDFeatureBlock'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, DualPathSequential from .resnet import ResInitBlock class MultiOutputSequential(nn.HybridSequential): """ A sequential container for blocks. Blocks will be executed in the order they are added. Output value contains results from all blocks. """ def __init__(self, **kwargs): super(MultiOutputSequential, self).__init__(**kwargs) def hybrid_forward(self, F, x): outs = [] for block in self._children.values(): x = block(x) outs.append(x) return outs class MultiBlockSequential(nn.HybridSequential): """ A sequential container for blocks. Blocks will be executed in the order they are added. Input is a list with length equal to number of blocks. """ def __init__(self, **kwargs): super(MultiBlockSequential, self).__init__(**kwargs) def hybrid_forward(self, F, x0, x_rest): outs = [] for block, x_i in zip(self._children.values(), [x0] + x_rest): y = block(x_i) outs.append(y) return outs class MSDBaseBlock(HybridBlock): """ MSDNet base block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. use_bottleneck : bool Whether to use a bottleneck. bottleneck_factor : int Bottleneck factor. """ def __init__(self, in_channels, out_channels, strides, use_bottleneck, bottleneck_factor, **kwargs): super(MSDBaseBlock, self).__init__(**kwargs) self.use_bottleneck = use_bottleneck mid_channels = min(in_channels, bottleneck_factor * out_channels) if use_bottleneck else in_channels with self.name_scope(): if self.use_bottleneck: self.bn_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, strides=strides) def hybrid_forward(self, F, x): if self.use_bottleneck: x = self.bn_conv(x) x = self.conv(x) return x class MSDFirstScaleBlock(HybridBlock): """ MSDNet first scale dense block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. use_bottleneck : bool Whether to use a bottleneck. bottleneck_factor : int Bottleneck factor. """ def __init__(self, in_channels, out_channels, use_bottleneck, bottleneck_factor, **kwargs): super(MSDFirstScaleBlock, self).__init__(**kwargs) assert (out_channels > in_channels) inc_channels = out_channels - in_channels with self.name_scope(): self.block = MSDBaseBlock( in_channels=in_channels, out_channels=inc_channels, strides=1, use_bottleneck=use_bottleneck, bottleneck_factor=bottleneck_factor) def hybrid_forward(self, F, x): y = self.block(x) y = F.concat(x, y, dim=1) return y class MSDScaleBlock(HybridBlock): """ MSDNet ordinary scale dense block. Parameters: ---------- in_channels_prev : int Number of input channels for the previous scale. in_channels : int Number of input channels for the current scale. out_channels : int Number of output channels. use_bottleneck : bool Whether to use a bottleneck. bottleneck_factor_prev : int Bottleneck factor for the previous scale. bottleneck_factor : int Bottleneck factor for the current scale. """ def __init__(self, in_channels_prev, in_channels, out_channels, use_bottleneck, bottleneck_factor_prev, bottleneck_factor, **kwargs): super(MSDScaleBlock, self).__init__(**kwargs) assert (out_channels > in_channels) assert (out_channels % 2 == 0) inc_channels = out_channels - in_channels mid_channels = inc_channels // 2 with self.name_scope(): self.down_block = MSDBaseBlock( in_channels=in_channels_prev, out_channels=mid_channels, strides=2, use_bottleneck=use_bottleneck, bottleneck_factor=bottleneck_factor_prev) self.curr_block = MSDBaseBlock( in_channels=in_channels, out_channels=mid_channels, strides=1, use_bottleneck=use_bottleneck, bottleneck_factor=bottleneck_factor) def hybrid_forward(self, F, x_prev, x): y_prev = self.down_block(x_prev) y = self.curr_block(x) x = F.concat(x, y_prev, y, dim=1) return x class MSDInitLayer(HybridBlock): """ MSDNet initial (so-called first) layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : list/tuple of int Number of output channels for each scale. """ def __init__(self, in_channels, out_channels, **kwargs): super(MSDInitLayer, self).__init__(**kwargs) with self.name_scope(): self.scale_blocks = MultiOutputSequential() for i, out_channels_per_scale in enumerate(out_channels): if i == 0: self.scale_blocks.add(ResInitBlock( in_channels=in_channels, out_channels=out_channels_per_scale)) else: self.scale_blocks.add(conv3x3_block( in_channels=in_channels, out_channels=out_channels_per_scale, strides=2)) in_channels = out_channels_per_scale def hybrid_forward(self, F, x): y = self.scale_blocks(x) return y class MSDLayer(HybridBlock): """ MSDNet ordinary layer. Parameters: ---------- in_channels : list/tuple of int Number of input channels for each input scale. out_channels : list/tuple of int Number of output channels for each output scale. use_bottleneck : bool Whether to use a bottleneck. bottleneck_factors : list/tuple of int Bottleneck factor for each input scale. """ def __init__(self, in_channels, out_channels, use_bottleneck, bottleneck_factors, **kwargs): super(MSDLayer, self).__init__(**kwargs) in_scales = len(in_channels) out_scales = len(out_channels) self.dec_scales = in_scales - out_scales assert (self.dec_scales >= 0) with self.name_scope(): self.scale_blocks = nn.HybridSequential(prefix="") for i in range(out_scales): if (i == 0) and (self.dec_scales == 0): self.scale_blocks.add(MSDFirstScaleBlock( in_channels=in_channels[self.dec_scales + i], out_channels=out_channels[i], use_bottleneck=use_bottleneck, bottleneck_factor=bottleneck_factors[self.dec_scales + i])) else: self.scale_blocks.add(MSDScaleBlock( in_channels_prev=in_channels[self.dec_scales + i - 1], in_channels=in_channels[self.dec_scales + i], out_channels=out_channels[i], use_bottleneck=use_bottleneck, bottleneck_factor_prev=bottleneck_factors[self.dec_scales + i - 1], bottleneck_factor=bottleneck_factors[self.dec_scales + i])) def hybrid_forward(self, F, x0, x_rest): x = [x0] + x_rest outs = [] for i in range(len(self.scale_blocks)): if (i == 0) and (self.dec_scales == 0): y = self.scale_blocks[i](x[i]) else: y = self.scale_blocks[i]( x[self.dec_scales + i - 1], x[self.dec_scales + i]) outs.append(y) return outs[0], outs[1:] class MSDTransitionLayer(HybridBlock): """ MSDNet transition layer. Parameters: ---------- in_channels : list/tuple of int Number of input channels for each scale. out_channels : list/tuple of int Number of output channels for each scale. """ def __init__(self, in_channels, out_channels, **kwargs): super(MSDTransitionLayer, self).__init__(**kwargs) assert (len(in_channels) == len(out_channels)) with self.name_scope(): self.scale_blocks = MultiBlockSequential() for i in range(len(out_channels)): self.scale_blocks.add(conv1x1_block( in_channels=in_channels[i], out_channels=out_channels[i])) def hybrid_forward(self, F, x0, x_rest): y = self.scale_blocks(x0, x_rest) return y[0], y[1:] class MSDFeatureBlock(HybridBlock): """ MSDNet feature block (stage of cascade, so-called block). Parameters: ---------- in_channels : list of list of int Number of input channels for each layer and for each input scale. out_channels : list of list of int Number of output channels for each layer and for each output scale. use_bottleneck : bool Whether to use a bottleneck. bottleneck_factors : list of list of int Bottleneck factor for each layer and for each input scale. """ def __init__(self, in_channels, out_channels, use_bottleneck, bottleneck_factors, **kwargs): super(MSDFeatureBlock, self).__init__(**kwargs) with self.name_scope(): self.blocks = DualPathSequential(prefix="") for i, out_channels_per_layer in enumerate(out_channels): if len(bottleneck_factors[i]) == 0: self.blocks.add(MSDTransitionLayer( in_channels=in_channels, out_channels=out_channels_per_layer)) else: self.blocks.add(MSDLayer( in_channels=in_channels, out_channels=out_channels_per_layer, use_bottleneck=use_bottleneck, bottleneck_factors=bottleneck_factors[i])) in_channels = out_channels_per_layer def hybrid_forward(self, F, x0, x_rest): x0, x_rest = self.blocks(x0, x_rest) return [x0] + x_rest class MSDClassifier(HybridBlock): """ MSDNet classifier. Parameters: ---------- in_channels : int Number of input channels. classes : int Number of classification classes. """ def __init__(self, in_channels, classes, **kwargs): super(MSDClassifier, self).__init__(**kwargs) with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3_block( in_channels=in_channels, out_channels=in_channels, strides=2)) self.features.add(conv3x3_block( in_channels=in_channels, out_channels=in_channels, strides=2)) self.features.add(nn.AvgPool2D( pool_size=2, strides=2)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x class MSDNet(HybridBlock): """ MSDNet model from 'Multi-Scale Dense Networks for Resource Efficient Image Classification,' https://arxiv.org/abs/1703.09844. Parameters: ---------- channels : list of list of list of int Number of output channels for each unit. init_layer_channels : list of int Number of output channels for the initial layer. num_feature_blocks : int Number of subnets. use_bottleneck : bool Whether to use a bottleneck. bottleneck_factors : list of list of int Bottleneck factor for each layers and for each input scale. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_layer_channels, num_feature_blocks, use_bottleneck, bottleneck_factors, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(MSDNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.init_layer = MSDInitLayer( in_channels=in_channels, out_channels=init_layer_channels) in_channels = init_layer_channels self.feature_blocks = nn.HybridSequential(prefix="") self.classifiers = nn.HybridSequential(prefix="") for i in range(num_feature_blocks): self.feature_blocks.add(MSDFeatureBlock( in_channels=in_channels, out_channels=channels[i], use_bottleneck=use_bottleneck, bottleneck_factors=bottleneck_factors[i])) in_channels = channels[i][-1] self.classifiers.add(MSDClassifier( in_channels=in_channels[-1], classes=classes)) def hybrid_forward(self, F, x, only_last=True): x = self.init_layer(x) outs = [] for feature_block, classifier in zip(self.feature_blocks, self.classifiers): x = feature_block(x[0], x[1:]) y = classifier(x[-1]) outs.append(y) if only_last: return outs[-1] else: return outs def get_msdnet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create MSDNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (blocks == 22) num_scales = 4 num_feature_blocks = 10 base = 4 step = 2 reduction_rate = 0.5 growth = 6 growth_factor = [1, 2, 4, 4] use_bottleneck = True bottleneck_factor_per_scales = [1, 2, 4, 4] assert (reduction_rate > 0.0) init_layer_channels = [64 * c for c in growth_factor[:num_scales]] step_mode = "even" layers_per_subnets = [base] for i in range(num_feature_blocks - 1): layers_per_subnets.append(step if step_mode == 'even' else step * i + 1) total_layers = sum(layers_per_subnets) interval = math.ceil(total_layers / num_scales) global_layer_ind = 0 channels = [] bottleneck_factors = [] in_channels_tmp = init_layer_channels in_scales = num_scales for i in range(num_feature_blocks): layers_per_subnet = layers_per_subnets[i] scales_i = [] channels_i = [] bottleneck_factors_i = [] for j in range(layers_per_subnet): out_scales = int(num_scales - math.floor(global_layer_ind / interval)) global_layer_ind += 1 scales_i += [out_scales] scale_offset = num_scales - out_scales in_dec_scales = num_scales - len(in_channels_tmp) out_channels = [in_channels_tmp[scale_offset - in_dec_scales + k] + growth * growth_factor[scale_offset + k] for k in range(out_scales)] in_dec_scales = num_scales - len(in_channels_tmp) bottleneck_factors_ij = bottleneck_factor_per_scales[in_dec_scales:][:len(in_channels_tmp)] in_channels_tmp = out_channels channels_i += [out_channels] bottleneck_factors_i += [bottleneck_factors_ij] if in_scales > out_scales: assert (in_channels_tmp[0] % growth_factor[scale_offset] == 0) out_channels1 = int(math.floor(in_channels_tmp[0] / growth_factor[scale_offset] * reduction_rate)) out_channels = [out_channels1 * growth_factor[scale_offset + k] for k in range(out_scales)] in_channels_tmp = out_channels channels_i += [out_channels] bottleneck_factors_i += [[]] in_scales = out_scales in_scales = scales_i[-1] channels += [channels_i] bottleneck_factors += [bottleneck_factors_i] net = MSDNet( channels=channels, init_layer_channels=init_layer_channels, num_feature_blocks=num_feature_blocks, use_bottleneck=use_bottleneck, bottleneck_factors=bottleneck_factors, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def msdnet22(**kwargs): """ MSDNet-22 model from 'Multi-Scale Dense Networks for Resource Efficient Image Classification,' https://arxiv.org/abs/1703.09844. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_msdnet(blocks=22, model_name="msdnet22", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ msdnet22, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != msdnet22 or weight_count == 20106676) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
20,581
31.566456
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/zfnet.py
""" ZFNet for ImageNet-1K, implemented in Gluon. Original paper: 'Visualizing and Understanding Convolutional Networks,' https://arxiv.org/abs/1311.2901. """ __all__ = ['zfnet', 'zfnetb'] import os from mxnet import cpu from .alexnet import AlexNet def get_zfnet(version="a", model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ZFNet model with specific parameters. Parameters: ---------- version : str, default 'a' Version of ZFNet ('a' or 'b'). model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if version == "a": channels = [[96], [256], [384, 384, 256]] kernel_sizes = [[7], [5], [3, 3, 3]] strides = [[2], [2], [1, 1, 1]] paddings = [[1], [0], [1, 1, 1]] use_lrn = True elif version == "b": channels = [[96], [256], [512, 1024, 512]] kernel_sizes = [[7], [5], [3, 3, 3]] strides = [[2], [2], [1, 1, 1]] paddings = [[1], [0], [1, 1, 1]] use_lrn = True else: raise ValueError("Unsupported ZFNet version {}".format(version)) net = AlexNet( channels=channels, kernel_sizes=kernel_sizes, strides=strides, paddings=paddings, use_lrn=use_lrn, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def zfnet(**kwargs): """ ZFNet model from 'Visualizing and Understanding Convolutional Networks,' https://arxiv.org/abs/1311.2901. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_zfnet(model_name="zfnet", **kwargs) def zfnetb(**kwargs): """ ZFNet-b model from 'Visualizing and Understanding Convolutional Networks,' https://arxiv.org/abs/1311.2901. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_zfnet(version="b", model_name="zfnetb", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ zfnet, zfnetb, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != zfnet or weight_count == 62357608) assert (model != zfnetb or weight_count == 107627624) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
4,058
28.201439
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/peleenet.py
""" PeleeNet for ImageNet-1K, implemented in Gluon. Original paper: 'Pelee: A Real-Time Object Detection System on Mobile Devices,' https://arxiv.org/abs/1804.06882. """ __all__ = ['PeleeNet', 'peleenet'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import conv1x1_block, conv3x3_block class PeleeBranch1(HybridBlock): """ PeleeNet branch type 1 block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of intermediate channels. strides : int or tuple/list of 2 int, default 1 Strides of the second convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, mid_channels, strides=1, bn_use_global_stats=False, **kwargs): super(PeleeBranch1, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class PeleeBranch2(HybridBlock): """ PeleeNet branch type 2 block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of intermediate channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, mid_channels, bn_use_global_stats, **kwargs): super(PeleeBranch2, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class StemBlock(HybridBlock): """ PeleeNet stem block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(StemBlock, self).__init__(**kwargs) mid1_channels = out_channels // 2 mid2_channels = out_channels * 2 with self.name_scope(): self.first_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, strides=2) self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(PeleeBranch1( in_channels=out_channels, out_channels=out_channels, mid_channels=mid1_channels, strides=2, bn_use_global_stats=bn_use_global_stats)) self.branches.add(nn.MaxPool2D( pool_size=2, strides=2, padding=0)) self.last_conv = conv1x1_block( in_channels=mid2_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.first_conv(x) x = self.branches(x) x = self.last_conv(x) return x class DenseBlock(HybridBlock): """ PeleeNet dense block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bottleneck_size : int Bottleneck width. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bottleneck_size, bn_use_global_stats, **kwargs): super(DenseBlock, self).__init__(**kwargs) inc_channels = (out_channels - in_channels) // 2 mid_channels = inc_channels * bottleneck_size with self.name_scope(): self.branch1 = PeleeBranch1( in_channels=in_channels, out_channels=inc_channels, mid_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.branch2 = PeleeBranch2( in_channels=in_channels, out_channels=inc_channels, mid_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x1 = self.branch1(x) x2 = self.branch2(x) x = F.concat(x, x1, x2, dim=1) return x class TransitionBlock(HybridBlock): """ PeleeNet's transition block, like in DensNet, but with ordinary convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, **kwargs): super(TransitionBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.pool = nn.AvgPool2D( pool_size=2, strides=2, padding=0) def hybrid_forward(self, F, x): x = self.conv(x) x = self.pool(x) return x class PeleeNet(HybridBlock): """ PeleeNet model from 'Pelee: A Real-Time Object Detection System on Mobile Devices,' https://arxiv.org/abs/1804.06882. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck_sizes : list of int Bottleneck sizes for each stage. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. dropout_rate : float, default 0.5 Parameter of Dropout layer. Faction of the input units to drop. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck_sizes, bn_use_global_stats=False, dropout_rate=0.5, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(PeleeNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(StemBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): bottleneck_size = bottleneck_sizes[i] stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): if i != 0: stage.add(TransitionBlock( in_channels=in_channels, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) for j, out_channels in enumerate(channels_per_stage): stage.add(DenseBlock( in_channels=in_channels, out_channels=out_channels, bottleneck_size=bottleneck_size, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(conv1x1_block( in_channels=in_channels, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dropout(rate=dropout_rate)) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_peleenet(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create PeleeNet model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ init_block_channels = 32 growth_rate = 32 layers = [3, 4, 8, 6] bottleneck_sizes = [1, 2, 4, 4] from functools import reduce channels = reduce( lambda xi, yi: xi + [reduce( lambda xj, yj: xj + [xj[-1] + yj], [growth_rate] * yi, [xi[-1][-1]])[1:]], layers, [[init_block_channels]])[1:] net = PeleeNet( channels=channels, init_block_channels=init_block_channels, bottleneck_sizes=bottleneck_sizes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def peleenet(**kwargs): """ PeleeNet model from 'Pelee: A Real-Time Object Detection System on Mobile Devices,' https://arxiv.org/abs/1804.06882. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_peleenet(model_name="peleenet", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ peleenet, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != peleenet or weight_count == 2802248) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
13,545
31.252381
117
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/regnetv.py
""" RegNetV for ImageNet-1K, implemented in Gluon. Original paper: 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. """ __all__ = ['RegNetV', 'regnetv002', 'regnetv004', 'regnetv006', 'regnetv008', 'regnetv016', 'regnetv032', 'regnetv040', 'regnetv064', 'regnetv080', 'regnetv120', 'regnetv160', 'regnetv320'] import os import numpy as np from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, dwsconv3x3_block class DownBlock(HybridBlock): """ ResNet(A)-like downsample block for the identity branch of a residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(DownBlock, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=strides, strides=strides, ceil_mode=True, count_include_pad=False) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) def hybrid_forward(self, F, x): x = self.pool(x) x = self.conv(x) return x class RegNetVUnit(HybridBlock): """ RegNetV unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. downscale : bool Whether to downscale tensor. dw_use_bn : bool Whether to use BatchNorm layer (depthwise convolution block). dw_activation : function or str or None Activation function after the depthwise convolution block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, downscale, dw_use_bn, dw_activation, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(RegNetVUnit, self).__init__(**kwargs) self.downscale = downscale with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) if self.downscale: self.pool = nn.AvgPool2D( pool_size=3, strides=2, padding=1) self.conv2 = dwsconv3x3_block( in_channels=out_channels, out_channels=out_channels, dw_use_bn=dw_use_bn, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dw_activation=dw_activation, pw_activation=None) if self.downscale: self.identity_block = DownBlock( in_channels=in_channels, out_channels=out_channels, strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.downscale: identity = self.identity_block(x) else: identity = x x = self.conv1(x) if self.downscale: x = self.pool(x) x = self.conv2(x) x = x + identity x = self.activ(x) return x class RegNetVInitBlock(HybridBlock): """ RegNetV specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(RegNetVInitBlock, self).__init__(**kwargs) mid_channels = out_channels // 2 with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.pool(x) x = self.conv2(x) return x class RegNetV(HybridBlock): """ RegNet model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. dw_use_bn : bool, default True Whether to use BatchNorm layer (depthwise convolution block). dw_activation : function or str or None, default nn.Activation('relu') Activation function after the depthwise convolution block. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, dw_use_bn=True, dw_activation=(lambda: nn.Activation("relu")), bn_use_global_stats=False, bn_cudnn_off=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(RegNetV, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(RegNetVInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): downscale = (j == 0) stage.add(RegNetVUnit( in_channels=in_channels, out_channels=out_channels, downscale=downscale, dw_use_bn=dw_use_bn, dw_activation=dw_activation, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_regnet(channels_init, channels_slope, channels_mult, depth, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create RegNet model with specific parameters. Parameters: ---------- channels_init : float Initial value for channels/widths. channels_slope : float Slope value for channels/widths. width_mult : float Width multiplier value. depth : int Depth value. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ divisor = 8 assert (channels_slope >= 0) and (channels_init > 0) and (channels_mult > 1) and (channels_init % divisor == 0) # Generate continuous per-block channels/widths: channels_cont = np.arange(depth) * channels_slope + channels_init # Generate quantized per-block channels/widths: channels_exps = np.round(np.log(channels_cont / channels_init) / np.log(channels_mult)) channels = channels_init * np.power(channels_mult, channels_exps) channels = (np.round(channels / divisor) * divisor).astype(np.int) # Generate per stage channels/widths and layers/depths: channels_per_stage, layers = np.unique(channels, return_counts=True) channels = [[ci] * li for (ci, li) in zip(channels_per_stage, layers)] init_block_channels = 32 dws_simplified = True if dws_simplified: dw_use_bn = False dw_activation = None else: dw_use_bn = True dw_activation = (lambda: nn.Activation("relu")) net = RegNetV( channels=channels, init_block_channels=init_block_channels, dw_use_bn=dw_use_bn, dw_activation=dw_activation, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def regnetv002(**kwargs): """ RegNetV-200MF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=24, channels_slope=36.44, channels_mult=2.49, depth=13, model_name="regnetv002", **kwargs) def regnetv004(**kwargs): """ RegNetV-400MF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=24, channels_slope=24.48, channels_mult=2.54, depth=22, model_name="regnetv004", **kwargs) def regnetv006(**kwargs): """ RegNetV-600MF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=48, channels_slope=36.97, channels_mult=2.24, depth=16, model_name="regnetv006", **kwargs) def regnetv008(**kwargs): """ RegNetV-800MF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=56, channels_slope=35.73, channels_mult=2.28, depth=16, model_name="regnetv008", **kwargs) def regnetv016(**kwargs): """ RegNetV-1.6GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=80, channels_slope=34.01, channels_mult=2.25, depth=18, model_name="regnetv016", **kwargs) def regnetv032(**kwargs): """ RegNetV-3.2GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=88, channels_slope=26.31, channels_mult=2.25, depth=25, model_name="regnetv032", **kwargs) def regnetv040(**kwargs): """ RegNetV-4.0GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=96, channels_slope=38.65, channels_mult=2.43, depth=23, model_name="regnetv040", **kwargs) def regnetv064(**kwargs): """ RegNetV-6.4GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=184, channels_slope=60.83, channels_mult=2.07, depth=17, model_name="regnetv064", **kwargs) def regnetv080(**kwargs): """ RegNetV-8.0GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=80, channels_slope=49.56, channels_mult=2.88, depth=23, model_name="regnetv080", **kwargs) def regnetv120(**kwargs): """ RegNetV-12GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=168, channels_slope=73.36, channels_mult=2.37, depth=19, model_name="regnetv120", **kwargs) def regnetv160(**kwargs): """ RegNetV-16GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=216, channels_slope=55.59, channels_mult=2.1, depth=22, model_name="regnetv160", **kwargs) def regnetv320(**kwargs): """ RegNetV-32GF model from 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_regnet(channels_init=320, channels_slope=69.86, channels_mult=2.0, depth=23, model_name="regnetv320", **kwargs) def _test(): import numpy as np import mxnet as mx dws_simplified = True pretrained = False models = [ regnetv002, regnetv004, regnetv006, regnetv008, regnetv016, regnetv032, regnetv040, regnetv064, regnetv080, regnetv120, regnetv160, regnetv320, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) if dws_simplified: assert (model != regnetv002 or weight_count == 2476840) assert (model != regnetv004 or weight_count == 4467080) assert (model != regnetv006 or weight_count == 5242936) assert (model != regnetv008 or weight_count == 6353000) assert (model != regnetv016 or weight_count == 7824440) assert (model != regnetv032 or weight_count == 11540536) assert (model != regnetv040 or weight_count == 18323824) assert (model != regnetv064 or weight_count == 20854680) assert (model != regnetv080 or weight_count == 21930224) assert (model != regnetv120 or weight_count == 32833720) assert (model != regnetv160 or weight_count == 36213360) assert (model != regnetv320 or weight_count == 64659576) else: assert (model != regnetv002 or weight_count == 2479160) assert (model != regnetv004 or weight_count == 4474712) assert (model != regnetv006 or weight_count == 5249352) assert (model != regnetv008 or weight_count == 6360344) assert (model != regnetv016 or weight_count == 7833768) assert (model != regnetv032 or weight_count == 11556520) assert (model != regnetv040 or weight_count == 18343728) assert (model != regnetv064 or weight_count == 20873384) assert (model != regnetv080 or weight_count == 21952400) assert (model != regnetv120 or weight_count == 32859432) assert (model != regnetv160 or weight_count == 36244240) assert (model != regnetv320 or weight_count == 64704008) batch = 14 size = 224 x = mx.nd.zeros((batch, 3, size, size), ctx=ctx) y = net(x) assert (y.shape == (batch, 1000)) if __name__ == "__main__": _test()
21,904
34.444984
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/sharesnet.py
""" ShaResNet for ImageNet-1K, implemented in Gluon. Original paper: 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. """ __all__ = ['ShaResNet', 'sharesnet18', 'sharesnet34', 'sharesnet50', 'sharesnet50b', 'sharesnet101', 'sharesnet101b', 'sharesnet152', 'sharesnet152b'] import os from inspect import isfunction from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import ReLU6, conv1x1_block, conv3x3_block from .resnet import ResInitBlock class ShaConvBlock(HybridBlock): """ Shared convolution block with Batch normalization and ReLU/ReLU6 activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activation : function or str or None, default nn.Activation("relu") Activation function or name of activation function. activate : bool, default True Whether activate the convolution block. shared_conv : HybridBlock, default None Shared convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, bn_use_global_stats=False, activation=(lambda: nn.Activation("relu")), activate=True, shared_conv=None, **kwargs): super(ShaConvBlock, self).__init__(**kwargs) self.activate = activate with self.name_scope(): if shared_conv is None: self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, in_channels=in_channels) else: self.conv = shared_conv self.bn = nn.BatchNorm( in_channels=out_channels, use_global_stats=bn_use_global_stats) if self.activate: assert (activation is not None) if isfunction(activation): self.activ = activation() elif isinstance(activation, str): if activation == "relu6": self.activ = ReLU6() else: self.activ = nn.Activation(activation) else: self.activ = activation def hybrid_forward(self, F, x): x = self.conv(x) x = self.bn(x) if self.activate: x = self.activ(x) return x def sha_conv3x3_block(in_channels, out_channels, strides=1, padding=1, dilation=1, groups=1, use_bias=False, bn_use_global_stats=False, activation=(lambda: nn.Activation("relu")), activate=True, shared_conv=None, **kwargs): """ 3x3 version of the shared convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activation : function or str or None, default nn.Activation("relu") Activation function or name of activation function. activate : bool, default True Whether activate the convolution block. shared_conv : HybridBlock, default None Shared convolution layer. """ return ShaConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, bn_use_global_stats=bn_use_global_stats, activation=activation, activate=activate, shared_conv=shared_conv, **kwargs) class ShaResBlock(HybridBlock): """ Simple ShaResNet block for residual path in ShaResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. shared_conv : HybridBlock, default None Shared convolution layer. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, shared_conv=None, **kwargs): super(ShaResBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) self.conv2 = sha_conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None, activate=False, shared_conv=shared_conv) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class ShaResBottleneck(HybridBlock): """ ShaResNet bottleneck block for residual path in ShaResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck_factor : int, default 4 Bottleneck factor. conv1_stride : bool, default False Whether to use stride in the first or the second convolution layer of the block. shared_conv : HybridBlock, default None Shared convolution layer. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, conv1_stride=False, bottleneck_factor=4, shared_conv=None, **kwargs): super(ShaResBottleneck, self).__init__(**kwargs) assert (conv1_stride or not ((strides > 1) and (shared_conv is not None))) mid_channels = out_channels // bottleneck_factor with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, strides=(strides if conv1_stride else 1), bn_use_global_stats=bn_use_global_stats) self.conv2 = sha_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=(1 if conv1_stride else strides), bn_use_global_stats=bn_use_global_stats, shared_conv=shared_conv) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class ShaResUnit(HybridBlock): """ ShaResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer of the block. shared_conv : HybridBlock, default None Shared convolution layer. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats, bottleneck, conv1_stride, shared_conv=None, **kwargs): super(ShaResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if bottleneck: self.body = ShaResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, conv1_stride=conv1_stride, shared_conv=shared_conv) else: self.body = ShaResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, shared_conv=shared_conv) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class ShaResNet(HybridBlock): """ ShaResNet model from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(ShaResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) shared_conv = None with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 unit = ShaResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=conv1_stride, shared_conv=shared_conv) if (shared_conv is None) and not (bottleneck and not conv1_stride and strides > 1): shared_conv = unit.body.conv2.conv stage.add(unit) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_sharesnet(blocks, conv1_stride=True, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create ShaResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 18: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] else: raise ValueError("Unsupported ShaResNet with number of blocks: {}".format(blocks)) init_block_channels = 64 if blocks < 50: channels_per_layers = [64, 128, 256, 512] bottleneck = False else: channels_per_layers = [256, 512, 1024, 2048] bottleneck = True channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = ShaResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def sharesnet18(**kwargs): """ ShaResNet-18 model from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=18, model_name="sharesnet18", **kwargs) def sharesnet34(**kwargs): """ ShaResNet-34 model from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=34, model_name="sharesnet34", **kwargs) def sharesnet50(**kwargs): """ ShaResNet-50 model from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=50, model_name="sharesnet50", **kwargs) def sharesnet50b(**kwargs): """ ShaResNet-50b model with stride at the second convolution in bottleneck block from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=50, conv1_stride=False, model_name="sharesnet50b", **kwargs) def sharesnet101(**kwargs): """ ShaResNet-101 model from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=101, model_name="sharesnet101", **kwargs) def sharesnet101b(**kwargs): """ ShaResNet-101b model with stride at the second convolution in bottleneck block from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=101, conv1_stride=False, model_name="sharesnet101b", **kwargs) def sharesnet152(**kwargs): """ ShaResNet-152 model from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=152, model_name="sharesnet152", **kwargs) def sharesnet152b(**kwargs): """ ShaResNet-152b model with stride at the second convolution in bottleneck block from 'ShaResNet: reducing residual network parameter number by sharing weights,' https://arxiv.org/abs/1702.08782. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sharesnet(blocks=152, conv1_stride=False, model_name="sharesnet152b", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ sharesnet18, sharesnet34, sharesnet50, sharesnet50b, sharesnet101, sharesnet101b, sharesnet152, sharesnet152b, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sharesnet18 or weight_count == 8556072) assert (model != sharesnet34 or weight_count == 13613864) assert (model != sharesnet50 or weight_count == 17373224) assert (model != sharesnet50b or weight_count == 20469800) assert (model != sharesnet101 or weight_count == 26338344) assert (model != sharesnet101b or weight_count == 29434920) assert (model != sharesnet152 or weight_count == 33724456) assert (model != sharesnet152b or weight_count == 36821032) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
23,240
33.73991
117
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/ibppose_coco.py
""" IBPPose for COCO Keypoint, implemented in Gluon. Original paper: 'Simple Pose: Rethinking and Improving a Bottom-up Approach for Multi-Person Pose Estimation,' https://arxiv.org/abs/1911.10529. """ __all__ = ['IbpPose', 'ibppose_coco'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import get_activation_layer, conv1x1_block, conv3x3_block, conv7x7_block, SEBlock, Hourglass,\ InterpolationBlock class IbpResBottleneck(HybridBlock): """ Bottleneck block for residual path in the residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. use_bias : bool, default False Whether the layer uses a bias vector. bottleneck_factor : int, default 2 Bottleneck factor. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, strides, use_bias=False, bottleneck_factor=2, activation=(lambda: nn.Activation("relu")), **kwargs): super(IbpResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // bottleneck_factor with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_bias=use_bias, activation=activation) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, use_bias=use_bias, activation=activation) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, use_bias=use_bias, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class IbpResUnit(HybridBlock): """ ResNet-like residual unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. use_bias : bool, default False Whether the layer uses a bias vector. bottleneck_factor : int, default 2 Bottleneck factor. activation : function or str or None, default nn.Activation('relu') Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, strides=1, use_bias=False, bottleneck_factor=2, activation=(lambda: nn.Activation("relu")), **kwargs): super(IbpResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = IbpResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, use_bias=use_bias, bottleneck_factor=bottleneck_factor, activation=activation) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, use_bias=use_bias, activation=None) self.activ = get_activation_layer(activation) def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class IbpBackbone(HybridBlock): """ IBPPose backbone. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. activation : function or str or None Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, activation, **kwargs): super(IbpBackbone, self).__init__(**kwargs) dilations = (3, 3, 4, 4, 5, 5) mid1_channels = out_channels // 4 mid2_channels = out_channels // 2 with self.name_scope(): self.conv1 = conv7x7_block( in_channels=in_channels, out_channels=mid1_channels, strides=2, activation=activation) self.res1 = IbpResUnit( in_channels=mid1_channels, out_channels=mid2_channels, activation=activation) self.pool = nn.MaxPool2D( pool_size=2, strides=2) self.res2 = IbpResUnit( in_channels=mid2_channels, out_channels=mid2_channels, activation=activation) self.dilation_branch = nn.HybridSequential(prefix="") for dilation in dilations: self.dilation_branch.add(conv3x3_block( in_channels=mid2_channels, out_channels=mid2_channels, padding=dilation, dilation=dilation, activation=activation)) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.res1(x) x = self.pool(x) x = self.res2(x) y = self.dilation_branch(x) x = F.concat(x, y, dim=1) return x class IbpDownBlock(HybridBlock): """ IBPPose down block for the hourglass. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. activation : function or str or None Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, activation, **kwargs): super(IbpDownBlock, self).__init__(**kwargs) with self.name_scope(): self.down = nn.MaxPool2D( pool_size=2, strides=2) self.res = IbpResUnit( in_channels=in_channels, out_channels=out_channels, activation=activation) def hybrid_forward(self, F, x): x = self.down(x) x = self.res(x) return x class IbpUpBlock(HybridBlock): """ IBPPose up block for the hourglass. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. use_bn : bool Whether to use BatchNorm layer. activation : function or str or None Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, use_bn, activation, **kwargs): super(IbpUpBlock, self).__init__(**kwargs) with self.name_scope(): self.res = IbpResUnit( in_channels=in_channels, out_channels=out_channels, activation=activation) self.up = InterpolationBlock( scale_factor=2, bilinear=False) self.conv = conv3x3_block( in_channels=out_channels, out_channels=out_channels, use_bias=(not use_bn), use_bn=use_bn, activation=activation) def hybrid_forward(self, F, x): x = self.res(x) x = self.up(x) x = self.conv(x) return x class MergeBlock(HybridBlock): """ IBPPose merge block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. use_bn : bool Whether to use BatchNorm layer. """ def __init__(self, in_channels, out_channels, use_bn, **kwargs): super(MergeBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, use_bias=(not use_bn), use_bn=use_bn, activation=None) def hybrid_forward(self, F, x): return self.conv(x) class IbpPreBlock(HybridBlock): """ IBPPose preliminary decoder block. Parameters: ---------- out_channels : int Number of output channels. use_bn : bool Whether to use BatchNorm layer. activation : function or str or None Activation function or name of activation function. """ def __init__(self, out_channels, use_bn, activation, **kwargs): super(IbpPreBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, use_bias=(not use_bn), use_bn=use_bn, activation=activation) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, use_bias=(not use_bn), use_bn=use_bn, activation=activation) self.se = SEBlock( channels=out_channels, use_conv=False, mid_activation=activation) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.se(x) return x class IbpPass(HybridBlock): """ IBPPose single pass decoder block. Parameters: ---------- channels : int Number of input/output channels. mid_channels : int Number of middle channels. depth : int Depth of hourglass. growth_rate : int Addition for number of channel for each level. use_bn : bool Whether to use BatchNorm layer. activation : function or str or None Activation function or name of activation function. """ def __init__(self, channels, mid_channels, depth, growth_rate, merge, use_bn, activation, **kwargs): super(IbpPass, self).__init__(**kwargs) self.merge = merge with self.name_scope(): down_seq = nn.HybridSequential(prefix="") up_seq = nn.HybridSequential(prefix="") skip_seq = nn.HybridSequential(prefix="") top_channels = channels bottom_channels = channels for i in range(depth + 1): skip_seq.add(IbpResUnit( in_channels=top_channels, out_channels=top_channels, activation=activation)) bottom_channels += growth_rate if i < depth: down_seq.add(IbpDownBlock( in_channels=top_channels, out_channels=bottom_channels, activation=activation)) up_seq.add(IbpUpBlock( in_channels=bottom_channels, out_channels=top_channels, use_bn=use_bn, activation=activation)) top_channels = bottom_channels self.hg = Hourglass( down_seq=down_seq, up_seq=up_seq, skip_seq=skip_seq) self.pre_block = IbpPreBlock( out_channels=channels, use_bn=use_bn, activation=activation) self.post_block = conv1x1_block( in_channels=channels, out_channels=mid_channels, use_bias=True, use_bn=False, activation=None) if self.merge: self.pre_merge_block = MergeBlock( in_channels=channels, out_channels=channels, use_bn=use_bn) self.post_merge_block = MergeBlock( in_channels=mid_channels, out_channels=channels, use_bn=use_bn) def hybrid_forward(self, F, x, x_prev): x = self.hg(x) if x_prev is not None: x = x + x_prev y = self.pre_block(x) z = self.post_block(y) if self.merge: z = self.post_merge_block(z) + self.pre_merge_block(y) return z class IbpPose(HybridBlock): """ IBPPose model from 'Simple Pose: Rethinking and Improving a Bottom-up Approach for Multi-Person Pose Estimation,' https://arxiv.org/abs/1911.10529. Parameters: ---------- passes : int Number of passes. backbone_out_channels : int Number of output channels for the backbone. outs_channels : int Number of output channels for the backbone. depth : int Depth of hourglass. growth_rate : int Addition for number of channel for each level. use_bn : bool Whether to use BatchNorm layer. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (256, 256) Spatial size of the expected input image. """ def __init__(self, passes, backbone_out_channels, outs_channels, depth, growth_rate, use_bn, in_channels=3, in_size=(256, 256), **kwargs): super(IbpPose, self).__init__(**kwargs) self.in_size = in_size activation = (lambda: nn.LeakyReLU(alpha=0.01)) with self.name_scope(): self.backbone = IbpBackbone( in_channels=in_channels, out_channels=backbone_out_channels, activation=activation) self.decoder = nn.HybridSequential(prefix="") for i in range(passes): merge = (i != passes - 1) self.decoder.add(IbpPass( channels=backbone_out_channels, mid_channels=outs_channels, depth=depth, growth_rate=growth_rate, merge=merge, use_bn=use_bn, activation=activation)) def hybrid_forward(self, F, x): x = self.backbone(x) x_prev = None for block in self.decoder._children.values(): if x_prev is not None: x = x + x_prev x_prev = block(x, x_prev) return x_prev def get_ibppose(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create IBPPose model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ passes = 4 backbone_out_channels = 256 outs_channels = 50 depth = 4 growth_rate = 128 use_bn = True net = IbpPose( passes=passes, backbone_out_channels=backbone_out_channels, outs_channels=outs_channels, depth=depth, growth_rate=growth_rate, use_bn=use_bn, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def ibppose_coco(**kwargs): """ IBPPose model for COCO Keypoint from 'Simple Pose: Rethinking and Improving a Bottom-up Approach for Multi-Person Pose Estimation,' https://arxiv.org/abs/1911.10529. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ibppose(model_name="ibppose_coco", **kwargs) def _test(): import numpy as np import mxnet as mx in_size = (256, 256) pretrained = False models = [ ibppose_coco, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ibppose_coco or weight_count == 95827784) batch = 14 x = mx.nd.random.normal(shape=(batch, 3, in_size[0], in_size[1]), ctx=ctx) y = net(x) assert (y.shape == (batch, 50, in_size[0] // 4, in_size[0] // 4)) if __name__ == "__main__": _test()
18,529
29.883333
117
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/xception.py
