repo
stringlengths
2
99
file
stringlengths
13
225
code
stringlengths
0
18.3M
file_length
int64
0
18.3M
avg_line_length
float64
0
1.36M
max_line_length
int64
0
4.26M
extension_type
stringclasses
1 value
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/fishnet.py
""" FishNet for ImageNet-1K, implemented in PyTorch. Original paper: 'FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction,' http://papers.nips.cc/paper/7356-fishnet-a-versatile-backbone-for-image-region-and-pixel-level-prediction.pdf. """ __all__ = ['FishNet', 'fishnet99', 'fishnet150', 'ChannelSqueeze'] import os import torch.nn as nn import torch.nn.init as init from .common import pre_conv1x1_block, pre_conv3x3_block, conv1x1, SesquialteralHourglass, Identity, InterpolationBlock from .preresnet import PreResActivation from .senet import SEInitBlock def channel_squeeze(x, groups): """ Channel squeeze operation. Parameters: ---------- x : Tensor Input tensor. groups : int Number of groups. Returns: ------- Tensor Resulted tensor. """ batch, channels, height, width = x.size() channels_per_group = channels // groups x = x.view(batch, channels_per_group, groups, height, width).sum(dim=2) return x class ChannelSqueeze(nn.Module): """ Channel squeeze 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): super(ChannelSqueeze, self).__init__() if channels % groups != 0: raise ValueError("channels must be divisible by groups") self.groups = groups def forward(self, x): return channel_squeeze(x, self.groups) class PreSEAttBlock(nn.Module): """ FishNet specific Squeeze-and-Excitation attention block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. reduction : int, default 16 Squeeze reduction value. """ def __init__(self, in_channels, out_channels, reduction=16): super(PreSEAttBlock, self).__init__() mid_cannels = out_channels // reduction self.bn = nn.BatchNorm2d(num_features=in_channels) self.relu = nn.ReLU(inplace=True) self.pool = nn.AdaptiveAvgPool2d(output_size=1) self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_cannels, bias=True) self.conv2 = conv1x1( in_channels=mid_cannels, out_channels=out_channels, bias=True) self.sigmoid = nn.Sigmoid() def forward(self, x): x = self.bn(x) x = self.relu(x) x = self.pool(x) x = self.conv1(x) x = self.relu(x) x = self.conv2(x) x = self.sigmoid(x) return x class FishBottleneck(nn.Module): """ FishNet bottleneck block for residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. dilation : int or tuple/list of 2 int Dilation value for convolution layer. """ def __init__(self, in_channels, out_channels, stride, dilation): super(FishBottleneck, self).__init__() mid_channels = out_channels // 4 self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, padding=dilation, dilation=dilation) self.conv3 = pre_conv1x1_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class FishBlock(nn.Module): """ FishNet block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. squeeze : bool, default False Whether to use a channel squeeze operation. """ def __init__(self, in_channels, out_channels, stride=1, dilation=1, squeeze=False): super(FishBlock, self).__init__() self.squeeze = squeeze self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = FishBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilation) if self.squeeze: assert (in_channels // 2 == out_channels) self.c_squeeze = ChannelSqueeze( channels=in_channels, groups=2) elif self.resize_identity: self.identity_conv = pre_conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride) def forward(self, x): if self.squeeze: identity = self.c_squeeze(x) elif self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity return x class DownUnit(nn.Module): """ FishNet down unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of output channels for each block. """ def __init__(self, in_channels, out_channels_list): super(DownUnit, self).__init__() self.blocks = nn.Sequential() for i, out_channels in enumerate(out_channels_list): self.blocks.add_module("block{}".format(i + 1), FishBlock( in_channels=in_channels, out_channels=out_channels)) in_channels = out_channels self.pool = nn.MaxPool2d( kernel_size=2, stride=2) def forward(self, x): x = self.blocks(x) x = self.pool(x) return x class UpUnit(nn.Module): """ FishNet up unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of output channels for each block. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. """ def __init__(self, in_channels, out_channels_list, dilation=1): super(UpUnit, self).__init__() self.blocks = nn.Sequential() for i, out_channels in enumerate(out_channels_list): squeeze = (dilation > 1) and (i == 0) self.blocks.add_module("block{}".format(i + 1), FishBlock( in_channels=in_channels, out_channels=out_channels, dilation=dilation, squeeze=squeeze)) in_channels = out_channels self.upsample = InterpolationBlock(scale_factor=2, mode="nearest", align_corners=None) def forward(self, x): x = self.blocks(x) x = self.upsample(x) return x class SkipUnit(nn.Module): """ FishNet skip connection unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of output channels for each block. """ def __init__(self, in_channels, out_channels_list): super(SkipUnit, self).__init__() self.blocks = nn.Sequential() for i, out_channels in enumerate(out_channels_list): self.blocks.add_module("block{}".format(i + 1), FishBlock( in_channels=in_channels, out_channels=out_channels)) in_channels = out_channels def forward(self, x): x = self.blocks(x) return x class SkipAttUnit(nn.Module): """ FishNet skip connection unit with attention block. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int Number of output channels for each block. """ def __init__(self, in_channels, out_channels_list): super(SkipAttUnit, self).__init__() mid_channels1 = in_channels // 2 mid_channels2 = 2 * in_channels self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels1) self.conv2 = pre_conv1x1_block( in_channels=mid_channels1, out_channels=mid_channels2, bias=True) in_channels = mid_channels2 self.se = PreSEAttBlock( in_channels=mid_channels2, out_channels=out_channels_list[-1]) self.blocks = nn.Sequential() for i, out_channels in enumerate(out_channels_list): self.blocks.add_module("block{}".format(i + 1), FishBlock( in_channels=in_channels, out_channels=out_channels)) in_channels = out_channels def forward(self, x): x = self.conv1(x) x = self.conv2(x) w = self.se(x) x = self.blocks(x) x = x * w + w return x class FishFinalBlock(nn.Module): """ FishNet final block. Parameters: ---------- in_channels : int Number of input channels. """ def __init__(self, in_channels): super(FishFinalBlock, self).__init__() mid_channels = in_channels // 2 self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.preactiv = PreResActivation( in_channels=mid_channels) def forward(self, x): x = self.conv1(x) x = self.preactiv(x) return x class FishNet(nn.Module): """ FishNet model from 'FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction,' http://papers.nips.cc/paper/7356-fishnet-a-versatile-backbone-for-image-region-and-pixel-level-prediction.pdf. Parameters: ---------- direct_channels : list of list of list of int Number of output channels for each unit along the straight path. skip_channels : list of list of list of int Number of output channels for each skip connection unit. init_block_channels : int Number of output channels for the initial unit. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, direct_channels, skip_channels, init_block_channels, in_channels=3, in_size=(224, 224), num_classes=1000): super(FishNet, self).__init__() self.in_size = in_size self.num_classes = num_classes depth = len(direct_channels[0]) down1_channels = direct_channels[0] up_channels = direct_channels[1] down2_channels = direct_channels[2] skip1_channels = skip_channels[0] skip2_channels = skip_channels[1] self.features = nn.Sequential() self.features.add_module("init_block", SEInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels down1_seq = nn.Sequential() skip1_seq = nn.Sequential() for i in range(depth + 1): skip1_channels_list = skip1_channels[i] if i < depth: skip1_seq.add_module("unit{}".format(i + 1), SkipUnit( in_channels=in_channels, out_channels_list=skip1_channels_list)) down1_channels_list = down1_channels[i] down1_seq.add_module("unit{}".format(i + 1), DownUnit( in_channels=in_channels, out_channels_list=down1_channels_list)) in_channels = down1_channels_list[-1] else: skip1_seq.add_module("unit{}".format(i + 1), SkipAttUnit( in_channels=in_channels, out_channels_list=skip1_channels_list)) in_channels = skip1_channels_list[-1] up_seq = nn.Sequential() skip2_seq = nn.Sequential() for i in range(depth + 1): skip2_channels_list = skip2_channels[i] if i > 0: in_channels += skip1_channels[depth - i][-1] if i < depth: skip2_seq.add_module("unit{}".format(i + 1), SkipUnit( in_channels=in_channels, out_channels_list=skip2_channels_list)) up_channels_list = up_channels[i] dilation = 2 ** i up_seq.add_module("unit{}".format(i + 1), UpUnit( in_channels=in_channels, out_channels_list=up_channels_list, dilation=dilation)) in_channels = up_channels_list[-1] else: skip2_seq.add_module("unit{}".format(i + 1), Identity()) down2_seq = nn.Sequential() for i in range(depth): down2_channels_list = down2_channels[i] down2_seq.add_module("unit{}".format(i + 1), DownUnit( in_channels=in_channels, out_channels_list=down2_channels_list)) in_channels = down2_channels_list[-1] + skip2_channels[depth - 1 - i][-1] self.features.add_module("hg", SesquialteralHourglass( down1_seq=down1_seq, skip1_seq=skip1_seq, up_seq=up_seq, skip2_seq=skip2_seq, down2_seq=down2_seq)) self.features.add_module("final_block", FishFinalBlock(in_channels=in_channels)) in_channels = in_channels // 2 self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Sequential() self.output.add_module("final_conv", conv1x1( in_channels=in_channels, out_channels=num_classes, bias=True)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_fishnet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create FishNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 99: direct_layers = [[2, 2, 6], [1, 1, 1], [1, 2, 2]] skip_layers = [[1, 1, 1, 2], [4, 1, 1, 0]] elif blocks == 150: direct_layers = [[2, 4, 8], [2, 2, 2], [2, 2, 4]] skip_layers = [[2, 2, 2, 4], [4, 2, 2, 0]] else: raise ValueError("Unsupported FishNet with number of blocks: {}".format(blocks)) direct_channels_per_layers = [[128, 256, 512], [512, 384, 256], [320, 832, 1600]] skip_channels_per_layers = [[64, 128, 256, 512], [512, 768, 512, 0]] direct_channels = [[[b] * c for (b, c) in zip(*a)] for a in ([(ci, li) for (ci, li) in zip(direct_channels_per_layers, direct_layers)])] skip_channels = [[[b] * c for (b, c) in zip(*a)] for a in ([(ci, li) for (ci, li) in zip(skip_channels_per_layers, skip_layers)])] init_block_channels = 64 net = FishNet( direct_channels=direct_channels, skip_channels=skip_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def fishnet99(**kwargs): """ FishNet-99 model from 'FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction,' http://papers.nips.cc/paper/7356-fishnet-a-versatile-backbone-for-image-region-and-pixel-level-prediction.pdf. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_fishnet(blocks=99, model_name="fishnet99", **kwargs) def fishnet150(**kwargs): """ FishNet-150 model from 'FishNet: A Versatile Backbone for Image, Region, and Pixel Level Prediction,' http://papers.nips.cc/paper/7356-fishnet-a-versatile-backbone-for-image-region-and-pixel-level-prediction.pdf. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_fishnet(blocks=150, model_name="fishnet150", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ fishnet99, fishnet150, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fishnet99 or weight_count == 16628904) assert (model != fishnet150 or weight_count == 24959400) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
19,302
30.033762
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/hrnet.py
""" HRNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. """ __all__ = ['hrnet_w18_small_v1', 'hrnet_w18_small_v2', 'hrnetv2_w18', 'hrnetv2_w30', 'hrnetv2_w32', 'hrnetv2_w40', 'hrnetv2_w44', 'hrnetv2_w48', 'hrnetv2_w64'] import os import torch.nn as nn from .common import conv1x1_block, conv3x3_block, Identity from .resnet import ResUnit class UpSamplingBlock(nn.Module): """ HFNet specific upsampling block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. scale_factor : int Multiplier for spatial size. """ def __init__(self, in_channels, out_channels, scale_factor): super(UpSamplingBlock, self).__init__() self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=1, activation=None) self.upsample = nn.Upsample( scale_factor=scale_factor, mode="nearest") def forward(self, x): x = self.conv(x) x = self.upsample(x) return x class HRBlock(nn.Module): """ HFNet block. Parameters: ---------- in_channels_list : list of int Number of input channels. out_channels_list : list of int Number of output channels. num_branches : int Number of branches. num_subblocks : list of int Number of subblock. """ def __init__(self, in_channels_list, out_channels_list, num_branches, num_subblocks): super(HRBlock, self).__init__() self.in_channels_list = in_channels_list self.num_branches = num_branches self.branches = nn.Sequential() for i in range(num_branches): layers = nn.Sequential() in_channels_i = self.in_channels_list[i] out_channels_i = out_channels_list[i] for j in range(num_subblocks[i]): layers.add_module("unit{}".format(j + 1), ResUnit( in_channels=in_channels_i, out_channels=out_channels_i, stride=1, bottleneck=False)) in_channels_i = out_channels_i self.in_channels_list[i] = out_channels_i self.branches.add_module("branch{}".format(i + 1), layers) if num_branches > 1: self.fuse_layers = nn.Sequential() for i in range(num_branches): fuse_layer = nn.Sequential() for j in range(num_branches): if j > i: fuse_layer.add_module("block{}".format(j + 1), UpSamplingBlock( in_channels=in_channels_list[j], out_channels=in_channels_list[i], scale_factor=2 ** (j - i))) elif j == i: fuse_layer.add_module("block{}".format(j + 1), Identity()) else: conv3x3_seq = nn.Sequential() for k in range(i - j): if k == i - j - 1: conv3x3_seq.add_module("subblock{}".format(k + 1), conv3x3_block( in_channels=in_channels_list[j], out_channels=in_channels_list[i], stride=2, activation=None)) else: conv3x3_seq.add_module("subblock{}".format(k + 1), conv3x3_block( in_channels=in_channels_list[j], out_channels=in_channels_list[j], stride=2)) fuse_layer.add_module("block{}".format(j + 1), conv3x3_seq) self.fuse_layers.add_module("layer{}".format(i + 1), fuse_layer) self.activ = nn.ReLU(True) def forward(self, x): for i in range(self.num_branches): x[i] = self.branches[i](x[i]) if self.num_branches == 1: return x x_fuse = [] for i in range(len(self.fuse_layers)): y = x[0] if i == 0 else self.fuse_layers[i][0](x[0]) for j in range(1, self.num_branches): if i == j: y = y + x[j] else: y = y + self.fuse_layers[i][j](x[j]) x_fuse.append(self.activ(y)) return x_fuse class HRStage(nn.Module): """ HRNet stage block. Parameters: ---------- in_channels_list : list of int Number of output channels from the previous layer. out_channels_list : list of int Number of output channels in the current layer. num_modules : int Number of modules. num_branches : int Number of branches. num_subblocks : list of int Number of subblocks. """ def __init__(self, in_channels_list, out_channels_list, num_modules, num_branches, num_subblocks): super(HRStage, self).