""" Xception for ImageNet-1K, implemented in Gluon. Original paper: 'Xception: Deep Learning with Depthwise Separable Convolutions,' https://arxiv.org/abs/1610.02357. """ __all__ = ['Xception', 'xception'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block class DwsConv(HybridBlock): """ Depthwise separable convolution layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 0 Padding value for convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, strides=1, padding=0, **kwargs): super(DwsConv, self).__init__(**kwargs) with self.name_scope(): self.dw_conv = nn.Conv2D( channels=in_channels, kernel_size=kernel_size, strides=strides, padding=padding, groups=in_channels, use_bias=False, in_channels=in_channels) self.pw_conv = nn.Conv2D( channels=out_channels, kernel_size=1, use_bias=False, in_channels=in_channels) def hybrid_forward(self, F, x): x = self.dw_conv(x) x = self.pw_conv(x) return x class DwsConvBlock(HybridBlock): """ Depthwise separable convolution block with batchnorm and ReLU pre-activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activate : bool Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, bn_use_global_stats, activate, **kwargs): super(DwsConvBlock, self).__init__(**kwargs) self.activate = activate with self.name_scope(): if self.activate: self.activ = nn.Activation("relu") self.conv = DwsConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding) self.bn = nn.BatchNorm( in_channels=out_channels, use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): if self.activate: x = self.activ(x) x = self.conv(x) x = self.bn(x) return x def dws_conv3x3_block(in_channels, out_channels, bn_use_global_stats, activate): """ 3x3 version of the depthwise separable convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activate : bool Whether activate the convolution block. """ return DwsConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=1, padding=1, bn_use_global_stats=bn_use_global_stats, activate=activate) class XceptionUnit(HybridBlock): """ Xception unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the downsample polling. reps : int Number of repetitions. start_with_relu : bool, default True Whether start with ReLU activation. grow_first : bool, default True Whether start from growing. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, reps, start_with_relu=True, grow_first=True, bn_use_global_stats=False, **kwargs): super(XceptionUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.body = nn.HybridSequential(prefix="") for i in range(reps): if (grow_first and (i == 0)) or ((not grow_first) and (i == reps - 1)): in_channels_i = in_channels out_channels_i = out_channels else: if grow_first: in_channels_i = out_channels out_channels_i = out_channels else: in_channels_i = in_channels out_channels_i = in_channels activate = start_with_relu if (i == 0) else True self.body.add(dws_conv3x3_block( in_channels=in_channels_i, out_channels=out_channels_i, bn_use_global_stats=bn_use_global_stats, activate=activate)) if strides != 1: self.body.add(nn.MaxPool2D( pool_size=3, strides=strides, padding=1)) def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = F.identity(x) x = self.body(x) x = x + identity return x class XceptionInitBlock(HybridBlock): """ Xception specific initial block. Parameters: ---------- in_channels : int Number of input channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, bn_use_global_stats, **kwargs): super(XceptionInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=32, strides=2, padding=0, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=32, out_channels=64, strides=1, padding=0, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class XceptionFinalBlock(HybridBlock): """ Xception specific final block. Parameters: ---------- bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, bn_use_global_stats, **kwargs): super(XceptionFinalBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = dws_conv3x3_block( in_channels=1024, out_channels=1536, bn_use_global_stats=bn_use_global_stats, activate=False) self.conv2 = dws_conv3x3_block( in_channels=1536, out_channels=2048, bn_use_global_stats=bn_use_global_stats, activate=True) self.activ = nn.Activation("relu") self.pool = nn.AvgPool2D( pool_size=10, strides=1) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.activ(x) x = self.pool(x) return x class Xception(HybridBlock): """ Xception model from 'Xception: Deep Learning with Depthwise Separable Convolutions,' https://arxiv.org/abs/1610.02357. Parameters: ---------- channels : list of list of int Number of output channels for each unit. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, bn_use_global_stats=False, in_channels=3, in_size=(299, 299), classes=1000, **kwargs): super(Xception, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(XceptionInitBlock( in_channels=in_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = 64 for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): stage.add(XceptionUnit( in_channels=in_channels, out_channels=out_channels, strides=(2 if (j == 0) else 1), reps=(2 if (j == 0) else 3), start_with_relu=((i != 0) or (j != 0)), grow_first=((i != len(channels) - 1) or (j != len(channels_per_stage) - 1)), bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(XceptionFinalBlock(bn_use_global_stats=bn_use_global_stats)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=2048)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_xception(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Xception model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ channels = [[128], [256], [728] * 9, [1024]] net = Xception( channels=channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def xception(**kwargs): """ Xception model from 'Xception: Deep Learning with Depthwise Separable Convolutions,' https://arxiv.org/abs/1610.02357. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_xception(model_name="xception", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ xception, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != xception or weight_count == 22855952) x = mx.nd.zeros((1, 3, 299, 299), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
14,182
30.87191
118
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/darknet53.py
""" DarkNet-53 for ImageNet-1K, implemented in Gluon. Original source: 'YOLOv3: An Incremental Improvement,' https://arxiv.org/abs/1804.02767. """ __all__ = ['DarkNet53', 'darknet53'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block class DarkUnit(HybridBlock): """ DarkNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. alpha : float Slope coefficient for Leaky ReLU activation. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, alpha, **kwargs): super(DarkUnit, self).__init__(**kwargs) assert (out_channels % 2 == 0) mid_channels = out_channels // 2 with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, activation=nn.LeakyReLU(alpha=alpha)) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=nn.LeakyReLU(alpha=alpha)) def hybrid_forward(self, F, x): identity = x x = self.conv1(x) x = self.conv2(x) return x + identity class DarkNet53(HybridBlock): """ DarkNet-53 model from 'YOLOv3: An Incremental Improvement,' https://arxiv.org/abs/1804.02767. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. alpha : float, default 0.1 Slope coefficient for Leaky ReLU activation. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, alpha=0.1, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(DarkNet53, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats, activation=nn.LeakyReLU(alpha=alpha))) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): if j == 0: stage.add(conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=2, bn_use_global_stats=bn_use_global_stats, activation=nn.LeakyReLU(alpha=alpha))) else: stage.add(DarkUnit( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, alpha=alpha)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_darknet53(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DarkNet model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ init_block_channels = 32 layers = [2, 3, 9, 9, 5] channels_per_layers = [64, 128, 256, 512, 1024] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = DarkNet53( channels=channels, init_block_channels=init_block_channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def darknet53(**kwargs): """ DarkNet-53 'Reference' model from 'YOLOv3: An Incremental Improvement,' https://arxiv.org/abs/1804.02767. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_darknet53(model_name="darknet53", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ darknet53, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != darknet53 or weight_count == 41609928) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
7,537
31.917031
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/mobilenet.py
""" MobileNet for ImageNet-1K, implemented in Gluon. Original paper: 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. """ __all__ = ['MobileNet', 'mobilenet_w1', 'mobilenet_w3d4', 'mobilenet_wd2', 'mobilenet_wd4', 'get_mobilenet'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv3x3_block, dwsconv3x3_block class MobileNet(HybridBlock): """ MobileNet model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- channels : list of list of int Number of output channels for each unit. first_stage_stride : bool Whether stride is used at the first stage. dw_use_bn : bool, default True Whether to use BatchNorm layer (depthwise convolution block). bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. dw_activation : function or str or None, default nn.Activation('relu') Activation function after the depthwise convolution block. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, first_stage_stride, dw_use_bn=True, bn_use_global_stats=False, bn_cudnn_off=False, dw_activation=(lambda: nn.Activation("relu")), in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(MobileNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") init_block_channels = channels[0][0] self.features.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels[1:]): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and ((i != 0) or first_stage_stride) else 1 stage.add(dwsconv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=strides, dw_use_bn=dw_use_bn, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, dw_activation=dw_activation)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_mobilenet(width_scale, dws_simplified=False, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create MobileNet model with specific parameters. Parameters: ---------- width_scale : float Scale factor for width of layers. dws_simplified : bool, default False Whether to use simplified depthwise separable convolution block. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ channels = [[32], [64], [128, 128], [256, 256], [512, 512, 512, 512, 512, 512], [1024, 1024]] first_stage_stride = False if width_scale != 1.0: channels = [[int(cij * width_scale) for cij in ci] for ci in channels] if dws_simplified: dw_use_bn = False dw_activation = None else: dw_use_bn = True dw_activation = (lambda: nn.Activation("relu")) net = MobileNet( channels=channels, first_stage_stride=first_stage_stride, dw_use_bn=dw_use_bn, dw_activation=dw_activation, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def mobilenet_w1(**kwargs): """ 1.0 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=1.0, model_name="mobilenet_w1", **kwargs) def mobilenet_w3d4(**kwargs): """ 0.75 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=0.75, model_name="mobilenet_w3d4", **kwargs) def mobilenet_wd2(**kwargs): """ 0.5 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=0.5, model_name="mobilenet_wd2", **kwargs) def mobilenet_wd4(**kwargs): """ 0.25 MobileNet-224 model from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=0.25, model_name="mobilenet_wd4", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ mobilenet_w1, mobilenet_w3d4, mobilenet_wd2, mobilenet_wd4, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mobilenet_w1 or weight_count == 4231976) assert (model != mobilenet_w3d4 or weight_count == 2585560) assert (model != mobilenet_wd2 or weight_count == 1331592) assert (model != mobilenet_wd4 or weight_count == 470072) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
9,265
34.098485
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/dpn.py
""" DPN for ImageNet-1K, implemented in Gluon. Original paper: 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. """ __all__ = ['DPN', 'dpn68', 'dpn68b', 'dpn98', 'dpn107', 'dpn131'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, DualPathSequential class GlobalAvgMaxPool2D(HybridBlock): """ Global average+max pooling operation for spatial data. """ def __init__(self, **kwargs): super(GlobalAvgMaxPool2D, self).__init__(**kwargs) with self.name_scope(): self.avg_pool = nn.GlobalAvgPool2D() self.max_pool = nn.GlobalMaxPool2D() def hybrid_forward(self, F, x): x_avg = self.avg_pool(x) x_max = self.max_pool(x) x = 0.5 * (x_avg + x_max) return x def dpn_batch_norm(channels): """ DPN specific Batch normalization layer. Parameters: ---------- channels : int Number of channels in input data. """ return nn.BatchNorm( epsilon=0.001, in_channels=channels) class PreActivation(HybridBlock): """ DPN specific block, which performs the preactivation like in RreResNet. Parameters: ---------- channels : int Number of channels. """ def __init__(self, channels, **kwargs): super(PreActivation, self).__init__(**kwargs) with self.name_scope(): self.bn = dpn_batch_norm(channels=channels) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) return x class DPNConv(HybridBlock): """ DPN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. groups : int Number of groups. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, groups, **kwargs): super(DPNConv, self).__init__(**kwargs) with self.name_scope(): self.bn = dpn_batch_norm(channels=in_channels) self.activ = nn.Activation("relu") self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, groups=groups, use_bias=False, in_channels=in_channels) def hybrid_forward(self, F, x): x = self.bn(x) x = self.activ(x) x = self.conv(x) return x def dpn_conv1x1(in_channels, out_channels, strides=1): """ 1x1 version of the DPN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. """ return DPNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, groups=1) def dpn_conv3x3(in_channels, out_channels, strides, groups): """ 3x3 version of the DPN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. groups : int Number of groups. """ return DPNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=1, groups=groups) class DPNUnit(HybridBlock): """ DPN unit. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of intermediate channels. bw : int Number of residual channels. inc : int Incrementing step for channels. groups : int Number of groups in the units. has_proj : bool Whether to use projection. key_strides : int Key strides of the convolutions. b_case : bool, default False Whether to use B-case model. """ def __init__(self, in_channels, mid_channels, bw, inc, groups, has_proj, key_strides, b_case=False, **kwargs): super(DPNUnit, self).__init__(**kwargs) self.bw = bw self.has_proj = has_proj self.b_case = b_case with self.name_scope(): if self.has_proj: self.conv_proj = dpn_conv1x1( in_channels=in_channels, out_channels=bw + 2 * inc, strides=key_strides) self.conv1 = dpn_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.conv2 = dpn_conv3x3( in_channels=mid_channels, out_channels=mid_channels, strides=key_strides, groups=groups) if b_case: self.preactiv = PreActivation(channels=mid_channels) self.conv3a = conv1x1( in_channels=mid_channels, out_channels=bw) self.conv3b = conv1x1( in_channels=mid_channels, out_channels=inc) else: self.conv3 = dpn_conv1x1( in_channels=mid_channels, out_channels=bw + inc) def hybrid_forward(self, F, x1, x2=None): x_in = F.concat(x1, x2, dim=1) if x2 is not None else x1 if self.has_proj: x_s = self.conv_proj(x_in) x_s1 = F.slice_axis(x_s, axis=1, begin=0, end=self.bw) x_s2 = F.slice_axis(x_s, axis=1, begin=self.bw, end=None) else: assert (x2 is not None) x_s1 = x1 x_s2 = x2 x_in = self.conv1(x_in) x_in = self.conv2(x_in) if self.b_case: x_in = self.preactiv(x_in) y1 = self.conv3a(x_in) y2 = self.conv3b(x_in) else: x_in = self.conv3(x_in) y1 = F.slice_axis(x_in, axis=1, begin=0, end=self.bw) y2 = F.slice_axis(x_in, axis=1, begin=self.bw, end=None) residual = x_s1 + y1 dense = F.concat(x_s2, y2, dim=1) return residual, dense class DPNInitBlock(HybridBlock): """ DPN specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. padding : int or tuple/list of 2 int Padding value for convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, padding, **kwargs): super(DPNInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=2, padding=padding, use_bias=False, in_channels=in_channels) self.bn = dpn_batch_norm(channels=out_channels) self.activ = nn.Activation("relu") self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) x = self.pool(x) return x class DPNFinalBlock(HybridBlock): """ DPN final block, which performs the preactivation with cutting. Parameters: ---------- channels : int Number of channels. """ def __init__(self, channels, **kwargs): super(DPNFinalBlock, self).__init__(**kwargs) with self.name_scope(): self.activ = PreActivation(channels=channels) def hybrid_forward(self, F, x1, x2): assert (x2 is not None) x = F.concat(x1, x2, dim=1) x = self.activ(x) return x, None class DPN(HybridBlock): """ DPN model from 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. init_block_kernel_size : int or tuple/list of 2 int Convolution window size for the initial unit. init_block_padding : int or tuple/list of 2 int Padding value for convolution layer in the initial unit. rs : list f int Number of intermediate channels for each unit. bws : list f int Number of residual channels for each unit. incs : list f int Incrementing step for channels for each unit. groups : int Number of groups in the units. b_case : bool Whether to use B-case model. for_training : bool Whether to use model for training. test_time_pool : bool Whether to use the avg-max pooling in the inference mode. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, init_block_kernel_size, init_block_padding, rs, bws, incs, groups, b_case, for_training, test_time_pool, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(DPN, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = DualPathSequential( return_two=False, first_ordinals=1, last_ordinals=0, prefix="") self.features.add(DPNInitBlock( in_channels=in_channels, out_channels=init_block_channels, kernel_size=init_block_kernel_size, padding=init_block_padding)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential(prefix="stage{}_".format(i + 1)) r = rs[i] bw = bws[i] inc = incs[i] with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): has_proj = (j == 0) key_strides = 2 if (j == 0) and (i != 0) else 1 stage.add(DPNUnit( in_channels=in_channels, mid_channels=r, bw=bw, inc=inc, groups=groups, has_proj=has_proj, key_strides=key_strides, b_case=b_case)) in_channels = out_channels self.features.add(stage) self.features.add(DPNFinalBlock(channels=in_channels)) self.output = nn.HybridSequential(prefix="") if for_training or not test_time_pool: self.output.add(nn.GlobalAvgPool2D()) self.output.add(conv1x1( in_channels=in_channels, out_channels=classes, use_bias=True)) self.output.add(nn.Flatten()) else: self.output.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output.add(conv1x1( in_channels=in_channels, out_channels=classes, use_bias=True)) self.output.add(GlobalAvgMaxPool2D()) self.output.add(nn.Flatten()) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_dpn(num_layers, b_case=False, for_training=False, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DPN model with specific parameters. Parameters: ---------- num_layers : int Number of layers. b_case : bool, default False Whether to use B-case model. for_training : bool Whether to use model for training. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if num_layers == 68: init_block_channels = 10 init_block_kernel_size = 3 init_block_padding = 1 bw_factor = 1 k_r = 128 groups = 32 k_sec = (3, 4, 12, 3) incs = (16, 32, 32, 64) test_time_pool = True elif num_layers == 98: init_block_channels = 96 init_block_kernel_size = 7 init_block_padding = 3 bw_factor = 4 k_r = 160 groups = 40 k_sec = (3, 6, 20, 3) incs = (16, 32, 32, 128) test_time_pool = True elif num_layers == 107: init_block_channels = 128 init_block_kernel_size = 7 init_block_padding = 3 bw_factor = 4 k_r = 200 groups = 50 k_sec = (4, 8, 20, 3) incs = (20, 64, 64, 128) test_time_pool = True elif num_layers == 131: init_block_channels = 128 init_block_kernel_size = 7 init_block_padding = 3 bw_factor = 4 k_r = 160 groups = 40 k_sec = (4, 8, 28, 3) incs = (16, 32, 32, 128) test_time_pool = True else: raise ValueError("Unsupported DPN version with number of layers {}".format(num_layers)) channels = [[0] * li for li in k_sec] rs = [0 * li for li in k_sec] bws = [0 * li for li in k_sec] for i in range(len(k_sec)): rs[i] = (2 ** i) * k_r bws[i] = (2 ** i) * 64 * bw_factor inc = incs[i] channels[i][0] = bws[i] + 3 * inc for j in range(1, k_sec[i]): channels[i][j] = channels[i][j - 1] + inc net = DPN( channels=channels, init_block_channels=init_block_channels, init_block_kernel_size=init_block_kernel_size, init_block_padding=init_block_padding, rs=rs, bws=bws, incs=incs, groups=groups, b_case=b_case, for_training=for_training, test_time_pool=test_time_pool, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def dpn68(**kwargs): """ DPN-68 model from 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dpn(num_layers=68, b_case=False, model_name="dpn68", **kwargs) def dpn68b(**kwargs): """ DPN-68b model from 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dpn(num_layers=68, b_case=True, model_name="dpn68b", **kwargs) def dpn98(**kwargs): """ DPN-98 model from 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dpn(num_layers=98, b_case=False, model_name="dpn98", **kwargs) def dpn107(**kwargs): """ DPN-107 model from 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dpn(num_layers=107, b_case=False, model_name="dpn107", **kwargs) def dpn131(**kwargs): """ DPN-131 model from 'Dual Path Networks,' https://arxiv.org/abs/1707.01629. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dpn(num_layers=131, b_case=False, model_name="dpn131", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False for_training = False models = [ dpn68, # dpn68b, dpn98, # dpn107, dpn131, ] for model in models: net = model(pretrained=pretrained, for_training=for_training) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != dpn68 or weight_count == 12611602) assert (model != dpn68b or weight_count == 12611602) assert (model != dpn98 or weight_count == 61570728) assert (model != dpn107 or weight_count == 86917800) assert (model != dpn131 or weight_count == 79254504) # net.hybridize() x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
20,220
28.957037
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/sknet.py
""" SKNet for ImageNet-1K, implemented in Gluon. Original paper: 'Selective Kernel Networks,' https://arxiv.org/abs/1903.06586. """ __all__ = ['SKNet', 'sknet50', 'sknet101', 'sknet152'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv1x1_block, conv3x3_block, Concurrent from .resnet import ResInitBlock class SKConvBlock(HybridBlock): """ SKNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. groups : int, default 32 Number of groups in branches. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. num_branches : int, default 2 Number of branches (`M` parameter in the paper). reduction : int, default 16 Reduction value for intermediate channels (`r` parameter in the paper). min_channels : int, default 32 Minimal number of intermediate channels (`L` parameter in the paper). """ def __init__(self, in_channels, out_channels, strides, groups=32, bn_use_global_stats=False, num_branches=2, reduction=16, min_channels=32, **kwargs): super(SKConvBlock, self).__init__(**kwargs) self.num_branches = num_branches self.out_channels = out_channels mid_channels = max(in_channels // reduction, min_channels) with self.name_scope(): self.branches = Concurrent(stack=True, prefix="") for i in range(num_branches): dilation = 1 + i self.branches.add(conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=strides, padding=dilation, dilation=dilation, groups=groups, bn_use_global_stats=bn_use_global_stats)) self.fc1 = conv1x1_block( in_channels=out_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.fc2 = conv1x1( in_channels=mid_channels, out_channels=(out_channels * num_branches)) def hybrid_forward(self, F, x): y = self.branches(x) u = y.sum(axis=1) s = F.contrib.AdaptiveAvgPooling2D(u, output_size=1) z = self.fc1(s) w = self.fc2(z) w = w.reshape((0, self.num_branches, self.out_channels)) w = F.softmax(w, axis=1) w = w.expand_dims(3).expand_dims(4) y = F.broadcast_mul(y, w) y = y.sum(axis=1) return y class SKNetBottleneck(HybridBlock): """ SKNet bottleneck block for residual path in SKNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck_factor : int, default 2 Bottleneck factor. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, bottleneck_factor=2, **kwargs): super(SKNetBottleneck, self).__init__(**kwargs) mid_channels = out_channels // bottleneck_factor with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.conv2 = SKConvBlock( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class SKNetUnit(HybridBlock): """ SKNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, **kwargs): super(SKNetUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = SKNetBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class SKNet(HybridBlock): """ SKNet model from 'Selective Kernel Networks,' https://arxiv.org/abs/1903.06586. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(SKNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(SKNetUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_sknet(blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create SKNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] else: raise ValueError("Unsupported SKNet with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = SKNet( channels=channels, init_block_channels=init_block_channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def sknet50(**kwargs): """ SKNet-50 model from 'Selective Kernel Networks,' https://arxiv.org/abs/1903.06586. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sknet(blocks=50, model_name="sknet50", **kwargs) def sknet101(**kwargs): """ SKNet-101 model from 'Selective Kernel Networks,' https://arxiv.org/abs/1903.06586. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sknet(blocks=101, model_name="sknet101", **kwargs) def sknet152(**kwargs): """ SKNet-152 model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_sknet(blocks=152, model_name="sknet152", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ sknet50, sknet101, sknet152, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sknet50 or weight_count == 27479784) assert (model != sknet101 or weight_count == 48736040) assert (model != sknet152 or weight_count == 66295656) x = mx.nd.zeros((14, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (14, 1000)) if __name__ == "__main__": _test()
12,972
31.595477
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/spnasnet.py
""" Single-Path NASNet for ImageNet-1K, implemented in Gluon. Original paper: 'Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours,' https://arxiv.org/abs/1904.02877. """ __all__ = ['SPNASNet', 'spnasnet'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block class SPNASUnit(HybridBlock): """ Single-Path NASNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the second convolution layer. use_kernel3 : bool Whether to use 3x3 (instead of 5x5) kernel. exp_factor : int Expansion factor for each unit. use_skip : bool, default True Whether to use skip connection. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. activation : str, default 'relu' Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, strides, use_kernel3, exp_factor, use_skip=True, bn_use_global_stats=False, activation="relu", **kwargs): super(SPNASUnit, self).__init__(**kwargs) assert (exp_factor >= 1) self.residual = (in_channels == out_channels) and (strides == 1) and use_skip self.use_exp_conv = exp_factor > 1 mid_channels = exp_factor * in_channels with self.name_scope(): if self.use_exp_conv: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, activation=activation) if use_kernel3: self.conv1 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=activation) else: self.conv1 = dwconv5x5_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=activation) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) def hybrid_forward(self, F, x): if self.residual: identity = x if self.use_exp_conv: x = self.exp_conv(x) x = self.conv1(x) x = self.conv2(x) if self.residual: x = x + identity return x class SPNASInitBlock(HybridBlock): """ Single-Path NASNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of middle channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. """ def __init__(self, in_channels, out_channels, mid_channels, bn_use_global_stats=False, **kwargs): super(SPNASInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, strides=2, bn_use_global_stats=bn_use_global_stats) self.conv2 = SPNASUnit( in_channels=mid_channels, out_channels=out_channels, strides=1, use_kernel3=True, exp_factor=1, use_skip=False, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class SPNASFinalBlock(HybridBlock): """ Single-Path NASNet specific final block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of middle channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. """ def __init__(self, in_channels, out_channels, mid_channels, bn_use_global_stats=False, **kwargs): super(SPNASFinalBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = SPNASUnit( in_channels=in_channels, out_channels=mid_channels, strides=1, use_kernel3=True, exp_factor=6, use_skip=False, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class SPNASNet(HybridBlock): """ Single-Path NASNet model from 'Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours,' https://arxiv.org/abs/1904.02877. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : list of 2 int Number of output channels for the initial unit. final_block_channels : list of 2 int Number of output channels for the final block of the feature extractor. kernels3 : list of list of int/bool Using 3x3 (instead of 5x5) kernel for each unit. exp_factors : list of list of int Expansion factor for each unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, kernels3, exp_factors, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(SPNASNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(SPNASInitBlock( in_channels=in_channels, out_channels=init_block_channels[1], mid_channels=init_block_channels[0], bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels[1] for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if ((j == 0) and (i != 3)) or\ ((j == len(channels_per_stage) // 2) and (i == 3)) else 1 use_kernel3 = kernels3[i][j] == 1 exp_factor = exp_factors[i][j] stage.add(SPNASUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, use_kernel3=use_kernel3, exp_factor=exp_factor, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(SPNASFinalBlock( in_channels=in_channels, out_channels=final_block_channels[1], mid_channels=final_block_channels[0], bn_use_global_stats=bn_use_global_stats)) in_channels = final_block_channels[1] self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_spnasnet(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Single-Path NASNet model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ init_block_channels = [32, 16] final_block_channels = [320, 1280] channels = [[24, 24, 24], [40, 40, 40, 40], [80, 80, 80, 80], [96, 96, 96, 96, 192, 192, 192, 192]] kernels3 = [[1, 1, 1], [0, 1, 1, 1], [0, 1, 1, 1], [0, 0, 0, 0, 0, 0, 0, 0]] exp_factors = [[3, 3, 3], [6, 3, 3, 3], [6, 3, 3, 3], [6, 3, 3, 3, 6, 6, 6, 6]] net = SPNASNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, kernels3=kernels3, exp_factors=exp_factors, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def spnasnet(**kwargs): """ Single-Path NASNet model from 'Single-Path NAS: Designing Hardware-Efficient ConvNets in less than 4 Hours,' https://arxiv.org/abs/1904.02877. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_spnasnet(model_name="spnasnet", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ spnasnet, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != spnasnet or weight_count == 4421616) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
12,626
33.405995
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/fastscnn.py
""" Fast-SCNN for image segmentation, implemented in Gluon. Original paper: 'Fast-SCNN: Fast Semantic Segmentation Network,' https://arxiv.org/abs/1902.04502. """ __all__ = ['FastSCNN', 'fastscnn_cityscapes'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import Identity from .common import conv1x1, conv1x1_block, conv3x3_block, dwconv3x3_block, dwsconv3x3_block, Concurrent,\ InterpolationBlock class Stem(HybridBlock): """ Fast-SCNN specific stem block. Parameters: ---------- in_channels : int Number of input channels. channels : tuple/list of 3 int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(Stem, self).__init__(**kwargs) assert (len(channels) == 3) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=channels[0], strides=2, padding=0, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv2 = dwsconv3x3_block( in_channels=channels[0], out_channels=channels[1], strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv3 = dwsconv3x3_block( in_channels=channels[1], out_channels=channels[2], strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class LinearBottleneck(HybridBlock): """ Fast-SCNN specific Linear Bottleneck layer from MobileNetV2. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the second convolution layer. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, strides, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(LinearBottleneck, self).__init__(**kwargs) self.residual = (in_channels == out_channels) and (strides == 1) mid_channels = in_channels * 6 with self.name_scope(): self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv2 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) def hybrid_forward(self, F, x): if self.residual: identity = x x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) if self.residual: x = x + identity return x class FeatureExtractor(HybridBlock): """ Fast-SCNN specific feature extractor/encoder. Parameters: ---------- in_channels : int Number of input channels. channels : list of list of int Number of output channels for each unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(FeatureExtractor, self).__init__(**kwargs) with self.name_scope(): self.features = nn.HybridSequential(prefix="") for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != len(channels) - 1) else 1 stage.add(LinearBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = out_channels self.features.add(stage) def hybrid_forward(self, F, x): x = self.features(x) return x class PoolingBranch(HybridBlock): """ Fast-SCNN specific pooling branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of 2 int or None Spatial size of input image. down_size : int Spatial size of downscaled image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, in_size, down_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(PoolingBranch, self).__init__(**kwargs) self.in_size = in_size self.down_size = down_size with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.up = InterpolationBlock( scale_factor=None, out_size=in_size) def hybrid_forward(self, F, x): in_size = self.in_size if self.in_size is not None else x.shape[2:] x = F.contrib.AdaptiveAvgPooling2D(x, output_size=self.down_size) x = self.conv(x) x = self.up(x, in_size) return x class FastPyramidPooling(HybridBlock): """ Fast-SCNN specific fast pyramid pooling block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of 2 int or None Spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, in_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(FastPyramidPooling, self).__init__(**kwargs) down_sizes = [1, 2, 3, 6] mid_channels = in_channels // 4 with self.name_scope(): self.branches = Concurrent() self.branches.add(Identity()) for down_size in down_sizes: self.branches.add(PoolingBranch( in_channels=in_channels, out_channels=mid_channels, in_size=in_size, down_size=down_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) self.conv = conv1x1_block( in_channels=(in_channels * 2), out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): x = self.branches(x) x = self.conv(x) return x class FeatureFusion(HybridBlock): """ Fast-SCNN specific feature fusion block. Parameters: ---------- x_in_channels : int Number of high resolution (x) input channels. y_in_channels : int Number of low resolution (y) input channels. out_channels : int Number of output channels. x_in_size : tuple of 2 int or None Spatial size of high resolution (x) input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, x_in_channels, y_in_channels, out_channels, x_in_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(FeatureFusion, self).__init__(**kwargs) self.x_in_size = x_in_size with self.name_scope(): self.up = InterpolationBlock( scale_factor=None, out_size=x_in_size) self.low_dw_conv = dwconv3x3_block( in_channels=y_in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.low_pw_conv = conv1x1_block( in_channels=out_channels, out_channels=out_channels, use_bias=True, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) self.high_conv = conv1x1_block( in_channels=x_in_channels, out_channels=out_channels, use_bias=True, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x, y): x_in_size = self.x_in_size if self.x_in_size is not None else x.shape[2:] y = self.up(y, x_in_size) y = self.low_dw_conv(y) y = self.low_pw_conv(y) x = self.high_conv(x) out = x + y return self.activ(out) class Head(HybridBlock): """ Fast-SCNN head (classifier) block. Parameters: ---------- in_channels : int Number of input channels. classes : int Number of classification classes. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, classes, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(Head, self).__init__(**kwargs) with self.name_scope(): self.conv1 = dwsconv3x3_block( in_channels=in_channels, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv2 = dwsconv3x3_block( in_channels=in_channels, out_channels=in_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.dropout = nn.Dropout(rate=0.1) self.conv3 = conv1x1( in_channels=in_channels, out_channels=classes, use_bias=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.dropout(x) x = self.conv3(x) return x class AuxHead(HybridBlock): """ Fast-SCNN auxiliary (after stem) head (classifier) block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. classes : int Number of classification classes. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, mid_channels, classes, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(AuxHead, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.dropout = nn.Dropout(rate=0.1) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=classes, use_bias=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.dropout(x) x = self.conv2(x) return x class FastSCNN(HybridBlock): """ Fast-SCNN from 'Fast-SCNN: Fast Semantic Segmentation Network,' https://arxiv.org/abs/1902.04502. Parameters: ---------- aux : bool, default False Whether to output an auxiliary result. fixed_size : bool, default True Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (1024, 1024) Spatial size of the expected input image. classes : int, default 19 Number of segmentation classes. """ def __init__(self, aux=False, fixed_size=True, bn_use_global_stats=False, bn_cudnn_off=False, in_channels=3, in_size=(1024, 1024), classes=19, **kwargs): super(FastSCNN, self).__init__(**kwargs) assert (in_channels > 0) assert ((in_size[0] % 32 == 0) and (in_size[1] % 32 == 0)) self.in_size = in_size self.classes = classes self.aux = aux self.fixed_size = fixed_size with self.name_scope(): steam_channels = [32, 48, 64] self.stem = Stem( in_channels=in_channels, channels=steam_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) in_channels = steam_channels[-1] feature_channels = [[64, 64, 64], [96, 96, 96], [128, 128, 128]] self.features = FeatureExtractor( in_channels=in_channels, channels=feature_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) pool_out_size = (in_size[0] // 32, in_size[1] // 32) if fixed_size else None self.pool = FastPyramidPooling( in_channels=feature_channels[-1][-1], out_channels=feature_channels[-1][-1], in_size=pool_out_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) fusion_out_size = (in_size[0] // 8, in_size[1] // 8) if fixed_size else None fusion_out_channels = 128 self.fusion = FeatureFusion( x_in_channels=steam_channels[-1], y_in_channels=feature_channels[-1][-1], out_channels=fusion_out_channels, x_in_size=fusion_out_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.head = Head( in_channels=fusion_out_channels, classes=classes, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.up = InterpolationBlock( scale_factor=None, out_size=in_size) if self.aux: self.aux_head = AuxHead( in_channels=64, mid_channels=64, classes=classes, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): in_size = self.in_size if self.fixed_size else x.shape[2:] x = self.stem(x) y = self.features(x) y = self.pool(y) y = self.fusion(x, y) y = self.head(y) y = self.up(y, in_size) if self.aux: x = self.aux_head(x) x = self.up(x, in_size) return y, x return y def get_fastscnn(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Fast-SCNN model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ net = FastSCNN( **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx, ignore_extra=True) return net def fastscnn_cityscapes(classes=19, aux=True, **kwargs): """ Fast-SCNN model for Cityscapes from 'Fast-SCNN: Fast Semantic Segmentation Network,' https://arxiv.org/abs/1902.04502. Parameters: ---------- classes : int, default 19 Number of segmentation classes. aux : bool, default True Whether to output an auxiliary result. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fastscnn(classes=classes, aux=aux, model_name="fastscnn_cityscapes", **kwargs) def _test(): import numpy as np import mxnet as mx # in_size = (1024, 1024) in_size = (1024, 2048) aux = True pretrained = False fixed_size = False models = [ (fastscnn_cityscapes, 19), ] for model, classes in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size, aux=aux) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) if aux: assert (model != fastscnn_cityscapes or weight_count == 1176278) else: assert (model != fastscnn_cityscapes or weight_count == 1138051) x = mx.nd.zeros((1, 3, in_size[0], in_size[1]), ctx=ctx) ys = net(x) y = ys[0] if aux else ys assert ((y.shape[0] == x.shape[0]) and (y.shape[1] == classes) and (y.shape[2] == x.shape[2]) and (y.shape[3] == x.shape[3])) if __name__ == "__main__": _test()
21,678
33.520701
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/res2net.py
""" Res2Net for ImageNet-1K, implemented in Gluon. Original paper: 'Res2Net: A New Multi-scale Backbone Architecture,' https://arxiv.org/abs/1904.01169. """ __all__ = ['Res2Net', 'res2net50_w14_s8', 'res2net50_w26_s8'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import Identity from .common import conv1x1, conv3x3, conv1x1_block from .resnet import ResInitBlock from .preresnet import PreResActivation class HierarchicalConcurrent(nn.HybridSequential): """ A container for hierarchical concatenation of blocks with parameters. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. multi_input : bool, default False Whether input is multiple. """ def __init__(self, axis=1, multi_input=False, **kwargs): super(HierarchicalConcurrent, self).__init__(**kwargs) self.axis = axis self.multi_input = multi_input def hybrid_forward(self, F, x): out = [] y_prev = None if self.multi_input: xs = F.split(x, axis=self.axis, num_outputs=len(self._children.values())) for i, block in enumerate(self._children.values()): if self.multi_input: y = block(xs[i]) else: y = block(x) if y_prev is not None: y = y + y_prev out.append(y) y_prev = y out = F.concat(*out, dim=self.axis) return out class Res2NetUnit(HybridBlock): """ Res2Net unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the branch convolution layers. width : int Width of filters. scale : int Number of scale. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, width, scale, bn_use_global_stats, **kwargs): super(Res2NetUnit, self).__init__(**kwargs) self.scale = scale downsample = (strides != 1) self.resize_identity = (in_channels != out_channels) or downsample mid_channels = width * scale brn_channels = width with self.name_scope(): self.reduce_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_use_global_stats=bn_use_global_stats) self.branches = HierarchicalConcurrent(axis=1, multi_input=True, prefix="") if downsample: self.branches.add(conv1x1( in_channels=brn_channels, out_channels=brn_channels, strides=strides)) else: self.branches.add(Identity()) for i in range(scale - 1): self.branches.add(conv3x3( in_channels=brn_channels, out_channels=brn_channels, strides=strides)) self.preactiv = PreResActivation(in_channels=mid_channels) self.merge_conv = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x y = self.reduce_conv(x) y = self.branches(y) y = self.preactiv(y) y = self.merge_conv(y) y = y + identity y = self.activ(y) return y class Res2Net(HybridBlock): """ Res2Net model from 'Res2Net: A New Multi-scale Backbone Architecture,' https://arxiv.org/abs/1904.01169. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. width : int Width of filters. scale : int Number of scale. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, width, scale, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(Res2Net, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(Res2NetUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, width=width, scale=scale, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_res2net(blocks, width, scale, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Res2Net model with specific parameters. Parameters: ---------- blocks : int Number of blocks. width : int Width of filters. scale : int Number of scale. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ bottleneck = True if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] else: raise ValueError("Unsupported Res2Net with number of blocks: {}".format(blocks)) assert (sum(layers) * 3 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = Res2Net( channels=channels, init_block_channels=init_block_channels, width=width, scale=scale, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def res2net50_w14_s8(**kwargs): """ Res2Net-50 (14wx8s) model from 'Res2Net: A New Multi-scale Backbone Architecture,' https://arxiv.org/abs/1904.01169. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_res2net(blocks=50, width=14, scale=8, model_name="res2net50_w14_s8", **kwargs) def res2net50_w26_s8(**kwargs): """ Res2Net-50 (26wx8s) model from 'Res2Net: A New Multi-scale Backbone Architecture,' https://arxiv.org/abs/1904.01169. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_res2net(blocks=50, width=26, scale=8, model_name="res2net50_w14_s8", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ res2net50_w14_s8, res2net50_w26_s8, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != res2net50_w14_s8 or weight_count == 8231732) assert (model != res2net50_w26_s8 or weight_count == 11432660) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
11,307
31.401146
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/darknet.py
""" DarkNet for ImageNet-1K, implemented in Gluon. Original source: 'Darknet: Open source neural networks in c,' https://github.com/pjreddie/darknet. """ __all__ = ['DarkNet', 'darknet_ref', 'darknet_tiny', 'darknet19'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block def dark_convYxY(in_channels, out_channels, bn_use_global_stats, alpha, pointwise): """ DarkNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. alpha : float Slope coefficient for Leaky ReLU activation. pointwise : bool Whether use 1x1 (pointwise) convolution or 3x3 convolution. """ if pointwise: return conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=nn.LeakyReLU(alpha=alpha)) else: return conv3x3_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=nn.LeakyReLU(alpha=alpha)) class DarkNet(HybridBlock): """ DarkNet model from 'Darknet: Open source neural networks in c,' https://github.com/pjreddie/darknet. Parameters: ---------- channels : list of list of int Number of output channels for each unit. odd_pointwise : bool Whether pointwise convolution layer is used for each odd unit. avg_pool_size : int Window size of the final average pooling. cls_activ : bool Whether classification convolution layer uses an activation. alpha : float, default 0.1 Slope coefficient for Leaky ReLU activation. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, odd_pointwise, avg_pool_size, cls_activ, alpha=0.1, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(DarkNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): stage.add(dark_convYxY( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, alpha=alpha, pointwise=(len(channels_per_stage) > 1) and not(((j + 1) % 2 == 1) ^ odd_pointwise))) in_channels = out_channels if i != len(channels) - 1: stage.add(nn.MaxPool2D( pool_size=2, strides=2)) self.features.add(stage) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Conv2D( channels=classes, kernel_size=1, in_channels=in_channels)) if cls_activ: self.output.add(nn.LeakyReLU(alpha=alpha)) self.output.add(nn.AvgPool2D( pool_size=avg_pool_size, strides=1)) self.output.add(nn.Flatten()) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_darknet(version, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DarkNet model with specific parameters. Parameters: ---------- version : str Version of SqueezeNet ('ref', 'tiny' or '19'). model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if version == 'ref': channels = [[16], [32], [64], [128], [256], [512], [1024]] odd_pointwise = False avg_pool_size = 3 cls_activ = True elif version == 'tiny': channels = [[16], [32], [16, 128, 16, 128], [32, 256, 32, 256], [64, 512, 64, 512, 128]] odd_pointwise = True avg_pool_size = 14 cls_activ = False elif version == '19': channels = [[32], [64], [128, 64, 128], [256, 128, 256], [512, 256, 512, 256, 512], [1024, 512, 1024, 512, 1024]] odd_pointwise = False avg_pool_size = 7 cls_activ = False else: raise ValueError("Unsupported DarkNet version {}".format(version)) net = DarkNet( channels=channels, odd_pointwise=odd_pointwise, avg_pool_size=avg_pool_size, cls_activ=cls_activ, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def darknet_ref(**kwargs): """ DarkNet 'Reference' model from 'Darknet: Open source neural networks in c,' https://github.com/pjreddie/darknet. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_darknet(version="ref", model_name="darknet_ref", **kwargs) def darknet_tiny(**kwargs): """ DarkNet Tiny model from 'Darknet: Open source neural networks in c,' https://github.com/pjreddie/darknet. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_darknet(version="tiny", model_name="darknet_tiny", **kwargs) def darknet19(**kwargs): """ DarkNet-19 model from 'Darknet: Open source neural networks in c,' https://github.com/pjreddie/darknet. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_darknet(version="19", model_name="darknet19", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ darknet_ref, darknet_tiny, darknet19, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != darknet_ref or weight_count == 7319416) assert (model != darknet_tiny or weight_count == 1042104) assert (model != darknet19 or weight_count == 20842376) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