__init__() self.branches = num_branches self.in_channels_list = out_channels_list in_branches = len(in_channels_list) out_branches = len(out_channels_list) self.transition = nn.Sequential() for i in range(out_branches): if i < in_branches: if out_channels_list[i] != in_channels_list[i]: self.transition.add_module("block{}".format(i + 1), conv3x3_block( in_channels=in_channels_list[i], out_channels=out_channels_list[i], stride=1)) else: self.transition.add_module("block{}".format(i + 1), Identity()) else: conv3x3_seq = nn.Sequential() for j in range(i + 1 - in_branches): in_channels_i = in_channels_list[-1] out_channels_i = out_channels_list[i] if j == i - in_branches else in_channels_i conv3x3_seq.add_module("subblock{}".format(j + 1), conv3x3_block( in_channels=in_channels_i, out_channels=out_channels_i, stride=2)) self.transition.add_module("block{}".format(i + 1), conv3x3_seq) self.layers = nn.Sequential() for i in range(num_modules): self.layers.add_module("block{}".format(i + 1), HRBlock( in_channels_list=self.in_channels_list, out_channels_list=out_channels_list, num_branches=num_branches, num_subblocks=num_subblocks)) self.in_channels_list = self.layers[-1].in_channels_list def forward(self, x): x_list = [] for j in range(self.branches): if not isinstance(self.transition[j], Identity): x_list.append(self.transition[j](x[-1] if type(x) is list else x)) else: x_list_j = x[j] if type(x) is list else x x_list.append(x_list_j) y_list = self.layers(x_list) return y_list class HRInitBlock(nn.Module): """ HRNet 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. num_subblocks : int Number of subblocks. """ def __init__(self, in_channels, out_channels, mid_channels, num_subblocks): super(HRInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=2) in_channels = mid_channels self.subblocks = nn.Sequential() for i in range(num_subblocks): self.subblocks.add_module("block{}".format(i + 1), ResUnit( in_channels=in_channels, out_channels=out_channels, stride=1, bottleneck=True)) in_channels = out_channels def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.subblocks(x) return x class HRFinalBlock(nn.Module): """ HRNet specific final block. Parameters: ---------- in_channels_list : list of int Number of input channels per stage. out_channels_list : list of int Number of output channels per stage. """ def __init__(self, in_channels_list, out_channels_list): super(HRFinalBlock, self).__init__() self.inc_blocks = nn.Sequential() for i, in_channels_i in enumerate(in_channels_list): self.inc_blocks.add_module("block{}".format(i + 1), ResUnit( in_channels=in_channels_i, out_channels=out_channels_list[i], stride=1, bottleneck=True)) self.down_blocks = nn.Sequential() for i in range(len(in_channels_list) - 1): self.down_blocks.add_module("block{}".format(i + 1), conv3x3_block( in_channels=out_channels_list[i], out_channels=out_channels_list[i + 1], stride=2, bias=True)) self.final_layer = conv1x1_block( in_channels=1024, out_channels=2048, stride=1, bias=True) def forward(self, x): y = self.inc_blocks[0](x[0]) for i in range(len(self.down_blocks)): y = self.inc_blocks[i + 1](x[i + 1]) + self.down_blocks[i](y) y = self.final_layer(y) return y class HRNet(nn.Module): """ HRNet model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- channels : list of int Number of output channels for each unit. init_block_channels : int Number of output channels for the initial unit. init_num_subblocks : int Number of subblocks in the initial unit. num_modules : int Number of modules per stage. num_subblocks : list of int Number of subblocks per stage. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, init_num_subblocks, num_modules, num_subblocks, in_channels=3, in_size=(224, 224), num_classes=1000): super(HRNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.branches = [2, 3, 4] self.features = nn.Sequential() self.features.add_module("init_block", HRInitBlock( in_channels=in_channels, out_channels=init_block_channels, mid_channels=64, num_subblocks=init_num_subblocks)) in_channels_list = [init_block_channels] for i in range(len(self.branches)): self.features.add_module("stage{}".format(i + 1), HRStage( in_channels_list=in_channels_list, out_channels_list=channels[i], num_modules=num_modules[i], num_branches=self.branches[i], num_subblocks=num_subblocks[i])) in_channels_list = self.features[-1].in_channels_list self.features.add_module("final_block", HRFinalBlock( in_channels_list=in_channels_list, out_channels_list=[128, 256, 512, 1024])) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=2048, out_features=num_classes) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight, mode="fan_out", nonlinearity="relu") if module.bias is not None: nn.init.constant_(module.bias, 0) elif isinstance(module, nn.BatchNorm2d): nn.init.constant_(module.weight, 1) nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_hrnet(version, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create HRNet model with specific parameters. Parameters: ---------- version : str Version of MobileNetV3 ('s' or 'm'). 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if version == "w18s1": init_block_channels = 128 init_num_subblocks = 1 channels = [[16, 32], [16, 32, 64], [16, 32, 64, 128]] num_modules = [1, 1, 1] elif version == "w18s2": init_block_channels = 256 init_num_subblocks = 2 channels = [[18, 36], [18, 36, 72], [18, 36, 72, 144]] num_modules = [1, 3, 2] elif version == "w18": init_block_channels = 256 init_num_subblocks = 4 channels = [[18, 36], [18, 36, 72], [18, 36, 72, 144]] num_modules = [1, 4, 3] elif version == "w30": init_block_channels = 256 init_num_subblocks = 4 channels = [[30, 60], [30, 60, 120], [30, 60, 120, 240]] num_modules = [1, 4, 3] elif version == "w32": init_block_channels = 256 init_num_subblocks = 4 channels = [[32, 64], [32, 64, 128], [32, 64, 128, 256]] num_modules = [1, 4, 3] elif version == "w40": init_block_channels = 256 init_num_subblocks = 4 channels = [[40, 80], [40, 80, 160], [40, 80, 160, 320]] num_modules = [1, 4, 3] elif version == "w44": init_block_channels = 256 init_num_subblocks = 4 channels = [[44, 88], [44, 88, 176], [44, 88, 176, 352]] num_modules = [1, 4, 3] elif version == "w48": init_block_channels = 256 init_num_subblocks = 4 channels = [[48, 96], [48, 96, 192], [48, 96, 192, 384]] num_modules = [1, 4, 3] elif version == "w64": init_block_channels = 256 init_num_subblocks = 4 channels = [[64, 128], [64, 128, 256], [64, 128, 256, 512]] num_modules = [1, 4, 3] else: raise ValueError("Unsupported HRNet version {}".format(version)) num_subblocks = [[max(2, init_num_subblocks)] * len(ci) for ci in channels] net = HRNet( channels=channels, init_block_channels=init_block_channels, init_num_subblocks=init_num_subblocks, num_modules=num_modules, num_subblocks=num_subblocks, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def hrnet_w18_small_v1(**kwargs): """ HRNet-W18 Small V1 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w18s1", model_name="hrnet_w18_small_v1", **kwargs) def hrnet_w18_small_v2(**kwargs): """ HRNet-W18 Small V2 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w18s2", model_name="hrnet_w18_small_v2", **kwargs) def hrnetv2_w18(**kwargs): """ HRNetV2-W18 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w18", model_name="hrnetv2_w18", **kwargs) def hrnetv2_w30(**kwargs): """ HRNetV2-W30 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w30", model_name="hrnetv2_w30", **kwargs) def hrnetv2_w32(**kwargs): """ HRNetV2-W32 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w32", model_name="hrnetv2_w32", **kwargs) def hrnetv2_w40(**kwargs): """ HRNetV2-W40 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w40", model_name="hrnetv2_w40", **kwargs) def hrnetv2_w44(**kwargs): """ HRNetV2-W44 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w44", model_name="hrnetv2_w44", **kwargs) def hrnetv2_w48(**kwargs): """ HRNetV2-W48 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w48", model_name="hrnetv2_w48", **kwargs) def hrnetv2_w64(**kwargs): """ HRNetV2-W64 model from 'Deep High-Resolution Representation Learning for Visual Recognition,' https://arxiv.org/abs/1908.07919. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hrnet(version="w64", model_name="hrnetv2_w64", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ hrnet_w18_small_v1, hrnet_w18_small_v2, hrnetv2_w18, hrnetv2_w30, hrnetv2_w32, hrnetv2_w40, hrnetv2_w44, hrnetv2_w48, hrnetv2_w64, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != hrnet_w18_small_v1 or weight_count == 13187464) assert (model != hrnet_w18_small_v2 or weight_count == 15597464) assert (model != hrnetv2_w18 or weight_count == 21299004) assert (model != hrnetv2_w30 or weight_count == 37712220) assert (model != hrnetv2_w32 or weight_count == 41232680) assert (model != hrnetv2_w40 or weight_count == 57557160) assert (model != hrnetv2_w44 or weight_count == 67064984) assert (model != hrnetv2_w48 or weight_count == 77469864) assert (model != hrnetv2_w64 or weight_count == 128059944) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