9,154
32.17029
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/ror_cifar.py
""" RoR-3 for CIFAR/SVHN, implemented in Gluon. Original paper: 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. """ __all__ = ['CIFARRoR', 'ror3_56_cifar10', 'ror3_56_cifar100', 'ror3_56_svhn', 'ror3_110_cifar10', 'ror3_110_cifar100', 'ror3_110_svhn', 'ror3_164_cifar10', 'ror3_164_cifar100', 'ror3_164_svhn'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1_block, conv3x3_block class RoRBlock(HybridBlock): """ RoR-3 block for residual path in residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, dropout_rate, **kwargs): super(RoRBlock, self).__init__(**kwargs) self.use_dropout = (dropout_rate != 0.0) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) if self.use_dropout: self.dropout = nn.Dropout(rate=dropout_rate) def hybrid_forward(self, F, x): x = self.conv1(x) if self.use_dropout: x = self.dropout(x) x = self.conv2(x) return x class RoRResUnit(HybridBlock): """ RoR-3 residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. last_activate : bool, default True Whether activate output. """ def __init__(self, in_channels, out_channels, bn_use_global_stats, dropout_rate, last_activate=True, **kwargs): super(RoRResUnit, self).__init__(**kwargs) self.last_activate = last_activate self.resize_identity = (in_channels != out_channels) with self.name_scope(): self.body = RoRBlock( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, dropout_rate=dropout_rate) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, activation=None) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity if self.last_activate: x = self.activ(x) return x class RoRResStage(HybridBlock): """ RoR-3 residual stage. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of output channels for each unit. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. downsample : bool, default True Whether downsample output. """ def __init__(self, in_channels, out_channels_list, bn_use_global_stats, dropout_rate, downsample=True, **kwargs): super(RoRResStage, self).__init__(**kwargs) self.downsample = downsample with self.name_scope(): self.shortcut = conv1x1_block( in_channels=in_channels, out_channels=out_channels_list[-1], bn_use_global_stats=bn_use_global_stats, activation=None) self.units = nn.HybridSequential(prefix="") with self.units.name_scope(): for i, out_channels in enumerate(out_channels_list): last_activate = (i != len(out_channels_list) - 1) self.units.add(RoRResUnit( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, dropout_rate=dropout_rate, last_activate=last_activate)) in_channels = out_channels if self.downsample: self.activ = nn.Activation("relu") self.pool = nn.MaxPool2D( pool_size=2, strides=2, padding=0) def hybrid_forward(self, F, x): identity = self.shortcut(x) x = self.units(x) x = x + identity if self.downsample: x = self.activ(x) x = self.pool(x) return x class RoRResBody(HybridBlock): """ RoR-3 residual body (main feature path). Parameters: ---------- in_channels : int Number of input channels. out_channels_lists : list of list of int Number of output channels for each stage. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels_lists, bn_use_global_stats, dropout_rate, **kwargs): super(RoRResBody, self).__init__(**kwargs) with self.name_scope(): self.shortcut = conv1x1_block( in_channels=in_channels, out_channels=out_channels_lists[-1][-1], strides=4, bn_use_global_stats=bn_use_global_stats, activation=None) self.stages = nn.HybridSequential(prefix="") with self.stages.name_scope(): for i, channels_per_stage in enumerate(out_channels_lists): downsample = (i != len(out_channels_lists) - 1) self.stages.add(RoRResStage( in_channels=in_channels, out_channels_list=channels_per_stage, bn_use_global_stats=bn_use_global_stats, dropout_rate=dropout_rate, downsample=downsample)) in_channels = channels_per_stage[-1] self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): identity = self.shortcut(x) x = self.stages(x) x = x + identity x = self.activ(x) return x class CIFARRoR(HybridBlock): """ RoR-3 model for CIFAR from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. dropout_rate : float, default 0.0 Parameter of Dropout layer. Faction of the input units to drop. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (32, 32) Spatial size of the expected input image. classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bn_use_global_stats=False, dropout_rate=0.0, in_channels=3, in_size=(32, 32), classes=10, **kwargs): super(CIFARRoR, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels self.features.add(RoRResBody( in_channels=in_channels, out_channels_lists=channels, bn_use_global_stats=bn_use_global_stats, dropout_rate=dropout_rate)) in_channels = channels[-1][-1] self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_ror_cifar(classes, blocks, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create RoR-3 model for CIFAR with specific parameters. Parameters: ---------- classes : int Number of classification classes. blocks : int Number of blocks. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ assert (classes in [10, 100]) assert ((blocks - 8) % 6 == 0) layers = [(blocks - 8) // 6] * 3 channels_per_layers = [16, 32, 64] init_block_channels = 16 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = CIFARRoR( channels=channels, init_block_channels=init_block_channels, classes=classes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def ror3_56_cifar10(classes=10, **kwargs): """ RoR-3-56 model for CIFAR-10 from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=56, model_name="ror3_56_cifar10", **kwargs) def ror3_56_cifar100(classes=100, **kwargs): """ RoR-3-56 model for CIFAR-100 from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=56, model_name="ror3_56_cifar100", **kwargs) def ror3_56_svhn(classes=10, **kwargs): """ RoR-3-56 model for SVHN from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=56, model_name="ror3_56_svhn", **kwargs) def ror3_110_cifar10(classes=10, **kwargs): """ RoR-3-110 model for CIFAR-10 from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=110, model_name="ror3_110_cifar10", **kwargs) def ror3_110_cifar100(classes=100, **kwargs): """ RoR-3-110 model for CIFAR-100 from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=110, model_name="ror3_110_cifar100", **kwargs) def ror3_110_svhn(classes=10, **kwargs): """ RoR-3-110 model for SVHN from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=110, model_name="ror3_110_svhn", **kwargs) def ror3_164_cifar10(classes=10, **kwargs): """ RoR-3-164 model for CIFAR-10 from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=164, model_name="ror3_164_cifar10", **kwargs) def ror3_164_cifar100(classes=100, **kwargs): """ RoR-3-164 model for CIFAR-100 from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=164, model_name="ror3_164_cifar100", **kwargs) def ror3_164_svhn(classes=10, **kwargs): """ RoR-3-164 model for SVHN from 'Residual Networks of Residual Networks: Multilevel Residual Networks,' https://arxiv.org/abs/1608.02908. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_ror_cifar(classes=classes, blocks=164, model_name="ror3_164_svhn", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ (ror3_56_cifar10, 10), (ror3_56_cifar100, 100), (ror3_56_svhn, 10), (ror3_110_cifar10, 10), (ror3_110_cifar100, 100), (ror3_110_svhn, 10), (ror3_164_cifar10, 10), (ror3_164_cifar100, 100), (ror3_164_svhn, 10), ] for model, classes in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ror3_56_cifar10 or weight_count == 762746) assert (model != ror3_56_cifar100 or weight_count == 768596) assert (model != ror3_56_svhn or weight_count == 762746) assert (model != ror3_110_cifar10 or weight_count == 1637690) assert (model != ror3_110_cifar100 or weight_count == 1643540) assert (model != ror3_110_svhn or weight_count == 1637690) assert (model != ror3_164_cifar10 or weight_count == 2512634) assert (model != ror3_164_cifar100 or weight_count == 2518484) assert (model != ror3_164_svhn or weight_count == 2512634) x = mx.nd.zeros((1, 3, 32, 32), ctx=ctx) y = net(x) assert (y.shape == (1, classes)) if __name__ == "__main__": _test()
19,573
33.522046
118
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/dicenet.py
""" DiCENet for ImageNet-1K, implemented in Gluon. Original paper: 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. """ __all__ = ['DiceNet', 'dicenet_wd5', 'dicenet_wd2', 'dicenet_w3d4', 'dicenet_w1', 'dicenet_w5d4', 'dicenet_w3d2', 'dicenet_w7d8', 'dicenet_w2'] import os import math from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import conv1x1, conv3x3, conv1x1_block, conv3x3_block, NormActivation, ChannelShuffle, Concurrent, PReLU2 class SpatialDiceBranch(HybridBlock): """ Spatial element of DiCE block for selected dimension. Parameters: ---------- sp_size : int Desired size for selected spatial dimension. is_height : bool Is selected dimension height. fixed_size : bool Whether to expect fixed spatial size of input image. """ def __init__(self, sp_size, is_height, fixed_size, **kwargs): super(SpatialDiceBranch, self).__init__(**kwargs) self.is_height = is_height self.fixed_size = fixed_size self.index = 2 if is_height else 3 self.base_sp_size = sp_size with self.name_scope(): self.conv = conv3x3( in_channels=self.base_sp_size, out_channels=self.base_sp_size, groups=self.base_sp_size) def hybrid_forward(self, F, x): if not self.fixed_size: height, width = x.shape[2:] if self.is_height: real_sp_size = height real_in_size = (real_sp_size, width) base_in_size = (self.base_sp_size, width) else: real_sp_size = width real_in_size = (height, real_sp_size) base_in_size = (height, self.base_sp_size) if real_sp_size != self.base_sp_size: if real_sp_size < self.base_sp_size: x = F.contrib.BilinearResize2D(x, height=base_in_size[0], width=base_in_size[1]) else: x = F.contrib.AdaptiveAvgPooling2D(x, output_size=base_in_size) x = x.swapaxes(1, self.index) x = self.conv(x) x = x.swapaxes(1, self.index) if not self.fixed_size: changed_sp_size = x.shape[self.index] if real_sp_size != changed_sp_size: if changed_sp_size < real_sp_size: x = F.contrib.BilinearResize2D(x, height=real_in_size[0], width=real_in_size[1]) else: x = F.contrib.AdaptiveAvgPooling2D(x, output_size=real_in_size) return x class DiceBaseBlock(HybridBlock): """ Base part of DiCE block (without attention). Parameters: ---------- channels : int Number of input/output channels. in_size : tuple of two ints Spatial size of the expected input image. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, channels, in_size, fixed_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(DiceBaseBlock, self).__init__(**kwargs) mid_channels = 3 * channels with self.name_scope(): self.convs = Concurrent() self.convs.add(conv3x3( in_channels=channels, out_channels=channels, groups=channels)) self.convs.add(SpatialDiceBranch( sp_size=in_size[0], is_height=True, fixed_size=fixed_size)) self.convs.add(SpatialDiceBranch( sp_size=in_size[1], is_height=False, fixed_size=fixed_size)) self.norm_activ = NormActivation( in_channels=mid_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=mid_channels))) self.shuffle = ChannelShuffle( channels=mid_channels, groups=3) self.squeeze_conv = conv1x1_block( in_channels=mid_channels, out_channels=channels, groups=channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=channels))) def hybrid_forward(self, F, x): x = self.convs(x) x = self.norm_activ(x) x = self.shuffle(x) x = self.squeeze_conv(x) return x class DiceAttBlock(HybridBlock): """ Pure attention part of DiCE block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. reduction : int, default 4 Squeeze reduction value. """ def __init__(self, in_channels, out_channels, reduction=4, **kwargs): super(DiceAttBlock, self).__init__(**kwargs) mid_channels = in_channels // reduction with self.name_scope(): self.pool = nn.GlobalAvgPool2D() self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels, use_bias=False) self.activ = nn.Activation("relu") self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, use_bias=False) self.sigmoid = nn.Activation("sigmoid") def hybrid_forward(self, F, x): w = self.pool(x) w = self.conv1(w) w = self.activ(w) w = self.conv2(w) w = self.sigmoid(w) return w class DiceBlock(HybridBlock): """ DiCE block (volume-wise separable convolutions). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of two ints Spatial size of the expected input image. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, in_size, fixed_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(DiceBlock, self).__init__(**kwargs) proj_groups = math.gcd(in_channels, out_channels) with self.name_scope(): self.base_block = DiceBaseBlock( channels=in_channels, in_size=in_size, fixed_size=fixed_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.att = DiceAttBlock( in_channels=in_channels, out_channels=out_channels) # assert (in_channels == out_channels) self.proj_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, groups=proj_groups, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=out_channels))) def hybrid_forward(self, F, x): x = self.base_block(x) w = self.att(x) x = self.proj_conv(x) x = F.broadcast_mul(x, w) return x class StridedDiceLeftBranch(HybridBlock): """ Left branch of the strided DiCE block. Parameters: ---------- channels : int Number of input/output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(StridedDiceLeftBranch, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=channels, out_channels=channels, strides=2, groups=channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=channels))) self.conv2 = conv1x1_block( in_channels=channels, out_channels=channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=channels))) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) return x class StridedDiceRightBranch(HybridBlock): """ Right branch of the strided DiCE block. Parameters: ---------- channels : int Number of input/output channels. in_size : tuple of two ints Spatial size of the expected input image. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, channels, in_size, fixed_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(StridedDiceRightBranch, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=3, strides=2, padding=1) self.dice = DiceBlock( in_channels=channels, out_channels=channels, in_size=(in_size[0] // 2, in_size[1] // 2), fixed_size=fixed_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.conv = conv1x1_block( in_channels=channels, out_channels=channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=channels))) def hybrid_forward(self, F, x): x = self.pool(x) x = self.dice(x) x = self.conv(x) return x class StridedDiceBlock(HybridBlock): """ Strided DiCE block (strided volume-wise separable convolutions). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of two ints Spatial size of the expected input image. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, in_size, fixed_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(StridedDiceBlock, self).__init__(**kwargs) assert (out_channels == 2 * in_channels) with self.name_scope(): self.branches = Concurrent() self.branches.add(StridedDiceLeftBranch( channels=in_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) self.branches.add(StridedDiceRightBranch( channels=in_channels, in_size=in_size, fixed_size=fixed_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) self.shuffle = ChannelShuffle( channels=out_channels, groups=2) def hybrid_forward(self, F, x): x = self.branches(x) x = self.shuffle(x) return x class ShuffledDiceRightBranch(HybridBlock): """ Right branch of the shuffled DiCE block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of two ints Spatial size of the expected input image. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, in_size, fixed_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(ShuffledDiceRightBranch, self).__init__(**kwargs) with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=out_channels))) self.dice = DiceBlock( in_channels=out_channels, out_channels=out_channels, in_size=in_size, fixed_size=fixed_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) def hybrid_forward(self, F, x): x = self.conv(x) x = self.dice(x) return x class ShuffledDiceBlock(HybridBlock): """ Shuffled DiCE block (shuffled volume-wise separable convolutions). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of two ints Spatial size of the expected input image. fixed_size : bool Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, in_size, fixed_size, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(ShuffledDiceBlock, self).__init__(**kwargs) self.left_part = in_channels - in_channels // 2 right_in_channels = in_channels - self.left_part right_out_channels = out_channels - self.left_part with self.name_scope(): self.right_branch = ShuffledDiceRightBranch( in_channels=right_in_channels, out_channels=right_out_channels, in_size=in_size, fixed_size=fixed_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off) self.shuffle = ChannelShuffle( channels=(2 * right_out_channels), groups=2) def hybrid_forward(self, F, x): x1, x2 = F.split(x, axis=1, num_outputs=2) x2 = self.right_branch(x2) x = F.concat(x1, x2, dim=1) x = self.shuffle(x) return x class DiceInitBlock(HybridBlock): """ DiceNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. """ def __init__(self, in_channels, out_channels, bn_use_global_stats=False, bn_cudnn_off=False, **kwargs): super(DiceInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=2, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off, activation=(lambda: PReLU2(in_channels=out_channels))) self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv(x) x = self.pool(x) return x class DiceClassifier(HybridBlock): """ DiceNet specific classifier block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. classes : int, default 1000 Number of classification classes. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, mid_channels, classes, dropout_rate, **kwargs): super(DiceClassifier, self).__init__(**kwargs) with self.name_scope(): self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels, groups=4) self.dropout = nn.Dropout(rate=dropout_rate) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=classes, use_bias=True) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.dropout(x) x = self.conv2(x) return x class DiceNet(HybridBlock): """ DiCENet model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. classifier_mid_channels : int Number of middle channels for classifier. dropout_rate : float Parameter of Dropout layer in classifier. Faction of the input units to drop. fixed_size : bool, default True Whether to expect fixed spatial size of input image. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. bn_cudnn_off : bool, default False Whether to disable CUDNN batch normalization operator. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, classifier_mid_channels, dropout_rate, fixed_size=True, bn_use_global_stats=False, bn_cudnn_off=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(DiceNet, self).__init__(**kwargs) assert ((in_size[0] % 32 == 0) and (in_size[1] % 32 == 0)) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(DiceInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = init_block_channels in_size = (in_size[0] // 4, in_size[1] // 4) for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): unit_class = StridedDiceBlock if j == 0 else ShuffledDiceBlock stage.add(unit_class( in_channels=in_channels, out_channels=out_channels, in_size=in_size, fixed_size=fixed_size, bn_use_global_stats=bn_use_global_stats, bn_cudnn_off=bn_cudnn_off)) in_channels = out_channels in_size = (in_size[0] // 2, in_size[1] // 2) if j == 0 else in_size self.features.add(stage) self.features.add(nn.GlobalAvgPool2D()) self.output = nn.HybridSequential(prefix="") self.output.add(DiceClassifier( in_channels=in_channels, mid_channels=classifier_mid_channels, classes=classes, dropout_rate=dropout_rate)) self.output.add(nn.Flatten()) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_dicenet(width_scale, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create DiCENet model with specific parameters. Parameters: ---------- width_scale : float Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ channels_per_layers_dict = { 0.2: [32, 64, 128], 0.5: [48, 96, 192], 0.75: [86, 172, 344], 1.0: [116, 232, 464], 1.25: [144, 288, 576], 1.5: [176, 352, 704], 1.75: [210, 420, 840], 2.0: [244, 488, 976], 2.4: [278, 556, 1112], } if width_scale not in channels_per_layers_dict.keys(): raise ValueError("Unsupported DiceNet with width scale: {}".format(width_scale)) channels_per_layers = channels_per_layers_dict[width_scale] layers = [3, 7, 3] if width_scale > 0.2: init_block_channels = 24 else: init_block_channels = 16 channels = [[ci] * li for i, (ci, li) in enumerate(zip(channels_per_layers, layers))] for i in range(len(channels)): pred_channels = channels[i - 1][-1] if i != 0 else init_block_channels channels[i] = [pred_channels * 2] + channels[i] if width_scale > 2.0: classifier_mid_channels = 1280 else: classifier_mid_channels = 1024 if width_scale > 1.0: dropout_rate = 0.2 else: dropout_rate = 0.1 net = DiceNet( channels=channels, init_block_channels=init_block_channels, classifier_mid_channels=classifier_mid_channels, dropout_rate=dropout_rate, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def dicenet_wd5(**kwargs): """ DiCENet x0.2 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=0.2, model_name="dicenet_wd5", **kwargs) def dicenet_wd2(**kwargs): """ DiCENet x0.5 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=0.5, model_name="dicenet_wd2", **kwargs) def dicenet_w3d4(**kwargs): """ DiCENet x0.75 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=0.75, model_name="dicenet_w3d4", **kwargs) def dicenet_w1(**kwargs): """ DiCENet x1.0 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=1.0, model_name="dicenet_w1", **kwargs) def dicenet_w5d4(**kwargs): """ DiCENet x1.25 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=1.25, model_name="dicenet_w5d4", **kwargs) def dicenet_w3d2(**kwargs): """ DiCENet x1.5 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=1.5, model_name="dicenet_w3d2", **kwargs) def dicenet_w7d8(**kwargs): """ DiCENet x1.75 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=1.75, model_name="dicenet_w7d8", **kwargs) def dicenet_w2(**kwargs): """ DiCENet x2.0 model from 'DiCENet: Dimension-wise Convolutions for Efficient Networks,' https://arxiv.org/abs/1906.03516. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_dicenet(width_scale=2.0, model_name="dicenet_w2", **kwargs) def _calc_width(net): import numpy as np net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) return weight_count def _test(): import mxnet as mx pretrained = False fixed_size = True models = [ dicenet_wd5, dicenet_wd2, dicenet_w3d4, dicenet_w1, dicenet_w5d4, dicenet_w3d2, dicenet_w7d8, dicenet_w2, ] for model in models: net = model(pretrained=pretrained, fixed_size=fixed_size) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != dicenet_wd5 or weight_count == 1130704) assert (model != dicenet_wd2 or weight_count == 1214120) assert (model != dicenet_w3d4 or weight_count == 1495676) assert (model != dicenet_w1 or weight_count == 1805604) assert (model != dicenet_w5d4 or weight_count == 2162888) assert (model != dicenet_w3d2 or weight_count == 2652200) assert (model != dicenet_w7d8 or weight_count == 3264932) assert (model != dicenet_w2 or weight_count == 3979044) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
31,317
32.820734
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/nvpattexp.py
""" Neural Voice Puppetry Audio-to-Expression net for speech-driven facial animation, implemented in Gluon. Original paper: 'Neural Voice Puppetry: Audio-driven Facial Reenactment,' https://arxiv.org/abs/1912.05566. """ __all__ = ['NvpAttExp', 'nvpattexp116bazel76'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import Softmax, DenseBlock, ConvBlock, ConvBlock1d, SelectableDense class NvpAttExpEncoder(HybridBlock): """ Neural Voice Puppetry Audio-to-Expression encoder. Parameters: ---------- audio_features : int Number of audio features (characters/sounds). audio_window_size : int Size of audio window (for time related audio features). seq_len : int, default Size of feature window. encoder_features : int Number of encoder features. """ def __init__(self, audio_features, audio_window_size, seq_len, encoder_features, **kwargs): super(NvpAttExpEncoder, self).__init__(**kwargs) self.audio_features = audio_features self.audio_window_size = audio_window_size self.seq_len = seq_len conv_channels = (32, 32, 64, 64) conv_slopes = (0.02, 0.02, 0.2, 0.2) fc_channels = (128, 64, encoder_features) fc_slopes = (0.02, 0.02, None) att_conv_channels = (16, 8, 4, 2, 1) att_conv_slopes = 0.02 with self.name_scope(): in_channels = audio_features self.conv_branch = nn.HybridSequential(prefix="") with self.conv_branch.name_scope(): for i, (out_channels, slope) in enumerate(zip(conv_channels, conv_slopes)): self.conv_branch.add(ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 1), strides=(2, 1), padding=(1, 0), use_bias=True, use_bn=False, activation=nn.LeakyReLU(alpha=slope))) in_channels = out_channels self.fc_branch = nn.HybridSequential(prefix="") with self.fc_branch.name_scope(): for i, (out_channels, slope) in enumerate(zip(fc_channels, fc_slopes)): activation = nn.LeakyReLU(alpha=slope) if slope is not None else nn.Activation("tanh") self.fc_branch.add(DenseBlock( in_channels=in_channels, out_channels=out_channels, use_bias=True, use_bn=False, activation=activation)) in_channels = out_channels self.att_conv_branch = nn.HybridSequential(prefix="") with self.att_conv_branch.name_scope(): for i, out_channels, in enumerate(att_conv_channels): self.att_conv_branch.add(ConvBlock1d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=1, padding=1, use_bias=True, use_bn=False, activation=nn.LeakyReLU(alpha=att_conv_slopes))) in_channels = out_channels self.att_fc = DenseBlock( in_channels=seq_len, out_channels=seq_len, use_bias=True, use_bn=False, activation=Softmax(axis=1)) def hybrid_forward(self, F, x): x = x.reshape((-3, 1, self.audio_window_size, self.audio_features)) x = x.swapaxes(1, 3) x = self.conv_branch(x) x = x.reshape((0, 1, -1)) x = self.fc_branch(x) x = x.reshape((-4, -1, self.seq_len, 0)) x = x.swapaxes(1, 2) y = x.slice_axis(axis=-1, begin=(self.seq_len // 2), end=(self.seq_len // 2) + 1).squeeze(axis=-1) w = self.att_conv_branch(x) w = w.reshape((0, -1)) w = self.att_fc(w) w = w.expand_dims(axis=-1) x = F.batch_dot(x, w) x = x.squeeze(axis=-1) return x, y class NvpAttExp(HybridBlock): """ Neural Voice Puppetry Audio-to-Expression model from 'Neural Voice Puppetry: Audio-driven Facial Reenactment,' https://arxiv.org/abs/1912.05566. Parameters: ---------- audio_features : int, default 29 Number of audio features (characters/sounds). audio_window_size : int, default 16 Size of audio window (for time related audio features). seq_len : int, default 8 Size of feature window. base_persons : int, default 116 Number of base persons (identities). blendshapes : int, default 76 Number of 3D model blendshapes. encoder_features : int, default 32 Number of encoder features. """ def __init__(self, audio_features=29, audio_window_size=16, seq_len=8, base_persons=116, blendshapes=76, encoder_features=32, **kwargs): super(NvpAttExp, self).__init__(**kwargs) self.base_persons = base_persons with self.name_scope(): self.encoder = NvpAttExpEncoder( audio_features=audio_features, audio_window_size=audio_window_size, seq_len=seq_len, encoder_features=encoder_features) self.decoder = SelectableDense( in_channels=encoder_features, out_channels=blendshapes, use_bias=False, num_options=base_persons) def hybrid_forward(self, F, x, pid): x, y = self.encoder(x) x = self.decoder(x, pid) y = self.decoder(y, pid) return x, y def get_nvpattexp(base_persons, blendshapes, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Neural Voice Puppetry Audio-to-Expression model with specific parameters. Parameters: ---------- base_persons : int Number of base persons (subjects). blendshapes : int Number of 3D model blendshapes. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ net = NvpAttExp( base_persons=base_persons, blendshapes=blendshapes, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def nvpattexp116bazel76(**kwargs): """ Neural Voice Puppetry Audio-to-Expression model for 116 base persons and Bazel topology with 76 blendshapes from 'Neural Voice Puppetry: Audio-driven Facial Reenactment,' https://arxiv.org/abs/1912.05566. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_nvpattexp(base_persons=116, blendshapes=76, model_name="nvpattexp116bazel76", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ nvpattexp116bazel76, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != nvpattexp116bazel76 or weight_count == 327397) batch = 14 seq_len = 8 audio_window_size = 16 audio_features = 29 blendshapes = 76 x = mx.nd.random.normal(shape=(batch, seq_len, audio_window_size, audio_features), ctx=ctx) pid = mx.nd.array(np.full(shape=(batch,), fill_value=3), ctx=ctx) y1, y2 = net(x, pid) assert (y1.shape == y2.shape == (batch, blendshapes)) if __name__ == "__main__": _test()
9,294
33.682836
116
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/octresnet.py
""" Oct-ResNet for ImageNet-1K, implemented in Gluon. Original paper: 'Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution,' https://arxiv.org/abs/1904.05049. """ __all__ = ['OctResNet', 'octresnet10_ad2', 'octresnet50b_ad2', 'OctResUnit'] import os from inspect import isfunction from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import ReLU6, DualPathSequential from .resnet import ResInitBlock class OctConv(nn.Conv2D): """ Octave convolution layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. oct_value : int, default 2 Octave value. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding=1, dilation=1, groups=1, use_bias=False, oct_alpha=0.0, oct_mode="std", oct_value=2, **kwargs): if isinstance(strides, int): strides = (strides, strides) self.downsample = (strides[0] > 1) or (strides[1] > 1) assert (strides[0] in [1, oct_value]) and (strides[1] in [1, oct_value]) strides = (1, 1) if oct_mode == "first": in_alpha = 0.0 out_alpha = oct_alpha elif oct_mode == "norm": in_alpha = oct_alpha out_alpha = oct_alpha elif oct_mode == "last": in_alpha = oct_alpha out_alpha = 0.0 elif oct_mode == "std": in_alpha = 0.0 out_alpha = 0.0 else: raise ValueError("Unsupported octave convolution mode: {}".format(oct_mode)) self.h_in_channels = int(in_channels * (1.0 - in_alpha)) self.h_out_channels = int(out_channels * (1.0 - out_alpha)) self.l_out_channels = out_channels - self.h_out_channels self.oct_alpha = oct_alpha self.oct_mode = oct_mode self.oct_value = oct_value super(OctConv, self).__init__( in_channels=in_channels, channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, **kwargs) self.conv_kwargs = self._kwargs.copy() del self.conv_kwargs["num_filter"] def hybrid_forward(self, F, hx, lx=None, weight=None, bias=None): if self.oct_mode == "std": return super(OctConv, self).hybrid_forward(F, hx, weight=weight, bias=bias), None if self.downsample: hx = F.Pooling( hx, kernel=(self.oct_value, self.oct_value), stride=(self.oct_value, self.oct_value), pool_type="avg") hhy = F.Convolution( hx, weight=weight.slice(begin=(None, None), end=(self.h_out_channels, self.h_in_channels)), bias=bias.slice(begin=(None,), end=(self.h_out_channels,)) if bias is not None else None, num_filter=self.h_out_channels, **self.conv_kwargs) if self.oct_mode != "first": hlx = F.Convolution( lx, weight=weight.slice(begin=(None, self.h_in_channels), end=(self.h_out_channels, None)), bias=bias.slice(begin=(None,), end=(self.h_out_channels,)) if bias is not None else None, num_filter=self.h_out_channels, **self.conv_kwargs) if self.oct_mode == "last": hy = hhy + hlx ly = None return hy, ly lhx = F.Pooling( hx, kernel=(self.oct_value, self.oct_value), stride=(self.oct_value, self.oct_value), pool_type="avg") lhy = F.Convolution( lhx, weight=weight.slice(begin=(self.h_out_channels, None), end=(None, self.h_in_channels)), bias=bias.slice(begin=(self.h_out_channels,), end=(None,)) if bias is not None else None, num_filter=self.l_out_channels, **self.conv_kwargs) if self.oct_mode == "first": hy = hhy ly = lhy return hy, ly if self.downsample: hly = hlx llx = F.Pooling( lx, kernel=(self.oct_value, self.oct_value), stride=(self.oct_value, self.oct_value), pool_type="avg") else: hly = F.UpSampling(hlx, scale=self.oct_value, sample_type="nearest") llx = lx lly = F.Convolution( llx, weight=weight.slice(begin=(self.h_out_channels, self.h_in_channels), end=(None, None)), bias=bias.slice(begin=(self.h_out_channels,), end=(None,)) if bias is not None else None, num_filter=self.l_out_channels, **self.conv_kwargs) hy = hhy + hly ly = lhy + lly return hy, ly def __repr__(self): s = '{name}({mapping}, kernel_size={kernel}, stride={stride}' len_kernel_size = len(self._kwargs['kernel']) if self._kwargs['pad'] != (0,) * len_kernel_size: s += ', padding={pad}' if self._kwargs['dilate'] != (1,) * len_kernel_size: s += ', dilation={dilate}' if hasattr(self, 'out_pad') and self.out_pad != (0,) * len_kernel_size: s += ', output_padding={out_pad}'.format(out_pad=self.out_pad) if self._kwargs['num_group'] != 1: s += ', groups={num_group}' if self.bias is None: s += ', bias=False' if self.act: s += ', {}'.format(self.act) s += ', oct_alpha={}'.format(self.oct_alpha) s += ', oct_mode={}'.format(self.oct_mode) s += ')' shape = self.weight.shape return s.format(name=self.__class__.__name__, mapping='{0} -> {1}'.format(shape[1] if shape[1] else None, shape[0]), **self._kwargs) class OctConvBlock(HybridBlock): """ Octave convolution block with Batch normalization and ReLU/ReLU6 activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activation : function or str or None, default nn.Activation("relu") Activation function or name of activation function. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, dilation=1, groups=1, use_bias=False, oct_alpha=0.0, oct_mode="std", bn_epsilon=1e-5, bn_use_global_stats=False, activation=(lambda: nn.Activation("relu")), activate=True, **kwargs): super(OctConvBlock, self).__init__(**kwargs) self.activate = activate self.last = (oct_mode == "last") or (oct_mode == "std") out_alpha = 0.0 if self.last else oct_alpha h_out_channels = int(out_channels * (1.0 - out_alpha)) l_out_channels = out_channels - h_out_channels with self.name_scope(): self.conv = OctConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, oct_alpha=oct_alpha, oct_mode=oct_mode) self.h_bn = nn.BatchNorm( in_channels=h_out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats) if not self.last: self.l_bn = nn.BatchNorm( in_channels=l_out_channels, epsilon=bn_epsilon, use_global_stats=bn_use_global_stats) if self.activate: assert (activation is not None) if isfunction(activation): self.activ = activation() elif isinstance(activation, str): if activation == "relu6": self.activ = ReLU6() else: self.activ = nn.Activation(activation) else: self.activ = activation def hybrid_forward(self, F, hx, lx=None): hx, lx = self.conv(hx, lx) hx = self.h_bn(hx) if self.activate: hx = self.activ(hx) if not self.last: lx = self.l_bn(lx) if self.activate: lx = self.activ(lx) return hx, lx def oct_conv1x1_block(in_channels, out_channels, strides=1, groups=1, use_bias=False, oct_alpha=0.0, oct_mode="std", bn_epsilon=1e-5, bn_use_global_stats=False, activation=(lambda: nn.Activation("relu")), activate=True, **kwargs): """ 1x1 version of the octave convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activation : function or str or None, default nn.Activation("relu") Activation function or name of activation function. activate : bool, default True Whether activate the convolution block. """ return OctConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, groups=groups, use_bias=use_bias, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, activation=activation, activate=activate, **kwargs) def oct_conv3x3_block(in_channels, out_channels, strides=1, padding=1, dilation=1, groups=1, use_bias=False, oct_alpha=0.0, oct_mode="std", bn_epsilon=1e-5, bn_use_global_stats=False, activation=(lambda: nn.Activation("relu")), activate=True, **kwargs): """ 3x3 version of the octave convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. use_bias : bool, default False Whether the layer uses a bias vector. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. activation : function or str or None, default nn.Activation("relu") Activation function or name of activation function. activate : bool, default True Whether activate the convolution block. """ return OctConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=padding, dilation=dilation, groups=groups, use_bias=use_bias, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats, activation=activation, activate=activate, **kwargs) class OctResBlock(HybridBlock): """ Simple Oct-ResNet block for residual path in Oct-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, strides, oct_alpha=0.0, oct_mode="std", bn_use_global_stats=False, **kwargs): super(OctResBlock, self).__init__(**kwargs) with self.name_scope(): self.conv1 = oct_conv3x3_block( in_channels=in_channels, out_channels=out_channels, strides=strides, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_use_global_stats=bn_use_global_stats) self.conv2 = oct_conv3x3_block( in_channels=out_channels, out_channels=out_channels, oct_alpha=oct_alpha, oct_mode=("std" if oct_mode == "last" else (oct_mode if oct_mode != "first" else "norm")), bn_use_global_stats=bn_use_global_stats, activation=None, activate=False) def hybrid_forward(self, F, hx, lx=None): hx, lx = self.conv1(hx, lx) hx, lx = self.conv2(hx, lx) return hx, lx class OctResBottleneck(HybridBlock): """ Oct-ResNet bottleneck block for residual path in Oct-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for the second convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for the second convolution layer. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. conv1_stride : bool, default False Whether to use stride in the first or the second convolution layer of the block. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, strides, padding=1, dilation=1, oct_alpha=0.0, oct_mode="std", bn_use_global_stats=False, conv1_stride=False, bottleneck_factor=4, **kwargs): super(OctResBottleneck, self).__init__(**kwargs) mid_channels = out_channels // bottleneck_factor with self.name_scope(): self.conv1 = oct_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, strides=(strides if conv1_stride else 1), oct_alpha=oct_alpha, oct_mode=(oct_mode if oct_mode != "last" else "norm"), bn_use_global_stats=bn_use_global_stats) self.conv2 = oct_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, strides=(1 if conv1_stride else strides), padding=padding, dilation=dilation, oct_alpha=oct_alpha, oct_mode=(oct_mode if oct_mode != "first" else "norm"), bn_use_global_stats=bn_use_global_stats) self.conv3 = oct_conv1x1_block( in_channels=mid_channels, out_channels=out_channels, oct_alpha=oct_alpha, oct_mode=("std" if oct_mode == "last" else (oct_mode if oct_mode != "first" else "norm")), bn_use_global_stats=bn_use_global_stats, activation=None, activate=False) def hybrid_forward(self, F, hx, lx=None): hx, lx = self.conv1(hx, lx) hx, lx = self.conv2(hx, lx) hx, lx = self.conv3(hx, lx) return hx, lx class OctResUnit(HybridBlock): """ Oct-ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int, default 1 Padding value for the second convolution layer in bottleneck. dilation : int or tuple/list of 2 int, default 1 Dilation value for the second convolution layer in bottleneck. oct_alpha : float, default 0.0 Octave alpha coefficient. oct_mode : str, default 'std' Octave convolution mode. It can be 'first', 'norm', 'last', or 'std'. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. bottleneck : bool, default True Whether to use a bottleneck or simple block in units. conv1_stride : bool, default False Whether to use stride in the first or the second convolution layer of the block. """ def __init__(self, in_channels, out_channels, strides, padding=1, dilation=1, oct_alpha=0.0, oct_mode="std", bn_use_global_stats=False, bottleneck=True, conv1_stride=False, **kwargs): super(OctResUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) or\ ((oct_mode == "first") and (oct_alpha != 0.0)) with self.name_scope(): if bottleneck: self.body = OctResBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, padding=padding, dilation=dilation, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_use_global_stats=bn_use_global_stats, conv1_stride=conv1_stride) else: self.body = OctResBlock( in_channels=in_channels, out_channels=out_channels, strides=strides, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_use_global_stats=bn_use_global_stats) if self.resize_identity: self.identity_conv = oct_conv1x1_block( in_channels=in_channels, out_channels=out_channels, strides=strides, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_use_global_stats=bn_use_global_stats, activation=None, activate=False) self.activ = nn.Activation("relu") def hybrid_forward(self, F, hx, lx=None): if self.resize_identity: h_identity, l_identity = self.identity_conv(hx, lx) else: h_identity, l_identity = hx, lx hx, lx = self.body(hx, lx) hx = hx + h_identity hx = self.activ(hx) if lx is not None: lx = lx + l_identity lx = self.activ(lx) return hx, lx class OctResNet(HybridBlock): """ Oct-ResNet model from 'Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution,' https://arxiv.org/abs/1904.05049. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. bottleneck : bool Whether to use a bottleneck or simple block in units. conv1_stride : bool Whether to use stride in the first or the second convolution layer in units. oct_alpha : float, default 0.5 Octave alpha coefficient. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, oct_alpha=0.5, bn_use_global_stats=False, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(OctResNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = DualPathSequential( return_two=False, first_ordinals=1, last_ordinals=1, prefix="") self.features.add(ResInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 if (i == 0) and (j == 0): oct_mode = "first" elif (i == len(channels) - 1) and (j == 0): oct_mode = "last" elif (i == len(channels) - 1) and (j != 0): oct_mode = "std" else: oct_mode = "norm" stage.add(OctResUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, oct_alpha=oct_alpha, oct_mode=oct_mode, bn_use_global_stats=bn_use_global_stats, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_octresnet(blocks, bottleneck=None, conv1_stride=True, oct_alpha=0.5, width_scale=1.0, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create Oct-ResNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bottleneck : bool, default None Whether to use a bottleneck or simple block in units. conv1_stride : bool, default True Whether to use stride in the first or the second convolution layer in units. oct_alpha : float, default 0.5 Octave alpha coefficient. width_scale : float, default 1.0 Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if bottleneck is None: bottleneck = (blocks >= 50) if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14 and not bottleneck: layers = [2, 2, 1, 1] elif (blocks == 14) and bottleneck: layers = [1, 1, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif blocks == 18: layers = [2, 2, 2, 2] elif (blocks == 26) and not bottleneck: layers = [3, 3, 3, 3] elif (blocks == 26) and bottleneck: layers = [2, 2, 2, 2] elif blocks == 34: layers = [3, 4, 6, 3] elif blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] elif blocks == 269: layers = [3, 30, 48, 8] else: raise ValueError("Unsupported Oct-ResNet with number of blocks: {}".format(blocks)) if bottleneck: assert (sum(layers) * 3 + 2 == blocks) else: assert (sum(layers) * 2 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] if bottleneck: bottleneck_factor = 4 channels_per_layers = [ci * bottleneck_factor for ci in channels_per_layers] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1.0: channels = [[int(cij * width_scale) if (i != len(channels) - 1) or (j != len(ci) - 1) else cij for j, cij in enumerate(ci)] for i, ci in enumerate(channels)] init_block_channels = int(init_block_channels * width_scale) net = OctResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, conv1_stride=conv1_stride, oct_alpha=oct_alpha, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def octresnet10_ad2(**kwargs): """ Oct-ResNet-10 (alpha=1/2) model from 'Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution,' https://arxiv.org/abs/1904.05049. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_octresnet(blocks=10, oct_alpha=0.5, model_name="octresnet10_ad2", **kwargs) def octresnet50b_ad2(**kwargs): """ Oct-ResNet-50b (alpha=1/2) model from 'Drop an Octave: Reducing Spatial Redundancy in Convolutional Neural Networks with Octave Convolution,' https://arxiv.org/abs/1904.05049. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_octresnet(blocks=50, conv1_stride=False, oct_alpha=0.5, model_name="octresnet50b_ad2", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ octresnet10_ad2, octresnet50b_ad2, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) # net.hybridize() net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != octresnet10_ad2 or weight_count == 5423016) assert (model != octresnet50b_ad2 or weight_count == 25557032) x = mx.nd.zeros((14, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (14, 1000)) if __name__ == "__main__": _test()
32,656
35.245283
119
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/alexnet.py
""" AlexNet for ImageNet-1K, implemented in Gluon. Original paper: 'One weird trick for parallelizing convolutional neural networks,' https://arxiv.org/abs/1404.5997. """ __all__ = ['AlexNet', 'alexnet', 'alexnetb'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from .common import ConvBlock class AlexConv(ConvBlock): """ AlexNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. use_lrn : bool Whether to use LRN layer. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, use_lrn, **kwargs): super(AlexConv, self).__init__( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, use_bias=True, use_bn=False, **kwargs) self.use_lrn = use_lrn def hybrid_forward(self, F, x): x = super(AlexConv, self).hybrid_forward(F, x) if self.use_lrn: x = F.LRN(x, nsize=5) return x class AlexDense(HybridBlock): """ AlexNet specific dense block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(AlexDense, self).__init__(**kwargs) with self.name_scope(): self.fc = nn.Dense( units=out_channels, in_units=in_channels) self.activ = nn.Activation("relu") self.dropout = nn.Dropout(rate=0.5) def hybrid_forward(self, F, x): x = self.fc(x) x = self.activ(x) x = self.dropout(x) return x class AlexOutputBlock(HybridBlock): """ AlexNet specific output block. Parameters: ---------- in_channels : int Number of input channels. classes : int Number of classification classes. """ def __init__(self, in_channels, classes, **kwargs): super(AlexOutputBlock, self).__init__(**kwargs) mid_channels = 4096 with self.name_scope(): self.fc1 = AlexDense( in_channels=in_channels, out_channels=mid_channels) self.fc2 = AlexDense( in_channels=mid_channels, out_channels=mid_channels) self.fc3 = nn.Dense( units=classes, in_units=mid_channels) def hybrid_forward(self, F, x): x = self.fc1(x) x = self.fc2(x) x = self.fc3(x) return x class AlexNet(HybridBlock): """ AlexNet model from 'One weird trick for parallelizing convolutional neural networks,' https://arxiv.org/abs/1404.5997. Parameters: ---------- channels : list of list of int Number of output channels for each unit. kernel_sizes : list of list of int Convolution window sizes for each unit. strides : list of list of int or tuple/list of 2 int Strides of the convolution for each unit. paddings : list of list of int or tuple/list of 2 int Padding value for convolution layer for each unit. use_lrn : bool Whether to use LRN layer. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, kernel_sizes, strides, paddings, use_lrn, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(AlexNet, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") for i, channels_per_stage in enumerate(channels): use_lrn_i = use_lrn and (i in [0, 1]) stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): stage.add(AlexConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_sizes[i][j], strides=strides[i][j], padding=paddings[i][j], use_lrn=use_lrn_i)) in_channels = out_channels stage.add(nn.MaxPool2D( pool_size=3, strides=2, padding=0, ceil_mode=True)) self.features.add(stage) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) in_channels = in_channels * 6 * 6 self.output.add(AlexOutputBlock( in_channels=in_channels, classes=classes)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_alexnet(version="a", model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create AlexNet model with specific parameters. Parameters: ---------- version : str, default 'a' Version of AlexNet ('a' or 'b'). model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if version == "a": channels = [[96], [256], [384, 384, 256]] kernel_sizes = [[11], [5], [3, 3, 3]] strides = [[4], [1], [1, 1, 1]] paddings = [[0], [2], [1, 1, 1]] use_lrn = True elif version == "b": channels = [[64], [192], [384, 256, 256]] kernel_sizes = [[11], [5], [3, 3, 3]] strides = [[4], [1], [1, 1, 1]] paddings = [[2], [2], [1, 1, 1]] use_lrn = False else: raise ValueError("Unsupported AlexNet version {}".format(version)) net = AlexNet( channels=channels, kernel_sizes=kernel_sizes, strides=strides, paddings=paddings, use_lrn=use_lrn, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def alexnet(**kwargs): """ AlexNet model from 'One weird trick for parallelizing convolutional neural networks,' https://arxiv.org/abs/1404.5997. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_alexnet(model_name="alexnet", **kwargs) def alexnetb(**kwargs): """ AlexNet-b model from 'One weird trick for parallelizing convolutional neural networks,' https://arxiv.org/abs/1404.5997. Non-standard version. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_alexnet(version="b", model_name="alexnetb", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ alexnet, alexnetb, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != alexnet or weight_count == 62378344) assert (model != alexnetb or weight_count == 61100840) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