22,226
32.83105
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/fcn8sd.py
""" FCN-8s(d) for image segmentation, implemented in PyTorch. Original paper: 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. """ __all__ = ['FCN8sd', 'fcn8sd_resnetd50b_voc', 'fcn8sd_resnetd101b_voc', 'fcn8sd_resnetd50b_coco', 'fcn8sd_resnetd101b_coco', 'fcn8sd_resnetd50b_ade20k', 'fcn8sd_resnetd101b_ade20k', 'fcn8sd_resnetd50b_cityscapes', 'fcn8sd_resnetd101b_cityscapes'] import os import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from .common import conv1x1, conv3x3_block from .resnetd import resnetd50b, resnetd101b class FCNFinalBlock(nn.Module): """ FCN-8s(d) 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): super(FCNFinalBlock, self).__init__() assert (in_channels % bottleneck_factor == 0) mid_channels = in_channels // bottleneck_factor self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels) self.dropout = nn.Dropout(p=0.1, inplace=False) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) def forward(self, x, out_size): x = self.conv1(x) x = self.dropout(x) x = self.conv2(x) x = F.interpolate(x, size=out_size, mode="bilinear", align_corners=True) return x class FCN8sd(nn.Module): """ FCN-8s(d) model from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. It is an experimental model mixed FCN-8s and PSPNet. 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. num_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), num_classes=21): super(FCN8sd, self).__init__() assert (in_channels > 0) self.in_size = in_size self.num_classes = num_classes self.aux = aux self.fixed_size = fixed_size self.backbone = backbone pool_out_channels = backbone_out_channels self.final_block = FCNFinalBlock( in_channels=pool_out_channels, out_channels=num_classes) if self.aux: aux_out_channels = backbone_out_channels // 2 self.aux_block = FCNFinalBlock( in_channels=aux_out_channels, out_channels=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): in_size = self.in_size if self.fixed_size else x.shape[2:] x, y = self.backbone(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_fcn8sd(backbone, num_classes, aux=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create FCN-8s(d) model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = FCN8sd( backbone=backbone, num_classes=num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def fcn8sd_resnetd50b_voc(pretrained_backbone=False, num_classes=21, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-50b for Pascal VOC from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd50b_voc", **kwargs) def fcn8sd_resnetd101b_voc(pretrained_backbone=False, num_classes=21, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-101b for Pascal VOC from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd101b_voc", **kwargs) def fcn8sd_resnetd50b_coco(pretrained_backbone=False, num_classes=21, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-50b for COCO from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd50b_coco", **kwargs) def fcn8sd_resnetd101b_coco(pretrained_backbone=False, num_classes=21, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-101b for COCO from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd101b_coco", **kwargs) def fcn8sd_resnetd50b_ade20k(pretrained_backbone=False, num_classes=150, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-50b for ADE20K from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd50b_ade20k", **kwargs) def fcn8sd_resnetd101b_ade20k(pretrained_backbone=False, num_classes=150, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-101b for ADE20K from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd101b_ade20k", **kwargs) def fcn8sd_resnetd50b_cityscapes(pretrained_backbone=False, num_classes=19, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-50b for Cityscapes from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd50b_cityscapes", **kwargs) def fcn8sd_resnetd101b_cityscapes(pretrained_backbone=False, num_classes=19, aux=True, **kwargs): """ FCN-8s(d) model on the base of ResNet(D)-101b for Cityscapes from 'Fully Convolutional Networks for Semantic Segmentation,' https://arxiv.org/abs/1411.4038. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_fcn8sd(backbone=backbone, num_classes=num_classes, aux=aux, model_name="fcn8sd_resnetd101b_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch in_size = (480, 480) aux = True pretrained = False models = [ (fcn8sd_resnetd50b_voc, 21), (fcn8sd_resnetd101b_voc, 21), (fcn8sd_resnetd50b_coco, 21), (fcn8sd_resnetd101b_coco, 21), (fcn8sd_resnetd50b_ade20k, 150), (fcn8sd_resnetd101b_ade20k, 150), (fcn8sd_resnetd50b_cityscapes, 19), (fcn8sd_resnetd101b_cityscapes, 19), ] for model, num_classes in models: net = model(pretrained=pretrained, in_size=in_size, aux=aux) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) if aux: assert (model != fcn8sd_resnetd50b_voc or weight_count == 35445994) assert (model != fcn8sd_resnetd101b_voc or weight_count == 54438122) assert (model != fcn8sd_resnetd50b_coco or weight_count == 35445994) assert (model != fcn8sd_resnetd101b_coco or weight_count == 54438122) assert (model != fcn8sd_resnetd50b_ade20k or weight_count == 35545324) assert (model != fcn8sd_resnetd101b_ade20k or weight_count == 54537452) assert (model != fcn8sd_resnetd50b_cityscapes or weight_count == 35444454) assert (model != fcn8sd_resnetd101b_cityscapes or weight_count == 54436582) else: assert (model != fcn8sd_resnetd50b_voc or weight_count == 33080789) assert (model != fcn8sd_resnetd101b_voc or weight_count == 52072917) assert (model != fcn8sd_resnetd50b_coco or weight_count == 33080789) assert (model != fcn8sd_resnetd101b_coco or weight_count == 52072917) assert (model != fcn8sd_resnetd50b_ade20k or weight_count == 33146966) assert (model != fcn8sd_resnetd101b_ade20k or weight_count == 52139094) assert (model != fcn8sd_resnetd50b_cityscapes or weight_count == 33079763) assert (model != fcn8sd_resnetd101b_cityscapes or weight_count == 52071891) x = torch.randn(1, 3, in_size[0], in_size[1]) ys = net(x) y = ys[0] if aux else ys y.sum().backward() assert ((y.size(0) == x.size(0)) and (y.size(1) == num_classes) and (y.size(2) == x.size(2)) and (y.size(3) == x.size(3))) if __name__ == "__main__": _test()
16,126
37.125296
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/selecsls.py
""" SelecSLS for ImageNet-1K, implemented in PyTorch. Original paper: 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. """ __all__ = ['SelecSLS', 'selecsls42', 'selecsls42b', 'selecsls60', 'selecsls60b', 'selecsls84'] import os import torch import torch.nn as nn from .common import conv1x1_block, conv3x3_block, DualPathSequential class SelecSLSBlock(nn.Module): """ SelecSLS block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(SelecSLSBlock, self).__init__() mid_channels = 2 * out_channels self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class SelecSLSUnit(nn.Module): """ SelecSLS unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. skip_channels : int Number of skipped channels. mid_channels : int Number of middle channels. stride : int or tuple/list of 2 int Strides of the branch convolution layers. """ def __init__(self, in_channels, out_channels, skip_channels, mid_channels, stride): super(SelecSLSUnit, self).