9,854
29.137615
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/mobilenet_cub.py
""" MobileNet & FD-MobileNet for CUB-200-2011, implemented in Gluon. Original papers: - 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. - 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. """ __all__ = ['mobilenet_w1_cub', 'mobilenet_w3d4_cub', 'mobilenet_wd2_cub', 'mobilenet_wd4_cub', 'fdmobilenet_w1_cub', 'fdmobilenet_w3d4_cub', 'fdmobilenet_wd2_cub', 'fdmobilenet_wd4_cub'] from .mobilenet import get_mobilenet from .fdmobilenet import get_fdmobilenet def mobilenet_w1_cub(classes=200, **kwargs): """ 1.0 MobileNet-224 model for CUB-200-2011 from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(classes=classes, width_scale=1.0, model_name="mobilenet_w1_cub", **kwargs) def mobilenet_w3d4_cub(classes=200, **kwargs): """ 0.75 MobileNet-224 model for CUB-200-2011 from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(classes=classes, width_scale=0.75, model_name="mobilenet_w3d4_cub", **kwargs) def mobilenet_wd2_cub(classes=200, **kwargs): """ 0.5 MobileNet-224 model for CUB-200-2011 from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(classes=classes, width_scale=0.5, model_name="mobilenet_wd2_cub", **kwargs) def mobilenet_wd4_cub(classes=200, **kwargs): """ 0.25 MobileNet-224 model for CUB-200-2011 from 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_mobilenet(classes=classes, width_scale=0.25, model_name="mobilenet_wd4_cub", **kwargs) def fdmobilenet_w1_cub(classes=200, **kwargs): """ FD-MobileNet 1.0x model for CUB-200-2011 from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(classes=classes, width_scale=1.0, model_name="fdmobilenet_w1_cub", **kwargs) def fdmobilenet_w3d4_cub(classes=200, **kwargs): """ FD-MobileNet 0.75x model for CUB-200-2011 from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(classes=classes, width_scale=0.75, model_name="fdmobilenet_w3d4_cub", **kwargs) def fdmobilenet_wd2_cub(classes=200, **kwargs): """ FD-MobileNet 0.5x model for CUB-200-2011 from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(classes=classes, width_scale=0.5, model_name="fdmobilenet_wd2_cub", **kwargs) def fdmobilenet_wd4_cub(classes=200, **kwargs): """ FD-MobileNet 0.25x model for CUB-200-2011 from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(classes=classes, width_scale=0.25, model_name="fdmobilenet_wd4_cub", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ mobilenet_w1_cub, mobilenet_w3d4_cub, mobilenet_wd2_cub, mobilenet_wd4_cub, fdmobilenet_w1_cub, fdmobilenet_w3d4_cub, fdmobilenet_wd2_cub, fdmobilenet_wd4_cub, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mobilenet_w1_cub or weight_count == 3411976) assert (model != mobilenet_w3d4_cub or weight_count == 1970360) assert (model != mobilenet_wd2_cub or weight_count == 921192) assert (model != mobilenet_wd4_cub or weight_count == 264472) assert (model != fdmobilenet_w1_cub or weight_count == 2081288) assert (model != fdmobilenet_w3d4_cub or weight_count == 1218104) assert (model != fdmobilenet_wd2_cub or weight_count == 583528) assert (model != fdmobilenet_wd4_cub or weight_count == 177560) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 200)) if __name__ == "__main__": _test()
7,904
35.597222
120
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/wrn.py
""" WRN for ImageNet-1K, implemented in Gluon. Original paper: 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. """ __all__ = ['WRN', 'wrn50_2'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock class WRNConv(HybridBlock): """ WRN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of 2 int Convolution window size. strides : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. activate : bool Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, strides, padding, activate, **kwargs): super(WRNConv, self).__init__(**kwargs) self.activate = activate with self.name_scope(): self.conv = nn.Conv2D( channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, use_bias=True, in_channels=in_channels) if self.activate: self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): x = self.conv(x) if self.activate: x = self.activ(x) return x def wrn_conv1x1(in_channels, out_channels, strides, activate): """ 1x1 version of the WRN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. activate : bool Whether activate the convolution block. """ return WRNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, strides=strides, padding=0, activate=activate) def wrn_conv3x3(in_channels, out_channels, strides, activate): """ 3x3 version of the WRN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. activate : bool Whether activate the convolution block. """ return WRNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, strides=strides, padding=1, activate=activate) class WRNBottleneck(HybridBlock): """ WRN bottleneck block for residual path in WRN unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. width_factor : float Wide scale factor for width of layers. """ def __init__(self, in_channels, out_channels, strides, width_factor, **kwargs): super(WRNBottleneck, self).__init__(**kwargs) mid_channels = int(round(out_channels // 4 * width_factor)) with self.name_scope(): self.conv1 = wrn_conv1x1( in_channels=in_channels, out_channels=mid_channels, strides=1, activate=True) self.conv2 = wrn_conv3x3( in_channels=mid_channels, out_channels=mid_channels, strides=strides, activate=True) self.conv3 = wrn_conv1x1( in_channels=mid_channels, out_channels=out_channels, strides=1, activate=False) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class WRNUnit(HybridBlock): """ WRN unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. strides : int or tuple/list of 2 int Strides of the convolution. width_factor : float Wide scale factor for width of layers. """ def __init__(self, in_channels, out_channels, strides, width_factor, **kwargs): super(WRNUnit, self).__init__(**kwargs) self.resize_identity = (in_channels != out_channels) or (strides != 1) with self.name_scope(): self.body = WRNBottleneck( in_channels=in_channels, out_channels=out_channels, strides=strides, width_factor=width_factor) if self.resize_identity: self.identity_conv = wrn_conv1x1( in_channels=in_channels, out_channels=out_channels, strides=strides, activate=False) self.activ = nn.Activation("relu") def hybrid_forward(self, F, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity x = self.activ(x) return x class WRNInitBlock(HybridBlock): """ WRN specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels, **kwargs): super(WRNInitBlock, self).__init__(**kwargs) with self.name_scope(): self.conv = WRNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=7, strides=2, padding=3, activate=True) self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=1) def hybrid_forward(self, F, x): x = self.conv(x) x = self.pool(x) return x class WRN(HybridBlock): """ WRN model from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. width_factor : float Wide scale factor for width of layers. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, width_factor, in_channels=3, in_size=(224, 224), classes=1000, **kwargs): super(WRN, self).__init__(**kwargs) self.in_size = in_size self.classes = classes with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(WRNInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) and (i != 0) else 1 stage.add(WRNUnit( in_channels=in_channels, out_channels=out_channels, strides=strides, width_factor=width_factor)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=7, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_wrn(blocks, width_factor, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create WRN model with specific parameters. Parameters: ---------- blocks : int Number of blocks. width_factor : float Wide scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] else: raise ValueError("Unsupported WRN with number of blocks: {}".format(blocks)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = WRN( channels=channels, init_block_channels=init_block_channels, width_factor=width_factor, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def wrn50_2(**kwargs): """ WRN-50-2 model from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_wrn(blocks=50, width_factor=2.0, model_name="wrn50_2", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ wrn50_2, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != wrn50_2 or weight_count == 68849128) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
12,149
27.723404
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/inceptionv3.py
""" InceptionV3 for ImageNet-1K, implemented in Gluon. Original paper: 'Rethinking the Inception Architecture for Computer Vision,' https://arxiv.org/abs/1512.00567. """ __all__ = ['InceptionV3', 'inceptionv3', 'inceptionv3_gl', 'MaxPoolBranch', 'AvgPoolBranch', 'Conv1x1Branch', 'ConvSeqBranch'] import os from mxnet import cpu from mxnet.gluon import nn, HybridBlock from mxnet.gluon.contrib.nn import HybridConcurrent from .common import ConvBlock, conv1x1_block, conv3x3_block class MaxPoolBranch(HybridBlock): """ Inception specific max pooling branch block. """ def __init__(self, **kwargs): super(MaxPoolBranch, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.MaxPool2D( pool_size=3, strides=2, padding=0) def hybrid_forward(self, F, x): x = self.pool(x) return x class AvgPoolBranch(HybridBlock): """ Inception specific average pooling branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. count_include_pad : bool, default True Whether to include the zero-padding in the averaging calculation. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, count_include_pad=True, **kwargs): super(AvgPoolBranch, self).__init__(**kwargs) with self.name_scope(): self.pool = nn.AvgPool2D( pool_size=3, strides=1, padding=1, count_include_pad=count_include_pad) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.pool(x) x = self.conv(x) return x class Conv1x1Branch(HybridBlock): """ Inception specific convolutional 1x1 branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(Conv1x1Branch, self).__init__(**kwargs) with self.name_scope(): self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv(x) return x class ConvSeqBranch(HybridBlock): """ Inception specific convolutional sequence branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of tuple of int List of numbers of output channels. kernel_size_list : list of tuple of int or tuple of tuple/list of 2 int List of convolution window sizes. strides_list : list of tuple of int or tuple of tuple/list of 2 int List of strides of the convolution. padding_list : list of tuple of int or tuple of tuple/list of 2 int List of padding values for convolution layers. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels_list, kernel_size_list, strides_list, padding_list, bn_epsilon, bn_use_global_stats, **kwargs): super(ConvSeqBranch, self).__init__(**kwargs) assert (len(out_channels_list) == len(kernel_size_list)) assert (len(out_channels_list) == len(strides_list)) assert (len(out_channels_list) == len(padding_list)) with self.name_scope(): self.conv_list = nn.HybridSequential(prefix="") for out_channels, kernel_size, strides, padding in zip( out_channels_list, kernel_size_list, strides_list, padding_list): self.conv_list.add(ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels def hybrid_forward(self, F, x): x = self.conv_list(x) return x class ConvSeq3x3Branch(HybridBlock): """ InceptionV3 specific convolutional sequence branch block with splitting by 3x3. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of tuple of int List of numbers of output channels. kernel_size_list : list of tuple of int or tuple of tuple/list of 2 int List of convolution window sizes. strides_list : list of tuple of int or tuple of tuple/list of 2 int List of strides of the convolution. padding_list : list of tuple of int or tuple of tuple/list of 2 int List of padding values for convolution layers. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels_list, kernel_size_list, strides_list, padding_list, bn_epsilon, bn_use_global_stats, **kwargs): super(ConvSeq3x3Branch, self).__init__(**kwargs) with self.name_scope(): self.conv_list = nn.HybridSequential(prefix="") for out_channels, kernel_size, strides, padding in zip( out_channels_list, kernel_size_list, strides_list, padding_list): self.conv_list.add(ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, strides=strides, padding=padding, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.conv1x3 = ConvBlock( in_channels=in_channels, out_channels=in_channels, kernel_size=(1, 3), strides=1, padding=(0, 1), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.conv3x1 = ConvBlock( in_channels=in_channels, out_channels=in_channels, kernel_size=(3, 1), strides=1, padding=(1, 0), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) def hybrid_forward(self, F, x): x = self.conv_list(x) y1 = self.conv1x3(x) y2 = self.conv3x1(x) x = F.concat(y1, y2, dim=1) return x class InceptionAUnit(HybridBlock): """ InceptionV3 type Inception-A unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(InceptionAUnit, self).__init__(**kwargs) assert (out_channels > 224) pool_out_channels = out_channels - 224 with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(Conv1x1Branch( in_channels=in_channels, out_channels=64, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(48, 64), kernel_size_list=(1, 5), strides_list=(1, 1), padding_list=(0, 2), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(64, 96, 96), kernel_size_list=(1, 3, 3), strides_list=(1, 1, 1), padding_list=(0, 1, 1), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(AvgPoolBranch( in_channels=in_channels, out_channels=pool_out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class ReductionAUnit(HybridBlock): """ InceptionV3 type Reduction-A unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(ReductionAUnit, self).__init__(**kwargs) assert (in_channels == 288) assert (out_channels == 768) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(384,), kernel_size_list=(3,), strides_list=(2,), padding_list=(0,), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(64, 96, 96), kernel_size_list=(1, 3, 3), strides_list=(1, 1, 2), padding_list=(0, 1, 0), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(MaxPoolBranch()) def hybrid_forward(self, F, x): x = self.branches(x) return x class InceptionBUnit(HybridBlock): """ InceptionV3 type Inception-B unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of output channels in the 7x7 branches. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, mid_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(InceptionBUnit, self).__init__(**kwargs) assert (in_channels == 768) assert (out_channels == 768) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(Conv1x1Branch( in_channels=in_channels, out_channels=192, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(mid_channels, mid_channels, 192), kernel_size_list=(1, (1, 7), (7, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 3), (3, 0)), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(mid_channels, mid_channels, mid_channels, mid_channels, 192), kernel_size_list=(1, (7, 1), (1, 7), (7, 1), (1, 7)), strides_list=(1, 1, 1, 1, 1), padding_list=(0, (3, 0), (0, 3), (3, 0), (0, 3)), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(AvgPoolBranch( in_channels=in_channels, out_channels=192, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class ReductionBUnit(HybridBlock): """ InceptionV3 type Reduction-B unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(ReductionBUnit, self).__init__(**kwargs) assert (in_channels == 768) assert (out_channels == 1280) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 320), kernel_size_list=(1, 3), strides_list=(1, 2), padding_list=(0, 0), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 192, 192, 192), kernel_size_list=(1, (1, 7), (7, 1), 3), strides_list=(1, 1, 1, 2), padding_list=(0, (0, 3), (3, 0), 0), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(MaxPoolBranch()) def hybrid_forward(self, F, x): x = self.branches(x) return x class InceptionCUnit(HybridBlock): """ InceptionV3 type Inception-C unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(InceptionCUnit, self).__init__(**kwargs) assert (out_channels == 2048) with self.name_scope(): self.branches = HybridConcurrent(axis=1, prefix="") self.branches.add(Conv1x1Branch( in_channels=in_channels, out_channels=320, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeq3x3Branch( in_channels=in_channels, out_channels_list=(384,), kernel_size_list=(1,), strides_list=(1,), padding_list=(0,), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(ConvSeq3x3Branch( in_channels=in_channels, out_channels_list=(448, 384), kernel_size_list=(1, 3), strides_list=(1, 1), padding_list=(0, 1), bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) self.branches.add(AvgPoolBranch( in_channels=in_channels, out_channels=192, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) def hybrid_forward(self, F, x): x = self.branches(x) return x class InceptInitBlock(HybridBlock): """ InceptionV3 specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_epsilon : float Small float added to variance in Batch norm. bn_use_global_stats : bool Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. """ def __init__(self, in_channels, out_channels, bn_epsilon, bn_use_global_stats, **kwargs): super(InceptInitBlock, self).__init__(**kwargs) assert (out_channels == 192) with self.name_scope(): self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=32, strides=2, padding=0, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.conv2 = conv3x3_block( in_channels=32, out_channels=32, strides=1, padding=0, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.conv3 = conv3x3_block( in_channels=32, out_channels=64, strides=1, padding=1, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.pool1 = nn.MaxPool2D( pool_size=3, strides=2, padding=0) self.conv4 = conv1x1_block( in_channels=64, out_channels=80, strides=1, padding=0, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.conv5 = conv3x3_block( in_channels=80, out_channels=192, strides=1, padding=0, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats) self.pool2 = nn.MaxPool2D( pool_size=3, strides=2, padding=0) def hybrid_forward(self, F, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.pool1(x) x = self.conv4(x) x = self.conv5(x) x = self.pool2(x) return x class InceptionV3(HybridBlock): """ InceptionV3 model from 'Rethinking the Inception Architecture for Computer Vision,' https://arxiv.org/abs/1512.00567. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. b_mid_channels : list of int Number of middle channels for each Inception-B unit. dropout_rate : float, default 0.0 Fraction of the input units to drop. Must be a number between 0 and 1. bn_epsilon : float, default 1e-5 Small float added to variance in Batch norm. bn_use_global_stats : bool, default False Whether global moving statistics is used instead of local batch-norm for BatchNorm layers. Useful for fine-tuning. in_channels : int, default 3 Number of input channels. in_size : tuple of two ints, default (299, 299) Spatial size of the expected input image. classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, b_mid_channels, dropout_rate=0.5, bn_epsilon=1e-5, bn_use_global_stats=False, in_channels=3, in_size=(299, 299), classes=1000, **kwargs): super(InceptionV3, self).__init__(**kwargs) self.in_size = in_size self.classes = classes normal_units = [InceptionAUnit, InceptionBUnit, InceptionCUnit] reduction_units = [ReductionAUnit, ReductionBUnit] with self.name_scope(): self.features = nn.HybridSequential(prefix="") self.features.add(InceptInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.HybridSequential(prefix="stage{}_".format(i + 1)) with stage.name_scope(): for j, out_channels in enumerate(channels_per_stage): if (j == 0) and (i != 0): unit = reduction_units[i - 1] else: unit = normal_units[i] if unit == InceptionBUnit: stage.add(unit( in_channels=in_channels, out_channels=out_channels, mid_channels=b_mid_channels[j - 1], bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) else: stage.add(unit( in_channels=in_channels, out_channels=out_channels, bn_epsilon=bn_epsilon, bn_use_global_stats=bn_use_global_stats)) in_channels = out_channels self.features.add(stage) self.features.add(nn.AvgPool2D( pool_size=8, strides=1)) self.output = nn.HybridSequential(prefix="") self.output.add(nn.Flatten()) self.output.add(nn.Dropout(rate=dropout_rate)) self.output.add(nn.Dense( units=classes, in_units=in_channels)) def hybrid_forward(self, F, x): x = self.features(x) x = self.output(x) return x def get_inceptionv3(model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create InceptionV3 model with specific parameters. Parameters: ---------- model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ init_block_channels = 192 channels = [[256, 288, 288], [768, 768, 768, 768, 768], [1280, 2048, 2048]] b_mid_channels = [128, 160, 160, 192] net = InceptionV3( channels=channels, init_block_channels=init_block_channels, b_mid_channels=b_mid_channels, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def inceptionv3(**kwargs): """ InceptionV3 model from 'Rethinking the Inception Architecture for Computer Vision,' https://arxiv.org/abs/1512.00567. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_inceptionv3(model_name="inceptionv3", bn_epsilon=1e-3, **kwargs) def inceptionv3_gl(**kwargs): """ InceptionV3 model (Gluon-like) from 'Rethinking the Inception Architecture for Computer Vision,' https://arxiv.org/abs/1512.00567. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_inceptionv3(model_name="inceptionv3_gl", bn_epsilon=1e-5, **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ inceptionv3, inceptionv3_gl, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != inceptionv3 or weight_count == 23834568) assert (model != inceptionv3_gl or weight_count == 23834568) x = mx.nd.zeros((1, 3, 299, 299), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
27,784
33.644638
115
py
imgclsmob
imgclsmob-master/gluon/gluoncv2/models/fdmobilenet.py
""" FD-MobileNet for ImageNet-1K, implemented in Gluon. Original paper: 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. """ __all__ = ['fdmobilenet_w1', 'fdmobilenet_w3d4', 'fdmobilenet_wd2', 'fdmobilenet_wd4', 'get_fdmobilenet'] import os from mxnet import cpu from .mobilenet import MobileNet def get_fdmobilenet(width_scale, model_name=None, pretrained=False, ctx=cpu(), root=os.path.join("~", ".mxnet", "models"), **kwargs): """ Create FD-MobileNet model with specific parameters. Parameters: ---------- width_scale : float Scale factor for width of layers. model_name : str or None, default None Model name for loading pretrained model. pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ channels = [[32], [64], [128, 128], [256, 256], [512, 512, 512, 512, 512, 1024]] first_stage_stride = True if width_scale != 1.0: channels = [[int(cij * width_scale) for cij in ci] for ci in channels] net = MobileNet( channels=channels, first_stage_stride=first_stage_stride, **kwargs) if pretrained: if (model_name is None) or (not model_name): raise ValueError("Parameter `model_name` should be properly initialized for loading pretrained model.") from .model_store import get_model_file net.load_parameters( filename=get_model_file( model_name=model_name, local_model_store_dir_path=root), ctx=ctx) return net def fdmobilenet_w1(**kwargs): """ FD-MobileNet 1.0x model from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(width_scale=1.0, model_name="fdmobilenet_w1", **kwargs) def fdmobilenet_w3d4(**kwargs): """ FD-MobileNet 0.75x model from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(width_scale=0.75, model_name="fdmobilenet_w3d4", **kwargs) def fdmobilenet_wd2(**kwargs): """ FD-MobileNet 0.5x model from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(width_scale=0.5, model_name="fdmobilenet_wd2", **kwargs) def fdmobilenet_wd4(**kwargs): """ FD-MobileNet 0.25x model from 'FD-MobileNet: Improved MobileNet with A Fast Downsampling Strategy,' https://arxiv.org/abs/1802.03750. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. ctx : Context, default CPU The context in which to load the pretrained weights. root : str, default '~/.mxnet/models' Location for keeping the model parameters. """ return get_fdmobilenet(width_scale=0.25, model_name="fdmobilenet_wd4", **kwargs) def _test(): import numpy as np import mxnet as mx pretrained = False models = [ fdmobilenet_w1, fdmobilenet_w3d4, fdmobilenet_wd2, fdmobilenet_wd4, ] for model in models: net = model(pretrained=pretrained) ctx = mx.cpu() if not pretrained: net.initialize(ctx=ctx) net_params = net.collect_params() weight_count = 0 for param in net_params.values(): if (param.shape is None) or (not param._differentiable): continue weight_count += np.prod(param.shape) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fdmobilenet_w1 or weight_count == 2901288) assert (model != fdmobilenet_w3d4 or weight_count == 1833304) assert (model != fdmobilenet_wd2 or weight_count == 993928) assert (model != fdmobilenet_wd4 or weight_count == 383160) x = mx.nd.zeros((1, 3, 224, 224), ctx=ctx) y = net(x) assert (y.shape == (1, 1000)) if __name__ == "__main__": _test()
5,360
30.910714
115
py
imgclsmob
imgclsmob-master/gluon/metrics/seg_metrics_nd.py
""" Routines for segmentation metrics on mx.ndarray. """ import numpy as np import mxnet as mx __all__ = ['seg_pixel_accuracy_nd', 'segm_mean_accuracy', 'segm_mean_iou', 'seg_mean_iou2_nd', 'segm_fw_iou', 'segm_fw_iou2'] def seg_pixel_accuracy_nd(label_imask, pred_imask, vague_idx=-1, use_vague=False, macro_average=True, empty_result=0.0): """ The segmentation pixel accuracy (for MXNet nd-arrays). Parameters: ---------- label_imask : mx.nd.array Ground truth index mask (maybe batch of). pred_imask : mx.nd.array Predicted index mask (maybe batch of). vague_idx : int, default -1 Index of masked pixels. use_vague : bool, default False Whether to use pixel masking. macro_average : bool, default True Whether to use micro or macro averaging. empty_result : float, default 0.0 Result value for an image without any classes. Returns: ------- float or tuple of two floats PA metric value. """ assert (label_imask.shape == pred_imask.shape) if use_vague: mask = (label_imask != vague_idx) sum_u_ij = mask.sum().asscalar() if sum_u_ij == 0: if macro_average: return empty_result else: return 0, 0 sum_u_ii = ((label_imask == pred_imask) * mask).sum().asscalar() else: sum_u_ii = (label_imask == pred_imask).sum().asscalar() sum_u_ij = pred_imask.size if macro_average: return float(sum_u_ii) / sum_u_ij else: return sum_u_ii, sum_u_ij def segm_mean_accuracy(label_hmask, pred_imask): """ The segmentation mean accuracy. Parameters: ---------- label_hmask : nd.array Ground truth one-hot mask. pred_imask : nd.array Predicted index mask. Returns: ------- float MA metric value. """ assert (len(label_hmask.shape) == 3) assert (len(pred_imask.shape) == 2) assert (pred_imask.shape == label_hmask.shape[1:]) n = label_hmask.shape[0] i_sum = 0 acc_sum = 0.0 for i in range(n): class_i_pred_mask = (pred_imask == i) class_i_label_mask = label_hmask[i, :, :] u_i = class_i_label_mask.sum().asscalar() if u_i == 0: continue u_ii = (class_i_pred_mask * class_i_label_mask).sum().asscalar() class_acc = float(u_ii) / u_i acc_sum += class_acc i_sum += 1 if i_sum > 0: mean_acc = acc_sum / i_sum else: mean_acc = 1.0 return mean_acc def segm_mean_iou(label_hmask, pred_imask): """ The segmentation mean intersection over union. Parameters: ---------- label_hmask : nd.array Ground truth one-hot mask. pred_imask : nd.array Predicted index mask. Returns: ------- float MIoU metric value. """ assert (len(label_hmask.shape) == 3) assert (len(pred_imask.shape) == 2) assert (pred_imask.shape == label_hmask.shape[1:]) n = label_hmask.shape[0] i_sum = 0 acc_iou = 0.0 for i in range(n): class_i_pred_mask = (pred_imask == i) class_i_label_mask = label_hmask[i, :, :] u_i = class_i_label_mask.sum().asscalar() u_ji_sj = class_i_pred_mask.sum().asscalar() if (u_i + u_ji_sj) == 0: continue u_ii = (class_i_pred_mask * class_i_label_mask).sum().asscalar() acc_iou += float(u_ii) / (u_i + u_ji_sj - u_ii) i_sum += 1 if i_sum > 0: mean_iou = acc_iou / i_sum else: mean_iou = 1.0 return mean_iou def seg_mean_iou2_nd(label_hmask, pred_hmask): """ The segmentation mean intersection over union. Parameters: ---------- label_hmask : nd.array Ground truth one-hot mask (batch of). pred_hmask : nd.array Predicted one-hot mask (batch of). Returns: ------- float MIoU metric value. """ assert (len(label_hmask.shape) == 4) assert (len(pred_hmask.shape) == 4) assert (pred_hmask.shape == label_hmask.shape) eps = np.finfo(np.float32).eps batch_axis = 0 # The axis that represents mini-batch class_axis = 1 # The axis that represents classes inter_hmask = label_hmask * pred_hmask u_i = label_hmask.sum(axis=[batch_axis, class_axis], exclude=True) u_ji_sj = pred_hmask.sum(axis=[batch_axis, class_axis], exclude=True) u_ii = inter_hmask.sum(axis=[batch_axis, class_axis], exclude=True) class_count = (u_i + u_ji_sj > 0.0).sum(axis=class_axis) + eps class_acc = u_ii / (u_i + u_ji_sj - u_ii + eps) acc_iou = class_acc.sum(axis=class_axis) + eps mean_iou = (acc_iou / class_count).mean().asscalar() return mean_iou def segm_fw_iou(label_hmask, pred_imask): """ The segmentation frequency weighted intersection over union. Parameters: ---------- label_hmask : nd.array Ground truth one-hot mask. pred_imask : nd.array Predicted index mask. Returns: ------- float FrIoU metric value. """ assert (len(label_hmask.shape) == 3) assert (len(pred_imask.shape) == 2) assert (pred_imask.shape == label_hmask.shape[1:]) n = label_hmask.shape[0] acc_iou = 0.0 for i in range(n): class_i_pred_mask = (pred_imask == i) class_i_label_mask = label_hmask[i, :, :] u_i = class_i_label_mask.sum().asscalar() u_ji_sj = class_i_pred_mask.sum().asscalar() if (u_i + u_ji_sj) == 0: continue u_ii = (class_i_pred_mask * class_i_label_mask).sum().asscalar() acc_iou += float(u_i) * float(u_ii) / (u_i + u_ji_sj - u_ii) fw_factor = pred_imask.size return acc_iou / fw_factor def segm_fw_iou2(label_hmask, pred_imask): """ The segmentation frequency weighted intersection over union. Parameters: ---------- label_hmask : nd.array Ground truth one-hot mask. pred_imask : nd.array Predicted index mask. Returns: ------- float FrIoU metric value. """ assert (len(label_hmask.shape) == 3) assert (len(pred_imask.shape) == 2) assert (pred_imask.shape == label_hmask.shape[1:]) n = label_hmask.shape[0] acc_iou = mx.nd.array([0.0], ctx=label_hmask.context) for i in range(n): class_i_pred_mask = (pred_imask == i) class_i_label_mask = label_hmask[i, :, :] u_i = class_i_label_mask.sum() u_ji_sj = class_i_pred_mask.sum() if (u_i + u_ji_sj).asscalar() == 0: continue u_ii = (class_i_pred_mask * class_i_label_mask).sum() acc_iou += mx.nd.cast(u_i, dtype=np.float32) *\ mx.nd.cast(u_ii, dtype=np.float32) / mx.nd.cast(u_i + u_ji_sj - u_ii, dtype=np.float32) fw_factor = pred_imask.size return acc_iou.asscalar() / fw_factor
7,161
25.924812
109
py
imgclsmob
imgclsmob-master/gluon/metrics/seg_metrics.py
""" Evaluation Metrics for Semantic Segmentation. """ __all__ = ['PixelAccuracyMetric', 'MeanIoUMetric'] import numpy as np import mxnet as mx from .seg_metrics_np import seg_pixel_accuracy_np, seg_mean_iou_imasks_np from .seg_metrics_nd import seg_pixel_accuracy_nd class PixelAccuracyMetric(mx.metric.EvalMetric): """ Computes the pixel-wise accuracy. Parameters: ---------- axis : int, default 1 The axis that represents classes. name : str, default 'pix_acc' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. on_cpu : bool, default True Calculate on CPU. sparse_label : bool, default True Whether label is an integer array instead of probability distribution. vague_idx : int, default -1 Index of masked pixels. use_vague : bool, default False Whether to use pixel masking. macro_average : bool, default True Whether to use micro or macro averaging. aux : bool, default False Whether to support auxiliary predictions. """ def __init__(self, axis=1, name="pix_acc", output_names=None, label_names=None, on_cpu=True, sparse_label=True, vague_idx=-1, use_vague=False, macro_average=True, aux=False): if name == "pix_acc": name = "{}-pix_acc".format("macro" if macro_average else "micro") self.macro_average = macro_average super(PixelAccuracyMetric, self).__init__( name, axis=axis, output_names=output_names, label_names=label_names) self.axis = axis self.on_cpu = on_cpu self.sparse_label = sparse_label self.vague_idx = vague_idx self.use_vague = use_vague self.aux = aux def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : list of `NDArray` The labels of the data. preds : list of `NDArray` Predicted values. """ if self.aux: preds = [p[0] for p in preds] assert (len(labels) == len(preds)) if self.on_cpu: for label, pred in zip(labels, preds): if self.sparse_label: label_imask = label.asnumpy().astype(np.int32) else: label_imask = mx.nd.argmax(label, axis=self.axis).asnumpy().astype(np.int32) pred_imask = mx.nd.argmax(pred, axis=self.axis).asnumpy().astype(np.int32) acc = seg_pixel_accuracy_np( label_imask=label_imask, pred_imask=pred_imask, vague_idx=self.vague_idx, use_vague=self.use_vague, macro_average=self.macro_average) if self.macro_average: self.sum_metric += acc self.num_inst += 1 else: self.sum_metric += acc[0] self.num_inst += acc[1] else: for label, pred in zip(labels, preds): if self.sparse_label: label_imask = mx.nd.cast(label, dtype=np.int32) else: label_imask = mx.nd.cast(mx.nd.argmax(label, axis=self.axis), dtype=np.int32) pred_imask = mx.nd.cast(mx.nd.argmax(pred, axis=self.axis), dtype=np.int32) acc = seg_pixel_accuracy_nd( label_imask=label_imask, pred_imask=pred_imask, vague_idx=self.vague_idx, use_vague=self.use_vague, macro_average=self.macro_average) if self.macro_average: self.sum_metric += acc self.num_inst += 1 else: self.sum_metric += acc[0] self.num_inst += acc[1] def reset(self): """ Resets the internal evaluation result to initial state. """ if self.macro_average: self.num_inst = 0 self.sum_metric = 0.0 else: self.num_inst = 0 self.sum_metric = 0 def get(self): """ Gets the current evaluation result. Returns: ------- names : list of str Name of the metrics. values : list of float Value of the evaluations. """ if self.macro_average: if self.num_inst == 0: return self.name, float("nan") else: return self.name, self.sum_metric / self.num_inst else: if self.num_inst == 0: return self.name, float("nan") else: return self.name, float(self.sum_metric) / self.num_inst class MeanIoUMetric(mx.metric.EvalMetric): """ Computes the mean intersection over union. Parameters: ---------- axis : int, default 1 The axis that represents classes name : str, default 'mean_iou' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. on_cpu : bool, default True Calculate on CPU. sparse_label : bool, default True Whether label is an integer array instead of probability distribution. num_classes : int Number of classes vague_idx : int, default -1 Index of masked pixels. use_vague : bool, default False Whether to use pixel masking. bg_idx : int, default -1 Index of background class. ignore_bg : bool, default False Whether to ignore background class. macro_average : bool, default True Whether to use micro or macro averaging. """ def __init__(self, axis=1, name="mean_iou", output_names=None, label_names=None, on_cpu=True, sparse_label=True, num_classes=None, vague_idx=-1, use_vague=False, bg_idx=-1, ignore_bg=False, macro_average=True): if name == "pix_acc": name = "{}-pix_acc".format("macro" if macro_average else "micro") self.macro_average = macro_average self.num_classes = num_classes self.ignore_bg = ignore_bg super(MeanIoUMetric, self).__init__( name, axis=axis, output_names=output_names, label_names=label_names) assert ((not ignore_bg) or (bg_idx in (0, num_classes - 1))) self.axis = axis self.on_cpu = on_cpu self.sparse_label = sparse_label self.vague_idx = vague_idx self.use_vague = use_vague self.bg_idx = bg_idx assert (on_cpu and sparse_label) def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : list of `NDArray` The labels of the data. preds : list of `NDArray` Predicted values. """ assert (len(labels) == len(preds)) if self.on_cpu: for label, pred in zip(labels, preds): if self.sparse_label: label_imask = label.asnumpy().astype(np.int32) # else: # label_hmask = label.asnumpy().astype(np.int32) pred_imask = mx.nd.argmax(pred, axis=self.axis).asnumpy().astype(np.int32) batch_size = label.shape[0] for k in range(batch_size): if self.sparse_label: acc = seg_mean_iou_imasks_np( label_imask=label_imask[k, :, :], pred_imask=pred_imask[k, :, :], num_classes=self.num_classes, vague_idx=self.vague_idx, use_vague=self.use_vague, bg_idx=self.bg_idx, ignore_bg=self.ignore_bg, macro_average=self.macro_average) # else: # acc = seg_mean_iou_np( # label_hmask=label_hmask[k, :, :, :], # pred_imask=pred_imask[k, :, :]) if self.macro_average: self.sum_metric += acc self.num_inst += 1 else: self.area_inter += acc[0] self.area_union += acc[1] # else: # for label, pred in zip(labels, preds): # if self.sparse_label: # label_imask = label # n = self.num_classes # label_hmask = mx.nd.one_hot(label_imask, depth=n).transpose((0, 3, 1, 2)) # else: # label_hmask = label # n = label_hmask.shape[1] # pred_imask = mx.nd.argmax(pred, axis=self.axis) # pred_hmask = mx.nd.one_hot(pred_imask, depth=n).transpose((0, 3, 1, 2)) # acc = seg_mean_iou2_nd( # label_hmask=label_hmask, # pred_hmask=pred_hmask) # self.sum_metric += acc # self.num_inst += 1 def reset(self): """ Resets the internal evaluation result to initial state. """ if self.macro_average: self.num_inst = 0 self.sum_metric = 0.0 else: class_count = self.num_classes - 1 if self.ignore_bg else self.num_classes self.area_inter = np.zeros((class_count,), np.uint64) self.area_union = np.zeros((class_count,), np.uint64) def get(self): """ Gets the current evaluation result. Returns: ------- names : list of str Name of the metrics. values : list of float Value of the evaluations. """ if self.macro_average: if self.num_inst == 0: return self.name, float("nan") else: return self.name, self.sum_metric / self.num_inst else: class_count = (self.area_union > 0).sum() if class_count == 0: return self.name, float("nan") eps = np.finfo(np.float32).eps area_union_eps = self.area_union + eps mean_iou = (self.area_inter / area_union_eps).sum() / class_count return self.name, mean_iou
11,492
35.485714
97
py
imgclsmob
imgclsmob-master/gluon/metrics/cls_metrics.py
""" Evaluation Metrics for Image Classification. """ import mxnet as mx __all__ = ['Top1Error', 'TopKError'] class Top1Error(mx.metric.Accuracy): """ Computes top-1 error (inverted accuracy classification score). Parameters: ---------- axis : int, default 1 The axis that represents classes. name : str, default 'top_1_error' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. """ def __init__(self, axis=1, name="top_1_error", output_names=None, label_names=None): super(Top1Error, self).__init__( axis=axis, name=name, output_names=output_names, label_names=label_names) def get(self): """ Gets the current evaluation result. Returns: ------- names : list of str Name of the metrics. values : list of float Value of the evaluations. """ if self.num_inst == 0: return self.name, float("nan") else: return self.name, 1.0 - self.sum_metric / self.num_inst class TopKError(mx.metric.TopKAccuracy): """ Computes top-k error (inverted top k predictions accuracy). Parameters: ---------- top_k : int Whether targets are out of top k predictions, default 1 name : str, default 'top_k_error' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. """ def __init__(self, top_k=1, name="top_k_error", output_names=None, label_names=None): name_ = name super(TopKError, self).__init__( top_k=top_k, name=name, output_names=output_names, label_names=label_names) self.name = name_.replace("_k_", "_{}_".format(top_k)) def get(self): """ Gets the current evaluation result. Returns: ------- names : list of str Name of the metrics. values : list of float Value of the evaluations. """ if self.num_inst == 0: return self.name, float("nan") else: return self.name, 1.0 - self.sum_metric / self.num_inst
2,977
28.78
79
py
imgclsmob
imgclsmob-master/gluon/metrics/metrics.py
""" Evaluation metrics for common tasks. """ import mxnet as mx if mx.__version__ < "2.0.0": from mxnet.metric import EvalMetric else: from mxnet.gluon.metric import EvalMetric __all__ = ['LossValue'] class LossValue(EvalMetric): """ Computes simple loss value fake metric. Parameters: ---------- name : str Name of this metric instance for display. output_names : list of str, or None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None Name of labels that should be used when updating with update_dict. By default include all labels. """ def __init__(self, name="loss", output_names=None, label_names=None): super(LossValue, self).__init__( name, output_names=output_names, label_names=label_names) def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : None Unused argument. preds : list of `NDArray` Loss values. """ loss = sum([ll.mean().asscalar() for ll in preds]) / len(preds) self.sum_metric += loss self.global_sum_metric += loss self.num_inst += 1 self.global_num_inst += 1
1,444
25.759259
79
py
imgclsmob
imgclsmob-master/gluon/metrics/det_metrics.py
""" Evaluation Metrics for Object Detection. """ import os import math import warnings import numpy as np import mxnet as mx from collections import defaultdict __all__ = ['CocoDetMApMetric', 'VOC07MApMetric', 'WiderfaceDetMetric'] class CocoDetMApMetric(mx.metric.EvalMetric): """ Detection metric for COCO bbox task. Parameters: ---------- img_height : int Processed image height. coco_annotations_file_path : str COCO anotation file path. contiguous_id_to_json : list of int Processed IDs. validation_ids : bool, default False Whether to use temporary file for estimation. use_file : bool, default False Whether to use temporary file for estimation. score_thresh : float, default 0.05 Detection results with confident scores smaller than `score_thresh` will be discarded before saving to results. data_shape : tuple of int, default is None If `data_shape` is provided as (height, width), we will rescale bounding boxes when saving the predictions. This is helpful when SSD/YOLO box predictions cannot be rescaled conveniently. Note that the data_shape must be fixed for all validation images. post_affine : a callable function with input signature (orig_w, orig_h, out_w, out_h) If not None, the bounding boxes will be affine transformed rather than simply scaled. name : str, default 'mAP' Name of this metric instance for display. """ def __init__(self, img_height, coco_annotations_file_path, contiguous_id_to_json, validation_ids=None, use_file=False, score_thresh=0.05, data_shape=None, post_affine=None, name="mAP"): super(CocoDetMApMetric, self).__init__(name=name) self.img_height = img_height self.coco_annotations_file_path = coco_annotations_file_path self.contiguous_id_to_json = contiguous_id_to_json self.validation_ids = validation_ids self.use_file = use_file self.score_thresh = score_thresh self.current_idx = 0 self.coco_result = [] if isinstance(data_shape, (tuple, list)): assert len(data_shape) == 2, "Data shape must be (height, width)" elif not data_shape: data_shape = None else: raise ValueError("data_shape must be None or tuple of int as (height, width)") self._data_shape = data_shape if post_affine is not None: assert self._data_shape is not None, "Using post affine transform requires data_shape" self._post_affine = post_affine else: self._post_affine = None from pycocotools.coco import COCO self.gt = COCO(self.coco_annotations_file_path) self._img_ids = sorted(self.gt.getImgIds()) def reset(self): self.current_idx = 0 self.coco_result = [] def get(self): """ Get evaluation metrics. """ if self.current_idx != len(self._img_ids): warnings.warn("Recorded {} out of {} validation images, incomplete results".format( self.current_idx, len(self._img_ids))) from pycocotools.coco import COCO gt = COCO(self.coco_annotations_file_path) import tempfile import json with tempfile.NamedTemporaryFile(mode="w", suffix=".json") as f: json.dump(self.coco_result, f) f.flush() pred = gt.loadRes(f.name) from pycocotools.cocoeval import COCOeval coco_eval = COCOeval(gt, pred, "bbox") if self.validation_ids is not None: coco_eval.params.imgIds = self.validation_ids coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize() return self.name, tuple(coco_eval.stats[:3]) def update2(self, pred_bboxes, pred_labels, pred_scores): """ Update internal buffer with latest predictions. Note that the statistics are not available until you call self.get() to return the metrics. Parameters: ---------- pred_bboxes : mxnet.NDArray or np.ndarray Prediction bounding boxes with shape `B, N, 4`. Where B is the size of mini-batch, N is the number of bboxes. pred_labels : mxnet.NDArray or np.ndarray Prediction bounding boxes labels with shape `B, N`. pred_scores : mxnet.NDArray or np.ndarray Prediction bounding boxes scores with shape `B, N`. """ def as_numpy(a): """ Convert a (list of) mx.NDArray into np.ndarray """ if isinstance(a, (list, tuple)): out = [x.asnumpy() if isinstance(x, mx.nd.NDArray) else x for x in a] return np.concatenate(out, axis=0) elif isinstance(a, mx.nd.NDArray): a = a.asnumpy() return a for pred_bbox, pred_label, pred_score in zip(*[as_numpy(x) for x in [pred_bboxes, pred_labels, pred_scores]]): valid_pred = np.where(pred_label.flat >= 0)[0] pred_bbox = pred_bbox[valid_pred, :].astype(np.float) pred_label = pred_label.flat[valid_pred].astype(int) pred_score = pred_score.flat[valid_pred].astype(np.float) imgid = self._img_ids[self.current_idx] self.current_idx += 1 affine_mat = None if self._data_shape is not None: entry = self.gt.loadImgs(imgid)[0] orig_height = entry["height"] orig_width = entry["width"] height_scale = float(orig_height) / self._data_shape[0] width_scale = float(orig_width) / self._data_shape[1] if self._post_affine is not None: affine_mat = self._post_affine(orig_width, orig_height, self._data_shape[1], self._data_shape[0]) else: height_scale, width_scale = (1.0, 1.0) # for each bbox detection in each image for bbox, label, score in zip(pred_bbox, pred_label, pred_score): if label not in self.contiguous_id_to_json: # ignore non-exist class continue if score < self.score_thresh: continue category_id = self.contiguous_id_to_json[label] # rescale bboxes/affine transform bboxes if affine_mat is not None: bbox[0:2] = self.affine_transform(bbox[0:2], affine_mat) bbox[2:4] = self.affine_transform(bbox[2:4], affine_mat) else: bbox[[0, 2]] *= width_scale bbox[[1, 3]] *= height_scale # convert [xmin, ymin, xmax, ymax] to [xmin, ymin, w, h] bbox[2:4] -= (bbox[:2] - 1) self.coco_result.append({"image_id": imgid, "category_id": category_id, "bbox": bbox[:4].tolist(), "score": score}) def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : list of `NDArray` The labels of the data. preds : list of `NDArray` Predicted values. """ det_bboxes = [] det_ids = [] det_scores = [] for x_rr, y in zip(preds, labels): bboxes = x_rr.slice_axis(axis=-1, begin=0, end=4) ids = x_rr.slice_axis(axis=-1, begin=4, end=5).squeeze(axis=2) scores = x_rr.slice_axis(axis=-1, begin=5, end=6).squeeze(axis=2) det_ids.append(ids) det_scores.append(scores) # clip to image size det_bboxes.append(bboxes.clip(0, self.img_height)) self.update2(det_bboxes, det_ids, det_scores) @staticmethod def affine_transform(pt, t): """ Apply affine transform to a bounding box given transform matrix t. Parameters: ---------- pt : np.ndarray Bounding box with shape (1, 2). t : np.ndarray Transformation matrix with shape (2, 3). Returns: ------- np.ndarray New bounding box with shape (1, 2). """ new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32).T new_pt = np.dot(t, new_pt) return new_pt[:2] class VOCMApMetric(mx.metric.EvalMetric): """ Calculate mean AP for object detection task Parameters: --------- iou_thresh : float IOU overlap threshold for TP class_names : list of str optional, if provided, will print out AP for each class name : str, default 'mAP' Name of this metric instance for display. """ def __init__(self, iou_thresh=0.5, class_names=None, name="mAP"): super(VOCMApMetric, self).__init__(name=name) if class_names is None: self.num = None else: assert isinstance(class_names, (list, tuple)) for name in class_names: assert isinstance(name, str), "must provide names as str" num = len(class_names) self.name = list(class_names) + ["mAP"] self.num = num + 1 self.reset() self.iou_thresh = iou_thresh self.class_names = class_names def reset(self): """ Clear the internal statistics to initial state. """ if getattr(self, 'num', None) is None: self.num_inst = 0 self.sum_metric = 0.0 else: self.num_inst = [0] * self.num self.sum_metric = [0.0] * self.num self._n_pos = defaultdict(int) self._score = defaultdict(list) self._match = defaultdict(list) def get(self): """ Get the current evaluation result. Returns: ------- name : str Name of the metric. value : float Value of the evaluation. """ self._update() # update metric at this time if self.num is None: if self.num_inst == 0: return self.name, float("nan") else: return self.name, self.sum_metric / self.num_inst else: names = ["%s" % self.name[i] for i in range(self.num)] values = [x / y if y != 0 else float("nan") for x, y in zip(self.sum_metric, self.num_inst)] return names, values def update(self, pred_bboxes, pred_labels, pred_scores, gt_bboxes, gt_labels, gt_difficults=None): """ Update internal buffer with latest prediction and gt pairs. Parameters: ---------- pred_bboxes : mxnet.NDArray or np.ndarray Prediction bounding boxes with shape `B, N, 4`. Where B is the size of mini-batch, N is the number of bboxes. pred_labels : mxnet.NDArray or np.ndarray Prediction bounding boxes labels with shape `B, N`. pred_scores : mxnet.NDArray or np.ndarray Prediction bounding boxes scores with shape `B, N`. gt_bboxes : mxnet.NDArray or np.ndarray Ground-truth bounding boxes with shape `B, M, 4`. Where B is the size of mini-batch, M is the number of ground-truths. gt_labels : mxnet.NDArray or np.ndarray Ground-truth bounding boxes labels with shape `B, M`. gt_difficults : mxnet.NDArray or np.ndarray, optional, default is None Ground-truth bounding boxes difficulty labels with shape `B, M`. """ def as_numpy(a): """ Convert a (list of) mx.NDArray into np.ndarray. """ if isinstance(a, (list, tuple)): out = [x.asnumpy() if isinstance(x, mx.nd.NDArray) else x for x in a] try: out = np.concatenate(out, axis=0) except ValueError: out = np.array(out) return out elif isinstance(a, mx.nd.NDArray): a = a.asnumpy() return a if gt_difficults is None: gt_difficults = [None for _ in as_numpy(gt_labels)] if isinstance(gt_labels, list): gt_diff_shape = gt_difficults[0].shape[0] if hasattr(gt_difficults[0], "shape") else 0 if len(gt_difficults) * gt_diff_shape != \ len(gt_labels) * gt_labels[0].shape[0]: gt_difficults = [None] * len(gt_labels) * gt_labels[0].shape[0] for pred_bbox, pred_label, pred_score, gt_bbox, gt_label, gt_difficult in zip( *[as_numpy(x) for x in [pred_bboxes, pred_labels, pred_scores, gt_bboxes, gt_labels, gt_difficults]]): # strip padding -1 for pred and gt valid_pred = np.where(pred_label.flat >= 0)[0] pred_bbox = pred_bbox[valid_pred, :] pred_label = pred_label.flat[valid_pred].astype(int) pred_score = pred_score.flat[valid_pred] valid_gt = np.where(gt_label.flat >= 0)[0] gt_bbox = gt_bbox[valid_gt, :] gt_label = gt_label.flat[valid_gt].astype(int) if gt_difficult is None: gt_difficult = np.zeros(gt_bbox.shape[0]) else: gt_difficult = gt_difficult.flat[valid_gt] for ll in np.unique(np.concatenate((pred_label, gt_label)).astype(int)): pred_mask_l = pred_label == ll pred_bbox_l = pred_bbox[pred_mask_l] pred_score_l = pred_score[pred_mask_l] # sort by score order = pred_score_l.argsort()[::-1] pred_bbox_l = pred_bbox_l[order] pred_score_l = pred_score_l[order] gt_mask_l = gt_label == ll gt_bbox_l = gt_bbox[gt_mask_l] gt_difficult_l = gt_difficult[gt_mask_l] self._n_pos[ll] += np.logical_not(gt_difficult_l).sum() self._score[ll].extend(pred_score_l) if len(pred_bbox_l) == 0: continue if len(gt_bbox_l) == 0: self._match[ll].extend((0,) * pred_bbox_l.shape[0]) continue # VOC evaluation follows integer typed bounding boxes. pred_bbox_l = pred_bbox_l.copy() pred_bbox_l[:, 2:] += 1 gt_bbox_l = gt_bbox_l.copy() gt_bbox_l[:, 2:] += 1 iou = self.bbox_iou(pred_bbox_l, gt_bbox_l) gt_index = iou.argmax(axis=1) # set -1 if there is no matching ground truth gt_index[iou.max(axis=1) < self.iou_thresh] = -1 del iou selec = np.zeros(gt_bbox_l.shape[0], dtype=bool) for gt_idx in gt_index: if gt_idx >= 0: if gt_difficult_l[gt_idx]: self._match[ll].append(-1) else: if not selec[gt_idx]: self._match[ll].append(1) else: self._match[ll].append(0) selec[gt_idx] = True else: self._match[ll].append(0) def _update(self): """ Update num_inst and sum_metric. """ aps = [] recall, precs = self._recall_prec() for ll, rec, prec in zip(range(len(precs)), recall, precs): ap = self._average_precision(rec, prec) aps.append(ap) if self.num is not None and ll < (self.num - 1): self.sum_metric[ll] = ap self.num_inst[ll] = 1 if self.num is None: self.num_inst = 1 self.sum_metric = np.nanmean(aps) else: self.num_inst[-1] = 1 self.sum_metric[-1] = np.nanmean(aps) def _recall_prec(self): """ Get recall and precision from internal records. """ n_fg_class = max(self._n_pos.keys()) + 1 prec = [None] * n_fg_class rec = [None] * n_fg_class for ll in self._n_pos.keys(): score_l = np.array(self._score[ll]) match_l = np.array(self._match[ll], dtype=np.int32) order = score_l.argsort()[::-1] match_l = match_l[order] tp = np.cumsum(match_l == 1) fp = np.cumsum(match_l == 0) # If an element of fp + tp is 0, # the corresponding element of prec[ll] is nan. with np.errstate(divide="ignore", invalid="ignore"): prec[ll] = tp / (fp + tp) # If n_pos[ll] is 0, rec[ll] is None. if self._n_pos[ll] > 0: rec[ll] = tp / self._n_pos[ll] return rec, prec def _average_precision(self, rec, prec): """ Calculate average precision. Params: ---------- rec : np.array cumulated recall prec : np.array cumulated precision Returns: ---------- float AP """ if rec is None or prec is None: return np.nan # append sentinel values at both ends mrec = np.concatenate(([0.0], rec, [1.0])) mpre = np.concatenate(([0.0], np.nan_to_num(prec), [0.0])) # compute precision integration ladder for i in range(mpre.size - 1, 0, -1): mpre[i - 1] = np.maximum(mpre[i - 1], mpre[i]) # look for recall value changes i = np.where(mrec[1:] != mrec[:-1])[0] # sum (\delta recall) * prec ap = np.sum((mrec[i + 1] - mrec[i]) * mpre[i + 1]) return ap @staticmethod def bbox_iou(bbox_a, bbox_b, offset=0): """ Calculate Intersection-Over-Union(IOU) of two bounding boxes. Parameters: ---------- bbox_a : np.ndarray An ndarray with shape :math:`(N, 4)`. bbox_b : np.ndarray An ndarray with shape :math:`(M, 4)`. offset : float or int, default is 0 The ``offset`` is used to control the whether the width(or height) is computed as (right - left + ``offset``). Note that the offset must be 0 for normalized bboxes, whose ranges are in ``[0, 1]``. Returns: ------- np.ndarray An ndarray with shape :math:`(N, M)` indicates IOU between each pairs of bounding boxes in `bbox_a` and `bbox_b`. """ if bbox_a.shape[1] < 4 or bbox_b.shape[1] < 4: raise IndexError("Bounding boxes axis 1 must have at least length 4") tl = np.maximum(bbox_a[:, None, :2], bbox_b[:, :2]) br = np.minimum(bbox_a[:, None, 2:4], bbox_b[:, 2:4]) area_i = np.prod(br - tl + offset, axis=2) * (tl < br).all(axis=2) area_a = np.prod(bbox_a[:, 2:4] - bbox_a[:, :2] + offset, axis=1) area_b = np.prod(bbox_b[:, 2:4] - bbox_b[:, :2] + offset, axis=1) return area_i / (area_a[:, None] + area_b - area_i) class VOC07MApMetric(VOCMApMetric): """ Mean average precision metric for PASCAL V0C 07 dataset. Parameters: --------- iou_thresh : float IOU overlap threshold for TP class_names : list of str optional, if provided, will print out AP for each class """ def __init__(self, *args, **kwargs): super(VOC07MApMetric, self).__init__(*args, **kwargs) def _average_precision(self, rec, prec): """ calculate average precision, override the default one, special 11-point metric Params: ---------- rec : np.array cumulated recall prec : np.array cumulated precision Returns: ---------- float AP """ if rec is None or prec is None: return np.nan ap = 0.0 for t in np.arange(0.0, 1.1, 0.1): if np.sum(rec >= t) == 0: p = 0 else: p = np.max(np.nan_to_num(prec)[rec >= t]) ap += p / 11.0 return ap class WiderfaceDetMetric(mx.metric.EvalMetric): """ Detection metric for WIDER FACE detection task. Parameters: ---------- receptive_field_center_starts : list of int The start location of the first receptive field of each scale. receptive_field_strides : list of int Receptive field stride for each scale. bbox_factors : list of float A half of bbox upper bound for each scale. output_dir_path : str Output file path. name : str, default 'WF' Name of this metric instance for display. """ def __init__(self, receptive_field_center_starts, receptive_field_strides, bbox_factors, output_dir_path, name="WF"): super(WiderfaceDetMetric, self).__init__(name=name) self.receptive_field_center_starts = receptive_field_center_starts self.receptive_field_strides = receptive_field_strides self.bbox_factors = bbox_factors self.output_dir_path = output_dir_path self.num_output_scales = len(self.bbox_factors) self.score_threshold = 0.11 self.nms_threshold = 0.4 self.top_k = 10000 def reset(self): pass def get(self): """ Get evaluation metrics. """ return self.name, 1.0 def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : list of `NDArray` The labels of the data. preds : list of `NDArray` Predicted values. """ for x_rr, label in zip(preds, labels): outputs = [] for output in x_rr: outputs.append(output.asnumpy()) label_split = label.split("/") resize_scale = float(label_split[2]) image_size = (int(label_split[3]), int(label_split[4])) bboxes, _ = self.predict(outputs, resize_scale, image_size) event_name = label_split[0] event_dir_name = os.path.join(self.output_dir_path, event_name) if not os.path.exists(event_dir_name): os.makedirs(event_dir_name) file_stem = label_split[1] fout = open(os.path.join(event_dir_name, file_stem + ".txt"), "w") fout.write(file_stem + "\n") fout.write(str(len(bboxes)) + "\n") for bbox in bboxes: fout.write("%d %d %d %d %.03f" % (math.floor(bbox[0]), math.floor(bbox[1]), math.ceil(bbox[2] - bbox[0]), math.ceil(bbox[3] - bbox[1]), bbox[4] if bbox[4] <= 1 else 1) + "\n") fout.close() def predict(self, outputs, resize_scale, image_size): bbox_collection = [] for i in range(self.num_output_scales): score_map = np.squeeze(outputs[i * 2], (0, 1)) bbox_map = np.squeeze(outputs[i * 2 + 1], 0) RF_center_Xs = np.array( [self.receptive_field_center_starts[i] + self.receptive_field_strides[i] * x for x in range(score_map.shape[1])]) RF_center_Xs_mat = np.tile(RF_center_Xs, [score_map.shape[0], 1]) RF_center_Ys = np.array( [self.receptive_field_center_starts[i] + self.receptive_field_strides[i] * y for y in range(score_map.shape[0])]) RF_center_Ys_mat = np.tile(RF_center_Ys, [score_map.shape[1], 1]).T x_lt_mat = RF_center_Xs_mat - bbox_map[0, :, :] * self.bbox_factors[i] y_lt_mat = RF_center_Ys_mat - bbox_map[1, :, :] * self.bbox_factors[i] x_rb_mat = RF_center_Xs_mat - bbox_map[2, :, :] * self.bbox_factors[i] y_rb_mat = RF_center_Ys_mat - bbox_map[3, :, :] * self.bbox_factors[i] x_lt_mat = x_lt_mat / resize_scale x_lt_mat[x_lt_mat < 0] = 0 y_lt_mat = y_lt_mat / resize_scale y_lt_mat[y_lt_mat < 0] = 0 x_rb_mat = x_rb_mat / resize_scale x_rb_mat[x_rb_mat > image_size[1]] = image_size[1] y_rb_mat = y_rb_mat / resize_scale y_rb_mat[y_rb_mat > image_size[0]] = image_size[0] select_index = np.where(score_map > self.score_threshold) for idx in range(select_index[0].size): bbox_collection.append((x_lt_mat[select_index[0][idx], select_index[1][idx]], y_lt_mat[select_index[0][idx], select_index[1][idx]], x_rb_mat[select_index[0][idx], select_index[1][idx]], y_rb_mat[select_index[0][idx], select_index[1][idx]], score_map[select_index[0][idx], select_index[1][idx]])) # NMS bbox_collection = sorted(bbox_collection, key=lambda item: item[-1], reverse=True) if len(bbox_collection) > self.top_k: bbox_collection = bbox_collection[0:self.top_k] bbox_collection_numpy = np.array(bbox_collection, dtype=np.float32) final_bboxes = self.nms(bbox_collection_numpy, self.nms_threshold) final_bboxes_ = [] for i in range(final_bboxes.shape[0]): final_bboxes_.append((final_bboxes[i, 0], final_bboxes[i, 1], final_bboxes[i, 2], final_bboxes[i, 3], final_bboxes[i, 4])) return final_bboxes_ @staticmethod def nms(boxes, overlap_threshold): if boxes.shape[0] == 0: return boxes if boxes.dtype != np.float32: boxes = boxes.astype(np.float32) pick = [] x1 = boxes[:, 0] y1 = boxes[:, 1] x2 = boxes[:, 2] y2 = boxes[:, 3] sc = boxes[:, 4] widths = x2 - x1 heights = y2 - y1 area = heights * widths idxs = np.argsort(sc) while len(idxs) > 0: last = len(idxs) - 1 i = idxs[last] pick.append(i) xx1 = np.maximum(x1[i], x1[idxs[:last]]) yy1 = np.maximum(y1[i], y1[idxs[:last]]) xx2 = np.minimum(x2[i], x2[idxs[:last]]) yy2 = np.minimum(y2[i], y2[idxs[:last]]) w = np.maximum(0, xx2 - xx1 + 1) h = np.maximum(0, yy2 - yy1 + 1) overlap = (w * h) / area[idxs[:last]] idxs = np.delete(idxs, np.concatenate(([last], np.where(overlap > overlap_threshold)[0]))) return boxes[pick]
27,536
36.11186
119
py
imgclsmob
imgclsmob-master/gluon/metrics/hpe_metrics.py
""" Evaluation Metrics for Human Pose Estimation. """ import mxnet as mx __all__ = ['CocoHpeOksApMetric'] class CocoHpeOksApMetric(mx.metric.EvalMetric): """ Detection metric for COCO bbox task. Parameters: ---------- coco_annotations_file_path : str COCO anotation file path. pose_postprocessing_fn : func An function for pose post-processing. validation_ids : bool, default False Whether to use temporary file for estimation. use_file : bool, default False Whether to use temporary file for estimation. name : str, default 'CocoOksAp' Name of this metric instance for display. """ def __init__(self, coco_annotations_file_path, pose_postprocessing_fn, validation_ids=None, use_file=False, name="CocoOksAp"): super(CocoHpeOksApMetric, self).__init__(name=name) self.coco_annotations_file_path = coco_annotations_file_path self.pose_postprocessing_fn = pose_postprocessing_fn self.validation_ids = validation_ids self.use_file = use_file self.coco_result = [] def reset(self): self.coco_result = [] def get(self): """ Get evaluation metrics. """ import copy from pycocotools.coco import COCO gt = COCO(self.coco_annotations_file_path) if self.use_file: import tempfile import json with tempfile.NamedTemporaryFile(mode="w", suffix=".json") as f: json.dump(self.coco_result, f) f.flush() pred = gt.loadRes(f.name) else: def calc_pred(coco, anns): import numpy as np import copy pred = COCO() pred.dataset["images"] = [img for img in coco.dataset["images"]] annsImgIds = [ann["image_id"] for ann in anns] assert set(annsImgIds) == (set(annsImgIds) & set(coco.getImgIds())) pred.dataset["categories"] = copy.deepcopy(coco.dataset["categories"]) for id, ann in enumerate(anns): s = ann["keypoints"] x = s[0::3] y = s[1::3] x0, x1, y0, y1 = np.min(x), np.max(x), np.min(y), np.max(y) ann["area"] = (x1 - x0) * (y1 - y0) ann["id"] = id + 1 ann["bbox"] = [x0, y0, x1 - x0, y1 - y0] pred.dataset["annotations"] = anns pred.createIndex() return pred pred = calc_pred(gt, copy.deepcopy(self.coco_result)) from pycocotools.cocoeval import COCOeval coco_eval = COCOeval(gt, pred, "keypoints") if self.validation_ids is not None: coco_eval.params.imgIds = self.validation_ids coco_eval.params.useSegm = None coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize() return self.name, tuple(coco_eval.stats[:3]) def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : list of `NDArray` The labels of the data. preds : list of `NDArray` Predicted values. """ for label, pred in zip(labels, preds): label = label.asnumpy() pred = pred.asnumpy() pred_pts_score, pred_person_score, label_img_id = self.pose_postprocessing_fn(pred, label) for idx in range(len(pred_pts_score)): image_id = int(label_img_id[idx]) kpt = pred_pts_score[idx].flatten().tolist() score = float(pred_person_score[idx]) self.coco_result.append({ "image_id": image_id, "category_id": 1, "keypoints": kpt, "score": score})
4,037
32.371901
102
py
imgclsmob
imgclsmob-master/gluon/metrics/asr_metrics.py
""" Evaluation Metrics for Automatic Speech Recognition (ASR). """ import mxnet as mx __all__ = ['WER'] class WER(mx.metric.EvalMetric): """ Computes Word Error Rate (WER) for Automatic Speech Recognition (ASR). Parameters: ---------- vocabulary : list of str Vocabulary of the dataset. name : str, default 'wer' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. """ def __init__(self, vocabulary, name="wer", output_names=None, label_names=None): super(WER, self).__init__( name=name, output_names=output_names, label_names=label_names, has_global_stats=True) self.vocabulary = vocabulary self.ctc_decoder = CtcDecoder(vocabulary=vocabulary) def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : list of `NDArray` The labels of the data. preds : list of `NDArray` Predicted values. """ import editdistance for labels_i, preds_i in zip(labels, preds): labels_code = labels_i.asnumpy() labels_i = [] for label_code in labels_code: label_text = "".join([self.ctc_decoder.labels_map[c] for c in label_code]) labels_i.append(label_text) preds_i = preds_i[0] greedy_predictions = preds_i.swapaxes(1, 2).log_softmax(axis=-1).argmax(axis=-1, keepdims=False).asnumpy() preds_i = self.ctc_decoder(greedy_predictions) assert (len(labels_i) == len(preds_i)) for pred, label in zip(preds_i, labels_i): pred = pred.split() label = label.split() word_error_count = editdistance.eval(label, pred) word_count = max(len(label), len(pred)) assert (word_error_count <= word_count) self.sum_metric += word_error_count self.global_sum_metric += word_error_count self.num_inst += word_count self.global_num_inst += word_count class CtcDecoder(object): """ CTC decoder (to decode a sequence of labels to words). Parameters: ---------- vocabulary : list of str Vocabulary of the dataset. """ def __init__(self, vocabulary): super().__init__() self.blank_id = len(vocabulary) self.labels_map = dict([(i, vocabulary[i]) for i in range(len(vocabulary))]) def __call__(self, predictions): """ Decode a sequence of labels to words. Parameters: ---------- predictions : np.array of int or list of list of int Tensor with predicted labels. Returns: ------- list of str Words. """ hypotheses = [] for prediction in predictions: decoded_prediction = [] previous = self.blank_id for p in prediction: if (p != previous or previous == self.blank_id) and p != self.blank_id: decoded_prediction.append(p) previous = p hypothesis = "".join([self.labels_map[c] for c in decoded_prediction]) hypotheses.append(hypothesis) return hypotheses
3,800
30.413223
118
py
imgclsmob
imgclsmob-master/gluon/datasets/imagenet1k_cls_dataset.py
""" ImageNet-1K classification dataset. """ import os import math import mxnet as mx from mxnet.gluon import HybridBlock from mxnet.gluon.data.vision import ImageFolderDataset from mxnet.gluon.data.vision import transforms from .dataset_metainfo import DatasetMetaInfo class ImageNet1K(ImageFolderDataset): """ ImageNet-1K classification dataset. Refer to MXNet documentation for the description of this dataset and how to prepare it. Parameters: ---------- root : str, default '~/.mxnet/datasets/imagenet' Path to the folder stored the dataset. mode : str, default 'train' 'train', 'val', or 'test'. transform : function, default None A function that takes data and label and transforms them. """ def __init__(self, root=os.path.join("~", ".mxnet", "datasets", "imagenet"), mode="train", transform=None): split = "train" if mode == "train" else "val" root = os.path.join(root, split) super(ImageNet1K, self).__init__(root=root, flag=1, transform=transform) class ImageNet1KMetaInfo(DatasetMetaInfo): """ Descriptor of ImageNet-1K dataset. """ def __init__(self): super(ImageNet1KMetaInfo, self).__init__() self.label = "ImageNet1K" self.short_label = "imagenet" self.root_dir_name = "imagenet" self.dataset_class = ImageNet1K self.num_training_samples = None self.in_channels = 3 self.num_classes = 1000 self.input_image_size = (224, 224) self.resize_inv_factor = 0.875 self.aug_type = "aug0" self.train_metric_capts = ["Train.Top1"] self.train_metric_names = ["Top1Error"] self.train_metric_extra_kwargs = [{"name": "err-top1"}] self.val_metric_capts = ["Val.Top1", "Val.Top5"] self.val_metric_names = ["Top1Error", "TopKError"] self.val_metric_extra_kwargs = [{"name": "err-top1"}, {"name": "err-top5", "top_k": 5}] self.saver_acc_ind = 1 self.train_transform = imagenet_train_transform self.val_transform = imagenet_val_transform self.test_transform = imagenet_val_transform self.ml_type = "imgcls" self.mean_rgb = (0.485, 0.456, 0.406) self.std_rgb = (0.229, 0.224, 0.225) self.interpolation = 1 self.loss_name = "SoftmaxCrossEntropy" def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for ImageNet-1K dataset metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(ImageNet1KMetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--input-size", type=int, default=self.input_image_size[0], help="size of the input for model") parser.add_argument( "--resize-inv-factor", type=float, default=self.resize_inv_factor, help="inverted ratio for input image crop") parser.add_argument( "--aug-type", type=str, default="aug0", help="augmentation type. options are aug0, aug1, aug2") parser.add_argument( "--mean-rgb", nargs=3, type=float, default=self.mean_rgb, help="Mean of RGB channels in the dataset") parser.add_argument( "--std-rgb", nargs=3, type=float, default=self.std_rgb, help="STD of RGB channels in the dataset") parser.add_argument( "--interpolation", type=int, default=self.interpolation, help="Preprocessing interpolation") def update(self, args): """ Update ImageNet-1K dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(ImageNet1KMetaInfo, self).update(args) self.input_image_size = (args.input_size, args.input_size) self.resize_inv_factor = args.resize_inv_factor self.aug_type = args.aug_type self.mean_rgb = args.mean_rgb self.std_rgb = args.std_rgb self.interpolation = args.interpolation class ImgAugTransform(HybridBlock): """ ImgAug-like transform (geometric, noise, and blur). """ def __init__(self): super(ImgAugTransform, self).__init__() from imgaug import augmenters as iaa from imgaug import parameters as iap self.seq = iaa.Sequential( children=[ iaa.Sequential( children=[ iaa.Sequential( children=[ iaa.OneOf( children=[ iaa.Sometimes( p=0.95, then_list=iaa.Affine( scale={"x": (0.9, 1.1), "y": (0.9, 1.1)}, translate_percent={"x": (-0.05, 0.05), "y": (-0.05, 0.05)}, rotate=(-30, 30), shear=(-15, 15), order=iap.Choice([0, 1, 3], p=[0.15, 0.80, 0.05]), mode="reflect", name="Affine")), iaa.Sometimes( p=0.05, then_list=iaa.PerspectiveTransform( scale=(0.01, 0.1)))], name="Blur"), iaa.Sometimes( p=0.01, then_list=iaa.PiecewiseAffine( scale=(0.0, 0.01), nb_rows=(4, 20), nb_cols=(4, 20), order=iap.Choice([0, 1, 3], p=[0.15, 0.80, 0.05]), mode="reflect", name="PiecewiseAffine"))], random_order=True, name="GeomTransform"), iaa.Sequential( children=[ iaa.Sometimes( p=0.75, then_list=iaa.Add( value=(-10, 10), per_channel=0.5, name="Brightness")), iaa.Sometimes( p=0.05, then_list=iaa.Emboss( alpha=(0.0, 0.5), strength=(0.5, 1.2), name="Emboss")), iaa.Sometimes( p=0.1, then_list=iaa.Sharpen( alpha=(0.0, 0.5), lightness=(0.5, 1.2), name="Sharpen")), iaa.Sometimes( p=0.25, then_list=iaa.ContrastNormalization( alpha=(0.5, 1.5), per_channel=0.5, name="ContrastNormalization")) ], random_order=True, name="ColorTransform"), iaa.Sequential( children=[ iaa.Sometimes( p=0.5, then_list=iaa.AdditiveGaussianNoise( loc=0, scale=(0.0, 10.0), per_channel=0.5, name="AdditiveGaussianNoise")), iaa.Sometimes( p=0.1, then_list=iaa.SaltAndPepper( p=(0, 0.001), per_channel=0.5, name="SaltAndPepper"))], random_order=True, name="Noise"), iaa.OneOf( children=[ iaa.Sometimes( p=0.05, then_list=iaa.MedianBlur( k=3, name="MedianBlur")), iaa.Sometimes( p=0.05, then_list=iaa.AverageBlur( k=(2, 4), name="AverageBlur")), iaa.Sometimes( p=0.5, then_list=iaa.GaussianBlur( sigma=(0.0, 2.0), name="GaussianBlur"))], name="Blur"), ], random_order=True, name="MainProcess")]) def hybrid_forward(self, F, x): img = x.asnumpy().copy() # cv2.imshow(winname="imgA", mat=img) img_aug = self.seq.augment_image(img) # cv2.imshow(winname="img_augA", mat=img_aug) # cv2.waitKey() x = mx.nd.array(img_aug, dtype=x.dtype, ctx=x.context) return x def imagenet_train_transform(ds_metainfo, jitter_param=0.4, lighting_param=0.1): """ Create image transform sequence for training subset. Parameters: ---------- ds_metainfo : DatasetMetaInfo ImageNet-1K dataset metainfo. jitter_param : float How much to jitter values. lighting_param : float How much to noise intensity of the image. Returns: ------- Sequential Image transform sequence. """ input_image_size = ds_metainfo.input_image_size if ds_metainfo.aug_type == "aug0": interpolation = ds_metainfo.interpolation transform_list = [] elif ds_metainfo.aug_type == "aug1": interpolation = 10 transform_list = [] elif ds_metainfo.aug_type == "aug2": interpolation = 10 transform_list = [ ImgAugTransform() ] else: raise RuntimeError("Unknown augmentation type: {}\n".format(ds_metainfo.aug_type)) transform_list += [ transforms.RandomResizedCrop( size=input_image_size, interpolation=interpolation), transforms.RandomFlipLeftRight(), transforms.RandomColorJitter( brightness=jitter_param, contrast=jitter_param, saturation=jitter_param), transforms.RandomLighting(lighting_param), transforms.ToTensor(), transforms.Normalize( mean=ds_metainfo.mean_rgb, std=ds_metainfo.std_rgb) ] return transforms.Compose(transform_list) def imagenet_val_transform(ds_metainfo): """ Create image transform sequence for validation subset. Parameters: ---------- ds_metainfo : DatasetMetaInfo ImageNet-1K dataset metainfo. Returns: ------- Sequential Image transform sequence. """ input_image_size = ds_metainfo.input_image_size resize_value = calc_val_resize_value( input_image_size=ds_metainfo.input_image_size, resize_inv_factor=ds_metainfo.resize_inv_factor) return transforms.Compose([ transforms.Resize( size=resize_value, keep_ratio=True, interpolation=ds_metainfo.interpolation), transforms.CenterCrop(size=input_image_size), transforms.ToTensor(), transforms.Normalize( mean=ds_metainfo.mean_rgb, std=ds_metainfo.std_rgb) ]) def calc_val_resize_value(input_image_size=(224, 224), resize_inv_factor=0.875): """ Calculate image resize value for validation subset. Parameters: ---------- input_image_size : tuple of 2 int Main script arguments. resize_inv_factor : float Resize inverted factor. Returns: ------- int Resize value. """ if isinstance(input_image_size, int): input_image_size = (input_image_size, input_image_size) resize_value = int(math.ceil(float(input_image_size[0]) / resize_inv_factor)) return resize_value
13,686
36.705234
107
py
imgclsmob
imgclsmob-master/gluon/datasets/coco_hpe1_dataset.py
""" COCO keypoint detection (2D single human pose estimation) dataset. """ import os import copy import cv2 import numpy as np import mxnet as mx from mxnet.gluon.data import dataset from .dataset_metainfo import DatasetMetaInfo class CocoHpe1Dataset(dataset.Dataset): """ COCO keypoint detection (2D single human pose estimation) dataset. Parameters: ---------- root : string Path to `annotations`, `train2017`, and `val2017` folders. mode : string, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. splits : list of str, default ['person_keypoints_val2017'] Json annotations name. Candidates can be: person_keypoints_val2017, person_keypoints_train2017. check_centers : bool, default is False If true, will force check centers of bbox and keypoints, respectively. If centers are far away from each other, remove this label. skip_empty : bool, default is False Whether skip entire image if no valid label is found. Use `False` if this dataset is for validation to avoid COCO metric error. """ def __init__(self, root, mode="train", transform=None, splits=("person_keypoints_val2017",), check_centers=False, skip_empty=True): super(CocoHpe1Dataset, self).__init__() self.root = os.path.expanduser(root) self.mode = mode self._transform = transform self.classes = ["person"] self.num_class = len(self.classes) self.keypoint_names = { 0: "nose", 1: "left_eye", 2: "right_eye", 3: "left_ear", 4: "right_ear", 5: "left_shoulder", 6: "right_shoulder", 7: "left_elbow", 8: "right_elbow", 9: "left_wrist", 10: "right_wrist", 11: "left_hip", 12: "right_hip", 13: "left_knee", 14: "right_knee", 15: "left_ankle", 16: "right_ankle" } self.skeleton = [ [16, 14], [14, 12], [17, 15], [15, 13], [12, 13], [6, 12], [7, 13], [6, 7], [6, 8], [7, 9], [8, 10], [9, 11], [2, 3], [1, 2], [1, 3], [2, 4], [3, 5], [4, 6], [5, 7]] # Joint pairs which defines the pairs of joint to be swapped when the image is flipped horizontally: self.joint_pairs = [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]] self.num_joints = 17 if isinstance(splits, mx.base.string_types): splits = [splits] self._splits = splits self._coco = [] self._check_centers = check_centers self._skip_empty = skip_empty self.index_map = dict(zip(self.classes, range(self.num_class))) self.json_id_to_contiguous = None self.contiguous_id_to_json = None self._items, self._labels = self._load_jsons() mode_name = "train" if mode == "train" else "val" annotations_dir_path = os.path.join(root, "annotations") annotations_file_path = os.path.join(annotations_dir_path, "person_keypoints_" + mode_name + "2017.json") self.annotations_file_path = annotations_file_path def __str__(self): detail = ",".join([str(s) for s in self._splits]) return self.__class__.__name__ + "(" + detail + ")" def __len__(self): return len(self._items) def __getitem__(self, idx): img_path = self._items[idx] img_id = int(os.path.splitext(os.path.basename(img_path))[0]) label = copy.deepcopy(self._labels[idx]) img = mx.image.imread(img_path, 1) if self._transform is not None: img, scale, center, score = self._transform(img, label) res_label = np.array([float(img_id)] + [float(score)] + list(center) + list(scale), np.float32) return img, res_label def _load_jsons(self): """ Load all image paths and labels from JSON annotation files into buffer. """ items = [] labels = [] from pycocotools.coco import COCO for split in self._splits: anno = os.path.join(self.root, "annotations", split) + ".json" _coco = COCO(anno) self._coco.append(_coco) classes = [c["name"] for c in _coco.loadCats(_coco.getCatIds())] if not classes == self.classes: raise ValueError("Incompatible category names with COCO: ") assert classes == self.classes json_id_to_contiguous = { v: k for k, v in enumerate(_coco.getCatIds())} if self.json_id_to_contiguous is None: self.json_id_to_contiguous = json_id_to_contiguous self.contiguous_id_to_json = { v: k for k, v in self.json_id_to_contiguous.items()} else: assert self.json_id_to_contiguous == json_id_to_contiguous # iterate through the annotations image_ids = sorted(_coco.getImgIds()) for entry in _coco.loadImgs(image_ids): dirname, filename = entry["coco_url"].split("/")[-2:] abs_path = os.path.join(self.root, dirname, filename) if not os.path.exists(abs_path): raise IOError("Image: {} not exists.".format(abs_path)) label = self._check_load_keypoints(_coco, entry) if not label: continue # num of items are relative to person, not image for obj in label: items.append(abs_path) labels.append(obj) return items, labels def _check_load_keypoints(self, coco, entry): """ Check and load ground-truth keypoints. """ ann_ids = coco.getAnnIds(imgIds=entry["id"], iscrowd=False) objs = coco.loadAnns(ann_ids) # check valid bboxes valid_objs = [] width = entry["width"] height = entry["height"] for obj in objs: contiguous_cid = self.json_id_to_contiguous[obj["category_id"]] if contiguous_cid >= self.num_class: # not class of interest continue if max(obj["keypoints"]) == 0: continue # convert from (x, y, w, h) to (xmin, ymin, xmax, ymax) and clip bound xmin, ymin, xmax, ymax = self.bbox_clip_xyxy(self.bbox_xywh_to_xyxy(obj["bbox"]), width, height) # require non-zero box area if obj['area'] <= 0 or xmax <= xmin or ymax <= ymin: continue # joints 3d: (num_joints, 3, 2); 3 is for x, y, z; 2 is for position, visibility joints_3d = np.zeros((self.num_joints, 3, 2), dtype=np.float32) for i in range(self.num_joints): joints_3d[i, 0, 0] = obj["keypoints"][i * 3 + 0] joints_3d[i, 1, 0] = obj["keypoints"][i * 3 + 1] # joints_3d[i, 2, 0] = 0 visible = min(1, obj["keypoints"][i * 3 + 2]) joints_3d[i, :2, 1] = visible # joints_3d[i, 2, 1] = 0 if np.sum(joints_3d[:, 0, 1]) < 1: # no visible keypoint continue if self._check_centers: bbox_center, bbox_area = self._get_box_center_area((xmin, ymin, xmax, ymax)) kp_center, num_vis = self._get_keypoints_center_count(joints_3d) ks = np.exp(-2 * np.sum(np.square(bbox_center - kp_center)) / bbox_area) if (num_vis / 80.0 + 47 / 80.0) > ks: continue valid_objs.append({ "bbox": (xmin, ymin, xmax, ymax), "joints_3d": joints_3d }) if not valid_objs: if not self._skip_empty: # dummy invalid labels if no valid objects are found valid_objs.append({ "bbox": np.array([-1, -1, 0, 0]), "joints_3d": np.zeros((self.num_joints, 3, 2), dtype=np.float32) }) return valid_objs @staticmethod def _get_box_center_area(bbox): """ Get bbox center. """ c = np.array([(bbox[0] + bbox[2]) / 2.0, (bbox[1] + bbox[3]) / 2.0]) area = (bbox[3] - bbox[1]) * (bbox[2] - bbox[0]) return c, area @staticmethod def _get_keypoints_center_count(keypoints): """ Get geometric center of all keypoints. """ keypoint_x = np.sum(keypoints[:, 0, 0] * (keypoints[:, 0, 1] > 0)) keypoint_y = np.sum(keypoints[:, 1, 0] * (keypoints[:, 1, 1] > 0)) num = float(np.sum(keypoints[:, 0, 1])) return np.array([keypoint_x / num, keypoint_y / num]), num @staticmethod def bbox_clip_xyxy(xyxy, width, height): """ Clip bounding box with format (xmin, ymin, xmax, ymax) to specified boundary. All bounding boxes will be clipped to the new region `(0, 0, width, height)`. Parameters: ---------- xyxy : list, tuple or numpy.ndarray The bbox in format (xmin, ymin, xmax, ymax). If numpy.ndarray is provided, we expect multiple bounding boxes with shape `(N, 4)`. width : int or float Boundary width. height : int or float Boundary height. Returns: ------- tuple or np.array Description of returned object. """ if isinstance(xyxy, (tuple, list)): if not len(xyxy) == 4: raise IndexError("Bounding boxes must have 4 elements, given {}".format(len(xyxy))) x1 = np.minimum(width - 1, np.maximum(0, xyxy[0])) y1 = np.minimum(height - 1, np.maximum(0, xyxy[1])) x2 = np.minimum(width - 1, np.maximum(0, xyxy[2])) y2 = np.minimum(height - 1, np.maximum(0, xyxy[3])) return x1, y1, x2, y2 elif isinstance(xyxy, np.ndarray): if not xyxy.size % 4 == 0: raise IndexError("Bounding boxes must have n * 4 elements, given {}".format(xyxy.shape)) x1 = np.minimum(width - 1, np.maximum(0, xyxy[:, 0])) y1 = np.minimum(height - 1, np.maximum(0, xyxy[:, 1])) x2 = np.minimum(width - 1, np.maximum(0, xyxy[:, 2])) y2 = np.minimum(height - 1, np.maximum(0, xyxy[:, 3])) return np.hstack((x1, y1, x2, y2)) else: raise TypeError("Expect input xywh a list, tuple or numpy.ndarray, given {}".format(type(xyxy))) @staticmethod def bbox_xywh_to_xyxy(xywh): """ Convert bounding boxes from format (xmin, ymin, w, h) to (xmin, ymin, xmax, ymax) Parameters: ---------- xywh : list, tuple or numpy.ndarray The bbox in format (x, y, w, h). If numpy.ndarray is provided, we expect multiple bounding boxes with shape `(N, 4)`. Returns: ------- tuple or np.ndarray The converted bboxes in format (xmin, ymin, xmax, ymax). If input is numpy.ndarray, return is numpy.ndarray correspondingly. """ if isinstance(xywh, (tuple, list)): if not len(xywh) == 4: raise IndexError("Bounding boxes must have 4 elements, given {}".format(len(xywh))) w, h = np.maximum(xywh[2] - 1, 0), np.maximum(xywh[3] - 1, 0) return xywh[0], xywh[1], xywh[0] + w, xywh[1] + h elif isinstance(xywh, np.ndarray): if not xywh.size % 4 == 0: raise IndexError("Bounding boxes must have n * 4 elements, given {}".format(xywh.shape)) xyxy = np.hstack((xywh[:, :2], xywh[:, :2] + np.maximum(0, xywh[:, 2:4] - 1))) return xyxy else: raise TypeError("Expect input xywh a list, tuple or numpy.ndarray, given {}".format(type(xywh))) # --------------------------------------------------------------------------------------------------------------------- class CocoHpeValTransform1(object): def __init__(self, ds_metainfo): self.ds_metainfo = ds_metainfo self.image_size = self.ds_metainfo.input_image_size height = self.image_size[0] width = self.image_size[1] self.aspect_ratio = float(width / height) self.mean = ds_metainfo.mean_rgb self.std = ds_metainfo.std_rgb def __call__(self, src, label): bbox = label["bbox"] assert len(bbox) == 4 xmin, ymin, xmax, ymax = bbox center, scale = _box_to_center_scale(xmin, ymin, xmax - xmin, ymax - ymin, self.aspect_ratio) score = label.get("score", 1) h, w = self.image_size trans = get_affine_transform(center, scale, 0, [w, h]) img = cv2.warpAffine(src.asnumpy(), trans, (int(w), int(h)), flags=cv2.INTER_LINEAR) img = mx.nd.image.to_tensor(mx.nd.array(img)) img = mx.nd.image.normalize(img, mean=self.mean, std=self.std) return img, scale, center, score def _box_to_center_scale(x, y, w, h, aspect_ratio=1.0, scale_mult=1.25): pixel_std = 1 center = np.zeros((2,), dtype=np.float32) center[0] = x + w * 0.5 center[1] = y + h * 0.5 if w > aspect_ratio * h: h = w / aspect_ratio elif w < aspect_ratio * h: w = h * aspect_ratio scale = np.array( [w * 1.0 / pixel_std, h * 1.0 / pixel_std], dtype=np.float32) if center[0] != -1: scale = scale * scale_mult return center, scale def get_dir(src_point, rot_rad): sn, cs = np.sin(rot_rad), np.cos(rot_rad) src_result = [0, 0] src_result[0] = src_point[0] * cs - src_point[1] * sn src_result[1] = src_point[0] * sn + src_point[1] * cs return src_result def crop(img, center, scale, output_size, rot=0): trans = get_affine_transform(center, scale, rot, output_size) dst_img = cv2.warpAffine( img, trans, (int(output_size[0]), int(output_size[1])), flags=cv2.INTER_LINEAR) return dst_img def get_3rd_point(a, b): direct = a - b return b + np.array([-direct[1], direct[0]], dtype=np.float32) def get_affine_transform(center, scale, rot, output_size, shift=np.array([0, 0], dtype=np.float32), inv=0): if not isinstance(scale, np.ndarray) and not isinstance(scale, list): scale = np.array([scale, scale]) scale_tmp = scale src_w = scale_tmp[0] dst_w = output_size[0] dst_h = output_size[1] rot_rad = np.pi * rot / 180 src_dir = get_dir([0, src_w * -0.5], rot_rad) dst_dir = np.array([0, dst_w * -0.5], np.float32) src = np.zeros((3, 2), dtype=np.float32) dst = np.zeros((3, 2), dtype=np.float32) src[0, :] = center + scale_tmp * shift src[1, :] = center + src_dir + scale_tmp * shift dst[0, :] = [dst_w * 0.5, dst_h * 0.5] dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5]) + dst_dir src[2:, :] = get_3rd_point(src[0, :], src[1, :]) dst[2:, :] = get_3rd_point(dst[0, :], dst[1, :]) if inv: trans = cv2.getAffineTransform(np.float32(dst), np.float32(src)) else: trans = cv2.getAffineTransform(np.float32(src), np.float32(dst)) return trans # --------------------------------------------------------------------------------------------------------------------- class CocoHpeValTransform2(object): def __init__(self, ds_metainfo): self.ds_metainfo = ds_metainfo self.image_size = self.ds_metainfo.input_image_size height = self.image_size[0] width = self.image_size[1] self.aspect_ratio = float(width / height) self.mean = ds_metainfo.mean_rgb self.std = ds_metainfo.std_rgb def __call__(self, src, label): # print(src.shape) src = src.asnumpy() bbox = label["bbox"] assert len(bbox) == 4 score = label.get('score', 1) img, scale_box = detector_to_alpha_pose( src, class_ids=mx.nd.array([[0.]]), scores=mx.nd.array([[1.]]), bounding_boxs=mx.nd.array(np.array([bbox])), output_shape=self.image_size) if scale_box.shape[0] == 1: pt1 = np.array(scale_box[0, (0, 1)], dtype=np.float32) pt2 = np.array(scale_box[0, (2, 3)], dtype=np.float32) else: assert scale_box.shape[0] == 4 pt1 = np.array(scale_box[(0, 1)], dtype=np.float32) pt2 = np.array(scale_box[(2, 3)], dtype=np.float32) return img[0], pt1, pt2, score def detector_to_alpha_pose(img, class_ids, scores, bounding_boxs, output_shape=(256, 192), thr=0.5): boxes, scores = alpha_pose_detection_processor( img=img, boxes=bounding_boxs, class_idxs=class_ids, scores=scores, thr=thr) pose_input, upscale_bbox = alpha_pose_image_cropper( source_img=img, boxes=boxes, output_shape=output_shape) return pose_input, upscale_bbox def alpha_pose_detection_processor(img, boxes, class_idxs, scores, thr=0.5): if len(boxes.shape) == 3: boxes = boxes.squeeze(axis=0) if len(class_idxs.shape) == 3: class_idxs = class_idxs.squeeze(axis=0) if len(scores.shape) == 3: scores = scores.squeeze(axis=0) # cilp coordinates boxes[:, [0, 2]] = mx.nd.clip(boxes[:, [0, 2]], 0., img.shape[1] - 1) boxes[:, [1, 3]] = mx.nd.clip(boxes[:, [1, 3]], 0., img.shape[0] - 1) # select boxes mask1 = (class_idxs == 0).asnumpy() mask2 = (scores > thr).asnumpy() picked_idxs = np.where((mask1 + mask2) > 1)[0] if picked_idxs.shape[0] == 0: return None, None else: return boxes[picked_idxs], scores[picked_idxs] def alpha_pose_image_cropper(source_img, boxes, output_shape=(256, 192)): if boxes is None: return None, boxes # crop person poses img_width, img_height = source_img.shape[1], source_img.shape[0] tensors = mx.nd.zeros([boxes.shape[0], 3, output_shape[0], output_shape[1]]) out_boxes = np.zeros([boxes.shape[0], 4]) for i, box in enumerate(boxes.asnumpy()): img = source_img.copy() box_width = box[2] - box[0] box_height = box[3] - box[1] if box_width > 100: scale_rate = 0.2 else: scale_rate = 0.3 # crop image left = int(max(0, box[0] - box_width * scale_rate / 2)) up = int(max(0, box[1] - box_height * scale_rate / 2)) right = int(min(img_width - 1, max(left + 5, box[2] + box_width * scale_rate / 2))) bottom = int(min(img_height - 1, max(up + 5, box[3] + box_height * scale_rate / 2))) crop_width = right - left if crop_width < 1: continue crop_height = bottom - up if crop_height < 1: continue ul = np.array((left, up)) br = np.array((right, bottom)) img = cv_cropBox(img, ul, br, output_shape[0], output_shape[1]) img = mx.nd.image.to_tensor(mx.nd.array(img)) # img = img.transpose((2, 0, 1)) img[0] = img[0] - 0.406 img[1] = img[1] - 0.457 img[2] = img[2] - 0.480 assert (img.shape[0] == 3) tensors[i] = img out_boxes[i] = (left, up, right, bottom) return tensors, out_boxes def cv_cropBox(img, ul, br, resH, resW, pad_val=0): ul = ul br = (br - 1) # br = br.int() lenH = max((br[1] - ul[1]).item(), (br[0] - ul[0]).item() * resH / resW) lenW = lenH * resW / resH if img.ndim == 2: img = img[:, np.newaxis] box_shape = [br[1] - ul[1], br[0] - ul[0]] pad_size = [(lenH - box_shape[0]) // 2, (lenW - box_shape[1]) // 2] # Padding Zeros img[:ul[1], :, :], img[:, :ul[0], :] = pad_val, pad_val img[br[1] + 1:, :, :], img[:, br[0] + 1:, :] = pad_val, pad_val src = np.zeros((3, 2), dtype=np.float32) dst = np.zeros((3, 2), dtype=np.float32) src[0, :] = np.array([ul[0] - pad_size[1], ul[1] - pad_size[0]], np.float32) src[1, :] = np.array([br[0] + pad_size[1], br[1] + pad_size[0]], np.float32) dst[0, :] = 0 dst[1, :] = np.array([resW - 1, resH - 1], np.float32) src[2:, :] = get_3rd_point(src[0, :], src[1, :]) dst[2:, :] = get_3rd_point(dst[0, :], dst[1, :]) trans = cv2.getAffineTransform(np.float32(src), np.float32(dst)) dst_img = cv2.warpAffine(img, trans, (resW, resH), flags=cv2.INTER_LINEAR) return dst_img # --------------------------------------------------------------------------------------------------------------------- def recalc_pose1(keypoints, bbs, image_size): def transform_preds(coords, center, scale, output_size): def affine_transform(pt, t): new_pt = np.array([pt[0], pt[1], 1.]).T new_pt = np.dot(t, new_pt) return new_pt[:2] target_coords = np.zeros(coords.shape) trans = get_affine_transform(center, scale, 0, output_size, inv=1) for p in range(coords.shape[0]): target_coords[p, 0:2] = affine_transform(coords[p, 0:2], trans) return target_coords center = bbs[:, :2] scale = bbs[:, 2:4] heatmap_height = image_size[0] // 4 heatmap_width = image_size[1] // 4 output_size = [heatmap_width, heatmap_height] preds = np.zeros_like(keypoints) for i in range(keypoints.shape[0]): preds[i] = transform_preds(keypoints[i], center[i], scale[i], output_size) return preds def recalc_pose1b(pred, label, image_size, visible_conf_threshold=0.0): label_img_id = label[:, 0].astype(np.int32) label_score = label[:, 1] label_bbs = label[:, 2:6] pred_keypoints = pred[:, :, :2] pred_score = pred[:, :, 2] pred[:, :, :2] = recalc_pose1(pred_keypoints, label_bbs, image_size) pred_person_score = [] batch = pred_keypoints.shape[0] num_joints = pred_keypoints.shape[1] for idx in range(batch): kpt_score = 0 count = 0 for i in range(num_joints): mval = float(pred_score[idx][i]) if mval > visible_conf_threshold: kpt_score += mval count += 1 if count > 0: kpt_score /= count kpt_score = kpt_score * float(label_score[idx]) pred_person_score.append(kpt_score) return pred, pred_person_score, label_img_id def recalc_pose2(keypoints, bbs, image_size): def transformBoxInvert(pt, ul, br, resH, resW): center = np.zeros(2) center[0] = (br[0] - 1 - ul[0]) / 2 center[1] = (br[1] - 1 - ul[1]) / 2 lenH = max(br[1] - ul[1], (br[0] - ul[0]) * resH / resW) lenW = lenH * resW / resH _pt = (pt * lenH) / resH if bool(((lenW - 1) / 2 - center[0]) > 0): _pt[0] = _pt[0] - ((lenW - 1) / 2 - center[0]) if bool(((lenH - 1) / 2 - center[1]) > 0): _pt[1] = _pt[1] - ((lenH - 1) / 2 - center[1]) new_point = np.zeros(2) new_point[0] = _pt[0] + ul[0] new_point[1] = _pt[1] + ul[1] return new_point pt2 = bbs[:, :2] pt1 = bbs[:, 2:4] heatmap_height = image_size[0] // 4 heatmap_width = image_size[1] // 4 preds = np.zeros_like(keypoints) for i in range(keypoints.shape[0]): for j in range(keypoints.shape[1]): preds[i, j] = transformBoxInvert(keypoints[i, j], pt1[i], pt2[i], heatmap_height, heatmap_width) return preds def recalc_pose2b(pred, label, image_size, visible_conf_threshold=0.0): label_img_id = label[:, 0].astype(np.int32) label_score = label[:, 1] label_bbs = label[:, 2:6] pred_keypoints = pred[:, :, :2] pred_score = pred[:, :, 2] pred[:, :, :2] = recalc_pose2(pred_keypoints, label_bbs, image_size) pred_person_score = [] batch = pred_keypoints.shape[0] num_joints = pred_keypoints.shape[1] for idx in range(batch): kpt_score = 0 count = 0 for i in range(num_joints): mval = float(pred_score[idx][i]) if mval > visible_conf_threshold: kpt_score += mval count += 1 if count > 0: kpt_score /= count kpt_score = kpt_score * float(label_score[idx]) pred_person_score.append(kpt_score) return pred, pred_person_score, label_img_id # --------------------------------------------------------------------------------------------------------------------- class CocoHpe1MetaInfo(DatasetMetaInfo): def __init__(self): super(CocoHpe1MetaInfo, self).__init__() self.label = "COCO" self.short_label = "coco" self.root_dir_name = "coco" self.dataset_class = CocoHpe1Dataset self.num_training_samples = None self.in_channels = 3 self.num_classes = 17 self.input_image_size = (256, 192) self.train_metric_capts = None self.train_metric_names = None self.train_metric_extra_kwargs = None self.val_metric_capts = None self.val_metric_names = None self.test_metric_capts = ["Val.CocoOksAp"] self.test_metric_names = ["CocoHpeOksApMetric"] self.test_metric_extra_kwargs = [ {"name": "OksAp", "coco_annotations_file_path": None, "use_file": False, "pose_postprocessing_fn": lambda x, y: recalc_pose1b(x, y, self.input_image_size)}] self.saver_acc_ind = 0 self.do_transform = True self.val_transform = CocoHpeValTransform1 self.test_transform = CocoHpeValTransform1 self.ml_type = "hpe" self.allow_hybridize = False self.test_net_extra_kwargs = {"fixed_size": False} self.mean_rgb = (0.485, 0.456, 0.406) self.std_rgb = (0.229, 0.224, 0.225) self.model_type = 1 def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for ImageNet-1K dataset metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(CocoHpe1MetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--input-size", type=int, nargs=2, default=self.input_image_size, help="size of the input for model") parser.add_argument( "--model-type", type=int, default=self.model_type, help="model type (1=SimplePose, 2=AlphaPose)") def update(self, args): """ Update ImageNet-1K dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(CocoHpe1MetaInfo, self).update(args) self.input_image_size = args.input_size self.model_type = args.model_type if self.model_type == 1: self.test_metric_extra_kwargs[0]["pose_postprocessing_fn"] =\ lambda x, y: recalc_pose1b(x, y, self.input_image_size) self.val_transform = CocoHpeValTransform1 self.test_transform = CocoHpeValTransform1 else: self.test_metric_extra_kwargs[0]["pose_postprocessing_fn"] =\ lambda x, y: recalc_pose2b(x, y, self.input_image_size) self.val_transform = CocoHpeValTransform2 self.test_transform = CocoHpeValTransform2 def update_from_dataset(self, dataset): """ Update dataset metainfo after a dataset class instance creation. Parameters: ---------- args : obj A dataset class instance. """ self.test_metric_extra_kwargs[0]["coco_annotations_file_path"] = dataset.annotations_file_path
28,936
34.289024
119
py
imgclsmob
imgclsmob-master/gluon/datasets/widerface_det_dataset.py