__init__() self.resize = (stride == 2) mid2_channels = mid_channels // 2 last_channels = 2 * mid_channels + (skip_channels if stride == 1 else 0) self.branch1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=stride) self.branch2 = SelecSLSBlock( in_channels=mid_channels, out_channels=mid2_channels) self.branch3 = SelecSLSBlock( in_channels=mid2_channels, out_channels=mid2_channels) self.last_conv = conv1x1_block( in_channels=last_channels, out_channels=out_channels) def forward(self, x, x0): x1 = self.branch1(x) x2 = self.branch2(x1) x3 = self.branch3(x2) if self.resize: y = torch.cat((x1, x2, x3), dim=1) y = self.last_conv(y) return y, y else: y = torch.cat((x1, x2, x3, x0), dim=1) y = self.last_conv(y) return y, x0 class SelecSLS(nn.Module): """ SelecSLS model from 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. Parameters: ---------- channels : list of list of int Number of output channels for each unit. skip_channels : list of list of int Number of skipped channels for each unit. mid_channels : list of list of int Number of middle channels for each unit. kernels3 : list of list of int/bool Using 3x3 (instead of 1x1) kernel for each head unit. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, skip_channels, mid_channels, kernels3, in_channels=3, in_size=(224, 224), num_classes=1000): super(SelecSLS, self).__init__() self.in_size = in_size self.num_classes = num_classes init_block_channels = 32 self.features = DualPathSequential( return_two=False, first_ordinals=1, last_ordinals=(1 + len(kernels3))) self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): k = i - len(skip_channels) stage = DualPathSequential() if k < 0 else nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if j == 0 else 1 if k < 0: unit = SelecSLSUnit( in_channels=in_channels, out_channels=out_channels, skip_channels=skip_channels[i][j], mid_channels=mid_channels[i][j], stride=stride) else: conv_block_class = conv3x3_block if kernels3[k][j] == 1 else conv1x1_block unit = conv_block_class( in_channels=in_channels, out_channels=out_channels, stride=stride) stage.add_module("unit{}".format(j + 1), unit) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=4, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight, mode="fan_out", nonlinearity="relu") if module.bias is not None: nn.init.constant_(module.bias, 0) elif isinstance(module, nn.BatchNorm2d): nn.init.constant_(module.weight, 1) nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_selecsls(version, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SelecSLS model with specific parameters. Parameters: ---------- version : str Version of SelecSLS. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if version in ("42", "42b"): channels = [[64, 128], [144, 288], [304, 480]] skip_channels = [[0, 64], [0, 144], [0, 304]] mid_channels = [[64, 64], [144, 144], [304, 304]] kernels3 = [[1, 1], [1, 0]] if version == "42": head_channels = [[960, 1024], [1024, 1280]] else: head_channels = [[960, 1024], [1280, 1024]] elif version in ("60", "60b"): channels = [[64, 128], [128, 128, 288], [288, 288, 288, 416]] skip_channels = [[0, 64], [0, 128, 128], [0, 288, 288, 288]] mid_channels = [[64, 64], [128, 128, 128], [288, 288, 288, 288]] kernels3 = [[1, 1], [1, 0]] if version == "60": head_channels = [[756, 1024], [1024, 1280]] else: head_channels = [[756, 1024], [1280, 1024]] elif version == "84": channels = [[64, 144], [144, 144, 144, 144, 304], [304, 304, 304, 304, 304, 512]] skip_channels = [[0, 64], [0, 144, 144, 144, 144], [0, 304, 304, 304, 304, 304]] mid_channels = [[64, 64], [144, 144, 144, 144, 144], [304, 304, 304, 304, 304, 304]] kernels3 = [[1, 1], [1, 1]] head_channels = [[960, 1024], [1024, 1280]] else: raise ValueError("Unsupported SelecSLS version {}".format(version)) channels += head_channels net = SelecSLS( channels=channels, skip_channels=skip_channels, mid_channels=mid_channels, kernels3=kernels3, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def selecsls42(**kwargs): """ SelecSLS-42 model from 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_selecsls(version="42", model_name="selecsls42", **kwargs) def selecsls42b(**kwargs): """ SelecSLS-42b model from 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_selecsls(version="42b", model_name="selecsls42b", **kwargs) def selecsls60(**kwargs): """ SelecSLS-60 model from 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_selecsls(version="60", model_name="selecsls60", **kwargs) def selecsls60b(**kwargs): """ SelecSLS-60b model from 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_selecsls(version="60b", model_name="selecsls60b", **kwargs) def selecsls84(**kwargs): """ SelecSLS-84 model from 'XNect: Real-time Multi-person 3D Human Pose Estimation with a Single RGB Camera,' https://arxiv.org/abs/1907.00837. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_selecsls(version="84", model_name="selecsls84", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ selecsls42, selecsls42b, selecsls60, selecsls60b, selecsls84, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != selecsls42 or weight_count == 30354952) assert (model != selecsls42b or weight_count == 32458248) assert (model != selecsls60 or weight_count == 30670768) assert (model != selecsls60b or weight_count == 32774064) assert (model != selecsls84 or weight_count == 50954600) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,347
31.580475
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/inceptionv4.py
""" InceptionV4 for ImageNet-1K, implemented in PyTorch. Original paper: 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. """ __all__ = ['InceptionV4', 'inceptionv4'] import os import torch import torch.nn as nn from .common import ConvBlock, conv3x3_block, Concurrent from .inceptionv3 import MaxPoolBranch, AvgPoolBranch, Conv1x1Branch, ConvSeqBranch class Conv3x3Branch(nn.Module): """ InceptionV4 specific convolutional 3x3 branch block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, bn_eps): super(Conv3x3Branch, self).__init__() self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, padding=0, bn_eps=bn_eps) def forward(self, x): x = self.conv(x) return x class ConvSeq3x3Branch(nn.Module): """ InceptionV4 specific convolutional sequence branch block with splitting by 3x3. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels_list : list of tuple of int List of numbers of output channels for middle layers. 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_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, mid_channels_list, kernel_size_list, strides_list, padding_list, bn_eps): super(ConvSeq3x3Branch, self).__init__() self.conv_list = nn.Sequential() for i, (mid_channels, kernel_size, strides, padding) in enumerate(zip( mid_channels_list, kernel_size_list, strides_list, padding_list)): self.conv_list.add_module("conv{}".format(i + 1), ConvBlock( in_channels=in_channels, out_channels=mid_channels, kernel_size=kernel_size, stride=strides, padding=padding, bn_eps=bn_eps)) in_channels = mid_channels self.conv1x3 = ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=(1, 3), stride=1, padding=(0, 1), bn_eps=bn_eps) self.conv3x1 = ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 1), stride=1, padding=(1, 0), bn_eps=bn_eps) def forward(self, x): x = self.conv_list(x) y1 = self.conv1x3(x) y2 = self.conv3x1(x) x = torch.cat((y1, y2), dim=1) return x class InceptionAUnit(nn.Module): """ InceptionV4 type Inception-A unit. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptionAUnit, self).__init__() in_channels = 384 self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=96, bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=(64, 96), kernel_size_list=(1, 3), strides_list=(1, 1), padding_list=(0, 1), bn_eps=bn_eps)) self.branches.add_module("branch3", 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_eps=bn_eps)) self.branches.add_module("branch4", AvgPoolBranch( in_channels=in_channels, out_channels=96, bn_eps=bn_eps, count_include_pad=False)) def forward(self, x): x = self.branches(x) return x class ReductionAUnit(nn.Module): """ InceptionV4 type Reduction-A unit. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(ReductionAUnit, self).