""" WIDER FACE detection dataset. """ import os import cv2 import mxnet as mx import numpy as np from mxnet.gluon.data import dataset from .dataset_metainfo import DatasetMetaInfo __all__ = ['WiderfaceDetMetaInfo'] class WiderfaceDetDataset(dataset.Dataset): """ WIDER FACE detection dataset. Parameters: ---------- root : str Path to folder storing the dataset. mode : string, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None): super(WiderfaceDetDataset, self).__init__() self.root = os.path.expanduser(root) self.mode = mode self._transform = transform self.synsets = [] self.items = [] image_dir_path = "{}/WIDER_{}/images".format(self.root, self.mode) for folder in sorted(os.listdir(image_dir_path)): path = os.path.join(root, folder) if not os.path.isdir(path): continue label = len(self.synsets) self.synsets.append(folder) for filename in sorted(os.listdir(path)): filename = os.path.join(path, filename) ext = os.path.splitext(filename)[1] if ext.lower() not in (".jpg",): continue self.items.append((filename, label)) def __len__(self): return len(self.items) def __getitem__(self, idx): img_path = self.items[idx][0] # image = cv2.imread(img_path, flags=cv2.IMREAD_COLOR) image = mx.image.imread(img_path, flag=1).asnumpy() image_size = image.shape[:2] shorter_side = min(image.shape[:2]) resize_scale = 1.0 if shorter_side < 128: resize_scale = 128.0 / shorter_side image = cv2.resize(image, (0, 0), fx=resize_scale, fy=resize_scale) image = image.transpose(2, 0, 1).astype(np.float32) image = mx.nd.array(image) label = "{}/{}/{}/{}/{}".format(self.synsets[self.items[idx][1]], (img_path.split("/")[1]).split(".")[0], resize_scale, image_size[0], image_size[1]) label = np.array(label).copy() if self._transform is not None: image, label = self._transform(image, label) return image, label # --------------------------------------------------------------------------------------------------------------------- class WiderfaceDetValTransform(object): def __init__(self, ds_metainfo): self.ds_metainfo = ds_metainfo def __call__(self, image, label): return image, label # --------------------------------------------------------------------------------------------------------------------- class WiderfaceDetMetaInfo(DatasetMetaInfo): def __init__(self): super(WiderfaceDetMetaInfo, self).__init__() self.label = "WiderFace" self.short_label = "widerface" self.root_dir_name = "WIDER_FACE" self.dataset_class = WiderfaceDetDataset self.num_training_samples = None self.in_channels = 3 self.input_image_size = (480, 640) self.train_metric_capts = None self.train_metric_names = None self.train_metric_extra_kwargs = None self.val_metric_capts = None self.val_metric_names = None self.test_metric_capts = ["WF"] self.test_metric_names = ["WiderfaceDetMetric"] self.test_metric_extra_kwargs = [ {"name": "WF"}] self.saver_acc_ind = 0 self.do_transform = True self.do_transform_first = False self.last_batch = "keep" self.val_transform = WiderfaceDetValTransform self.test_transform = WiderfaceDetValTransform self.ml_type = "det" self.allow_hybridize = False self.test_net_extra_kwargs = None self.model_type = 1 self.receptive_field_center_starts = None self.receptive_field_strides = None self.bbox_factors = None def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for ImageNet-1K dataset metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(WiderfaceDetMetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--model-type", type=int, default=self.model_type, help="model type (1=320, 2=560)") def update(self, args): """ Update ImageNet-1K dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(WiderfaceDetMetaInfo, self).update(args) self.model_type = args.model_type if self.model_type == 1: self.receptive_field_center_starts = [3, 7, 15, 31, 63] self.receptive_field_strides = [4, 8, 16, 32, 64] self.bbox_factors = [10.0, 20.0, 40.0, 80.0, 160.0] else: self.receptive_field_center_starts = [3, 3, 7, 7, 15, 31, 31, 31] self.receptive_field_strides = [4, 4, 8, 8, 16, 32, 32, 32] self.bbox_factors = [7.5, 10.0, 20.0, 35.0, 55.0, 125.0, 200.0, 280.0]
5,676
32.791667
119
py
imgclsmob
imgclsmob-master/gluon/datasets/coco_det_dataset.py
""" MS COCO object detection dataset. """ __all__ = ['CocoDetMetaInfo'] import os import cv2 import logging import mxnet as mx import numpy as np from PIL import Image from mxnet.gluon.data import dataset from .dataset_metainfo import DatasetMetaInfo class CocoDetDataset(dataset.Dataset): """ MS COCO detection dataset. Parameters: ---------- root : str Path to folder storing the dataset. mode : string, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. splits : list of str, default ['instances_val2017'] Json annotations name. Candidates can be: instances_val2017, instances_train2017. min_object_area : float Minimum accepted ground-truth area, if an object's area is smaller than this value, it will be ignored. skip_empty : bool, default is True Whether skip images with no valid object. This should be `True` in training, otherwise it will cause undefined behavior. use_crowd : bool, default is True Whether use boxes labeled as crowd instance. """ CLASSES = ['person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush'] def __init__(self, root, mode="train", transform=None, splits=('instances_val2017',), min_object_area=0, skip_empty=True, use_crowd=True): super(CocoDetDataset, self).__init__() self._root = os.path.expanduser(root) self.mode = mode self._transform = transform self.num_class = len(self.CLASSES) self._min_object_area = min_object_area self._skip_empty = skip_empty self._use_crowd = use_crowd if isinstance(splits, mx.base.string_types): splits = [splits] self._splits = splits self.index_map = dict(zip(type(self).CLASSES, range(self.num_class))) self.json_id_to_contiguous = None self.contiguous_id_to_json = None self._coco = [] self._items, self._labels, self._im_aspect_ratios = self._load_jsons() mode_name = "train" if mode == "train" else "val" annotations_dir_path = os.path.join(root, "annotations") annotations_file_path = os.path.join(annotations_dir_path, "instances_" + mode_name + "2017.json") self.annotations_file_path = annotations_file_path def __str__(self): detail = ','.join([str(s) for s in self._splits]) return self.__class__.__name__ + '(' + detail + ')' @property def coco(self): """ Return pycocotools object for evaluation purposes. """ if not self._coco: raise ValueError("No coco objects found, dataset not initialized.") if len(self._coco) > 1: raise NotImplementedError( "Currently we don't support evaluating {} JSON files. \ Please use single JSON dataset and evaluate one by one".format(len(self._coco))) return self._coco[0] @property def classes(self): """ Category names. """ return type(self).CLASSES @property def annotation_dir(self): """ The subdir for annotations. Default is 'annotations'(coco default) For example, a coco format json file will be searched as 'root/annotation_dir/xxx.json' You can override if custom dataset don't follow the same pattern """ return 'annotations' def get_im_aspect_ratio(self): """Return the aspect ratio of each image in the order of the raw data.""" if self._im_aspect_ratios is not None: return self._im_aspect_ratios self._im_aspect_ratios = [None] * len(self._items) for i, img_path in enumerate(self._items): with Image.open(img_path) as im: w, h = im.size self._im_aspect_ratios[i] = 1.0 * w / h return self._im_aspect_ratios def _parse_image_path(self, entry): """How to parse image dir and path from entry. Parameters: ---------- entry : dict COCO entry, e.g. including width, height, image path, etc.. Returns: ------- abs_path : str Absolute path for corresponding image. """ dirname, filename = entry["coco_url"].split("/")[-2:] abs_path = os.path.join(self._root, dirname, filename) return abs_path def __len__(self): return len(self._items) def __getitem__(self, idx): img_path = self._items[idx] label = self._labels[idx] img = mx.image.imread(img_path, 1) label = np.array(label).copy() if self._transform is not None: img, label = self._transform(img, label) return img, label def _load_jsons(self): """ Load all image paths and labels from JSON annotation files into buffer. """ items = [] labels = [] im_aspect_ratios = [] from pycocotools.coco import COCO for split in self._splits: anno = os.path.join(self._root, self.annotation_dir, split) + ".json" _coco = COCO(anno) self._coco.append(_coco) classes = [c["name"] for c in _coco.loadCats(_coco.getCatIds())] if not classes == self.classes: raise ValueError("Incompatible category names with COCO: ") assert classes == self.classes json_id_to_contiguous = { v: k for k, v in enumerate(_coco.getCatIds())} if self.json_id_to_contiguous is None: self.json_id_to_contiguous = json_id_to_contiguous self.contiguous_id_to_json = { v: k for k, v in self.json_id_to_contiguous.items()} else: assert self.json_id_to_contiguous == json_id_to_contiguous # iterate through the annotations image_ids = sorted(_coco.getImgIds()) for entry in _coco.loadImgs(image_ids): abs_path = self._parse_image_path(entry) if not os.path.exists(abs_path): raise IOError("Image: {} not exists.".format(abs_path)) label = self._check_load_bbox(_coco, entry) if not label: continue im_aspect_ratios.append(float(entry["width"]) / entry["height"]) items.append(abs_path) labels.append(label) return items, labels, im_aspect_ratios def _check_load_bbox(self, coco, entry): """ Check and load ground-truth labels. """ entry_id = entry['id'] # fix pycocotools _isArrayLike which don't work for str in python3 entry_id = [entry_id] if not isinstance(entry_id, (list, tuple)) else entry_id ann_ids = coco.getAnnIds(imgIds=entry_id, iscrowd=None) objs = coco.loadAnns(ann_ids) # check valid bboxes valid_objs = [] width = entry["width"] height = entry["height"] for obj in objs: if obj["area"] < self._min_object_area: continue if obj.get("ignore", 0) == 1: continue if not self._use_crowd and obj.get("iscrowd", 0): continue # convert from (x, y, w, h) to (xmin, ymin, xmax, ymax) and clip bound xmin, ymin, xmax, ymax = self.bbox_clip_xyxy(self.bbox_xywh_to_xyxy(obj["bbox"]), width, height) # require non-zero box area if obj["area"] > 0 and xmax > xmin and ymax > ymin: contiguous_cid = self.json_id_to_contiguous[obj["category_id"]] valid_objs.append([xmin, ymin, xmax, ymax, contiguous_cid]) if not valid_objs: if not self._skip_empty: # dummy invalid labels if no valid objects are found valid_objs.append([-1, -1, -1, -1, -1]) return valid_objs @staticmethod def bbox_clip_xyxy(xyxy, width, height): """ Clip bounding box with format (xmin, ymin, xmax, ymax) to specified boundary. All bounding boxes will be clipped to the new region `(0, 0, width, height)`. Parameters: ---------- xyxy : list, tuple or numpy.ndarray The bbox in format (xmin, ymin, xmax, ymax). If numpy.ndarray is provided, we expect multiple bounding boxes with shape `(N, 4)`. width : int or float Boundary width. height : int or float Boundary height. Returns: ------- tuple or np.array Description of returned object. """ if isinstance(xyxy, (tuple, list)): if not len(xyxy) == 4: raise IndexError("Bounding boxes must have 4 elements, given {}".format(len(xyxy))) x1 = np.minimum(width - 1, np.maximum(0, xyxy[0])) y1 = np.minimum(height - 1, np.maximum(0, xyxy[1])) x2 = np.minimum(width - 1, np.maximum(0, xyxy[2])) y2 = np.minimum(height - 1, np.maximum(0, xyxy[3])) return x1, y1, x2, y2 elif isinstance(xyxy, np.ndarray): if not xyxy.size % 4 == 0: raise IndexError("Bounding boxes must have n * 4 elements, given {}".format(xyxy.shape)) x1 = np.minimum(width - 1, np.maximum(0, xyxy[:, 0])) y1 = np.minimum(height - 1, np.maximum(0, xyxy[:, 1])) x2 = np.minimum(width - 1, np.maximum(0, xyxy[:, 2])) y2 = np.minimum(height - 1, np.maximum(0, xyxy[:, 3])) return np.hstack((x1, y1, x2, y2)) else: raise TypeError("Expect input xywh a list, tuple or numpy.ndarray, given {}".format(type(xyxy))) @staticmethod def bbox_xywh_to_xyxy(xywh): """ Convert bounding boxes from format (xmin, ymin, w, h) to (xmin, ymin, xmax, ymax) Parameters: ---------- xywh : list, tuple or numpy.ndarray The bbox in format (x, y, w, h). If numpy.ndarray is provided, we expect multiple bounding boxes with shape `(N, 4)`. Returns: ------- tuple or np.ndarray The converted bboxes in format (xmin, ymin, xmax, ymax). If input is numpy.ndarray, return is numpy.ndarray correspondingly. """ if isinstance(xywh, (tuple, list)): if not len(xywh) == 4: raise IndexError("Bounding boxes must have 4 elements, given {}".format(len(xywh))) w, h = np.maximum(xywh[2] - 1, 0), np.maximum(xywh[3] - 1, 0) return xywh[0], xywh[1], xywh[0] + w, xywh[1] + h elif isinstance(xywh, np.ndarray): if not xywh.size % 4 == 0: raise IndexError("Bounding boxes must have n * 4 elements, given {}".format(xywh.shape)) xyxy = np.hstack((xywh[:, :2], xywh[:, :2] + np.maximum(0, xywh[:, 2:4] - 1))) return xyxy else: raise TypeError("Expect input xywh a list, tuple or numpy.ndarray, given {}".format(type(xywh))) # --------------------------------------------------------------------------------------------------------------------- class CocoDetValTransform(object): def __init__(self, ds_metainfo): self.ds_metainfo = ds_metainfo self.image_size = self.ds_metainfo.input_image_size self._height = self.image_size[0] self._width = self.image_size[1] self._mean = np.array(ds_metainfo.mean_rgb, dtype=np.float32).reshape(1, 1, 3) self._std = np.array(ds_metainfo.std_rgb, dtype=np.float32).reshape(1, 1, 3) def __call__(self, src, label): # resize img, bbox = src.asnumpy(), label input_h, input_w = self._height, self._width h, w, _ = src.shape s = max(h, w) * 1.0 c = np.array([w / 2., h / 2.], dtype=np.float32) trans_input = self.get_affine_transform(c, s, 0, [input_w, input_h]) inp = cv2.warpAffine(img, trans_input, (input_w, input_h), flags=cv2.INTER_LINEAR) output_w = input_w output_h = input_h trans_output = self.get_affine_transform(c, s, 0, [output_w, output_h]) for i in range(bbox.shape[0]): bbox[i, :2] = self.affine_transform(bbox[i, :2], trans_output) bbox[i, 2:4] = self.affine_transform(bbox[i, 2:4], trans_output) bbox[:, :2] = np.clip(bbox[:, :2], 0, output_w - 1) bbox[:, 2:4] = np.clip(bbox[:, 2:4], 0, output_h - 1) img = inp # to tensor img = img.astype(np.float32) / 255.0 img = (img - self._mean) / self._std img = img.transpose(2, 0, 1).astype(np.float32) img = mx.nd.array(img) return img, bbox.astype(img.dtype) @staticmethod def get_affine_transform(center, scale, rot, output_size, shift=np.array([0, 0], dtype=np.float32), inv=0): """ Get affine transform matrix given center, scale and rotation. Parameters: ---------- center : tuple of float Center point. scale : float Scaling factor. rot : float Rotation degree. output_size : tuple of int (width, height) of the output size. shift : float Shift factor. inv : bool Whether inverse the computation. Returns: ------- numpy.ndarray Affine matrix. """ if not isinstance(scale, np.ndarray) and not isinstance(scale, list): scale = np.array([scale, scale], dtype=np.float32) scale_tmp = scale src_w = scale_tmp[0] dst_w = output_size[0] dst_h = output_size[1] rot_rad = np.pi * rot / 180 src_dir = CocoDetValTransform.get_rot_dir([0, src_w * -0.5], rot_rad) dst_dir = np.array([0, dst_w * -0.5], np.float32) src = np.zeros((3, 2), dtype=np.float32) dst = np.zeros((3, 2), dtype=np.float32) src[0, :] = center + scale_tmp * shift src[1, :] = center + src_dir + scale_tmp * shift dst[0, :] = [dst_w * 0.5, dst_h * 0.5] dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5], np.float32) + dst_dir src[2:, :] = CocoDetValTransform.get_3rd_point(src[0, :], src[1, :]) dst[2:, :] = CocoDetValTransform.get_3rd_point(dst[0, :], dst[1, :]) if inv: trans = cv2.getAffineTransform(np.float32(dst), np.float32(src)) else: trans = cv2.getAffineTransform(np.float32(src), np.float32(dst)) return trans @staticmethod def get_rot_dir(src_point, rot_rad): """ Get rotation direction. Parameters: ---------- src_point : tuple of float Original point. rot_rad : float Rotation radian. Returns: ------- tuple of float Rotation. """ sn, cs = np.sin(rot_rad), np.cos(rot_rad) src_result = [0, 0] src_result[0] = src_point[0] * cs - src_point[1] * sn src_result[1] = src_point[0] * sn + src_point[1] * cs return src_result @staticmethod def get_3rd_point(a, b): """ Get the 3rd point position given first two points. Parameters: ---------- a : tuple of float First point. b : tuple of float Second point. Returns: ------- tuple of float Third point. """ direct = a - b return b + np.array([-direct[1], direct[0]], dtype=np.float32) @staticmethod def affine_transform(pt, t): """ Apply affine transform to a bounding box given transform matrix t. Parameters: ---------- pt : numpy.ndarray Bounding box with shape (1, 2). t : numpy.ndarray Transformation matrix with shape (2, 3). Returns: ------- numpy.ndarray New bounding box with shape (1, 2). """ new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32).T new_pt = np.dot(t, new_pt) return new_pt[:2] class Tuple(object): """ Wrap multiple batchify functions to form a function apply each input function on each input fields respectively. """ def __init__(self, fn, *args): if isinstance(fn, (list, tuple)): self._fn = fn else: self._fn = (fn,) + args def __call__(self, data): """ Batchify the input data. Parameters: ---------- data : list The samples to batchfy. Each sample should contain N attributes. Returns: ------- tuple A tuple of length N. Contains the batchified result of each attribute in the input. """ ret = [] for i, ele_fn in enumerate(self._fn): ret.append(ele_fn([ele[i] for ele in data])) return tuple(ret) class Stack(object): """ Stack the input data samples to construct the batch. """ def __call__(self, data): """ Batchify the input data. Parameters: ---------- data : list The input data samples Returns: ------- NDArray Result. """ return self._stack_arrs(data, True) @staticmethod def _stack_arrs(arrs, use_shared_mem=False): """ Internal imple for stacking arrays. """ if isinstance(arrs[0], mx.nd.NDArray): if use_shared_mem: out = mx.nd.empty((len(arrs),) + arrs[0].shape, dtype=arrs[0].dtype, ctx=mx.Context("cpu_shared", 0)) return mx.nd.stack(*arrs, out=out) else: return mx.nd.stack(*arrs) else: out = np.asarray(arrs) if use_shared_mem: return mx.nd.array(out, ctx=mx.Context("cpu_shared", 0)) else: return mx.nd.array(out) class Pad(object): """ Pad the input ndarrays along the specific padding axis and stack them to get the output. """ def __init__(self, axis=0, pad_val=0, num_shards=1, ret_length=False): self._axis = axis self._pad_val = pad_val self._num_shards = num_shards self._ret_length = ret_length def __call__(self, data): """ Batchify the input data. Parameters: ---------- data : list A list of N samples. Each sample can be 1) ndarray or 2) a list/tuple of ndarrays Returns: ------- NDArray Data in the minibatch. Shape is (N, ...) NDArray, optional The sequences' original lengths at the padded axis. Shape is (N,). This will only be returned in `ret_length` is True. """ if isinstance(data[0], (mx.nd.NDArray, np.ndarray, list)): padded_arr, original_length = self._pad_arrs_to_max_length( data, self._axis, self._pad_val, self._num_shards, True) if self._ret_length: return padded_arr, original_length else: return padded_arr else: raise NotImplementedError @staticmethod def _pad_arrs_to_max_length(arrs, pad_axis, pad_val, num_shards=1, use_shared_mem=False): """ Inner Implementation of the Pad batchify. """ if not isinstance(arrs[0], (mx.nd.NDArray, np.ndarray)): arrs = [np.asarray(ele) for ele in arrs] if isinstance(pad_axis, tuple): original_length = [] for axis in pad_axis: original_length.append(np.array([ele.shape[axis] for ele in arrs])) original_length = np.stack(original_length).T else: original_length = np.array([ele.shape[pad_axis] for ele in arrs]) pad_axis = [pad_axis] if len(original_length) % num_shards != 0: logging.warning( 'Batch size cannot be evenly split. Trying to shard %d items into %d shards', len(original_length), num_shards) original_length = np.array_split(original_length, num_shards) max_lengths = [np.max(ll, axis=0, keepdims=len(pad_axis) == 1) for ll in original_length] # add batch dimension ret_shape = [[ll.shape[0], ] + list(arrs[0].shape) for ll in original_length] for i, shape in enumerate(ret_shape): for j, axis in enumerate(pad_axis): shape[1 + axis] = max_lengths[i][j] if use_shared_mem: ret = [mx.nd.full(shape=tuple(shape), val=pad_val, ctx=mx.Context('cpu_shared', 0), dtype=arrs[0].dtype) for shape in ret_shape] original_length = [mx.nd.array(ll, ctx=mx.Context('cpu_shared', 0), dtype=np.int32) for ll in original_length] else: ret = [mx.nd.full(shape=tuple(shape), val=pad_val, dtype=arrs[0].dtype) for shape in ret_shape] original_length = [mx.nd.array(ll, dtype=np.int32) for ll in original_length] for i, arr in enumerate(arrs): if ret[i // ret[0].shape[0]].shape[1:] == arr.shape: ret[i // ret[0].shape[0]][i % ret[0].shape[0]] = arr else: slices = [slice(0, ll) for ll in arr.shape] ret[i // ret[0].shape[0]][i % ret[0].shape[0]][tuple(slices)] = arr if len(ret) == len(original_length) == 1: return ret[0], original_length[0] return ret, original_length def get_post_transform(orig_w, orig_h, out_w, out_h): """Get the post prediction affine transforms. This will be used to adjust the prediction results according to original coco image resolutions. Parameters: ---------- orig_w : int Original width of the image. orig_h : int Original height of the image. out_w : int Width of the output image after prediction. out_h : int Height of the output image after prediction. Returns: ------- numpy.ndarray Affine transform matrix 3x2. """ s = max(orig_w, orig_h) * 1.0 c = np.array([orig_w / 2., orig_h / 2.], dtype=np.float32) trans_output = CocoDetValTransform.get_affine_transform(c, s, 0, [out_w, out_h], inv=True) return trans_output class CocoDetMetaInfo(DatasetMetaInfo): def __init__(self): super(CocoDetMetaInfo, self).__init__() self.label = "COCO" self.short_label = "coco" self.root_dir_name = "coco" self.dataset_class = CocoDetDataset self.num_training_samples = None self.in_channels = 3 self.num_classes = CocoDetDataset.classes self.input_image_size = (512, 512) self.train_metric_capts = None self.train_metric_names = None self.train_metric_extra_kwargs = None self.val_metric_capts = None self.val_metric_names = None self.test_metric_capts = ["Val.mAP"] self.test_metric_names = ["CocoDetMApMetric"] self.test_metric_extra_kwargs = [ {"name": "mAP", "img_height": 512, "coco_annotations_file_path": None, "contiguous_id_to_json": None, "data_shape": None, "post_affine": get_post_transform}] self.dataset_class_extra_kwargs = {"skip_empty": False} self.saver_acc_ind = 0 self.do_transform = True self.do_transform_first = False self.last_batch = "keep" self.batchify_fn = Tuple(Stack(), Pad(pad_val=-1)) self.val_transform = CocoDetValTransform self.test_transform = CocoDetValTransform self.ml_type = "det" self.allow_hybridize = False self.test_net_extra_kwargs = None self.mean_rgb = (0.485, 0.456, 0.406) self.std_rgb = (0.229, 0.224, 0.225) def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for ImageNet-1K dataset metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(CocoDetMetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--input-size", type=int, nargs=2, default=self.input_image_size, help="size of the input for model") def update(self, args): """ Update ImageNet-1K dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(CocoDetMetaInfo, self).update(args) self.input_image_size = args.input_size self.test_metric_extra_kwargs[0]["img_height"] = self.input_image_size[0] self.test_metric_extra_kwargs[0]["data_shape"] = self.input_image_size def update_from_dataset(self, dataset): """ Update dataset metainfo after a dataset class instance creation. Parameters: ---------- args : obj A dataset class instance. """ self.test_metric_extra_kwargs[0]["coco_annotations_file_path"] = dataset.annotations_file_path self.test_metric_extra_kwargs[0]["contiguous_id_to_json"] = dataset.contiguous_id_to_json
27,151
35.691892
119
py
imgclsmob
imgclsmob-master/gluon/datasets/ade20k_seg_dataset.py
""" ADE20K semantic segmentation dataset. """ import os import numpy as np import mxnet as mx from PIL import Image from .seg_dataset import SegDataset from .voc_seg_dataset import VOCMetaInfo class ADE20KSegDataset(SegDataset): """ ADE20K semantic segmentation dataset. Parameters: ---------- root : str Path to a folder with `ADEChallengeData2016` subfolder. mode : str, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None, **kwargs): super(ADE20KSegDataset, self).__init__( root=root, mode=mode, transform=transform, **kwargs) base_dir_path = os.path.join(root, "ADEChallengeData2016") assert os.path.exists(base_dir_path), "Please prepare dataset" image_dir_path = os.path.join(base_dir_path, "images") mask_dir_path = os.path.join(base_dir_path, "annotations") mode_dir_name = "training" if mode == "train" else "validation" image_dir_path = os.path.join(image_dir_path, mode_dir_name) mask_dir_path = os.path.join(mask_dir_path, mode_dir_name) self.images = [] self.masks = [] for image_file_name in os.listdir(image_dir_path): image_file_stem, _ = os.path.splitext(image_file_name) if image_file_name.endswith(".jpg"): image_file_path = os.path.join(image_dir_path, image_file_name) mask_file_name = image_file_stem + ".png" mask_file_path = os.path.join(mask_dir_path, mask_file_name) if os.path.isfile(mask_file_path): self.images.append(image_file_path) self.masks.append(mask_file_path) else: print("Cannot find the mask: {}".format(mask_file_path)) assert (len(self.images) == len(self.masks)) if len(self.images) == 0: raise RuntimeError("Found 0 images in subfolders of: {}\n".format(base_dir_path)) def __getitem__(self, index): image = Image.open(self.images[index]).convert("RGB") # image = mx.image.imread(self.images[index]) if self.mode == "demo": image = self._img_transform(image) if self.transform is not None: image = self.transform(image) return image, os.path.basename(self.images[index]) mask = Image.open(self.masks[index]) # mask = mx.image.imread(self.masks[index]) if self.mode == "train": image, mask = self._train_sync_transform(image, mask) elif self.mode == "val": image, mask = self._val_sync_transform(image, mask) else: assert (self.mode == "test") image = self._img_transform(image) mask = self._mask_transform(mask) if self.transform is not None: image = self.transform(image) return image, mask classes = 150 vague_idx = 150 use_vague = True background_idx = -1 ignore_bg = False @staticmethod def _mask_transform(mask): np_mask = np.array(mask).astype(np.int32) np_mask[np_mask == 0] = ADE20KSegDataset.vague_idx + 1 np_mask -= 1 return mx.nd.array(np_mask, mx.cpu()) def __len__(self): return len(self.images) class ADE20KMetaInfo(VOCMetaInfo): def __init__(self): super(ADE20KMetaInfo, self).__init__() self.label = "ADE20K" self.short_label = "voc" self.root_dir_name = "ade20k" self.dataset_class = ADE20KSegDataset self.num_classes = ADE20KSegDataset.classes self.test_metric_extra_kwargs = [ {"vague_idx": ADE20KSegDataset.vague_idx, "use_vague": ADE20KSegDataset.use_vague, "macro_average": False}, {"num_classes": ADE20KSegDataset.classes, "vague_idx": ADE20KSegDataset.vague_idx, "use_vague": ADE20KSegDataset.use_vague, "bg_idx": ADE20KSegDataset.background_idx, "ignore_bg": ADE20KSegDataset.ignore_bg, "macro_average": False}]
4,339
34
93
py
imgclsmob
imgclsmob-master/gluon/datasets/dataset_metainfo.py
""" Base dataset metainfo class. """ import os class DatasetMetaInfo(object): """ Base descriptor of dataset. """ def __init__(self): self.use_imgrec = False self.do_transform = False self.do_transform_first = True self.last_batch = None self.batchify_fn = None self.label = None self.root_dir_name = None self.root_dir_path = None self.dataset_class = None self.dataset_class_extra_kwargs = None self.num_training_samples = None self.in_channels = None self.num_classes = None self.input_image_size = None self.train_metric_capts = None self.train_metric_names = None self.train_metric_extra_kwargs = None self.train_use_weighted_sampler = False self.val_metric_capts = None self.val_metric_names = None self.val_metric_extra_kwargs = None self.test_metric_capts = None self.test_metric_names = None self.test_metric_extra_kwargs = None self.saver_acc_ind = None self.ml_type = None self.allow_hybridize = True self.train_net_extra_kwargs = {"root": os.path.join("~", ".mxnet", "models")} self.test_net_extra_kwargs = None self.load_ignore_extra = False self.loss_name = None self.loss_extra_kwargs = None def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for dataset specific metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ parser.add_argument( "--data-dir", type=str, default=os.path.join(work_dir_path, self.root_dir_name), help="path to directory with {} dataset".format(self.label)) parser.add_argument( "--num-classes", type=int, default=self.num_classes, help="number of classes") parser.add_argument( "--in-channels", type=int, default=self.in_channels, help="number of input channels") parser.add_argument( "--net-root", type=str, default=os.path.join("~", ".mxnet", "models"), help="root for pretrained net cache") def update(self, args): """ Update dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ self.root_dir_path = args.data_dir self.num_classes = args.num_classes self.in_channels = args.in_channels self.train_net_extra_kwargs["root"] = args.net_root def update_from_dataset(self, dataset): """ Update dataset metainfo after a dataset class instance creation. Parameters: ---------- args : obj A dataset class instance. """ pass
3,226
29.158879
85
py
imgclsmob
imgclsmob-master/gluon/datasets/seg_dataset.py
import random import numpy as np import mxnet as mx from PIL import Image, ImageOps, ImageFilter from mxnet.gluon.data import dataset class SegDataset(dataset.Dataset): """ Segmentation base dataset. Parameters: ---------- root : str Path to data folder. mode : str 'train', 'val', 'test', or 'demo'. transform : callable A function that transforms the image. """ def __init__(self, root, mode, transform, base_size=520, crop_size=480): assert (mode in ("train", "val", "test", "demo")) self.root = root self.mode = mode self.transform = transform self.base_size = base_size self.crop_size = crop_size def _val_sync_transform(self, image, mask, ctx=mx.cpu()): short_size = self.crop_size w, h = image.size if w > h: oh = short_size ow = int(float(w * oh) / h) else: ow = short_size oh = int(float(h * ow) / w) image = image.resize((ow, oh), Image.BILINEAR) mask = mask.resize((ow, oh), Image.NEAREST) # Center crop: outsize = self.crop_size x1 = int(round(0.5 * (ow - outsize))) y1 = int(round(0.5 * (oh - outsize))) image = image.crop((x1, y1, x1 + outsize, y1 + outsize)) mask = mask.crop((x1, y1, x1 + outsize, y1 + outsize)) # Final transform: image, mask = self._img_transform(image, ctx=ctx), self._mask_transform(mask, ctx=ctx) return image, mask def _train_sync_transform(self, image, mask, ctx=mx.cpu()): # Random mirror: if random.random() < 0.5: image = image.transpose(Image.FLIP_LEFT_RIGHT) mask = mask.transpose(Image.FLIP_LEFT_RIGHT) # Random scale (short edge): short_size = random.randint(int(self.base_size * 0.5), int(self.base_size * 2.0)) w, h = image.size if w > h: oh = short_size ow = int(float(w * oh) / h) else: ow = short_size oh = int(float(h * ow) / w) image = image.resize((ow, oh), Image.BILINEAR) mask = mask.resize((ow, oh), Image.NEAREST) # Pad crop: crop_size = self.crop_size if short_size < crop_size: padh = crop_size - oh if oh < crop_size else 0 padw = crop_size - ow if ow < crop_size else 0 image = ImageOps.expand(image, border=(0, 0, padw, padh), fill=0) mask = ImageOps.expand(mask, border=(0, 0, padw, padh), fill=0) # Random crop crop_size: w, h = image.size x1 = random.randint(0, w - crop_size) y1 = random.randint(0, h - crop_size) image = image.crop((x1, y1, x1 + crop_size, y1 + crop_size)) mask = mask.crop((x1, y1, x1 + crop_size, y1 + crop_size)) # Gaussian blur as in PSP: if random.random() < 0.5: image = image.filter(ImageFilter.GaussianBlur(radius=random.random())) # Final transform: image, mask = self._img_transform(image, ctx=ctx), self._mask_transform(mask, ctx=ctx) return image, mask @staticmethod def _img_transform(image, ctx=mx.cpu()): return mx.nd.array(np.array(image), ctx=ctx) @staticmethod def _mask_transform(mask, ctx=mx.cpu()): return mx.nd.array(np.array(mask), ctx=ctx, dtype=np.int32)
3,490
34.622449
94
py
imgclsmob
imgclsmob-master/gluon/datasets/coco_hpe2_dataset.py
""" COCO keypoint detection (2D multiple human pose estimation) dataset (for Lightweight OpenPose). """ import os import json import math import cv2 from operator import itemgetter import numpy as np from mxnet.gluon.data import dataset from .dataset_metainfo import DatasetMetaInfo class CocoHpe2Dataset(dataset.Dataset): """ COCO keypoint detection (2D multiple human pose estimation) dataset. Parameters: ---------- root : string Path to `annotations`, `train2017`, and `val2017` folders. mode : string, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None): super(CocoHpe2Dataset, self).__init__() self._root = os.path.expanduser(root) self.mode = mode self.transform = transform mode_name = "train" if mode == "train" else "val" annotations_dir_path = os.path.join(root, "annotations") annotations_file_path = os.path.join(annotations_dir_path, "person_keypoints_" + mode_name + "2017.json") with open(annotations_file_path, "r") as f: self.file_names = json.load(f)["images"] self.image_dir_path = os.path.join(root, mode_name + "2017") self.annotations_file_path = annotations_file_path def __str__(self): return self.__class__.__name__ + "(" + self._root + ")" def __len__(self): return len(self.file_names) def __getitem__(self, idx): file_name = self.file_names[idx]["file_name"] image_file_path = os.path.join(self.image_dir_path, file_name) image = cv2.imread(image_file_path, flags=cv2.IMREAD_COLOR) # image = cv2.cvtColor(img, code=cv2.COLOR_BGR2RGB) img_mean = (128, 128, 128) img_scale = 1.0 / 256 base_height = 368 stride = 8 pad_value = (0, 0, 0) height, width, _ = image.shape image = self.normalize(image, img_mean, img_scale) ratio = base_height / float(image.shape[0]) image = cv2.resize(image, (0, 0), fx=ratio, fy=ratio, interpolation=cv2.INTER_CUBIC) min_dims = [base_height, max(image.shape[1], base_height)] image, pad = self.pad_width( image, stride, pad_value, min_dims) image = image.astype(np.float32) image = image.transpose((2, 0, 1)) # image = torch.from_numpy(image) # if self.transform is not None: # image = self.transform(image) image_id = int(os.path.splitext(os.path.basename(file_name))[0]) label = np.array([image_id, 1.0] + pad + [height, width], np.float32) # label = torch.from_numpy(label) return image, label @staticmethod def normalize(img, img_mean, img_scale): img = np.array(img, dtype=np.float32) img = (img - img_mean) * img_scale return img @staticmethod def pad_width(img, stride, pad_value, min_dims): h, w, _ = img.shape h = min(min_dims[0], h) min_dims[0] = math.ceil(min_dims[0] / float(stride)) * stride min_dims[1] = max(min_dims[1], w) min_dims[1] = math.ceil(min_dims[1] / float(stride)) * stride top = int(math.floor((min_dims[0] - h) / 2.0)) left = int(math.floor((min_dims[1] - w) / 2.0)) bottom = int(min_dims[0] - h - top) right = int(min_dims[1] - w - left) pad = [top, left, bottom, right] padded_img = cv2.copyMakeBorder( src=img, top=top, bottom=bottom, left=left, right=right, borderType=cv2.BORDER_CONSTANT, value=pad_value) return padded_img, pad # --------------------------------------------------------------------------------------------------------------------- class CocoHpe2ValTransform(object): def __init__(self, ds_metainfo): self.ds_metainfo = ds_metainfo def __call__(self, src, label): return src, label def extract_keypoints(heatmap, all_keypoints, total_keypoint_num): heatmap[heatmap < 0.1] = 0 heatmap_with_borders = np.pad(heatmap, [(2, 2), (2, 2)], mode="constant") heatmap_center = heatmap_with_borders[1:heatmap_with_borders.shape[0] - 1, 1:heatmap_with_borders.shape[1] - 1] heatmap_left = heatmap_with_borders[1:heatmap_with_borders.shape[0] - 1, 2:heatmap_with_borders.shape[1]] heatmap_right = heatmap_with_borders[1:heatmap_with_borders.shape[0] - 1, 0:heatmap_with_borders.shape[1] - 2] heatmap_up = heatmap_with_borders[2:heatmap_with_borders.shape[0], 1:heatmap_with_borders.shape[1] - 1] heatmap_down = heatmap_with_borders[0:heatmap_with_borders.shape[0] - 2, 1:heatmap_with_borders.shape[1] - 1] heatmap_peaks = (heatmap_center > heatmap_left) &\ (heatmap_center > heatmap_right) &\ (heatmap_center > heatmap_up) &\ (heatmap_center > heatmap_down) heatmap_peaks = heatmap_peaks[1:heatmap_center.shape[0] - 1, 1:heatmap_center.shape[1] - 1] keypoints = list(zip(np.nonzero(heatmap_peaks)[1], np.nonzero(heatmap_peaks)[0])) # (w, h) keypoints = sorted(keypoints, key=itemgetter(0)) suppressed = np.zeros(len(keypoints), np.uint8) keypoints_with_score_and_id = [] keypoint_num = 0 for i in range(len(keypoints)): if suppressed[i]: continue for j in range(i + 1, len(keypoints)): if math.sqrt((keypoints[i][0] - keypoints[j][0]) ** 2 + (keypoints[i][1] - keypoints[j][1]) ** 2) < 6: suppressed[j] = 1 keypoint_with_score_and_id = ( keypoints[i][0], keypoints[i][1], heatmap[keypoints[i][1], keypoints[i][0]], total_keypoint_num + keypoint_num) keypoints_with_score_and_id.append(keypoint_with_score_and_id) keypoint_num += 1 all_keypoints.append(keypoints_with_score_and_id) return keypoint_num def group_keypoints(all_keypoints_by_type, pafs, pose_entry_size=20, min_paf_score=0.05): def linspace2d(start, stop, n=10): points = 1 / (n - 1) * (stop - start) return points[:, None] * np.arange(n) + start[:, None] BODY_PARTS_KPT_IDS = [[1, 2], [1, 5], [2, 3], [3, 4], [5, 6], [6, 7], [1, 8], [8, 9], [9, 10], [1, 11], [11, 12], [12, 13], [1, 0], [0, 14], [14, 16], [0, 15], [15, 17], [2, 16], [5, 17]] BODY_PARTS_PAF_IDS = ([12, 13], [20, 21], [14, 15], [16, 17], [22, 23], [24, 25], [0, 1], [2, 3], [4, 5], [6, 7], [8, 9], [10, 11], [28, 29], [30, 31], [34, 35], [32, 33], [36, 37], [18, 19], [26, 27]) pose_entries = [] all_keypoints = np.array([item for sublist in all_keypoints_by_type for item in sublist]) for part_id in range(len(BODY_PARTS_PAF_IDS)): part_pafs = pafs[:, :, BODY_PARTS_PAF_IDS[part_id]] kpts_a = all_keypoints_by_type[BODY_PARTS_KPT_IDS[part_id][0]] kpts_b = all_keypoints_by_type[BODY_PARTS_KPT_IDS[part_id][1]] num_kpts_a = len(kpts_a) num_kpts_b = len(kpts_b) kpt_a_id = BODY_PARTS_KPT_IDS[part_id][0] kpt_b_id = BODY_PARTS_KPT_IDS[part_id][1] if num_kpts_a == 0 and num_kpts_b == 0: # no keypoints for such body part continue elif num_kpts_a == 0: # body part has just 'b' keypoints for i in range(num_kpts_b): num = 0 for j in range(len(pose_entries)): # check if already in some pose, was added by another body part if pose_entries[j][kpt_b_id] == kpts_b[i][3]: num += 1 continue if num == 0: pose_entry = np.ones(pose_entry_size) * -1 pose_entry[kpt_b_id] = kpts_b[i][3] # keypoint idx pose_entry[-1] = 1 # num keypoints in pose pose_entry[-2] = kpts_b[i][2] # pose score pose_entries.append(pose_entry) continue elif num_kpts_b == 0: # body part has just 'a' keypoints for i in range(num_kpts_a): num = 0 for j in range(len(pose_entries)): if pose_entries[j][kpt_a_id] == kpts_a[i][3]: num += 1 continue if num == 0: pose_entry = np.ones(pose_entry_size) * -1 pose_entry[kpt_a_id] = kpts_a[i][3] pose_entry[-1] = 1 pose_entry[-2] = kpts_a[i][2] pose_entries.append(pose_entry) continue connections = [] for i in range(num_kpts_a): kpt_a = np.array(kpts_a[i][0:2]) for j in range(num_kpts_b): kpt_b = np.array(kpts_b[j][0:2]) mid_point = [(), ()] mid_point[0] = (int(round((kpt_a[0] + kpt_b[0]) * 0.5)), int(round((kpt_a[1] + kpt_b[1]) * 0.5))) mid_point[1] = mid_point[0] vec = [kpt_b[0] - kpt_a[0], kpt_b[1] - kpt_a[1]] vec_norm = math.sqrt(vec[0] ** 2 + vec[1] ** 2) if vec_norm == 0: continue vec[0] /= vec_norm vec[1] /= vec_norm cur_point_score = (vec[0] * part_pafs[mid_point[0][1], mid_point[0][0], 0] + vec[1] * part_pafs[mid_point[1][1], mid_point[1][0], 1]) height_n = pafs.shape[0] // 2 success_ratio = 0 point_num = 10 # number of points to integration over paf if cur_point_score > -100: passed_point_score = 0 passed_point_num = 0 x, y = linspace2d(kpt_a, kpt_b) for point_idx in range(point_num): px = int(round(x[point_idx])) py = int(round(y[point_idx])) paf = part_pafs[py, px, 0:2] cur_point_score = vec[0] * paf[0] + vec[1] * paf[1] if cur_point_score > min_paf_score: passed_point_score += cur_point_score passed_point_num += 1 success_ratio = passed_point_num / point_num ratio = 0 if passed_point_num > 0: ratio = passed_point_score / passed_point_num ratio += min(height_n / vec_norm - 1, 0) if ratio > 0 and success_ratio > 0.8: score_all = ratio + kpts_a[i][2] + kpts_b[j][2] connections.append([i, j, ratio, score_all]) if len(connections) > 0: connections = sorted(connections, key=itemgetter(2), reverse=True) num_connections = min(num_kpts_a, num_kpts_b) has_kpt_a = np.zeros(num_kpts_a, dtype=np.int32) has_kpt_b = np.zeros(num_kpts_b, dtype=np.int32) filtered_connections = [] for row in range(len(connections)): if len(filtered_connections) == num_connections: break i, j, cur_point_score = connections[row][0:3] if not has_kpt_a[i] and not has_kpt_b[j]: filtered_connections.append([kpts_a[i][3], kpts_b[j][3], cur_point_score]) has_kpt_a[i] = 1 has_kpt_b[j] = 1 connections = filtered_connections if len(connections) == 0: continue if part_id == 0: pose_entries = [np.ones(pose_entry_size) * -1 for _ in range(len(connections))] for i in range(len(connections)): pose_entries[i][BODY_PARTS_KPT_IDS[0][0]] = connections[i][0] pose_entries[i][BODY_PARTS_KPT_IDS[0][1]] = connections[i][1] pose_entries[i][-1] = 2 pose_entries[i][-2] = np.sum(all_keypoints[connections[i][0:2], 2]) + connections[i][2] elif part_id == 17 or part_id == 18: kpt_a_id = BODY_PARTS_KPT_IDS[part_id][0] kpt_b_id = BODY_PARTS_KPT_IDS[part_id][1] for i in range(len(connections)): for j in range(len(pose_entries)): if pose_entries[j][kpt_a_id] == connections[i][0] and pose_entries[j][kpt_b_id] == -1: pose_entries[j][kpt_b_id] = connections[i][1] elif pose_entries[j][kpt_b_id] == connections[i][1] and pose_entries[j][kpt_a_id] == -1: pose_entries[j][kpt_a_id] = connections[i][0] continue else: kpt_a_id = BODY_PARTS_KPT_IDS[part_id][0] kpt_b_id = BODY_PARTS_KPT_IDS[part_id][1] for i in range(len(connections)): num = 0 for j in range(len(pose_entries)): if pose_entries[j][kpt_a_id] == connections[i][0]: pose_entries[j][kpt_b_id] = connections[i][1] num += 1 pose_entries[j][-1] += 1 pose_entries[j][-2] += all_keypoints[connections[i][1], 2] + connections[i][2] if num == 0: pose_entry = np.ones(pose_entry_size) * -1 pose_entry[kpt_a_id] = connections[i][0] pose_entry[kpt_b_id] = connections[i][1] pose_entry[-1] = 2 pose_entry[-2] = np.sum(all_keypoints[connections[i][0:2], 2]) + connections[i][2] pose_entries.append(pose_entry) filtered_entries = [] for i in range(len(pose_entries)): if pose_entries[i][-1] < 3 or (pose_entries[i][-2] / pose_entries[i][-1] < 0.2): continue filtered_entries.append(pose_entries[i]) pose_entries = np.asarray(filtered_entries) return pose_entries, all_keypoints def convert_to_coco_format(pose_entries, all_keypoints): coco_keypoints = [] scores = [] for n in range(len(pose_entries)): if len(pose_entries[n]) == 0: continue keypoints = [0] * 17 * 3 to_coco_map = [0, -1, 6, 8, 10, 5, 7, 9, 12, 14, 16, 11, 13, 15, 2, 1, 4, 3] person_score = pose_entries[n][-2] position_id = -1 for keypoint_id in pose_entries[n][:-2]: position_id += 1 if position_id == 1: # no 'neck' in COCO continue cx, cy, score, visibility = 0, 0, 0, 0 # keypoint not found if keypoint_id != -1: cx, cy, score = all_keypoints[int(keypoint_id), 0:3] cx = cx + 0.5 cy = cy + 0.5 visibility = 1 keypoints[to_coco_map[position_id] * 3 + 0] = cx keypoints[to_coco_map[position_id] * 3 + 1] = cy keypoints[to_coco_map[position_id] * 3 + 2] = visibility coco_keypoints.append(keypoints) scores.append(person_score * max(0, (pose_entries[n][-1] - 1))) # -1 for 'neck' return coco_keypoints, scores def recalc_pose(pred, label): label_img_id = label[:, 0].astype(np.int32) # label_score = label[:, 1] pads = label[:, 2:6].astype(np.int32) heights = label[:, 6].astype(np.int32) widths = label[:, 7].astype(np.int32) keypoints = 19 stride = 8 heatmap2ds = pred[:, :keypoints] paf2ds = pred[:, keypoints:(3 * keypoints)] pred_pts_score = [] pred_person_score = [] label_img_id_ = [] batch = pred.shape[0] for batch_i in range(batch): label_img_id_i = label_img_id[batch_i] pad = list(pads[batch_i]) height = int(heights[batch_i]) width = int(widths[batch_i]) heatmap2d = heatmap2ds[batch_i] paf2d = paf2ds[batch_i] heatmaps = np.transpose(heatmap2d, (1, 2, 0)) heatmaps = cv2.resize(heatmaps, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC) heatmaps = heatmaps[pad[0]:heatmaps.shape[0] - pad[2], pad[1]:heatmaps.shape[1] - pad[3]:, :] heatmaps = cv2.resize(heatmaps, (width, height), interpolation=cv2.INTER_CUBIC) pafs = np.transpose(paf2d, (1, 2, 0)) pafs = cv2.resize(pafs, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC) pafs = pafs[pad[0]:pafs.shape[0] - pad[2], pad[1]:pafs.shape[1] - pad[3], :] pafs = cv2.resize(pafs, (width, height), interpolation=cv2.INTER_CUBIC) total_keypoints_num = 0 all_keypoints_by_type = [] for kpt_idx in range(18): # 19th for bg total_keypoints_num += extract_keypoints( heatmaps[:, :, kpt_idx], all_keypoints_by_type, total_keypoints_num) pose_entries, all_keypoints = group_keypoints( all_keypoints_by_type, pafs) coco_keypoints, scores = convert_to_coco_format( pose_entries, all_keypoints) pred_pts_score.append(coco_keypoints) pred_person_score.append(scores) label_img_id_.append([label_img_id_i] * len(scores)) return np.array(pred_pts_score).reshape((-1, 17, 3)), np.array(pred_person_score)[0], np.array(label_img_id_[0]) # --------------------------------------------------------------------------------------------------------------------- class CocoHpe2MetaInfo(DatasetMetaInfo): def __init__(self): super(CocoHpe2MetaInfo, self).__init__() self.label = "COCO" self.short_label = "coco" self.root_dir_name = "coco" self.dataset_class = CocoHpe2Dataset self.num_training_samples = None self.in_channels = 3 self.num_classes = 17 self.input_image_size = (368, 368) self.train_metric_capts = None self.train_metric_names = None self.train_metric_extra_kwargs = None self.val_metric_capts = None self.val_metric_names = None self.test_metric_capts = ["Val.CocoOksAp"] self.test_metric_names = ["CocoHpeOksApMetric"] self.test_metric_extra_kwargs = [ {"name": "OksAp", "coco_annotations_file_path": None, "use_file": False, "pose_postprocessing_fn": lambda x, y: recalc_pose(x, y)}] self.saver_acc_ind = 0 self.do_transform = True self.val_transform = CocoHpe2ValTransform self.test_transform = CocoHpe2ValTransform self.ml_type = "hpe" self.test_net_extra_kwargs = None self.mean_rgb = (0.485, 0.456, 0.406) self.std_rgb = (0.229, 0.224, 0.225) self.load_ignore_extra = False def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for ImageNet-1K dataset metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(CocoHpe2MetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--input-size", type=int, nargs=2, default=self.input_image_size, help="size of the input for model") parser.add_argument( "--load-ignore-extra", action="store_true", help="ignore extra layers in the source PyTroch model") def update(self, args): """ Update ImageNet-1K dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(CocoHpe2MetaInfo, self).update(args) self.input_image_size = args.input_size self.load_ignore_extra = args.load_ignore_extra def update_from_dataset(self, dataset): """ Update dataset metainfo after a dataset class instance creation. Parameters: ---------- args : obj A dataset class instance. """ self.test_metric_extra_kwargs[0]["coco_annotations_file_path"] = dataset.annotations_file_path
20,786
39.8389
119
py
imgclsmob
imgclsmob-master/gluon/datasets/svhn_cls_dataset.py
""" SVHN classification dataset. """ import os import numpy as np import mxnet as mx from mxnet import gluon from mxnet.gluon.utils import download, check_sha1 from .cifar10_cls_dataset import CIFAR10MetaInfo class SVHN(gluon.data.dataset._DownloadedDataset): """ SVHN image classification dataset from http://ufldl.stanford.edu/housenumbers/. Each sample is an image (in 3D NDArray) with shape (32, 32, 3). Note: The SVHN dataset assigns the label `10` to the digit `0`. However, in this Dataset, we assign the label `0` to the digit `0`. Parameters: ---------- root : str, default $MXNET_HOME/datasets/svhn Path to temp folder for storing data. mode : str, default 'train' 'train', 'val', or 'test'. transform : function, default None A user defined callback that transforms each sample. """ def __init__(self, root=os.path.join("~", ".mxnet", "datasets", "svhn"), mode="train", transform=None): self._mode = mode self._train_data = [("http://ufldl.stanford.edu/housenumbers/train_32x32.mat", "train_32x32.mat", "e6588cae42a1a5ab5efe608cc5cd3fb9aaffd674")] self._test_data = [("http://ufldl.stanford.edu/housenumbers/test_32x32.mat", "test_32x32.mat", "29b312382ca6b9fba48d41a7b5c19ad9a5462b20")] super(SVHN, self).__init__(root, transform) def _get_data(self): if any(not os.path.exists(path) or not check_sha1(path, sha1) for path, sha1 in ((os.path.join(self._root, name), sha1) for _, name, sha1 in self._train_data + self._test_data)): for url, _, sha1 in self._train_data + self._test_data: download(url=url, path=self._root, sha1_hash=sha1) if self._mode == "train": data_files = self._train_data[0] else: data_files = self._test_data[0] import scipy.io as sio loaded_mat = sio.loadmat(os.path.join(self._root, data_files[1])) data = loaded_mat["X"] data = np.transpose(data, (3, 0, 1, 2)) self._data = mx.nd.array(data, dtype=data.dtype) self._label = loaded_mat["y"].astype(np.int32).squeeze() np.place(self._label, self._label == 10, 0) class SVHNMetaInfo(CIFAR10MetaInfo): def __init__(self): super(SVHNMetaInfo, self).__init__() self.label = "SVHN" self.root_dir_name = "svhn" self.dataset_class = SVHN self.num_training_samples = 73257
2,574
35.785714
113
py
imgclsmob
imgclsmob-master/gluon/datasets/coco_hpe3_dataset.py
""" COCO keypoint detection (2D multiple human pose estimation) dataset (for IBPPose). """ import os # import json import math import cv2 import numpy as np from mxnet.gluon.data import dataset from .dataset_metainfo import DatasetMetaInfo class CocoHpe3Dataset(dataset.Dataset): """ COCO keypoint detection (2D multiple human pose estimation) dataset. Parameters: ---------- root : string Path to `annotations`, `train2017`, and `val2017` folders. mode : string, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None): super(CocoHpe3Dataset, self).__init__() self._root = os.path.expanduser(root) self.mode = mode self.transform = transform mode_name = "train" if mode == "train" else "val" annotations_dir_path = os.path.join(root, "annotations") annotations_file_path = os.path.join(annotations_dir_path, "person_keypoints_" + mode_name + "2017.json") # with open(annotations_file_path, "r") as f: # self.file_names = json.load(f)["images"] self.image_dir_path = os.path.join(root, mode_name + "2017") self.annotations_file_path = annotations_file_path from pycocotools.coco import COCO self.coco_gt = COCO(self.annotations_file_path) self.validation_ids = self.coco_gt.getImgIds()[:] def __str__(self): return self.__class__.__name__ + "(" + self._root + ")" def __len__(self): return len(self.validation_ids) def __getitem__(self, idx): # file_name = self.file_names[idx]["file_name"] image_id = self.validation_ids[idx] file_name = self.coco_gt.imgs[image_id]["file_name"] image_file_path = os.path.join(self.image_dir_path, file_name) image = cv2.imread(image_file_path, flags=cv2.IMREAD_COLOR) # image = cv2.cvtColor(img, code=cv2.COLOR_BGR2RGB) image_src_shape = image.shape[:2] boxsize = 512 max_downsample = 64 pad_value = 128 scale = boxsize / image.shape[0] if scale * image.shape[0] > 2600 or scale * image.shape[1] > 3800: scale = min(2600 / image.shape[0], 3800 / image.shape[1]) image = cv2.resize(image, (0, 0), fx=scale, fy=scale, interpolation=cv2.INTER_CUBIC) image, pad = self.pad_right_down_corner(image, max_downsample, pad_value) image = np.float32(image / 255) image = image.transpose((2, 0, 1)) # image_id = int(os.path.splitext(os.path.basename(file_name))[0]) label = np.array([image_id, 1.0] + pad + list(image_src_shape), np.float32) return image, label @staticmethod def pad_right_down_corner(img, stride, pad_value): h = img.shape[0] w = img.shape[1] pad = 4 * [None] pad[0] = 0 # up pad[1] = 0 # left pad[2] = 0 if (h % stride == 0) else stride - (h % stride) # down pad[3] = 0 if (w % stride == 0) else stride - (w % stride) # right img_padded = img pad_up = np.tile(img_padded[0:1, :, :] * 0 + pad_value, (pad[0], 1, 1)) img_padded = np.concatenate((pad_up, img_padded), axis=0) pad_left = np.tile(img_padded[:, 0:1, :] * 0 + pad_value, (1, pad[1], 1)) img_padded = np.concatenate((pad_left, img_padded), axis=1) pad_down = np.tile(img_padded[-2:-1, :, :] * 0 + pad_value, (pad[2], 1, 1)) img_padded = np.concatenate((img_padded, pad_down), axis=0) pad_right = np.tile(img_padded[:, -2:-1, :] * 0 + pad_value, (1, pad[3], 1)) img_padded = np.concatenate((img_padded, pad_right), axis=1) return img_padded, pad # --------------------------------------------------------------------------------------------------------------------- class CocoHpe2ValTransform(object): def __init__(self, ds_metainfo): self.ds_metainfo = ds_metainfo def __call__(self, src, label): return src, label def recalc_pose(pred, label): dt_gt_mapping = {0: 0, 1: None, 2: 6, 3: 8, 4: 10, 5: 5, 6: 7, 7: 9, 8: 12, 9: 14, 10: 16, 11: 11, 12: 13, 13: 15, 14: 2, 15: 1, 16: 4, 17: 3} parts = ["nose", "neck", "Rsho", "Relb", "Rwri", "Lsho", "Lelb", "Lwri", "Rhip", "Rkne", "Rank", "Lhip", "Lkne", "Lank", "Reye", "Leye", "Rear", "Lear"] num_parts = len(parts) parts_dict = dict(zip(parts, range(num_parts))) limb_from = ['neck', 'neck', 'neck', 'neck', 'neck', 'nose', 'nose', 'Reye', 'Leye', 'neck', 'Rsho', 'Relb', 'neck', 'Lsho', 'Lelb', 'neck', 'Rhip', 'Rkne', 'neck', 'Lhip', 'Lkne', 'nose', 'nose', 'Rsho', 'Rhip', 'Lsho', 'Lhip', 'Rear', 'Lear', 'Rhip'] limb_to = ['nose', 'Reye', 'Leye', 'Rear', 'Lear', 'Reye', 'Leye', 'Rear', 'Lear', 'Rsho', 'Relb', 'Rwri', 'Lsho', 'Lelb', 'Lwri', 'Rhip', 'Rkne', 'Rank', 'Lhip', 'Lkne', 'Lank', 'Rsho', 'Lsho', 'Rhip', 'Lkne', 'Lhip', 'Rkne', 'Rsho', 'Lsho', 'Lhip'] limb_from = [parts_dict[n] for n in limb_from] limb_to = [parts_dict[n] for n in limb_to] assert limb_from == [x for x in [ 1, 1, 1, 1, 1, 0, 0, 14, 15, 1, 2, 3, 1, 5, 6, 1, 8, 9, 1, 11, 12, 0, 0, 2, 8, 5, 11, 16, 17, 8]] assert limb_to == [x for x in [ 0, 14, 15, 16, 17, 14, 15, 16, 17, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 2, 5, 8, 12, 11, 9, 2, 5, 11]] limbs_conn = list(zip(limb_from, limb_to)) limb_seq = limbs_conn paf_layers = 30 num_layers = 50 stride = 4 label_img_id = label[:, 0].astype(np.int32) # label_score = label[:, 1] pads = label[:, 2:6].astype(np.int32) image_src_shapes = label[:, 6:8].astype(np.int32) pred_pts_score = [] pred_person_score = [] label_img_id_ = [] batch = pred.shape[0] for batch_i in range(batch): label_img_id_i = label_img_id[batch_i] pad = list(pads[batch_i]) image_src_shape = list(image_src_shapes[batch_i]) output_blob = pred[batch_i].transpose((1, 2, 0)) output_paf = output_blob[:, :, :paf_layers] output_heatmap = output_blob[:, :, paf_layers:num_layers] heatmap = cv2.resize(output_heatmap, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC) heatmap = heatmap[ pad[0]:(output_blob.shape[0] * stride - pad[2]), pad[1]:(output_blob.shape[1] * stride - pad[3]), :] heatmap = cv2.resize(heatmap, (image_src_shape[1], image_src_shape[0]), interpolation=cv2.INTER_CUBIC) paf = cv2.resize(output_paf, (0, 0), fx=stride, fy=stride, interpolation=cv2.INTER_CUBIC) paf = paf[ pad[0]:(output_blob.shape[0] * stride - pad[2]), pad[1]:(output_blob.shape[1] * stride - pad[3]), :] paf = cv2.resize(paf, (image_src_shape[1], image_src_shape[0]), interpolation=cv2.INTER_CUBIC) all_peaks = find_peaks(heatmap) connection_all, special_k = find_connections(all_peaks, paf, image_src_shape[0], limb_seq) subset, candidate = find_people(connection_all, special_k, all_peaks, limb_seq) for s in subset[..., 0]: keypoint_indexes = s[:18] person_keypoint_coordinates = [] for index in keypoint_indexes: if index == -1: X, Y, C = 0, 0, 0 else: X, Y, C = list(candidate[index.astype(int)][:2]) + [1] person_keypoint_coordinates.append([X, Y, C]) person_keypoint_coordinates_coco = [None] * 17 for dt_index, gt_index in dt_gt_mapping.items(): if gt_index is None: continue person_keypoint_coordinates_coco[gt_index] = person_keypoint_coordinates[dt_index] pred_pts_score.append(person_keypoint_coordinates_coco) pred_person_score.append(1 - 1.0 / s[18]) label_img_id_.append(label_img_id_i) return np.array(pred_pts_score).reshape((-1, 17, 3)), np.array(pred_person_score), np.array(label_img_id_) def find_peaks(heatmap_avg): import torch thre1 = 0.1 offset_radius = 2 all_peaks = [] peak_counter = 0 heatmap_avg = heatmap_avg.astype(np.float32) filter_map = heatmap_avg[:, :, :18].copy().transpose((2, 0, 1))[None, ...] filter_map = torch.from_numpy(filter_map).cuda() filter_map = keypoint_heatmap_nms(filter_map, kernel=3, thre=thre1) filter_map = filter_map.cpu().numpy().squeeze().transpose((1, 2, 0)) for part in range(18): map_ori = heatmap_avg[:, :, part] peaks_binary = filter_map[:, :, part] peaks = list(zip(np.nonzero(peaks_binary)[1], np.nonzero(peaks_binary)[0])) # note reverse refined_peaks_with_score = [refine_centroid(map_ori, anchor, offset_radius) for anchor in peaks] id = range(peak_counter, peak_counter + len(refined_peaks_with_score)) peaks_with_score_and_id = [refined_peaks_with_score[i] + (id[i],) for i in range(len(id))] all_peaks.append(peaks_with_score_and_id) peak_counter += len(peaks) return all_peaks def keypoint_heatmap_nms(heat, kernel=3, thre=0.1): from torch.nn import functional as F # keypoint NMS on heatmap (score map) pad = (kernel - 1) // 2 pad_heat = F.pad(heat, (pad, pad, pad, pad), mode="reflect") hmax = F.max_pool2d(pad_heat, (kernel, kernel), stride=1, padding=0) keep = (hmax == heat).float() * (heat >= thre).float() return heat * keep def refine_centroid(scorefmp, anchor, radius): """ Refine the centroid coordinate. It dose not affect the results after testing. :param scorefmp: 2-D numpy array, original regressed score map :param anchor: python tuple, (x,y) coordinates :param radius: int, range of considered scores :return: refined anchor, refined score """ x_c, y_c = anchor x_min = x_c - radius x_max = x_c + radius + 1 y_min = y_c - radius y_max = y_c + radius + 1 if y_max > scorefmp.shape[0] or y_min < 0 or x_max > scorefmp.shape[1] or x_min < 0: return anchor + (scorefmp[y_c, x_c], ) score_box = scorefmp[y_min:y_max, x_min:x_max] x_grid, y_grid = np.mgrid[-radius:radius + 1, -radius:radius + 1] offset_x = (score_box * x_grid).sum() / score_box.sum() offset_y = (score_box * y_grid).sum() / score_box.sum() x_refine = x_c + offset_x y_refine = y_c + offset_y refined_anchor = (x_refine, y_refine) return refined_anchor + (score_box.mean(),) def find_connections(all_peaks, paf_avg, image_width, limb_seq): mid_num_ = 20 thre2 = 0.1 connect_ration = 0.8 connection_all = [] special_k = [] for k in range(len(limb_seq)): score_mid = paf_avg[:, :, k] candA = all_peaks[limb_seq[k][0]] candB = all_peaks[limb_seq[k][1]] nA = len(candA) nB = len(candB) if nA != 0 and nB != 0: connection_candidate = [] for i in range(nA): for j in range(nB): vec = np.subtract(candB[j][:2], candA[i][:2]) norm = math.sqrt(vec[0] * vec[0] + vec[1] * vec[1]) mid_num = min(int(round(norm + 1)), mid_num_) if norm == 0: continue startend = list(zip(np.linspace(candA[i][0], candB[j][0], num=mid_num), np.linspace(candA[i][1], candB[j][1], num=mid_num))) limb_response = np.array([score_mid[int(round(startend[I][1])), int(round(startend[I][0]))] for I in range(len(startend))]) score_midpts = limb_response score_with_dist_prior = sum(score_midpts) / len(score_midpts) + min(0.5 * image_width / norm - 1, 0) criterion1 = len(np.nonzero(score_midpts > thre2)[0]) >= connect_ration * len(score_midpts) criterion2 = score_with_dist_prior > 0 if criterion1 and criterion2: connection_candidate.append([ i, j, score_with_dist_prior, norm, 0.5 * score_with_dist_prior + 0.25 * candA[i][2] + 0.25 * candB[j][2]]) connection_candidate = sorted(connection_candidate, key=lambda x: x[4], reverse=True) connection = np.zeros((0, 6)) for c in range(len(connection_candidate)): i, j, s, limb_len = connection_candidate[c][0:4] if i not in connection[:, 3] and j not in connection[:, 4]: connection = np.vstack([connection, [candA[i][3], candB[j][3], s, i, j, limb_len]]) if len(connection) >= min(nA, nB): break connection_all.append(connection) else: special_k.append(k) connection_all.append([]) return connection_all, special_k def find_people(connection_all, special_k, all_peaks, limb_seq): len_rate = 16.0 connection_tole = 0.7 remove_recon = 0 subset = -1 * np.ones((0, 20, 2)) candidate = np.array([item for sublist in all_peaks for item in sublist]) for k in range(len(limb_seq)): if k not in special_k: partAs = connection_all[k][:, 0] partBs = connection_all[k][:, 1] indexA, indexB = np.array(limb_seq[k]) for i in range(len(connection_all[k])): found = 0 subset_idx = [-1, -1] for j in range(len(subset)): if subset[j][indexA][0].astype(int) == (partAs[i]).astype(int) or subset[j][indexB][0].astype( int) == partBs[i].astype(int): if found >= 2: continue subset_idx[found] = j found += 1 if found == 1: j = subset_idx[0] if subset[j][indexB][0].astype(int) == -1 and\ len_rate * subset[j][-1][1] > connection_all[k][i][-1]: subset[j][indexB][0] = partBs[i] subset[j][indexB][1] = connection_all[k][i][2] subset[j][-1][0] += 1 subset[j][-2][0] += candidate[partBs[i].astype(int), 2] + connection_all[k][i][2] subset[j][-1][1] = max(connection_all[k][i][-1], subset[j][-1][1]) elif subset[j][indexB][0].astype(int) != partBs[i].astype(int): if subset[j][indexB][1] >= connection_all[k][i][2]: pass else: if len_rate * subset[j][-1][1] <= connection_all[k][i][-1]: continue subset[j][-2][0] -= candidate[subset[j][indexB][0].astype(int), 2] + subset[j][indexB][1] subset[j][indexB][0] = partBs[i] subset[j][indexB][1] = connection_all[k][i][2] subset[j][-2][0] += candidate[partBs[i].astype(int), 2] + connection_all[k][i][2] subset[j][-1][1] = max(connection_all[k][i][-1], subset[j][-1][1]) elif subset[j][indexB][0].astype(int) == partBs[i].astype(int) and\ subset[j][indexB][1] <= connection_all[k][i][2]: subset[j][-2][0] -= candidate[subset[j][indexB][0].astype(int), 2] + subset[j][indexB][1] subset[j][indexB][0] = partBs[i] subset[j][indexB][1] = connection_all[k][i][2] subset[j][-2][0] += candidate[partBs[i].astype(int), 2] + connection_all[k][i][2] subset[j][-1][1] = max(connection_all[k][i][-1], subset[j][-1][1]) else: pass elif found == 2: j1, j2 = subset_idx membership1 = ((subset[j1][..., 0] >= 0).astype(int))[:-2] membership2 = ((subset[j2][..., 0] >= 0).astype(int))[:-2] membership = membership1 + membership2 if len(np.nonzero(membership == 2)[0]) == 0: min_limb1 = np.min(subset[j1, :-2, 1][membership1 == 1]) min_limb2 = np.min(subset[j2, :-2, 1][membership2 == 1]) min_tolerance = min(min_limb1, min_limb2) if connection_all[k][i][2] < connection_tole * min_tolerance or\ len_rate * subset[j1][-1][1] <= connection_all[k][i][-1]: continue subset[j1][:-2][...] += (subset[j2][:-2][...] + 1) subset[j1][-2:][:, 0] += subset[j2][-2:][:, 0] subset[j1][-2][0] += connection_all[k][i][2] subset[j1][-1][1] = max(connection_all[k][i][-1], subset[j1][-1][1]) subset = np.delete(subset, j2, 0) else: if connection_all[k][i][0] in subset[j1, :-2, 0]: c1 = np.where(subset[j1, :-2, 0] == connection_all[k][i][0]) c2 = np.where(subset[j2, :-2, 0] == connection_all[k][i][1]) else: c1 = np.where(subset[j1, :-2, 0] == connection_all[k][i][1]) c2 = np.where(subset[j2, :-2, 0] == connection_all[k][i][0]) c1 = int(c1[0]) c2 = int(c2[0]) assert c1 != c2, "an candidate keypoint is used twice, shared by two people" if connection_all[k][i][2] < subset[j1][c1][1] and connection_all[k][i][2] < subset[j2][c2][1]: continue small_j = j1 remove_c = c1 if subset[j1][c1][1] > subset[j2][c2][1]: small_j = j2 remove_c = c2 if remove_recon > 0: subset[small_j][-2][0] -= candidate[subset[small_j][remove_c][0].astype(int), 2] + \ subset[small_j][remove_c][1] subset[small_j][remove_c][0] = -1 subset[small_j][remove_c][1] = -1 subset[small_j][-1][0] -= 1 elif not found and k < len(limb_seq): row = -1 * np.ones((20, 2)) row[indexA][0] = partAs[i] row[indexA][1] = connection_all[k][i][2] row[indexB][0] = partBs[i] row[indexB][1] = connection_all[k][i][2] row[-1][0] = 2 row[-1][1] = connection_all[k][i][-1] row[-2][0] = sum(candidate[connection_all[k][i, :2].astype(int), 2]) + connection_all[k][i][2] row = row[np.newaxis, :, :] subset = np.concatenate((subset, row), axis=0) deleteIdx = [] for i in range(len(subset)): if subset[i][-1][0] < 2 or subset[i][-2][0] / subset[i][-1][0] < 0.45: deleteIdx.append(i) subset = np.delete(subset, deleteIdx, axis=0) return subset, candidate # --------------------------------------------------------------------------------------------------------------------- class CocoHpe3MetaInfo(DatasetMetaInfo): def __init__(self): super(CocoHpe3MetaInfo, self).__init__() self.label = "COCO" self.short_label = "coco" self.root_dir_name = "coco" self.dataset_class = CocoHpe3Dataset self.num_training_samples = None self.in_channels = 3 self.num_classes = 17 self.input_image_size = (256, 256) self.train_metric_capts = None self.train_metric_names = None self.train_metric_extra_kwargs = None self.val_metric_capts = None self.val_metric_names = None self.test_metric_capts = ["Val.CocoOksAp"] self.test_metric_names = ["CocoHpeOksApMetric"] self.test_metric_extra_kwargs = [ {"name": "OksAp", "coco_annotations_file_path": None, "validation_ids": None, "use_file": False, "pose_postprocessing_fn": lambda x, y: recalc_pose(x, y)}] self.saver_acc_ind = 0 self.do_transform = True self.val_transform = CocoHpe2ValTransform self.test_transform = CocoHpe2ValTransform self.ml_type = "hpe" self.test_net_extra_kwargs = None self.mean_rgb = (0.485, 0.456, 0.406) self.std_rgb = (0.229, 0.224, 0.225) self.load_ignore_extra = False def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for ImageNet-1K dataset metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(CocoHpe3MetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--input-size", type=int, nargs=2, default=self.input_image_size, help="size of the input for model") parser.add_argument( "--load-ignore-extra", action="store_true", help="ignore extra layers in the source PyTroch model") def update(self, args): """ Update ImageNet-1K dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(CocoHpe3MetaInfo, self).update(args) self.input_image_size = args.input_size self.load_ignore_extra = args.load_ignore_extra def update_from_dataset(self, dataset): """ Update dataset metainfo after a dataset class instance creation. Parameters: ---------- args : obj A dataset class instance. """ self.test_metric_extra_kwargs[0]["coco_annotations_file_path"] = dataset.annotations_file_path # self.test_metric_extra_kwargs[0]["validation_ids"] = dataset.validation_ids
23,125
40.003546
120
py
imgclsmob
imgclsmob-master/gluon/datasets/imagenet1k_rec_cls_dataset.py
""" ImageNet-1K classification dataset (via MXNet image record iterators). """ import os import mxnet as mx from .imagenet1k_cls_dataset import ImageNet1KMetaInfo, calc_val_resize_value class ImageNet1KRecMetaInfo(ImageNet1KMetaInfo): def __init__(self): super(ImageNet1KRecMetaInfo, self).__init__() self.use_imgrec = True self.label = "ImageNet1K_rec" self.root_dir_name = "imagenet_rec" self.dataset_class = None self.num_training_samples = 1281167 self.train_imgrec_file_path = "train.rec" self.train_imgidx_file_path = "train.idx" self.val_imgrec_file_path = "val.rec" self.val_imgidx_file_path = "val.idx" self.train_imgrec_iter = imagenet_train_imgrec_iter self.val_imgrec_iter = imagenet_val_imgrec_iter def imagenet_train_imgrec_iter(ds_metainfo, batch_size, num_workers, mean_rgb=(123.68, 116.779, 103.939), std_rgb=(58.393, 57.12, 57.375), jitter_param=0.4, lighting_param=0.1): assert (isinstance(ds_metainfo.input_image_size, tuple) and len(ds_metainfo.input_image_size) == 2) imgrec_file_path = os.path.join(ds_metainfo.root_dir_path, ds_metainfo.train_imgrec_file_path) imgidx_file_path = os.path.join(ds_metainfo.root_dir_path, ds_metainfo.train_imgidx_file_path) data_shape = (ds_metainfo.in_channels,) + ds_metainfo.input_image_size kwargs = { "path_imgrec": imgrec_file_path, "path_imgidx": imgidx_file_path, "preprocess_threads": num_workers, "shuffle": True, "batch_size": batch_size, "data_shape": data_shape, "mean_r": mean_rgb[0], "mean_g": mean_rgb[1], "mean_b": mean_rgb[2], "std_r": std_rgb[0], "std_g": std_rgb[1], "std_b": std_rgb[2], "rand_mirror": True, "random_resized_crop": True, "max_aspect_ratio": (4.0 / 3.0), "min_aspect_ratio": (3.0 / 4.0), "max_random_area": 1, "min_random_area": 0.08, "brightness": jitter_param, "saturation": jitter_param, "contrast": jitter_param, "pca_noise": lighting_param } if ds_metainfo.aug_type == "aug0": pass elif ds_metainfo.aug_type == "aug1": kwargs["inter_method"] = 10 elif ds_metainfo.aug_type == "aug2": kwargs["inter_method"] = 10 kwargs["max_rotate_angle"] = 30 kwargs["max_shear_ratio"] = 0.05 else: raise RuntimeError("Unknown augmentation type: {}\n".format(ds_metainfo.aug_type)) return mx.io.ImageRecordIter(**kwargs) def imagenet_val_imgrec_iter(ds_metainfo, batch_size, num_workers, mean_rgb=(123.68, 116.779, 103.939), std_rgb=(58.393, 57.12, 57.375)): assert (isinstance(ds_metainfo.input_image_size, tuple) and len(ds_metainfo.input_image_size) == 2) imgrec_file_path = os.path.join(ds_metainfo.root_dir_path, ds_metainfo.val_imgrec_file_path) imgidx_file_path = os.path.join(ds_metainfo.root_dir_path, ds_metainfo.val_imgidx_file_path) data_shape = (ds_metainfo.in_channels,) + ds_metainfo.input_image_size resize_value = calc_val_resize_value( input_image_size=ds_metainfo.input_image_size, resize_inv_factor=ds_metainfo.resize_inv_factor) return mx.io.ImageRecordIter( path_imgrec=imgrec_file_path, path_imgidx=imgidx_file_path, preprocess_threads=num_workers, shuffle=False, batch_size=batch_size, resize=resize_value, data_shape=data_shape, mean_r=mean_rgb[0], mean_g=mean_rgb[1], mean_b=mean_rgb[2], std_r=std_rgb[0], std_g=std_rgb[1], std_b=std_rgb[2])
3,974
38.75
103
py
imgclsmob
imgclsmob-master/gluon/datasets/asr_dataset.py
""" Automatic Speech Recognition (ASR) abstract dataset. """ __all__ = ['AsrDataset', 'asr_test_transform'] from mxnet.gluon.data import dataset from mxnet.gluon.data.vision import transforms from gluon.gluoncv2.models.jasper import NemoAudioReader class AsrDataset(dataset.Dataset): """ Automatic Speech Recognition (ASR) abstract dataset. Parameters: ---------- root : str Path to the folder stored the dataset. mode : str 'train', 'val', 'test', or 'demo'. transform : func A function that takes data and transforms it. """ def __init__(self, root, mode, transform): super(AsrDataset, self).__init__() assert (mode in ("train", "val", "test", "demo")) self.root = root self.mode = mode self._transform = transform self.data = [] self.audio_reader = NemoAudioReader() def __getitem__(self, index): wav_file_path, label_text = self.data[index] audio_data = self.audio_reader.read_from_file(wav_file_path) audio_len = audio_data.shape[0] return (audio_data, audio_len), label_text def __len__(self): return len(self.data) def asr_test_transform(ds_metainfo): assert (ds_metainfo is not None) return transforms.Compose([])
1,358
26.18
68
py
imgclsmob
imgclsmob-master/gluon/datasets/cifar10_cls_dataset.py
""" CIFAR-10 classification dataset. """ import os import numpy as np import mxnet as mx from mxnet.gluon import Block from mxnet.gluon.data.vision import CIFAR10 from mxnet.gluon.data.vision import transforms from .dataset_metainfo import DatasetMetaInfo class CIFAR10Fine(CIFAR10): """ CIFAR-10 image classification dataset. Parameters: ---------- root : str, default $MXNET_HOME/datasets/cifar10 Path to temp folder for storing data. mode : str, default 'train' 'train', 'val', or 'test'. transform : function, default None A user defined callback that transforms each sample. """ def __init__(self, root=os.path.join("~", ".mxnet", "datasets", "cifar10"), mode="train", transform=None): super(CIFAR10Fine, self).__init__( root=root, train=(mode == "train"), transform=transform) class CIFAR10MetaInfo(DatasetMetaInfo): def __init__(self): super(CIFAR10MetaInfo, self).__init__() self.label = "CIFAR10" self.short_label = "cifar" self.root_dir_name = "cifar10" self.dataset_class = CIFAR10Fine self.num_training_samples = 50000 self.in_channels = 3 self.num_classes = 10 self.input_image_size = (32, 32) self.train_metric_capts = ["Train.Err"] self.train_metric_names = ["Top1Error"] self.train_metric_extra_kwargs = [{"name": "err"}] self.val_metric_capts = ["Val.Err"] self.val_metric_names = ["Top1Error"] self.val_metric_extra_kwargs = [{"name": "err"}] self.saver_acc_ind = 0 self.train_transform = cifar10_train_transform self.val_transform = cifar10_val_transform self.test_transform = cifar10_val_transform self.ml_type = "imgcls" self.loss_name = "SoftmaxCrossEntropy" class RandomCrop(Block): """ Randomly crop `src` with `size` (width, height). Padding is optional. Upsample result if `src` is smaller than `size`. Parameters: ---------- size : int or tuple of (W, H) Size of the final output. pad: int or tuple, default None if int, size of the zero-padding if tuple, number of values padded to the edges of each axis. ((before_1, after_1), ... (before_N, after_N)) unique pad widths for each axis. ((before, after),) yields same before and after pad for each axis. (pad,) or int is a shortcut for before = after = pad width for all axes. interpolation : int, default 2 Interpolation method for resizing. By default uses bilinear interpolation. See OpenCV's resize function for available choices. """ def __init__(self, size, pad=None, interpolation=2): super(RandomCrop, self).__init__() numeric_types = (float, int, np.generic) if isinstance(size, numeric_types): size = (size, size) self._args = (size, interpolation) if isinstance(pad, int): self.pad = ((pad, pad), (pad, pad), (0, 0)) else: self.pad = pad def forward(self, x): if self.pad: x_pad = np.pad(x.asnumpy(), self.pad, mode="constant", constant_values=0) return mx.image.random_crop(mx.nd.array(x_pad), *self._args)[0] def cifar10_train_transform(ds_metainfo, mean_rgb=(0.4914, 0.4822, 0.4465), std_rgb=(0.2023, 0.1994, 0.2010), jitter_param=0.4, lighting_param=0.1): assert (ds_metainfo is not None) assert (ds_metainfo.input_image_size[0] == 32) return transforms.Compose([ RandomCrop( size=32, pad=4), transforms.RandomFlipLeftRight(), transforms.RandomColorJitter( brightness=jitter_param, contrast=jitter_param, saturation=jitter_param), transforms.RandomLighting(lighting_param), transforms.ToTensor(), transforms.Normalize( mean=mean_rgb, std=std_rgb) ]) def cifar10_val_transform(ds_metainfo, mean_rgb=(0.4914, 0.4822, 0.4465), std_rgb=(0.2023, 0.1994, 0.2010)): assert (ds_metainfo is not None) return transforms.Compose([ transforms.ToTensor(), transforms.Normalize( mean=mean_rgb, std=std_rgb) ])
4,585
32.474453
91
py
imgclsmob
imgclsmob-master/gluon/datasets/cub200_2011_cls_dataset.py
""" CUB-200-2011 classification dataset. """ import os import numpy as np import pandas as pd import mxnet as mx from mxnet.gluon.data import dataset from .imagenet1k_cls_dataset import ImageNet1KMetaInfo class CUB200_2011(dataset.Dataset): """ CUB-200-2011 fine-grained classification dataset. Parameters: ---------- root : str, default '~/.mxnet/datasets/CUB_200_2011' Path to the folder stored the dataset. mode : str, default 'train' 'train', 'val', or 'test'. transform : function, default None A function that takes data and label and transforms them. """ def __init__(self, root=os.path.join("~", ".mxnet", "datasets", "CUB_200_2011"), mode="train", transform=None): super(CUB200_2011, self).__init__() root_dir_path = os.path.expanduser(root) assert os.path.exists(root_dir_path) images_file_name = "images.txt" images_file_path = os.path.join(root_dir_path, images_file_name) if not os.path.exists(images_file_path): raise Exception("Images file doesn't exist: {}".format(images_file_name)) class_file_name = "image_class_labels.txt" class_file_path = os.path.join(root_dir_path, class_file_name) if not os.path.exists(class_file_path): raise Exception("Image class file doesn't exist: {}".format(class_file_name)) split_file_name = "train_test_split.txt" split_file_path = os.path.join(root_dir_path, split_file_name) if not os.path.exists(split_file_path): raise Exception("Split file doesn't exist: {}".format(split_file_name)) images_df = pd.read_csv( images_file_path, sep="\s+", header=None, index_col=False, names=["image_id", "image_path"], dtype={"image_id": np.int32, "image_path": np.unicode}) class_df = pd.read_csv( class_file_path, sep="\s+", header=None, index_col=False, names=["image_id", "class_id"], dtype={"image_id": np.int32, "class_id": np.uint8}) split_df = pd.read_csv( split_file_path, sep="\s+", header=None, index_col=False, names=["image_id", "split_flag"], dtype={"image_id": np.int32, "split_flag": np.uint8}) df = images_df.join(class_df, rsuffix="_class_df").join(split_df, rsuffix="_split_df") split_flag = 1 if mode == "train" else 0 subset_df = df[df.split_flag == split_flag] self.image_ids = subset_df["image_id"].values.astype(np.int32) self.class_ids = subset_df["class_id"].values.astype(np.int32) - 1 self.image_file_names = subset_df["image_path"].values.astype(np.unicode) images_dir_name = "images" self.images_dir_path = os.path.join(root_dir_path, images_dir_name) assert os.path.exists(self.images_dir_path) self._transform = transform def __getitem__(self, index): image_file_name = self.image_file_names[index] image_file_path = os.path.join(self.images_dir_path, image_file_name) img = mx.image.imread(image_file_path, flag=1) label = int(self.class_ids[index]) if self._transform is not None: return self._transform(img, label) return img, label def __len__(self): return len(self.image_ids) class CUB200MetaInfo(ImageNet1KMetaInfo): def __init__(self): super(CUB200MetaInfo, self).__init__() self.label = "CUB200_2011" self.short_label = "cub" self.root_dir_name = "CUB_200_2011" self.dataset_class = CUB200_2011 self.num_training_samples = None self.num_classes = 200 self.train_metric_capts = ["Train.Err"] self.train_metric_names = ["Top1Error"] self.train_metric_extra_kwargs = [{"name": "err"}] self.val_metric_capts = ["Val.Err"] self.val_metric_names = ["Top1Error"] self.val_metric_extra_kwargs = [{"name": "err"}] self.saver_acc_ind = 0 self.test_net_extra_kwargs = {"aux": False} self.load_ignore_extra = True def add_dataset_parser_arguments(self, parser, work_dir_path): super(CUB200MetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--no-aux", dest="no_aux", action="store_true", help="no `aux` mode in model") def update(self, args): super(CUB200MetaInfo, self).update(args) if args.no_aux: self.test_net_extra_kwargs = None self.load_ignore_extra = False
4,865
35.586466
94
py
imgclsmob
imgclsmob-master/gluon/datasets/mcv_asr_dataset.py
""" Mozilla Common Voice ASR dataset. """ __all__ = ['McvDataset', 'McvMetaInfo'] import os import re import numpy as np import pandas as pd from .dataset_metainfo import DatasetMetaInfo from .asr_dataset import AsrDataset, asr_test_transform class McvDataset(AsrDataset): """ Mozilla Common Voice dataset for Automatic Speech Recognition (ASR). Parameters: ---------- root : str, default '~/.torch/datasets/mcv' Path to the folder stored the dataset. mode : str, default 'test' 'train', 'val', 'test', or 'demo'. lang : str, default 'en' Language. subset : str, default 'dev' Data subset. transform : function, default None A function that takes data and transforms it. """ def __init__(self, root=os.path.join("~", ".torch", "datasets", "mcv"), mode="test", lang="en", subset="dev", transform=None): super(McvDataset, self).__init__( root=root, mode=mode, transform=transform) assert (lang in ("en", "fr", "de", "it", "es", "ca", "pl", "ru", "ru34")) self.vocabulary = self.get_vocabulary_for_lang(lang=lang) desired_audio_sample_rate = 16000 vocabulary_dict = {c: i for i, c in enumerate(self.vocabulary)} import soundfile import librosa from librosa.core import resample as lr_resample import unicodedata import unidecode root_dir_path = os.path.expanduser(root) assert os.path.exists(root_dir_path) lang_ = lang if lang != "ru34" else "ru" data_dir_path = os.path.join(root_dir_path, lang_) assert os.path.exists(data_dir_path) metainfo_file_path = os.path.join(data_dir_path, subset + ".tsv") assert os.path.exists(metainfo_file_path) metainfo_df = pd.read_csv( metainfo_file_path, sep="\t", header=0, index_col=False) metainfo_df = metainfo_df[["path", "sentence"]] self.data_paths = metainfo_df["path"].values self.data_sentences = metainfo_df["sentence"].values clips_dir_path = os.path.join(data_dir_path, "clips") assert os.path.exists(clips_dir_path) for clip_file_name, sentence in zip(self.data_paths, self.data_sentences): mp3_file_path = os.path.join(clips_dir_path, clip_file_name) assert os.path.exists(mp3_file_path) wav_file_name = clip_file_name.replace(".mp3", ".wav") wav_file_path = os.path.join(clips_dir_path, wav_file_name) # print("==> {}".format(sentence)) text = sentence.lower() if lang == "en": text = re.sub("\.|-|–|—", " ", text) text = re.sub("&", " and ", text) text = re.sub("ō", "o", text) text = re.sub("â|á", "a", text) text = re.sub("é", "e", text) text = re.sub(",|;|:|!|\?|\"|“|”|‘|’|\(|\)", "", text) text = re.sub("\s+", " ", text) text = re.sub(" '", " ", text) text = re.sub("' ", " ", text) elif lang == "fr": text = "".join(c for c in text if unicodedata.combining(c) == 0) text = re.sub("\.|-|–|—|=|×|\*|†|/|ቀ|_|…", " ", text) text = re.sub(",|;|:|!|\?|ʻ|“|”|\"|„|«|»|\(|\)", "", text) text = re.sub("먹|삼|생|고|기|집|\$|ʔ|の|ひ", "", text) text = re.sub("’|´", "'", text) text = re.sub("&", " and ", text) text = re.sub("œ", "oe", text) text = re.sub("æ", "ae", text) text = re.sub("á|ā|ã|ä|ą|ă|å", "a", text) text = re.sub("ö|ō|ó|ð|ổ|ø", "o", text) text = re.sub("ē|ė|ę", "e", text) text = re.sub("í|ī", "i", text) text = re.sub("ú|ū", "u", text) text = re.sub("ý", "y", text) text = re.sub("š|ś|ș|ş", "s", text) text = re.sub("ž|ź|ż", "z", text) text = re.sub("ñ|ń|ṇ", "n", text) text = re.sub("ł|ľ", "l", text) text = re.sub("ć|č", "c", text) text = re.sub("я", "ya", text) text = re.sub("ř", "r", text) text = re.sub("đ", "d", text) text = re.sub("ț", "t", text) text = re.sub("þ", "th", text) text = re.sub("ğ", "g", text) text = re.sub("ß", "ss", text) text = re.sub("µ", "mu", text) text = re.sub("\s+", " ", text) elif lang == "de": text = re.sub("\.|-|–|—|/|_|…", " ", text) text = re.sub(",|;|:|!|\?|\"|'|‘|’|ʻ|ʿ|‚|“|”|\"|„|«|»|›|‹|\(|\)", "", text) text = re.sub("°|幺|乡|辶", "", text) text = re.sub("&", " and ", text) text = re.sub("ə", "a", text) text = re.sub("æ", "ae", text) text = re.sub("å|ā|á|ã|ă|â|ą", "a", text) text = re.sub("ó|ð|ø|ọ|ő|ō|ô", "o", text) text = re.sub("é|ë|ê|ě|ę", "e", text) text = re.sub("ū|ứ", "u", text) text = re.sub("í|ï|ı", "i", text) text = re.sub("š|ș|ś|ş", "s", text) text = re.sub("č|ć", "c", text) text = re.sub("đ", "d", text) text = re.sub("ğ", "g", text) text = re.sub("ł", "l", text) text = re.sub("ř", "r", text) text = re.sub("ñ", "n", text) text = re.sub("ț", "t", text) text = re.sub("ž|ź", "z", text) text = re.sub("\s+", " ", text) elif lang == "it": text = re.sub("\.|-|–|—|/|_|…", " ", text) text = re.sub(",|;|:|!|\?|\"|“|”|\"|„|«|»|›|‹|<|>|\(|\)", "", text) text = re.sub("\$|#|禅", "", text) text = re.sub("’|`", "'", text) text = re.sub("ə", "a", text) text = "".join((c if c in self.vocabulary else unidecode.unidecode(c)) for c in text) text = re.sub("\s+", " ", text) elif lang == "es": text = re.sub("\.|-|–|—|/|=|_|{|…", " ", text) text = re.sub(",|;|:|!|\?|\"|“|”|\"|„|«|»|›|‹|<|>|\(|\)|¿|¡", "", text) text = re.sub("蝦|夷", "", text) text = "".join((c if c in self.vocabulary else unidecode.unidecode(c)) for c in text) text = re.sub("\s+", " ", text) elif lang == "ca": text = re.sub("\.|-|–|—|/|=|_|·|@|\+|…", " ", text) text = re.sub(",|;|:|!|\?|\"|“|”|\"|„|«|»|›|‹|<|>|\(|\)|¿|¡", "", text) text = re.sub("ঃ|ং", "", text) text = "".join((c if c in self.vocabulary else unidecode.unidecode(c)) for c in text) text = re.sub("\s+", " ", text) elif lang == "pl": text = re.sub("\.|-|–|—|/|=|_|·|@|\+|…", " ", text) text = re.sub(",|;|:|!|\?|\"|“|”|\"|„|«|»|›|‹|<|>|\(|\)", "", text) text = re.sub("q", "k", text) text = re.sub("x", "ks", text) text = re.sub("v", "w", text) text = "".join((c if c in self.vocabulary else unidecode.unidecode(c)) for c in text) text = re.sub("\s+", " ", text) elif lang in ("ru", "ru34"): text = re.sub("по-", "по", text) text = re.sub("во-", "во", text) text = re.sub("-то", "то", text) text = re.sub("\.|−|-|–|—|…", " ", text) text = re.sub(",|;|:|!|\?|‘|’|\"|“|”|«|»|'", "", text) text = re.sub("m", "м", text) text = re.sub("o", "о", text) text = re.sub("z", "з", text) text = re.sub("i", "и", text) text = re.sub("l", "л", text) text = re.sub("a", "а", text) text = re.sub("f", "ф", text) text = re.sub("r", "р", text) text = re.sub("e", "е", text) text = re.sub("x", "кс", text) text = re.sub("h", "х", text) text = re.sub("\s+", " ", text) if lang == "ru34": text = re.sub("ё", "е", text) text = re.sub(" $", "", text) # print("<== {}".format(text)) text = np.array([vocabulary_dict[c] for c in text], dtype=np.long) self.data.append((wav_file_path, text)) # continue if os.path.exists(wav_file_path): continue # pass x, sr = librosa.load(path=mp3_file_path, sr=None) if desired_audio_sample_rate != sr: y = lr_resample(y=x, orig_sr=sr, target_sr=desired_audio_sample_rate) soundfile.write(file=wav_file_path, data=y, samplerate=desired_audio_sample_rate) @staticmethod def get_vocabulary_for_lang(lang="en"): """ Get the vocabulary for a language. Parameters: ---------- lang : str, default 'en' Language. Returns: ------- list of str Vocabulary set. """ assert (lang in ("en", "fr", "de", "it", "es", "ca", "pl", "ru", "ru34")) if lang == "en": return [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'"] elif lang == "fr": return [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'", 'ç', 'é', 'â', 'ê', 'î', 'ô', 'û', 'à', 'è', 'ù', 'ë', 'ï', 'ü', 'ÿ'] elif lang == "de": return [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', 'ä', 'ö', 'ü', 'ß'] elif lang == "it": return [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'", 'à', 'é', 'è', 'í', 'ì', 'î', 'ó', 'ò', 'ú', 'ù'] elif lang == "es": return [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'", 'á', 'é', 'í', 'ó', 'ú', 'ñ', 'ü'] elif lang == "ca": return [' ', 'a', 'b', 'c', 'd', 'e', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'm', 'n', 'o', 'p', 'q', 'r', 's', 't', 'u', 'v', 'w', 'x', 'y', 'z', "'", 'à', 'é', 'è', 'í', 'ï', 'ó', 'ò', 'ú', 'ü', 'ŀ'] elif lang == "pl": return [' ', 'a', 'ą', 'b', 'c', 'ć', 'd', 'e', 'ę', 'f', 'g', 'h', 'i', 'j', 'k', 'l', 'ł', 'm', 'n', 'ń', 'o', 'ó', 'p', 'r', 's', 'ś', 't', 'u', 'w', 'y', 'z', 'ź', 'ż'] elif lang == "ru": return [' ', 'а', 'б', 'в', 'г', 'д', 'е', 'ё', 'ж', 'з', 'и', 'й', 'к', 'л', 'м', 'н', 'о', 'п', 'р', 'с', 'т', 'у', 'ф', 'х', 'ц', 'ч', 'ш', 'щ', 'ъ', 'ы', 'ь', 'э', 'ю', 'я'] elif lang == "ru34": return [' ', 'а', 'б', 'в', 'г', 'д', 'е', 'ж', 'з', 'и', 'й', 'к', 'л', 'м', 'н', 'о', 'п', 'р', 'с', 'т', 'у', 'ф', 'х', 'ц', 'ч', 'ш', 'щ', 'ъ', 'ы', 'ь', 'э', 'ю', 'я'] else: return None class McvMetaInfo(DatasetMetaInfo): def __init__(self): super(McvMetaInfo, self).__init__() self.label = "MCV" self.short_label = "mcv" self.root_dir_name = "cv-corpus-6.1-2020-12-11" self.dataset_class = McvDataset self.lang = "en" self.dataset_class_extra_kwargs = { "lang": self.lang, "subset": "dev"} self.ml_type = "asr" self.num_classes = None self.val_metric_extra_kwargs = [{"vocabulary": None}] self.val_metric_capts = ["Val.WER"] self.val_metric_names = ["WER"] self.test_metric_extra_kwargs = [{"vocabulary": None}] self.test_metric_capts = ["Test.WER"] self.test_metric_names = ["WER"] self.val_transform = asr_test_transform self.test_transform = asr_test_transform self.saver_acc_ind = 0 def add_dataset_parser_arguments(self, parser, work_dir_path): """ Create python script parameters (for dataset specific metainfo). Parameters: ---------- parser : ArgumentParser ArgumentParser instance. work_dir_path : str Path to working directory. """ super(McvMetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--lang", type=str, default="en", help="language") parser.add_argument( "--subset", type=str, default="dev", help="data subset") def update(self, args): """ Update dataset metainfo after user customizing. Parameters: ---------- args : ArgumentParser Main script arguments. """ super(McvMetaInfo, self).update(args) self.lang = args.lang self.dataset_class_extra_kwargs["lang"] = args.lang self.dataset_class_extra_kwargs["subset"] = args.subset def update_from_dataset(self, dataset): """ Update dataset metainfo after a dataset class instance creation. Parameters: ---------- args : obj A dataset class instance. """ vocabulary = dataset._data.vocabulary self.num_classes = len(vocabulary) + 1 self.val_metric_extra_kwargs[0]["vocabulary"] = vocabulary self.test_metric_extra_kwargs[0]["vocabulary"] = vocabulary
14,293
41.924925
119
py
imgclsmob
imgclsmob-master/gluon/datasets/cityscapes_seg_dataset.py
""" Cityscapes semantic segmentation dataset. """ import os import numpy as np import mxnet as mx from PIL import Image from .seg_dataset import SegDataset from .voc_seg_dataset import VOCMetaInfo class CityscapesSegDataset(SegDataset): """ Cityscapes semantic segmentation dataset. Parameters: ---------- root : str Path to a folder with `leftImg8bit` and `gtFine` subfolders. mode : str, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None, **kwargs): super(CityscapesSegDataset, self).__init__( root=root, mode=mode, transform=transform, **kwargs) image_dir_path = os.path.join(root, "leftImg8bit") mask_dir_path = os.path.join(root, "gtFine") assert os.path.exists(image_dir_path) and os.path.exists(mask_dir_path), "Please prepare dataset" mode_dir_name = "train" if mode == "train" else "val" image_dir_path = os.path.join(image_dir_path, mode_dir_name) # mask_dir_path = os.path.join(mask_dir_path, mode_dir_name) self.images = [] self.masks = [] for image_subdir_path, _, image_file_names in os.walk(image_dir_path): for image_file_name in image_file_names: if image_file_name.endswith(".png"): image_file_path = os.path.join(image_subdir_path, image_file_name) mask_file_name = image_file_name.replace("leftImg8bit", "gtFine_labelIds") mask_subdir_path = image_subdir_path.replace("leftImg8bit", "gtFine") mask_file_path = os.path.join(mask_subdir_path, mask_file_name) if os.path.isfile(mask_file_path): self.images.append(image_file_path) self.masks.append(mask_file_path) else: print("Cannot find the mask: {}".format(mask_file_path)) assert (len(self.images) == len(self.masks)) if len(self.images) == 0: raise RuntimeError("Found 0 images in subfolders of: {}\n".format(image_dir_path)) def __getitem__(self, index): image = Image.open(self.images[index]).convert("RGB") if self.mode == "demo": image = self._img_transform(image) if self.transform is not None: image = self.transform(image) return image, os.path.basename(self.images[index]) mask = Image.open(self.masks[index]) if self.mode == "train": image, mask = self._train_sync_transform(image, mask) elif self.mode == "val": image, mask = self._val_sync_transform(image, mask) else: assert (self.mode == "test") image = self._img_transform(image) mask = self._mask_transform(mask) if self.transform is not None: image = self.transform(image) return image, mask classes = 19 vague_idx = 19 use_vague = True background_idx = -1 ignore_bg = False _key = np.array([-1, -1, -1, -1, -1, -1, -1, -1, 0, 1, -1, -1, 2, 3, 4, -1, -1, -1, 5, -1, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, -1, -1, 16, 17, 18]) _mapping = np.array(range(-1, len(_key) - 1)).astype(np.int32) @staticmethod def _class_to_index(mask): values = np.unique(mask) for value in values: assert(value in CityscapesSegDataset._mapping) index = np.digitize(mask.ravel(), CityscapesSegDataset._mapping, right=True) return CityscapesSegDataset._key[index].reshape(mask.shape) @staticmethod def _mask_transform(mask): np_mask = np.array(mask).astype(np.int32) np_mask = CityscapesSegDataset._class_to_index(np_mask) np_mask[np_mask == -1] = CityscapesSegDataset.vague_idx return mx.nd.array(np_mask, mx.cpu()) def __len__(self): return len(self.images) class CityscapesMetaInfo(VOCMetaInfo): def __init__(self): super(CityscapesMetaInfo, self).__init__() self.label = "Cityscapes" self.short_label = "voc" self.root_dir_name = "cityscapes" self.dataset_class = CityscapesSegDataset self.num_classes = CityscapesSegDataset.classes self.test_metric_extra_kwargs = [ {"vague_idx": CityscapesSegDataset.vague_idx, "use_vague": CityscapesSegDataset.use_vague, "macro_average": False}, {"num_classes": CityscapesSegDataset.classes, "vague_idx": CityscapesSegDataset.vague_idx, "use_vague": CityscapesSegDataset.use_vague, "bg_idx": CityscapesSegDataset.background_idx, "ignore_bg": CityscapesSegDataset.ignore_bg, "macro_average": False}]
5,110
36.306569
105
py
imgclsmob
imgclsmob-master/gluon/datasets/coco_seg_dataset.py
""" COCO semantic segmentation dataset. """ import os import logging import numpy as np import mxnet as mx from PIL import Image from tqdm import trange from .seg_dataset import SegDataset from .voc_seg_dataset import VOCMetaInfo class CocoSegDataset(SegDataset): """ COCO semantic segmentation dataset. Parameters: ---------- root : string Path to `annotations`, `train2017`, and `val2017` folders. mode : string, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None, **kwargs): super(CocoSegDataset, self).__init__( root=root, mode=mode, transform=transform, **kwargs) year = "2017" mode_name = "train" if mode == "train" else "val" annotations_dir_path = os.path.join(root, "annotations") annotations_file_path = os.path.join(annotations_dir_path, "instances_" + mode_name + year + ".json") idx_file_path = os.path.join(annotations_dir_path, mode_name + "_idx.npy") self.image_dir_path = os.path.join(root, mode_name + year) from pycocotools.coco import COCO from pycocotools import mask as coco_mask self.coco = COCO(annotations_file_path) self.coco_mask = coco_mask if os.path.exists(idx_file_path): self.idx = np.load(idx_file_path) else: idx_list = list(self.coco.imgs.keys()) self.idx = self._filter_idx(idx_list, idx_file_path) def __getitem__(self, index): image_id = int(self.idx[index]) image_metadata = self.coco.loadImgs(image_id)[0] image_file_name = image_metadata["file_name"] image_file_path = os.path.join(self.image_dir_path, image_file_name) image = Image.open(image_file_path).convert("RGB") if self.mode == "demo": image = self._img_transform(image) if self.transform is not None: image = self.transform(image) return image, os.path.basename(image_file_path) coco_target = self.coco.loadAnns(self.coco.getAnnIds(imgIds=image_id)) mask = Image.fromarray(self._gen_seg_mask( target=coco_target, height=image_metadata["height"], width=image_metadata["width"])) if self.mode == "train": image, mask = self._train_sync_transform(image, mask) elif self.mode == "val": image, mask = self._val_sync_transform(image, mask) else: assert (self.mode == "test") image, mask = self._img_transform(image), self._mask_transform(mask) if self.transform is not None: image = self.transform(image) return image, mask def _gen_seg_mask(self, target, height, width): cat_list = [0, 5, 2, 16, 9, 44, 6, 3, 17, 62, 21, 67, 18, 19, 4, 1, 64, 20, 63, 7, 72] mask = np.zeros((height, width), dtype=np.uint8) for instance in target: rle = self.coco_mask.frPyObjects(instance["segmentation"], height, width) m = self.coco_mask.decode(rle) cat = instance["category_id"] if cat in cat_list: c = cat_list.index(cat) else: continue if len(m.shape) < 3: mask[:, :] += (mask == 0) * (m * c) else: mask[:, :] += (mask == 0) * (((np.sum(m, axis=2)) > 0) * c).astype(np.uint8) return mask def _filter_idx(self, idx_list, idx_file_path, pixels_thr=1000): logging.info("Filtering mask index:") tbar = trange(len(idx_list)) filtered_idx = [] for i in tbar: img_id = idx_list[i] coco_target = self.coco.loadAnns(self.coco.getAnnIds(imgIds=img_id)) img_metadata = self.coco.loadImgs(img_id)[0] mask = self._gen_seg_mask( coco_target, img_metadata["height"], img_metadata["width"]) if (mask > 0).sum() > pixels_thr: filtered_idx.append(img_id) tbar.set_description("Doing: {}/{}, got {} qualified images".format(i, len(idx_list), len(filtered_idx))) logging.info("Found number of qualified images: {}".format(len(filtered_idx))) np.save(idx_file_path, np.array(filtered_idx, np.int32)) return filtered_idx classes = 21 vague_idx = -1 use_vague = False background_idx = 0 ignore_bg = True @staticmethod def _mask_transform(mask, ctx=mx.cpu()): np_mask = np.array(mask).astype(np.int32) # print("min={}, max={}".format(np_mask.min(), np_mask.max())) return mx.nd.array(np_mask, ctx=ctx) def __len__(self): return len(self.idx) class CocoSegMetaInfo(VOCMetaInfo): def __init__(self): super(CocoSegMetaInfo, self).__init__() self.label = "COCO" self.short_label = "coco" self.root_dir_name = "coco" self.dataset_class = CocoSegDataset self.num_classes = CocoSegDataset.classes self.train_metric_extra_kwargs = [ {"vague_idx": CocoSegDataset.vague_idx, "use_vague": CocoSegDataset.use_vague, "macro_average": False, "aux": self.train_aux}] self.val_metric_extra_kwargs = [ {"vague_idx": CocoSegDataset.vague_idx, "use_vague": CocoSegDataset.use_vague, "macro_average": False}, {"num_classes": CocoSegDataset.classes, "vague_idx": CocoSegDataset.vague_idx, "use_vague": CocoSegDataset.use_vague, "bg_idx": CocoSegDataset.background_idx, "ignore_bg": CocoSegDataset.ignore_bg, "macro_average": False}] self.test_metric_extra_kwargs = self.val_metric_extra_kwargs
6,102
35.54491
117
py
imgclsmob
imgclsmob-master/gluon/datasets/voc_seg_dataset.py