__init__() in_channels = 384 self.branches = Concurrent() self.branches.add_module("branch1", ConvSeqBranch( in_channels=in_channels, out_channels_list=(384,), kernel_size_list=(3,), strides_list=(2,), padding_list=(0,), bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 224, 256), kernel_size_list=(1, 3, 3), strides_list=(1, 1, 2), padding_list=(0, 1, 0), bn_eps=bn_eps)) self.branches.add_module("branch3", MaxPoolBranch()) def forward(self, x): x = self.branches(x) return x class InceptionBUnit(nn.Module): """ InceptionV4 type Inception-B unit. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptionBUnit, self).__init__() in_channels = 1024 self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=384, bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 224, 256), kernel_size_list=(1, (1, 7), (7, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 3), (3, 0)), bn_eps=bn_eps)) self.branches.add_module("branch3", ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 192, 224, 224, 256), 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_eps=bn_eps)) self.branches.add_module("branch4", AvgPoolBranch( in_channels=in_channels, out_channels=128, bn_eps=bn_eps, count_include_pad=False)) def forward(self, x): x = self.branches(x) return x class ReductionBUnit(nn.Module): """ InceptionV4 type Reduction-B unit. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(ReductionBUnit, self).__init__() in_channels = 1024 self.branches = Concurrent() self.branches.add_module("branch1", ConvSeqBranch( in_channels=in_channels, out_channels_list=(192, 192), kernel_size_list=(1, 3), strides_list=(1, 2), padding_list=(0, 0), bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=(256, 256, 320, 320), kernel_size_list=(1, (1, 7), (7, 1), 3), strides_list=(1, 1, 1, 2), padding_list=(0, (0, 3), (3, 0), 0), bn_eps=bn_eps)) self.branches.add_module("branch3", MaxPoolBranch()) def forward(self, x): x = self.branches(x) return x class InceptionCUnit(nn.Module): """ InceptionV4 type Inception-C unit. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptionCUnit, self).__init__() in_channels = 1536 self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=256, bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeq3x3Branch( in_channels=in_channels, out_channels=256, mid_channels_list=(384,), kernel_size_list=(1,), strides_list=(1,), padding_list=(0,), bn_eps=bn_eps)) self.branches.add_module("branch3", ConvSeq3x3Branch( in_channels=in_channels, out_channels=256, mid_channels_list=(384, 448, 512), kernel_size_list=(1, (3, 1), (1, 3)), strides_list=(1, 1, 1), padding_list=(0, (1, 0), (0, 1)), bn_eps=bn_eps)) self.branches.add_module("branch4", AvgPoolBranch( in_channels=in_channels, out_channels=256, bn_eps=bn_eps, count_include_pad=False)) def forward(self, x): x = self.branches(x) return x class InceptBlock3a(nn.Module): """ InceptionV4 type Mixed-3a block. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptBlock3a, self).__init__() self.branches = Concurrent() self.branches.add_module("branch1", MaxPoolBranch()) self.branches.add_module("branch2", Conv3x3Branch( in_channels=64, out_channels=96, bn_eps=bn_eps)) def forward(self, x): x = self.branches(x) return x class InceptBlock4a(nn.Module): """ InceptionV4 type Mixed-4a block. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptBlock4a, self).__init__() self.branches = Concurrent() self.branches.add_module("branch1", ConvSeqBranch( in_channels=160, out_channels_list=(64, 96), kernel_size_list=(1, 3), strides_list=(1, 1), padding_list=(0, 0), bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=160, out_channels_list=(64, 64, 64, 96), kernel_size_list=(1, (1, 7), (7, 1), 3), strides_list=(1, 1, 1, 1), padding_list=(0, (0, 3), (3, 0), 0), bn_eps=bn_eps)) def forward(self, x): x = self.branches(x) return x class InceptBlock5a(nn.Module): """ InceptionV4 type Mixed-5a block. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptBlock5a, self).__init__() self.branches = Concurrent() self.branches.add_module("branch1", Conv3x3Branch( in_channels=192, out_channels=192, bn_eps=bn_eps)) self.branches.add_module("branch2", MaxPoolBranch()) def forward(self, x): x = self.branches(x) return x class InceptInitBlock(nn.Module): """ InceptionV4 specific initial block. Parameters: ---------- in_channels : int Number of input channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, bn_eps): super(InceptInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=32, stride=2, padding=0, bn_eps=bn_eps) self.conv2 = conv3x3_block( in_channels=32, out_channels=32, stride=1, padding=0, bn_eps=bn_eps) self.conv3 = conv3x3_block( in_channels=32, out_channels=64, stride=1, padding=1, bn_eps=bn_eps) self.block1 = InceptBlock3a(bn_eps=bn_eps) self.block2 = InceptBlock4a(bn_eps=bn_eps) self.block3 = InceptBlock5a(bn_eps=bn_eps) def forward(self, 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 InceptionV4(nn.Module): """ InceptionV4 model from 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. Parameters: ---------- dropout_rate : float, default 0.0 Fraction of the input units to drop. Must be a number between 0 and 1. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, dropout_rate=0.0, bn_eps=1e-5, in_channels=3, in_size=(299, 299), num_classes=1000): super(InceptionV4, self).__init__() self.in_size = in_size self.num_classes = num_classes layers = [4, 8, 4] normal_units = [InceptionAUnit, InceptionBUnit, InceptionCUnit] reduction_units = [ReductionAUnit, ReductionBUnit] self.features = nn.Sequential() self.features.add_module("init_block", InceptInitBlock( in_channels=in_channels, bn_eps=bn_eps)) for i, layers_per_stage in enumerate(layers): stage = nn.Sequential() for j in range(layers_per_stage): if (j == 0) and (i != 0): unit = reduction_units[i - 1] else: unit = normal_units[i] stage.add_module("unit{}".format(j + 1), unit(bn_eps=bn_eps)) self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Sequential() if dropout_rate > 0.0: self.output.add_module("dropout", nn.Dropout(p=dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=1536, out_features=num_classes)) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_inceptionv4(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create InceptionV4 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = InceptionV4(**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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def inceptionv4(**kwargs): """ InceptionV4 model from 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_inceptionv4(model_name="inceptionv4", bn_eps=1e-3, **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ inceptionv4, ] for model in models: net = model(pretrained=pretrained) # net.train() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != InceptionV4 or weight_count == 42679816) x = torch.randn(1, 3, 299, 299) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
17,876
28.944724
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/regnet.py
""" RegNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Designing Network Design Spaces,' https://arxiv.org/abs/2003.13678. """ __all__ = ['RegNet', 'regnetx002', 'regnetx004', 'regnetx006', 'regnetx008', 'regnetx016', 'regnetx032', 'regnetx040', 'regnetx064', 'regnetx080', 'regnetx120', 'regnetx160', 'regnetx320', 'regnety002', 'regnety004', 'regnety006', 'regnety008', 'regnety016', 'regnety032', 'regnety040', 'regnety064', 'regnety080', 'regnety120', 'regnety160', 'regnety320'] import os import numpy as np import torch.nn as nn from .common import conv1x1_block, conv3x3_block, SEBlock class RegNetBottleneck(nn.Module): """ RegNet bottleneck block for residual path in RegNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. groups : int Number of groups. use_se : bool Whether to use SE-module. bottleneck_factor : int, default 1 Bottleneck factor. """ def __init__(self, in_channels, out_channels, stride, groups, use_se, bottleneck_factor=1): super(RegNetBottleneck, self).__init__() self.use_se = use_se mid_channels = out_channels // bottleneck_factor mid_groups = mid_channels // groups self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, groups=mid_groups) if self.use_se: self.se = SEBlock( channels=mid_channels, mid_channels=(in_channels // 4)) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) if self.use_se: x = self.se(x) x = self.conv3(x) return x class RegNetUnit(nn.Module): """ RegNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. groups : int Number of groups. use_se : bool Whether to use SE-module. """ def __init__(self, in_channels, out_channels, stride, groups, use_se): super(RegNetUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = RegNetBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, groups=groups, use_se=use_se) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 RegNet(nn.Module): """ 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. groups : list of int Number of groups for each stage. use_se : bool Whether to use SE-module. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, groups, use_se, in_channels=3, in_size=(224, 224), num_classes=1000): super(RegNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2, padding=1)) in_channels = init_block_channels for i, (channels_per_stage, groups_per_stage) in enumerate(zip(channels, groups)): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) else 1 stage.add_module("unit{}".format(j + 1), RegNetUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, groups=groups_per_stage, use_se=use_se)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_regnet(channels_init, channels_slope, channels_mult, depth, groups, use_se=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. groups : int Number of groups. depth : int Depth value. use_se : bool, default False Whether to use SE-module. 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. root : str, default '~/.torch/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) # Adjusts the compatibility of channels/widths and groups: groups_per_stage = [min(groups, c) for c in channels_per_stage] channels_per_stage = [int(round(c / g) * g) for c, g in zip(channels_per_stage, groups_per_stage)] channels = [[ci] * li for (ci, li) in zip(channels_per_stage, layers)] init_block_channels = 32 net = RegNet( channels=channels, init_block_channels=init_block_channels, groups=groups_per_stage, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def regnetx002(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=24, channels_slope=36.44, channels_mult=2.49, depth=13, groups=8, model_name="regnetx002", **kwargs) def regnetx004(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=24, channels_slope=24.48, channels_mult=2.54, depth=22, groups=16, model_name="regnetx004", **kwargs) def regnetx006(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=48, channels_slope=36.97, channels_mult=2.24, depth=16, groups=24, model_name="regnetx006", **kwargs) def regnetx008(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=56, channels_slope=35.73, channels_mult=2.28, depth=16, groups=16, model_name="regnetx008", **kwargs) def regnetx016(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=80, channels_slope=34.01, channels_mult=2.25, depth=18, groups=24, model_name="regnetx016", **kwargs) def regnetx032(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=88, channels_slope=26.31, channels_mult=2.25, depth=25, groups=48, model_name="regnetx032", **kwargs) def regnetx040(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=96, channels_slope=38.65, channels_mult=2.43, depth=23, groups=40, model_name="regnetx040", **kwargs) def regnetx064(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=184, channels_slope=60.83, channels_mult=2.07, depth=17, groups=56, model_name="regnetx064", **kwargs) def regnetx080(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=80, channels_slope=49.56, channels_mult=2.88, depth=23, groups=120, model_name="regnetx080", **kwargs) def regnetx120(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=168, channels_slope=73.36, channels_mult=2.37, depth=19, groups=112, model_name="regnetx120", **kwargs) def regnetx160(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=216, channels_slope=55.59, channels_mult=2.1, depth=22, groups=128, model_name="regnetx160", **kwargs) def regnetx320(**kwargs): """ RegNetX-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=320, channels_slope=69.86, channels_mult=2.0, depth=23, groups=168, model_name="regnetx320", **kwargs) def regnety002(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=24, channels_slope=36.44, channels_mult=2.49, depth=13, groups=8, use_se=True, model_name="regnety002", **kwargs) def regnety004(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=48, channels_slope=27.89, channels_mult=2.09, depth=16, groups=8, use_se=True, model_name="regnety004", **kwargs) def regnety006(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=48, channels_slope=32.54, channels_mult=2.32, depth=15, groups=16, use_se=True, model_name="regnety006", **kwargs) def regnety008(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=56, channels_slope=38.84, channels_mult=2.4, depth=14, groups=16, use_se=True, model_name="regnety008", **kwargs) def regnety016(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=48, channels_slope=20.71, channels_mult=2.65, depth=27, groups=24, use_se=True, model_name="regnety016", **kwargs) def regnety032(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=80, channels_slope=42.63, channels_mult=2.66, depth=21, groups=24, use_se=True, model_name="regnety032", **kwargs) def regnety040(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=96, channels_slope=31.41, channels_mult=2.24, depth=22, groups=64, use_se=True, model_name="regnety040", **kwargs) def regnety064(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=112, channels_slope=33.22, channels_mult=2.27, depth=25, groups=72, use_se=True, model_name="regnety064", **kwargs) def regnety080(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=192, channels_slope=76.82, channels_mult=2.19, depth=17, groups=56, use_se=True, model_name="regnety080", **kwargs) def regnety120(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=168, channels_slope=73.36, channels_mult=2.37, depth=19, groups=112, use_se=True, model_name="regnety120", **kwargs) def regnety160(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=200, channels_slope=106.23, channels_mult=2.48, depth=18, groups=112, use_se=True, model_name="regnety160", **kwargs) def regnety320(**kwargs): """ RegNetY-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_regnet(channels_init=232, channels_slope=115.89, channels_mult=2.53, depth=20, groups=232, use_se=True, model_name="regnety320", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ regnetx002, regnetx004, regnetx006, regnetx008, regnetx016, regnetx032, regnetx040, regnetx064, regnetx080, regnetx120, regnetx160, regnetx320, regnety002, regnety004, regnety006, regnety008, regnety016, regnety032, regnety040, regnety064, regnety080, regnety120, regnety160, regnety320, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != regnetx002 or weight_count == 2684792) assert (model != regnetx004 or weight_count == 5157512) assert (model != regnetx006 or weight_count == 6196040) assert (model != regnetx008 or weight_count == 7259656) assert (model != regnetx016 or weight_count == 9190136) assert (model != regnetx032 or weight_count == 15296552) assert (model != regnetx040 or weight_count == 22118248) assert (model != regnetx064 or weight_count == 26209256) assert (model != regnetx080 or weight_count == 39572648) assert (model != regnetx120 or weight_count == 46106056) assert (model != regnetx160 or weight_count == 54278536) assert (model != regnetx320 or weight_count == 107811560) assert (model != regnety002 or weight_count == 3162996) assert (model != regnety004 or weight_count == 4344144) assert (model != regnety006 or weight_count == 6055160) assert (model != regnety008 or weight_count == 6263168) assert (model != regnety016 or weight_count == 11202430) assert (model != regnety032 or weight_count == 19436338) assert (model != regnety040 or weight_count == 20646656) assert (model != regnety064 or weight_count == 30583252) assert (model != regnety080 or weight_count == 39180068) assert (model != regnety120 or weight_count == 51822544) assert (model != regnety160 or weight_count == 83590140) assert (model != regnety320 or weight_count == 145046770) batch = 14 size = 224 x = torch.randn(batch, 3, size, size) y = net(x) y.sum().backward() assert (tuple(y.size()) == (batch, 1000)) if __name__ == "__main__": _test()