""" Pascal VOC2012 semantic segmentation dataset. """ import os import numpy as np import mxnet as mx from PIL import Image from mxnet.gluon.data.vision import transforms from .seg_dataset import SegDataset from .dataset_metainfo import DatasetMetaInfo class VOCSegDataset(SegDataset): """ Pascal VOC2012 semantic segmentation dataset. Parameters: ---------- root : str Path to VOCdevkit folder. mode : str, default 'train' 'train', 'val', 'test', or 'demo'. transform : callable, optional A function that transforms the image. """ def __init__(self, root, mode="train", transform=None, **kwargs): super(VOCSegDataset, self).__init__( root=root, mode=mode, transform=transform, **kwargs) base_dir_path = os.path.join(root, "VOC2012") image_dir_path = os.path.join(base_dir_path, "JPEGImages") mask_dir_path = os.path.join(base_dir_path, "SegmentationClass") splits_dir_path = os.path.join(base_dir_path, "ImageSets", "Segmentation") if mode == "train": split_file_path = os.path.join(splits_dir_path, "train.txt") elif mode in ("val", "test", "demo"): split_file_path = os.path.join(splits_dir_path, "val.txt") else: raise RuntimeError("Unknown dataset splitting mode") self.images = [] self.masks = [] with open(os.path.join(split_file_path), "r") as lines: for line in lines: image_file_path = os.path.join(image_dir_path, line.rstrip('\n') + ".jpg") assert os.path.isfile(image_file_path) self.images.append(image_file_path) mask_file_path = os.path.join(mask_dir_path, line.rstrip('\n') + ".png") assert os.path.isfile(mask_file_path) self.masks.append(mask_file_path) assert (len(self.images) == len(self.masks)) def __getitem__(self, index): image = Image.open(self.images[index]).convert("RGB") if self.mode == "demo": image = self._img_transform(image) if self.transform is not None: image = self.transform(image) return image, os.path.basename(self.images[index]) mask = Image.open(self.masks[index]) if self.mode == "train": image, mask = self._train_sync_transform(image, mask) elif self.mode == "val": image, mask = self._val_sync_transform(image, mask) else: assert self.mode == "test" image, mask = self._img_transform(image), self._mask_transform(mask) if self.transform is not None: image = self.transform(image) return image, mask classes = 21 vague_idx = 255 use_vague = True background_idx = 0 ignore_bg = True @staticmethod def _mask_transform(mask, ctx=mx.cpu()): np_mask = np.array(mask).astype(np.int32) # np_mask[np_mask == 255] = VOCSegDataset.vague_idx return mx.nd.array(np_mask, ctx=ctx) def __len__(self): return len(self.images) def voc_transform(ds_metainfo, mean_rgb=(0.485, 0.456, 0.406), std_rgb=(0.229, 0.224, 0.225)): assert (ds_metainfo is not None) return transforms.Compose([ transforms.ToTensor(), transforms.Normalize( mean=mean_rgb, std=std_rgb) ]) class VOCMetaInfo(DatasetMetaInfo): def __init__(self): super(VOCMetaInfo, self).__init__() self.label = "VOC" self.short_label = "voc" self.root_dir_name = "voc" self.dataset_class = VOCSegDataset self.num_training_samples = None self.in_channels = 3 self.num_classes = VOCSegDataset.classes self.train_aux = False self.input_image_size = (480, 480) self.train_metric_capts = ["Train.PixAcc"] self.train_metric_names = ["PixelAccuracyMetric"] self.train_metric_extra_kwargs = [ {"vague_idx": VOCSegDataset.vague_idx, "use_vague": VOCSegDataset.use_vague, "macro_average": False, "aux": self.train_aux}] self.val_metric_capts = ["Val.PixAcc", "Val.IoU"] self.val_metric_names = ["PixelAccuracyMetric", "MeanIoUMetric"] self.val_metric_extra_kwargs = [ {"vague_idx": VOCSegDataset.vague_idx, "use_vague": VOCSegDataset.use_vague, "macro_average": False}, {"num_classes": VOCSegDataset.classes, "vague_idx": VOCSegDataset.vague_idx, "use_vague": VOCSegDataset.use_vague, "bg_idx": VOCSegDataset.background_idx, "ignore_bg": VOCSegDataset.ignore_bg, "macro_average": False}] self.test_metric_capts = ["Test.PixAcc", "Test.IoU"] self.test_metric_names = self.val_metric_names self.test_metric_extra_kwargs = self.val_metric_extra_kwargs self.saver_acc_ind = 1 self.do_transform = True self.train_transform = voc_transform self.val_transform = voc_transform self.test_transform = voc_transform self.ml_type = "imgseg" self.allow_hybridize = False self.train_net_extra_kwargs = {"aux": self.train_aux} self.test_net_extra_kwargs = {"aux": False, "fixed_size": False} self.load_ignore_extra = True self.image_base_size = 520 self.image_crop_size = 480 self.loss_name = "SegSoftmaxCrossEntropy" self.loss_extra_kwargs = None # self.loss_name = "MixSoftmaxCrossEntropy" # self.loss_extra_kwargs = {"aux": self.train_aux, "aux_weight": 0.5} def add_dataset_parser_arguments(self, parser, work_dir_path): super(VOCMetaInfo, self).add_dataset_parser_arguments(parser, work_dir_path) parser.add_argument( "--image-base-size", type=int, default=520, help="base image size") parser.add_argument( "--image-crop-size", type=int, default=480, help="crop image size") def update(self, args): super(VOCMetaInfo, self).update(args) self.image_base_size = args.image_base_size self.image_crop_size = args.image_crop_size
6,554
34.625
90
py
imgclsmob
imgclsmob-master/gluon/datasets/cifar100_cls_dataset.py
""" CIFAR-100 classification dataset. """ import os from mxnet.gluon.data.vision import CIFAR100 from .cifar10_cls_dataset import CIFAR10MetaInfo class CIFAR100Fine(CIFAR100): """ CIFAR-100 image classification dataset. Parameters: ---------- root : str, default $MXNET_HOME/datasets/cifar100 Path to temp folder for storing data. mode : str, default 'train' 'train', 'val', or 'test'. transform : function, default None A user defined callback that transforms each sample. """ def __init__(self, root=os.path.join("~", ".mxnet", "datasets", "cifar100"), mode="train", transform=None): super(CIFAR100Fine, self).__init__( root=root, fine_label=True, train=(mode == "train"), transform=transform) class CIFAR100MetaInfo(CIFAR10MetaInfo): def __init__(self): super(CIFAR100MetaInfo, self).__init__() self.label = "CIFAR100" self.root_dir_name = "cifar100" self.dataset_class = CIFAR100Fine self.num_classes = 100
1,133
26
74
py
imgclsmob
imgclsmob-master/gluon/datasets/hpatches_mch_dataset.py
""" HPatches image matching dataset. """ import os import cv2 import numpy as np import mxnet as mx from mxnet.gluon.data import dataset from mxnet.gluon.data.vision import transforms from .dataset_metainfo import DatasetMetaInfo class HPatches(dataset.Dataset): """ HPatches (full image sequences) image matching dataset. Info URL: https://github.com/hpatches/hpatches-dataset Data URL: http://icvl.ee.ic.ac.uk/vbalnt/hpatches/hpatches-sequences-release.tar.gz Parameters: ---------- root : str, default '~/.mxnet/datasets/hpatches' Path to the folder stored the dataset. mode : str, default 'train' 'train', 'val', or 'test'. alteration : str, default 'all' 'all', 'i' for illumination or 'v' for viewpoint. transform : function, default None A function that takes data and label and transforms them. """ def __init__(self, root=os.path.join("~", ".mxnet", "datasets", "hpatches"), mode="train", alteration="all", transform=None): super(HPatches, self).__init__() assert os.path.exists(root) num_images = 5 image_file_ext = ".ppm" self.mode = mode self.image_paths = [] self.warped_image_paths = [] self.homographies = [] subdir_names = [name for name in os.listdir(root) if os.path.isdir(os.path.join(root, name))] if alteration != "all": subdir_names = [name for name in subdir_names if name[0] == alteration] for subdir_name in subdir_names: subdir_path = os.path.join(root, subdir_name) for i in range(num_images): k = i + 2 self.image_paths.append(os.path.join(subdir_path, "1" + image_file_ext)) self.warped_image_paths.append(os.path.join(subdir_path, str(k) + image_file_ext)) self.homographies.append(np.loadtxt(os.path.join(subdir_path, "H_1_" + str(k)))) self.transform = transform def __getitem__(self, index): # image = cv2.imread(self.image_paths[index], flags=cv2.IMREAD_GRAYSCALE) # warped_image = cv2.imread(self.warped_image_paths[index], flags=cv2.IMREAD_GRAYSCALE) # image = mx.image.imread(self.image_paths[index], flag=0) # warped_image = mx.image.imread(self.warped_image_paths[index], flag=0) print("Image file name: {}, index: {}".format(self.image_paths[index], index)) image = cv2.imread(self.image_paths[index], flags=0) if image.shape[0] > 1500: image = cv2.resize( src=image, dsize=None, fx=0.5, fy=0.5, interpolation=cv2.INTER_AREA) image = mx.nd.array(np.expand_dims(image, axis=2)) print("Image shape: {}".format(image.shape)) warped_image = cv2.imread(self.warped_image_paths[index], flags=0) if warped_image.shape[0] > 1500: warped_image = cv2.resize( src=warped_image, dsize=None, fx=0.5, fy=0.5, interpolation=cv2.INTER_AREA) warped_image = mx.nd.array(np.expand_dims(warped_image, axis=2)) print("W-Image shape: {}".format(warped_image.shape)) homography = mx.nd.array(self.homographies[index]) if self.transform is not None: image = self.transform(image) warped_image = self.transform(warped_image) return image, warped_image, homography def __len__(self): return len(self.image_paths) class HPatchesMetaInfo(DatasetMetaInfo): def __init__(self): super(HPatchesMetaInfo, self).__init__() self.label = "hpatches" self.short_label = "hpatches" self.root_dir_name = "hpatches" self.dataset_class = HPatches self.ml_type = "imgmch" self.do_transform = True self.val_transform = hpatches_val_transform self.test_transform = hpatches_val_transform self.allow_hybridize = False self.test_net_extra_kwargs = {"hybridizable": False, "in_size": None} def hpatches_val_transform(ds_metainfo): assert (ds_metainfo is not None) return transforms.Compose([ transforms.ToTensor() ]) def _test(): dataset = HPatches( root="../imgclsmob_data/hpatches", mode="train", alteration="i", transform=None) scale_factor = 0.5 for image, warped_image, _ in dataset: cv2.imshow( winname="image", mat=cv2.resize( src=image, dsize=None, fx=scale_factor, fy=scale_factor, interpolation=cv2.INTER_NEAREST)) cv2.imshow( winname="warped_image", mat=cv2.resize( src=warped_image, dsize=None, fx=scale_factor, fy=scale_factor, interpolation=cv2.INTER_NEAREST)) cv2.waitKey(0) assert (dataset is not None) if __name__ == "__main__": _test()
5,163
33.198675
101
py
imgclsmob
imgclsmob-master/pytorch/dataset_utils.py
""" Dataset routines. """ __all__ = ['get_dataset_metainfo', 'get_train_data_source', 'get_val_data_source', 'get_test_data_source'] from .datasets.imagenet1k_cls_dataset import ImageNet1KMetaInfo from .datasets.cub200_2011_cls_dataset import CUB200MetaInfo from .datasets.cifar10_cls_dataset import CIFAR10MetaInfo from .datasets.cifar100_cls_dataset import CIFAR100MetaInfo from .datasets.svhn_cls_dataset import SVHNMetaInfo from .datasets.voc_seg_dataset import VOCMetaInfo from .datasets.ade20k_seg_dataset import ADE20KMetaInfo from .datasets.cityscapes_seg_dataset import CityscapesMetaInfo from .datasets.coco_seg_dataset import CocoSegMetaInfo from .datasets.coco_det_dataset import CocoDetMetaInfo from .datasets.coco_hpe1_dataset import CocoHpe1MetaInfo from .datasets.coco_hpe2_dataset import CocoHpe2MetaInfo from .datasets.coco_hpe3_dataset import CocoHpe3MetaInfo from .datasets.hpatches_mch_dataset import HPatchesMetaInfo from .datasets.librispeech_asr_dataset import LibriSpeechMetaInfo from .datasets.mcv_asr_dataset import McvMetaInfo from torch.utils.data import DataLoader from torch.utils.data.sampler import WeightedRandomSampler def get_dataset_metainfo(dataset_name): """ Get dataset metainfo by name of dataset. Parameters: ---------- dataset_name : str Dataset name. Returns: ------- DatasetMetaInfo Dataset metainfo. """ dataset_metainfo_map = { "ImageNet1K": ImageNet1KMetaInfo, "CUB200_2011": CUB200MetaInfo, "CIFAR10": CIFAR10MetaInfo, "CIFAR100": CIFAR100MetaInfo, "SVHN": SVHNMetaInfo, "VOC": VOCMetaInfo, "ADE20K": ADE20KMetaInfo, "Cityscapes": CityscapesMetaInfo, "CocoSeg": CocoSegMetaInfo, "CocoDet": CocoDetMetaInfo, "CocoHpe1": CocoHpe1MetaInfo, "CocoHpe2": CocoHpe2MetaInfo, "CocoHpe3": CocoHpe3MetaInfo, "HPatches": HPatchesMetaInfo, "LibriSpeech": LibriSpeechMetaInfo, "MCV": McvMetaInfo, } if dataset_name in dataset_metainfo_map.keys(): return dataset_metainfo_map[dataset_name]() else: raise Exception("Unrecognized dataset: {}".format(dataset_name)) def get_train_data_source(ds_metainfo, batch_size, num_workers): """ Get data source for training subset. Parameters: ---------- ds_metainfo : DatasetMetaInfo Dataset metainfo. batch_size : int Batch size. num_workers : int Number of background workers. Returns: ------- DataLoader Data source. """ transform_train = ds_metainfo.train_transform(ds_metainfo=ds_metainfo) kwargs = ds_metainfo.dataset_class_extra_kwargs if ds_metainfo.dataset_class_extra_kwargs is not None else {} dataset = ds_metainfo.dataset_class( root=ds_metainfo.root_dir_path, mode="train", transform=transform_train, **kwargs) ds_metainfo.update_from_dataset(dataset) if not ds_metainfo.train_use_weighted_sampler: return DataLoader( dataset=dataset, batch_size=batch_size, shuffle=True, num_workers=num_workers, pin_memory=True) else: sampler = WeightedRandomSampler( weights=dataset.sample_weights, num_samples=len(dataset)) return DataLoader( dataset=dataset, batch_size=batch_size, # shuffle=True, sampler=sampler, num_workers=num_workers, pin_memory=True) def get_val_data_source(ds_metainfo, batch_size, num_workers): """ Get data source for validation subset. Parameters: ---------- ds_metainfo : DatasetMetaInfo Dataset metainfo. batch_size : int Batch size. num_workers : int Number of background workers. Returns: ------- DataLoader Data source. """ transform_val = ds_metainfo.val_transform(ds_metainfo=ds_metainfo) kwargs = ds_metainfo.dataset_class_extra_kwargs if ds_metainfo.dataset_class_extra_kwargs is not None else {} dataset = ds_metainfo.dataset_class( root=ds_metainfo.root_dir_path, mode="val", transform=transform_val, **kwargs) ds_metainfo.update_from_dataset(dataset) return DataLoader( dataset=dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True) def get_test_data_source(ds_metainfo, batch_size, num_workers): """ Get data source for testing subset. Parameters: ---------- ds_metainfo : DatasetMetaInfo Dataset metainfo. batch_size : int Batch size. num_workers : int Number of background workers. Returns: ------- DataLoader Data source. """ transform_test = ds_metainfo.test_transform(ds_metainfo=ds_metainfo) kwargs = ds_metainfo.dataset_class_extra_kwargs if ds_metainfo.dataset_class_extra_kwargs is not None else {} dataset = ds_metainfo.dataset_class( root=ds_metainfo.root_dir_path, mode="test", transform=transform_test, **kwargs) ds_metainfo.update_from_dataset(dataset) return DataLoader( dataset=dataset, batch_size=batch_size, shuffle=False, num_workers=num_workers, pin_memory=True)
5,563
29.404372
113
py
imgclsmob
imgclsmob-master/pytorch/model_stats.py
""" Routines for model statistics calculation. """ import logging import numpy as np import torch import torch.nn as nn from torch.autograd import Variable from .pytorchcv.models.common import ChannelShuffle, ChannelShuffle2, Identity, Flatten, Swish, HSigmoid, HSwish,\ InterpolationBlock, HeatmapMaxDetBlock from .pytorchcv.models.fishnet import ChannelSqueeze from .pytorchcv.models.irevnet import IRevDownscale, IRevSplitBlock, IRevMergeBlock from .pytorchcv.models.rir_cifar import RiRFinalBlock from .pytorchcv.models.proxylessnas import ProxylessUnit from .pytorchcv.models.lwopenpose_cmupan import LwopDecoderFinalBlock from .pytorchcv.models.centernet import CenterNetHeatmapMaxDet from .pytorchcv.models.danet import ScaleBlock from .pytorchcv.models.jasper import MaskConv1d, NemoMelSpecExtractor __all__ = ['measure_model'] def calc_block_num_params2(net): """ Calculate number of trainable parameters in the block (not iterative). Parameters: ---------- net : Module Model/block. Returns: ------- int Number of parameters. """ net_params = filter(lambda p: p.requires_grad, net.parameters()) weight_count = 0 for param in net_params: weight_count += np.prod(param.size()) return weight_count def calc_block_num_params(module): """ Calculate number of trainable parameters in the block (iterative). Parameters: ---------- module : Module Model/block. Returns: ------- int Number of parameters. """ assert isinstance(module, nn.Module) net_params = filter(lambda p: isinstance(p[1], nn.parameter.Parameter) and p[1].requires_grad, module._parameters.items()) weight_count = 0 for param in net_params: weight_count += np.prod(param[1].size()) return weight_count def measure_model(model, in_shapes): """ Calculate model statistics. Parameters: ---------- model : HybridBlock Tested model. in_shapes : list of tuple of ints Shapes of the input tensors. """ global num_flops global num_macs global num_params # global names num_flops = 0 num_macs = 0 num_params = 0 # names = {} def call_hook(module, x, y): if not (isinstance(module, IRevSplitBlock) or isinstance(module, IRevMergeBlock) or isinstance(module, RiRFinalBlock) or isinstance(module, InterpolationBlock) or isinstance(module, MaskConv1d) or isinstance(module, NemoMelSpecExtractor)): assert (len(x) == 1) assert (len(module._modules) == 0) if isinstance(module, nn.Linear): batch = x[0].shape[0] in_units = module.in_features out_units = module.out_features extra_num_macs = in_units * out_units if module.bias is None: extra_num_flops = (2 * in_units - 1) * out_units else: extra_num_flops = 2 * in_units * out_units extra_num_flops *= batch extra_num_macs *= batch elif isinstance(module, nn.ReLU): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.ELU): extra_num_flops = 3 * x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.Sigmoid): extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.LeakyReLU): extra_num_flops = 2 * x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.ReLU6): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.PReLU): extra_num_flops = 3 * x[0].numel() extra_num_macs = 0 elif isinstance(module, Swish): extra_num_flops = 5 * x[0].numel() extra_num_macs = 0 elif isinstance(module, HSigmoid): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, HSwish): extra_num_flops = 2 * x[0].numel() extra_num_macs = 0 elif type(module) in [nn.ConvTranspose2d]: extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif type(module) in [nn.Conv2d]: batch = x[0].shape[0] x_h = x[0].shape[2] x_w = x[0].shape[3] kernel_size = module.kernel_size stride = module.stride dilation = module.dilation padding = module.padding groups = module.groups in_channels = module.in_channels out_channels = module.out_channels y_h = (x_h + 2 * padding[0] - dilation[0] * (kernel_size[0] - 1) - 1) // stride[0] + 1 y_w = (x_w + 2 * padding[1] - dilation[1] * (kernel_size[1] - 1) - 1) // stride[1] + 1 assert (out_channels == y.shape[1]) assert (y_h == y.shape[2]) assert (y_w == y.shape[3]) kernel_total_size = kernel_size[0] * kernel_size[1] y_size = y_h * y_w extra_num_macs = kernel_total_size * in_channels * y_size * out_channels // groups if module.bias is None: extra_num_flops = (2 * kernel_total_size * y_size - 1) * in_channels * out_channels // groups else: extra_num_flops = 2 * kernel_total_size * in_channels * y_size * out_channels // groups extra_num_flops *= batch extra_num_macs *= batch elif isinstance(module, nn.BatchNorm2d): extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.InstanceNorm2d): extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.BatchNorm1d): extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif type(module) in [nn.MaxPool2d, nn.AvgPool2d]: assert (x[0].shape[1] == y.shape[1]) batch = x[0].shape[0] kernel_size = module.kernel_size if isinstance(module.kernel_size, tuple) else\ (module.kernel_size, module.kernel_size) y_h = y.shape[2] y_w = y.shape[3] channels = x[0].shape[1] y_size = y_h * y_w pool_total_size = kernel_size[0] * kernel_size[1] extra_num_flops = channels * y_size * pool_total_size extra_num_macs = 0 extra_num_flops *= batch extra_num_macs *= batch elif type(module) in [nn.AdaptiveAvgPool2d, nn.AdaptiveMaxPool2d]: assert (x[0].shape[1] == y.shape[1]) batch = x[0].shape[0] x_h = x[0].shape[2] x_w = x[0].shape[3] y_h = y.shape[2] y_w = y.shape[3] channels = x[0].shape[1] y_size = y_h * y_w pool_total_size = x_h * x_w extra_num_flops = channels * y_size * pool_total_size extra_num_macs = 0 extra_num_flops *= batch extra_num_macs *= batch elif isinstance(module, nn.Dropout): extra_num_flops = 0 extra_num_macs = 0 elif isinstance(module, nn.Sequential): assert (len(module._modules) == 0) extra_num_flops = 0 extra_num_macs = 0 elif type(module) in [ChannelShuffle, ChannelShuffle2]: extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, nn.ZeroPad2d): extra_num_flops = 0 extra_num_macs = 0 elif isinstance(module, Identity): extra_num_flops = 0 extra_num_macs = 0 elif isinstance(module, nn.PixelShuffle): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, Flatten): extra_num_flops = 0 extra_num_macs = 0 elif isinstance(module, nn.Upsample): extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif isinstance(module, ChannelSqueeze): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, IRevDownscale): extra_num_flops = 5 * x[0].numel() extra_num_macs = 0 elif isinstance(module, IRevSplitBlock): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, IRevMergeBlock): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, RiRFinalBlock): extra_num_flops = x[0].numel() extra_num_macs = 0 elif isinstance(module, ProxylessUnit): extra_num_flops = x[0].numel() extra_num_macs = 0 elif type(module) in [nn.Softmax2d, nn.Softmax]: extra_num_flops = 4 * x[0].numel() extra_num_macs = 0 elif type(module) in [MaskConv1d, nn.Conv1d]: if isinstance(y, tuple): assert isinstance(module, MaskConv1d) y = y[0] batch = x[0].shape[0] x_h = x[0].shape[2] kernel_size = module.kernel_size stride = module.stride dilation = module.dilation padding = module.padding groups = module.groups in_channels = module.in_channels out_channels = module.out_channels y_h = (x_h + 2 * padding[0] - dilation[0] * (kernel_size[0] - 1) - 1) // stride[0] + 1 assert (out_channels == y.shape[1]) assert (y_h == y.shape[2]) kernel_total_size = kernel_size[0] y_size = y_h extra_num_macs = kernel_total_size * in_channels * y_size * out_channels // groups if module.bias is None: extra_num_flops = (2 * kernel_total_size * y_size - 1) * in_channels * out_channels // groups else: extra_num_flops = 2 * kernel_total_size * in_channels * y_size * out_channels // groups extra_num_flops *= batch extra_num_macs *= batch elif type(module) in [InterpolationBlock, HeatmapMaxDetBlock, CenterNetHeatmapMaxDet, ScaleBlock, NemoMelSpecExtractor]: extra_num_flops, extra_num_macs = module.calc_flops(x[0]) elif isinstance(module, LwopDecoderFinalBlock): if not module.calc_3d_features: extra_num_flops = 0 extra_num_macs = 0 else: raise TypeError("LwopDecoderFinalBlock!") else: raise TypeError("Unknown layer type: {}".format(type(module))) global num_flops global num_macs global num_params # global names num_flops += extra_num_flops num_macs += extra_num_macs # if module.name not in names: # names[module.name] = 1 # num_params += calc_block_num_params(module) num_params += calc_block_num_params(module) def register_forward_hooks(a_module): if len(a_module._modules) > 0: assert (calc_block_num_params(a_module) == 0) children_handles = [] for child_module in a_module._modules.values(): child_handles = register_forward_hooks(child_module) children_handles += child_handles return children_handles else: handle = a_module.register_forward_hook(call_hook) return [handle] hook_handles = register_forward_hooks(model) model.eval() if len(in_shapes) == 1: x = Variable(torch.zeros(*in_shapes[0])) model(x) elif len(in_shapes) == 2: x1 = Variable(torch.zeros(*in_shapes[0])) x2 = Variable(torch.zeros(*in_shapes[1])) model(x1, x2) else: raise NotImplementedError() num_params1 = calc_block_num_params2(model) if num_params != num_params1: logging.warning( "Calculated numbers of parameters are different: standard method: {},\tper-leaf method: {}".format( num_params1, num_params)) [h.remove() for h in hook_handles] return num_flops, num_macs, num_params1
12,391
37.01227
114
py
imgclsmob
imgclsmob-master/pytorch/setup.py
from setuptools import setup, find_packages from os import path from io import open here = path.abspath(path.dirname(__file__)) with open(path.join(here, 'README.md'), encoding='utf-8') as f: long_description = f.read() setup( name='pytorchcv', version='0.0.67', description='Image classification and segmentation models for PyTorch', license='MIT', long_description=long_description, long_description_content_type='text/markdown', url='https://github.com/osmr/imgclsmob', author='Oleg Sémery', author_email='osemery@gmail.com', classifiers=[ 'Development Status :: 3 - Alpha', 'Intended Audience :: Science/Research', 'License :: OSI Approved :: MIT License', 'Operating System :: OS Independent', 'Programming Language :: Python', 'Topic :: Scientific/Engineering :: Image Recognition', ], keywords='machine-learning deep-learning neuralnetwork image-classification pytorch imagenet cifar svhn vgg resnet ' 'pyramidnet diracnet densenet condensenet wrn drn dpn darknet fishnet espnetv2 xdensnet squeezenet ' 'squeezenext shufflenet menet mobilenet igcv3 mnasnet darts xception inception polynet nasnet pnasnet ror ' 'proxylessnas dianet efficientnet mixnet image-segmentation voc ade20k cityscapes coco pspnet deeplabv3 ' 'fcn', packages=find_packages(exclude=['datasets', 'metrics', 'others', '*.others', 'others.*', '*.others.*']), include_package_data=True, install_requires=['numpy', 'requests'], )
1,571
42.666667
120
py
imgclsmob
imgclsmob-master/pytorch/utils.py
""" Main routines shared between training and evaluation scripts. """ import logging import os import numpy as np import torch.utils.data from .pytorchcv.model_provider import get_model from .metrics.metric import EvalMetric, CompositeEvalMetric from .metrics.cls_metrics import Top1Error, TopKError from .metrics.seg_metrics import PixelAccuracyMetric, MeanIoUMetric from .metrics.det_metrics import CocoDetMApMetric from .metrics.hpe_metrics import CocoHpeOksApMetric from .metrics.asr_metrics import WER def prepare_pt_context(num_gpus, batch_size): """ Correct batch size. Parameters: ---------- num_gpus : int Number of GPU. batch_size : int Batch size for each GPU. Returns: ------- bool Whether to use CUDA. int Batch size for all GPUs. """ use_cuda = (num_gpus > 0) batch_size *= max(1, num_gpus) return use_cuda, batch_size def prepare_model(model_name, use_pretrained, pretrained_model_file_path, use_cuda, use_data_parallel=True, net_extra_kwargs=None, load_ignore_extra=False, num_classes=None, in_channels=None, remap_to_cpu=False, remove_module=False): """ Create and initialize model by name. Parameters: ---------- model_name : str Model name. use_pretrained : bool Whether to use pretrained weights. pretrained_model_file_path : str Path to file with pretrained weights. use_cuda : bool Whether to use CUDA. use_data_parallel : bool, default True Whether to use parallelization. net_extra_kwargs : dict, default None Extra parameters for model. load_ignore_extra : bool, default False Whether to ignore extra layers in pretrained model. num_classes : int, default None Number of classes. in_channels : int, default None Number of input channels. remap_to_cpu : bool, default False Whether to remape model to CPU during loading. remove_module : bool, default False Whether to remove module from loaded model. Returns: ------- Module Model. """ kwargs = {"pretrained": use_pretrained} if num_classes is not None: kwargs["num_classes"] = num_classes if in_channels is not None: kwargs["in_channels"] = in_channels if net_extra_kwargs is not None: kwargs.update(net_extra_kwargs) net = get_model(model_name, **kwargs) if pretrained_model_file_path: assert (os.path.isfile(pretrained_model_file_path)) logging.info("Loading model: {}".format(pretrained_model_file_path)) checkpoint = torch.load( pretrained_model_file_path, map_location=(None if use_cuda and not remap_to_cpu else "cpu")) if (type(checkpoint) == dict) and ("state_dict" in checkpoint): checkpoint = checkpoint["state_dict"] if load_ignore_extra: pretrained_state = checkpoint model_dict = net.state_dict() pretrained_state = {k: v for k, v in pretrained_state.items() if k in model_dict} net.load_state_dict(pretrained_state) else: if remove_module: net_tmp = torch.nn.DataParallel(net) net_tmp.load_state_dict(checkpoint) net.load_state_dict(net_tmp.module.cpu().state_dict()) else: net.load_state_dict(checkpoint) if use_data_parallel and use_cuda: net = torch.nn.DataParallel(net) if use_cuda: net = net.cuda() return net def calc_net_weight_count(net): """ Calculate number of model trainable parameters. Parameters: ---------- net : Module Model. Returns: ------- int Number of parameters. """ net.train() net_params = filter(lambda p: p.requires_grad, net.parameters()) weight_count = 0 for param in net_params: weight_count += np.prod(param.size()) return weight_count def validate(metric, net, val_data, use_cuda): """ Core validation/testing routine. Parameters: ---------- metric : EvalMetric Metric object instance. net : Module Model. val_data : DataLoader Data loader. use_cuda : bool Whether to use CUDA. Returns: ------- EvalMetric Metric object instance. """ net.eval() metric.reset() with torch.no_grad(): for data, target in val_data: if use_cuda: target = target.cuda(non_blocking=True) output = net(data) metric.update(target, output) return metric def report_accuracy(metric, extended_log=False): """ Make report string for composite metric. Parameters: ---------- metric : EvalMetric Metric object instance. extended_log : bool, default False Whether to log more precise accuracy values. Returns: ------- str Report string. """ def create_msg(name, value): if type(value) in [list, tuple]: if extended_log: return "{}={} ({})".format("{}", "/".join(["{:.4f}"] * len(value)), "/".join(["{}"] * len(value))).\ format(name, *(value + value)) else: return "{}={}".format("{}", "/".join(["{:.4f}"] * len(value))).format(name, *value) else: if extended_log: return "{name}={value:.4f} ({value})".format(name=name, value=value) else: return "{name}={value:.4f}".format(name=name, value=value) metric_info = metric.get() if isinstance(metric, CompositeEvalMetric): msg = ", ".join([create_msg(name=m[0], value=m[1]) for m in zip(*metric_info)]) elif isinstance(metric, EvalMetric): msg = create_msg(name=metric_info[0], value=metric_info[1]) else: raise Exception("Wrong metric type: {}".format(type(metric))) return msg def get_metric(metric_name, metric_extra_kwargs): """ Get metric by name. Parameters: ---------- metric_name : str Metric name. metric_extra_kwargs : dict Metric extra parameters. EvalMetric ------- EvalMetric Metric object instance. """ if metric_name == "Top1Error": return Top1Error(**metric_extra_kwargs) elif metric_name == "TopKError": return TopKError(**metric_extra_kwargs) elif metric_name == "PixelAccuracyMetric": return PixelAccuracyMetric(**metric_extra_kwargs) elif metric_name == "MeanIoUMetric": return MeanIoUMetric(**metric_extra_kwargs) elif metric_name == "CocoDetMApMetric": return CocoDetMApMetric(**metric_extra_kwargs) elif metric_name == "CocoHpeOksApMetric": return CocoHpeOksApMetric(**metric_extra_kwargs) elif metric_name == "WER": return WER(**metric_extra_kwargs) else: raise Exception("Wrong metric name: {}".format(metric_name)) def get_composite_metric(metric_names, metric_extra_kwargs): """ Get composite metric by list of metric names. Parameters: ---------- metric_names : list of str Metric name list. metric_extra_kwargs : list of dict Metric extra parameters list. Returns: ------- CompositeEvalMetric Metric object instance. """ if len(metric_names) == 1: metric = get_metric(metric_names[0], metric_extra_kwargs[0]) else: metric = CompositeEvalMetric() for name, extra_kwargs in zip(metric_names, metric_extra_kwargs): metric.add(get_metric(name, extra_kwargs)) return metric def get_metric_name(metric, index): """ Get metric name by index in the composite metric. Parameters: ---------- metric : CompositeEvalMetric or EvalMetric Metric object instance. index : int Index. Returns: ------- str Metric name. """ if isinstance(metric, CompositeEvalMetric): return metric.metrics[index].name elif isinstance(metric, EvalMetric): assert (index == 0) return metric.name else: raise Exception("Wrong metric type: {}".format(type(metric)))
8,538
26.996721
116
py
imgclsmob
imgclsmob-master/pytorch/metrics/seg_metrics.py
""" Evaluation Metrics for Semantic Segmentation. """ import numpy as np import torch from .metric import EvalMetric, check_label_shapes from .seg_metrics_np import seg_pixel_accuracy_np, seg_mean_iou_imasks_np __all__ = ['PixelAccuracyMetric', 'MeanIoUMetric'] class PixelAccuracyMetric(EvalMetric): """ Computes the pixel-wise accuracy. Parameters: ---------- axis : int, default 1 The axis that represents classes. name : str, default 'pix_acc' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. on_cpu : bool, default True Calculate on CPU. sparse_label : bool, default True Whether label is an integer array instead of probability distribution. vague_idx : int, default -1 Index of masked pixels. use_vague : bool, default False Whether to use pixel masking. macro_average : bool, default True Whether to use micro or macro averaging. """ def __init__(self, axis=1, name="pix_acc", output_names=None, label_names=None, on_cpu=True, sparse_label=True, vague_idx=-1, use_vague=False, macro_average=True): self.macro_average = macro_average super(PixelAccuracyMetric, self).__init__( name, axis=axis, output_names=output_names, label_names=label_names) self.axis = axis self.on_cpu = on_cpu self.sparse_label = sparse_label self.vague_idx = vague_idx self.use_vague = use_vague def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : torch.Tensor The labels of the data. preds : torch.Tensor Predicted values. """ with torch.no_grad(): check_label_shapes(labels, preds) if self.on_cpu: if self.sparse_label: label_imask = labels.cpu().numpy().astype(np.int32) else: label_imask = torch.argmax(labels, dim=self.axis).cpu().numpy().astype(np.int32) pred_imask = torch.argmax(preds, dim=self.axis).cpu().numpy().astype(np.int32) acc = seg_pixel_accuracy_np( label_imask=label_imask, pred_imask=pred_imask, vague_idx=self.vague_idx, use_vague=self.use_vague, macro_average=self.macro_average) if self.macro_average: self.sum_metric += acc self.num_inst += 1 else: self.sum_metric += acc[0] self.num_inst += acc[1] else: assert False def reset(self): """ Resets the internal evaluation result to initial state. """ if self.macro_average: self.num_inst = 0 self.sum_metric = 0.0 else: self.num_inst = 0 self.sum_metric = 0 def get(self): """ Gets the current evaluation result. Returns: ------- names : list of str Name of the metrics. values : list of float Value of the evaluations. """ if self.macro_average: if self.num_inst == 0: return self.name, float("nan") else: return self.name, self.sum_metric / self.num_inst else: if self.num_inst == 0: return self.name, float("nan") else: return self.name, float(self.sum_metric) / self.num_inst class MeanIoUMetric(EvalMetric): """ Computes the mean intersection over union. Parameters: ---------- axis : int, default 1 The axis that represents classes name : str, default 'mean_iou' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. on_cpu : bool, default True Calculate on CPU. sparse_label : bool, default True Whether label is an integer array instead of probability distribution. num_classes : int Number of classes vague_idx : int, default -1 Index of masked pixels. use_vague : bool, default False Whether to use pixel masking. bg_idx : int, default -1 Index of background class. ignore_bg : bool, default False Whether to ignore background class. macro_average : bool, default True Whether to use micro or macro averaging. """ def __init__(self, axis=1, name="mean_iou", output_names=None, label_names=None, on_cpu=True, sparse_label=True, num_classes=None, vague_idx=-1, use_vague=False, bg_idx=-1, ignore_bg=False, macro_average=True): self.macro_average = macro_average self.num_classes = num_classes self.ignore_bg = ignore_bg super(MeanIoUMetric, self).__init__( name, axis=axis, output_names=output_names, label_names=label_names) assert ((not ignore_bg) or (bg_idx in (0, num_classes - 1))) self.axis = axis self.on_cpu = on_cpu self.sparse_label = sparse_label self.vague_idx = vague_idx self.use_vague = use_vague self.bg_idx = bg_idx assert (on_cpu and sparse_label) def update(self, labels, preds): """ Updates the internal evaluation result. Parameters: ---------- labels : torch.Tensor The labels of the data. preds : torch.Tensor Predicted values. """ assert (len(labels) == len(preds)) with torch.no_grad(): if self.on_cpu: if self.sparse_label: label_imask = labels.cpu().numpy().astype(np.int32) else: assert False pred_imask = torch.argmax(preds, dim=self.axis).cpu().numpy().astype(np.int32) batch_size = labels.shape[0] for k in range(batch_size): if self.sparse_label: acc = seg_mean_iou_imasks_np( label_imask=label_imask[k, :, :], pred_imask=pred_imask[k, :, :], num_classes=self.num_classes, vague_idx=self.vague_idx, use_vague=self.use_vague, bg_idx=self.bg_idx, ignore_bg=self.ignore_bg, macro_average=self.macro_average) else: assert False if self.macro_average: self.sum_metric += acc self.num_inst += 1 else: self.area_inter += acc[0] self.area_union += acc[1] else: assert False def reset(self): """ Resets the internal evaluation result to initial state. """ if self.macro_average: self.num_inst = 0 self.sum_metric = 0.0 else: class_count = self.num_classes - 1 if self.ignore_bg else self.num_classes self.area_inter = np.zeros((class_count,), np.uint64) self.area_union = np.zeros((class_count,), np.uint64) def get(self): """ Gets the current evaluation result. Returns: ------- names : list of str Name of the metrics. values : list of float Value of the evaluations. """ if self.macro_average: if self.num_inst == 0: return self.name, float("nan") else: return self.name, self.sum_metric / self.num_inst else: class_count = (self.area_union > 0).sum() if class_count == 0: return self.name, float("nan") eps = np.finfo(np.float32).eps area_union_eps = self.area_union + eps mean_iou = (self.area_inter / area_union_eps).sum() / class_count return self.name, mean_iou
9,276
33.106618
100
py
imgclsmob
imgclsmob-master/pytorch/metrics/ret_metrics.py
""" Evaluation Metrics for Image Retrieval. """ import numpy as np import torch from .metric import EvalMetric __all__ = ['PointDetectionMatchRatio', 'PointDescriptionMatchRatio'] class PointDetectionMatchRatio(EvalMetric): """ Computes point detection match ratio (with mean residual). Parameters: ---------- pts_max_count : int Maximal count of points. axis : int, default 1 The axis that represents classes name : str, default 'accuracy' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. """ def __init__(self, pts_max_count, axis=1, name="pt_det_ratio", output_names=None, label_names=None): super(PointDetectionMatchRatio, self).__init__( name, axis=axis, output_names=output_names, label_names=label_names, has_global_stats=True) self.axis = axis self.pts_max_count = pts_max_count self.resudual_sum = 0.0 self.resudual_count = 0 def update_alt(self, homography, src_pts, dst_pts, src_confs, dst_confs, src_img_size, dst_img_size): """ Updates the internal evaluation result. Parameters: ---------- homography : torch.Tensor Homography (from source image to destination one). src_pts : torch.Tensor Detected points for the first (source) image. dst_pts : torch.Tensor Detected points for the second (destination) image. src_confs : torch.Tensor Confidences for detected points on the source image. dst_confs : torch.Tensor Confidences for detected points on the destination image. src_img_size : tuple of 2 int Size (H, W) of the source image. dst_img_size : tuple of 2 int Size (H, W) of the destination image. """ assert (src_confs.argsort(descending=True).cpu().detach().numpy() == np.arange(src_confs.shape[0])).all() assert (dst_confs.argsort(descending=True).cpu().detach().numpy() == np.arange(dst_confs.shape[0])).all() max_dist_sat_value = 1e5 eps = 1e-5 # print("src_img_size={}".format(src_img_size)) # print("dst_img_size={}".format(dst_img_size)) homography = homography.to(src_pts.device) self.normalize_homography(homography) homography_inv = self.calc_homography_inv(homography) # print("homography={}".format(homography)) # print("homography_inv={}".format(homography_inv)) # print("src_pts={}".format(src_pts[:10, :].int())) src_pts = src_pts.flip(dims=(1,)) dst_pts = dst_pts.flip(dims=(1,)) # print("src_pts={}".format(src_pts[:10, :].int())) # print("src_pts.shape={}".format(src_pts.shape)) # print("dst_pts.shape={}".format(dst_pts.shape)) # print("src_pts={}".format(src_pts[:10, :].int())) # print("dst_pts={}".format(dst_pts[:10, :].int())) # with torch.no_grad(): src_hmg_pts = self.calc_homogeneous_coords(src_pts.float()) dst_hmg_pts = self.calc_homogeneous_coords(dst_pts.float()) # print("src_hmg_pts={}".format(src_hmg_pts[:10, :].int())) # print("dst_hmg_pts={}".format(dst_hmg_pts[:10, :].int())) src_hmg_pts, src_confs = self.filter_inside_points( src_hmg_pts, src_confs, homography, dst_img_size) dst_hmg_pts, dst_confs = self.filter_inside_points( dst_hmg_pts, dst_confs, homography_inv, src_img_size) # print("src_hmg_pts.shape={}".format(src_hmg_pts.shape)) # print("dst_hmg_pts.shape={}".format(dst_hmg_pts.shape)) # # print("src_hmg_pts={}".format(src_hmg_pts[:10, :].int())) # print("dst_hmg_pts={}".format(dst_hmg_pts[:10, :].int())) src_pts_count = src_hmg_pts.shape[0] dst_pts_count = dst_hmg_pts.shape[0] src_pts_count2 = min(src_pts_count, self.pts_max_count) src_hmg_pts, conf_thr = self.filter_best_points( hmg_pts=src_hmg_pts, confs=src_confs, max_count=src_pts_count2, min_conf=None) dst_pts_count2 = min(dst_pts_count, self.pts_max_count) dst_hmg_pts, _ = self.filter_best_points( hmg_pts=dst_hmg_pts, confs=dst_confs, max_count=dst_pts_count2, min_conf=conf_thr) # print("src_hmg_pts.shape={}".format(src_hmg_pts.shape)) # print("dst_hmg_pts.shape={}".format(dst_hmg_pts.shape)) # print("src_hmg_pts={}".format(src_hmg_pts[:10, :].int())) # print("dst_hmg_pts={}".format(dst_hmg_pts[:10, :].int())) preds_dst_hmg_pts = self.transform_points( src_hmg_pts, homography) # print("preds_dst_hmg_pts={}".format(preds_dst_hmg_pts[:10, :].int())) cost = self.calc_pairwise_distances(x=preds_dst_hmg_pts, y=dst_hmg_pts).cpu().detach().numpy() self.saturate_distance_matrix( dist_mat=cost, max_dist_thr=8.0, max_dist_sat=max_dist_sat_value) # print("cost.shape={}".format(cost.shape)) from scipy.optimize import linear_sum_assignment row_ind, col_ind = linear_sum_assignment(cost) # print("row_ind.shape={}".format(row_ind.shape)) # print("col_ind.shape={}".format(col_ind.shape)) resuduals = cost[row_ind, col_ind] resuduals = resuduals[resuduals < (max_dist_sat_value - eps)] resudual_count = len(resuduals) self.sum_metric += resudual_count self.global_sum_metric += resudual_count self.num_inst += src_pts_count2 self.global_num_inst += src_pts_count2 print("ratio_resudual={}".format(float(resudual_count) / src_pts_count2)) if resudual_count != 0: self.resudual_sum += resuduals.sum() self.resudual_count += resudual_count @staticmethod def normalize_homography(homography): homography /= homography[2, 2] @staticmethod def calc_homography_inv(homography): homography_inv = homography.inverse() PointDetectionMatchRatio.normalize_homography(homography_inv) return homography_inv @staticmethod def calc_homogeneous_coords(pts): hmg_pts = torch.cat((pts, torch.ones((pts.shape[0], 1), dtype=pts.dtype, device=pts.device)), dim=1) return hmg_pts @staticmethod def calc_cartesian_coords(hmg_pts): pts = hmg_pts[:, :2] return pts @staticmethod def transform_points(src_hmg_pts, homography): # print("transform_points -> src_hmg_pts.shape={}".format(src_hmg_pts.shape)) # print("transform_points -> homography.shape={}".format(homography.shape)) # print("homography={}".format(homography)) # print("transform_points -> src_hmg_pts={}".format(src_hmg_pts[:10, :].int())) dst_hmg_pts = torch.matmul(src_hmg_pts, homography.t()) # print("transform_points -> dst_hmg_pts={}".format(dst_hmg_pts[:10, :].int())) # print("transform_points -> dst_hmg_pts.shape={}".format(dst_hmg_pts.shape)) dst_hmg_pts /= dst_hmg_pts[:, 2:] return dst_hmg_pts @staticmethod def calc_inside_pts_mask(pts, img_size): eps = 1e-3 border_size = 1.0 border = border_size - eps mask = (pts[:, 0] >= border) & (pts[:, 0] < img_size[0] - border) &\ (pts[:, 1] >= border) & (pts[:, 1] < img_size[1] - border) return mask @staticmethod def filter_inside_points(src_hmg_pts, src_confs, homography, dst_img_size): # print("fip->src_hmg_pts.shape={}".format(src_hmg_pts.shape)) # print("fip->src_hmg_pts={}".format(src_hmg_pts[:10, :].int())) # print("fip->src_confs.shape={}".format(src_confs.shape)) # print("fip->src_confs={}".format(src_confs[:10])) # print("homography_inv={}".format(homography)) dst_hmg_pts = PointDetectionMatchRatio.transform_points(src_hmg_pts, homography) # print("fip->dst_hmg_pts.shape={}".format(dst_hmg_pts.shape)) # print("fip->dst_hmg_pts={}".format(dst_hmg_pts[:10, :])) mask = PointDetectionMatchRatio.calc_inside_pts_mask(dst_hmg_pts, dst_img_size) # print("fip->mask={}".format(mask[:10])) # print("fip->mask.sum()={}".format(mask.sum())) return src_hmg_pts[mask], src_confs[mask] @staticmethod def filter_best_points(hmg_pts, confs, max_count, min_conf=None): if min_conf is not None: max_ind = (confs < min_conf).nonzero()[0, 0].item() max_count = max(max_count, max_ind) inds = confs.argsort(descending=True)[:max_count] return hmg_pts[inds], confs[inds][-1] @staticmethod def calc_pairwise_distances(x, y): diff = x.unsqueeze(1) - y.unsqueeze(0) return torch.sum(diff * diff, dim=-1).sqrt() @staticmethod def saturate_distance_matrix(dist_mat, max_dist_thr, max_dist_sat): dist_mat[dist_mat > max_dist_thr] = max_dist_sat class PointDescriptionMatchRatio(EvalMetric): """ Computes point description match ratio. Parameters: ---------- pts_max_count : int Maximal count of points. dist_ratio_thr : float, default 0.9 Distance ratio threshold for point filtering. axis : int, default 1 The axis that represents classes name : str, default 'accuracy' Name of this metric instance for display. output_names : list of str, or None, default None Name of predictions that should be used when updating with update_dict. By default include all predictions. label_names : list of str, or None, default None Name of labels that should be used when updating with update_dict. By default include all labels. """ def __init__(self, pts_max_count, dist_ratio_thr=0.95, axis=1, name="pt_desc_ratio", output_names=None, label_names=None): super(PointDescriptionMatchRatio, self).__init__( name, axis=axis, output_names=output_names, label_names=label_names, has_global_stats=True) self.axis = axis self.pts_max_count = pts_max_count self.dist_ratio_thr = dist_ratio_thr self.resudual_sum = 0.0 self.resudual_count = 0 def update_alt(self, homography, src_pts, dst_pts, src_descs, dst_descs, src_img_size, dst_img_size): """ Updates the internal evaluation result. Parameters: ---------- homography : torch.Tensor Homography (from source image to destination one). src_pts : torch.Tensor Detected points for the first (source) image. dst_pts : torch.Tensor Detected points for the second (destination) image. src_descs : torch.Tensor Descriptors for detected points on the source image. dst_descs : torch.Tensor Descriptors for detected points on the destination image. src_img_size : tuple of 2 int Size (H, W) of the source image. dst_img_size : tuple of 2 int Size (H, W) of the destination image. """ # max_dist_sat_value = 1e5 # eps = 1e-5 homography = homography.to(src_pts.device) self.normalize_homography(homography) homography_inv = self.calc_homography_inv(homography) src_pts = src_pts.flip(dims=(1,)) dst_pts = dst_pts.flip(dims=(1,)) src_hmg_pts = self.calc_homogeneous_coords(src_pts.float()) dst_hmg_pts = self.calc_homogeneous_coords(dst_pts.float()) src_hmg_pts = self.filter_inside_points( src_hmg_pts, homography, dst_img_size) dst_hmg_pts = self.filter_inside_points( dst_hmg_pts, homography_inv, src_img_size) src_pts_count = src_hmg_pts.shape[0] dst_pts_count = dst_hmg_pts.shape[0] src_pts_count2 = min(src_pts_count, self.pts_max_count * 10) src_hmg_pts, src_descs = self.filter_best_points( hmg_pts=src_hmg_pts, descs=src_descs, max_count=src_pts_count2) dst_pts_count2 = min(dst_pts_count, self.pts_max_count * 10) dst_hmg_pts, dst_descs = self.filter_best_points( hmg_pts=dst_hmg_pts, descs=dst_descs, max_count=dst_pts_count2) dist_mat = self.calc_pairwise_distances(x=src_descs, y=dst_descs) vals, inds = dist_mat.topk(k=2, dim=1, largest=True, sorted=True) inds = inds[:, 0][(vals[:, 1] / vals[:, 0]) < 0.95] src_hmg_pts = src_hmg_pts[inds] preds_dst_hmg_pts = self.transform_points( src_hmg_pts, homography) print(preds_dst_hmg_pts) # self.saturate_distance_matrix( # dist_mat=cost, # max_dist_thr=8.0, # max_dist_sat=max_dist_sat_value) # # # print("cost.shape={}".format(cost.shape)) # # from scipy.optimize import linear_sum_assignment # row_ind, col_ind = linear_sum_assignment(cost) # # # print("row_ind.shape={}".format(row_ind.shape)) # # print("col_ind.shape={}".format(col_ind.shape)) # # resuduals = cost[row_ind, col_ind] # resuduals = resuduals[resuduals < (max_dist_sat_value - eps)] # resudual_count = len(resuduals) resudual_count = 1 self.sum_metric += resudual_count self.global_sum_metric += resudual_count self.num_inst += src_pts_count2 self.global_num_inst += src_pts_count2 print("ratio_resudual={}".format(float(resudual_count) / src_pts_count2)) @staticmethod def normalize_homography(homography): homography /= homography[2, 2] @staticmethod def calc_homography_inv(homography): homography_inv = homography.inverse() PointDetectionMatchRatio.normalize_homography(homography_inv) return homography_inv @staticmethod def calc_homogeneous_coords(pts): hmg_pts = torch.cat((pts, torch.ones((pts.shape[0], 1), dtype=pts.dtype, device=pts.device)), dim=1) return hmg_pts @staticmethod def calc_cartesian_coords(hmg_pts): pts = hmg_pts[:, :2] return pts @staticmethod def transform_points(src_hmg_pts, homography): # print("transform_points -> src_hmg_pts.shape={}".format(src_hmg_pts.shape)) # print("transform_points -> homography.shape={}".format(homography.shape)) # print("homography={}".format(homography)) # print("transform_points -> src_hmg_pts={}".format(src_hmg_pts[:10, :].int())) dst_hmg_pts = torch.matmul(src_hmg_pts, homography.t()) # print("transform_points -> dst_hmg_pts={}".format(dst_hmg_pts[:10, :].int())) # print("transform_points -> dst_hmg_pts.shape={}".format(dst_hmg_pts.shape)) dst_hmg_pts /= dst_hmg_pts[:, 2:] return dst_hmg_pts @staticmethod def calc_inside_pts_mask(pts, img_size): eps = 1e-3 border_size = 1.0 border = border_size - eps mask = (pts[:, 0] >= border) & (pts[:, 0] < img_size[0] - border) &\ (pts[:, 1] >= border) & (pts[:, 1] < img_size[1] - border) return mask @staticmethod def filter_inside_points(src_hmg_pts, homography, dst_img_size): dst_hmg_pts = PointDetectionMatchRatio.transform_points(src_hmg_pts, homography) mask = PointDetectionMatchRatio.calc_inside_pts_mask(dst_hmg_pts, dst_img_size) return src_hmg_pts[mask] @staticmethod def filter_best_points(hmg_pts, descs, max_count): return hmg_pts[:max_count], descs[:max_count] @staticmethod def calc_pairwise_distances(x, y): diff = x.unsqueeze(1) - y.unsqueeze(0) return torch.sum(diff * diff, dim=-1).sqrt() @staticmethod def saturate_distance_matrix(dist_mat, max_dist_thr, max_dist_sat): dist_mat[dist_mat > max_dist_thr] = max_dist_sat
17,535
34.56998
113
py