24,321
32.874652
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/icnet.py
""" ICNet for image segmentation, implemented in PyTorch. Original paper: 'ICNet for Real-Time Semantic Segmentation on High-Resolution Images,' https://arxiv.org/abs/1704.08545. """ __all__ = ['ICNet', 'icnet_resnetd50b_cityscapes'] import os import torch.nn as nn from .common import conv1x1, conv1x1_block, conv3x3_block, InterpolationBlock, MultiOutputSequential from .pspnet import PyramidPooling from .resnetd import resnetd50b class ICInitBlock(nn.Module): """ ICNet 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): super(ICInitBlock, self).__init__() mid_channels = out_channels // 2 self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=2) self.conv3 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, stride=2) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class PSPBlock(nn.Module): """ ICNet specific PSPNet reduced head block. 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. bottleneck_factor : int Bottleneck factor. """ def __init__(self, in_channels, upscale_out_size, bottleneck_factor): super(PSPBlock, self).__init__() assert (in_channels % bottleneck_factor == 0) mid_channels = in_channels // bottleneck_factor self.pool = PyramidPooling( in_channels=in_channels, upscale_out_size=upscale_out_size) self.conv = conv3x3_block( in_channels=4096, out_channels=mid_channels) self.dropout = nn.Dropout(p=0.1, inplace=False) def forward(self, x): x = self.pool(x) x = self.conv(x) x = self.dropout(x) return x class CFFBlock(nn.Module): """ Cascade Feature Fusion block. Parameters: ---------- in_channels_low : int Number of input channels (low input). in_channels_high : int Number of input channels (low high). out_channels : int Number of output channels. num_classes : int Number of classification classes. """ def __init__(self, in_channels_low, in_channels_high, out_channels, num_classes): super(CFFBlock, self).__init__() self.up = InterpolationBlock(scale_factor=2) self.conv_low = conv3x3_block( in_channels=in_channels_low, out_channels=out_channels, padding=2, dilation=2, activation=None) self.conv_hign = conv1x1_block( in_channels=in_channels_high, out_channels=out_channels, activation=None) self.activ = nn.ReLU(inplace=True) self.conv_cls = conv1x1( in_channels=out_channels, out_channels=num_classes) def forward(self, xl, xh): xl = self.up(xl) xl = self.conv_low(xl) xh = self.conv_hign(xh) x = xl + xh x = self.activ(x) x_cls = self.conv_cls(xl) return x, x_cls class ICHeadBlock(nn.Module): """ ICNet head block. Parameters: ---------- num_classes : int Number of classification classes. """ def __init__(self, num_classes): super(ICHeadBlock, self).__init__() self.cff_12 = CFFBlock( in_channels_low=128, in_channels_high=64, out_channels=128, num_classes=num_classes) self.cff_24 = CFFBlock( in_channels_low=256, in_channels_high=256, out_channels=128, num_classes=num_classes) self.up_x2 = InterpolationBlock(scale_factor=2) self.up_x8 = InterpolationBlock(scale_factor=4) self.conv_cls = conv1x1( in_channels=128, out_channels=num_classes) def forward(self, x1, x2, x4): outputs = [] x_cff_24, x_24_cls = self.cff_24(x4, x2) outputs.append(x_24_cls) x_cff_12, x_12_cls = self.cff_12(x_cff_24, x1) outputs.append(x_12_cls) up_x2 = self.up_x2(x_cff_12) up_x2 = self.conv_cls(up_x2) outputs.append(up_x2) up_x8 = self.up_x8(up_x2) outputs.append(up_x8) # 1 -> 1/4 -> 1/8 -> 1/16 outputs.reverse() return tuple(outputs) class ICNet(nn.Module): """ ICNet model from 'ICNet for Real-Time Semantic Segmentation on High-Resolution Images,' https://arxiv.org/abs/1704.08545. Parameters: ---------- backbones : tuple of nn.Sequential Feature extractors. backbones_out_channels : tuple of int Number of output channels form each feature extractor. num_classes : tuple of int Number of output channels for each branch. 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. num_classes : int, default 21 Number of segmentation classes. """ def __init__(self, backbones, backbones_out_channels, channels, aux=False, fixed_size=True, in_channels=3, in_size=(480, 480), num_classes=21): super(ICNet, self).__init__() assert (in_channels > 0) assert ((in_size[0] % 8 == 0) and (in_size[1] % 8 == 0)) self.in_size = in_size self.num_classes = num_classes self.aux = aux self.fixed_size = fixed_size psp_pool_out_size = (self.in_size[0] // 32, self.in_size[1] // 32) if fixed_size else None psp_head_out_channels = 512 self.branch1 = ICInitBlock( in_channels=in_channels, out_channels=channels[0]) self.branch2 = MultiOutputSequential() self.branch2.add_module("down1", InterpolationBlock(scale_factor=2, up=False)) backbones[0].do_output = True self.branch2.add_module("backbones1", backbones[0]) self.branch2.add_module("down2", InterpolationBlock(scale_factor=2, up=False)) self.branch2.add_module("backbones2", backbones[1]) self.branch2.add_module("psp", PSPBlock( in_channels=backbones_out_channels[1], upscale_out_size=psp_pool_out_size, bottleneck_factor=4)) self.branch2.add_module("final_block", conv1x1_block( in_channels=psp_head_out_channels, out_channels=channels[2])) self.conv_y2 = conv1x1_block( in_channels=backbones_out_channels[0], out_channels=channels[1]) self.final_block = ICHeadBlock(num_classes=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): y1 = self.branch1(x) y3, y2 = self.branch2(x) y2 = self.conv_y2(y2) x = self.final_block(y1, y2, y3) if self.aux: return x else: return x[0] def get_icnet(backbones, backbones_out_channels, num_classes, aux=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create ICNet model with specific parameters. Parameters: ---------- backbones : tuple of nn.Sequential Feature extractors. backbones_out_channels : tuple of int Number of output channels form each feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = (64, 256, 256) backbones[0].multi_output = False backbones[1].multi_output = False net = ICNet( backbones=backbones, backbones_out_channels=backbones_out_channels, channels=channels, num_classes=num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def icnet_resnetd50b_cityscapes(pretrained_backbone=False, num_classes=19, aux=True, **kwargs): """ ICNet model on the base of ResNet(D)-50b for Cityscapes from 'ICNet for Real-Time Semantic Segmentation on High-Resolution Images,' https://arxiv.org/abs/1704.08545. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone1 = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=None).features for i in range(len(backbone1) - 3): del backbone1[-1] backbone2 = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=None).features del backbone2[-1] for i in range(3): del backbone2[0] backbones = (backbone1, backbone2) backbones_out_channels = (512, 2048) return get_icnet(backbones=backbones, backbones_out_channels=backbones_out_channels, num_classes=num_classes, aux=aux, model_name="icnet_resnetd50b_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch in_size = (480, 480) aux = False fixed_size = False pretrained = False models = [ (icnet_resnetd50b_cityscapes, 19), ] for model, num_classes in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size, aux=aux) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != icnet_resnetd50b_cityscapes or weight_count == 47489184) x = torch.randn(1, 3, in_size[0], in_size[1]) ys = net(x) y = ys[0] if aux else ys y.sum().backward() assert ((y.size(0) == x.size(0)) and (y.size(1) == num_classes) and (y.size(2) == x.size(2)) and (y.size(3) == x.size(3))) if __name__ == "__main__": _test()
12,295
29.894472
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/mobilenetb.py
""" MobileNet(B) with simplified depthwise separable convolution block for ImageNet-1K, implemented in Gluon. Original paper: 'MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications,' https://arxiv.org/abs/1704.04861. """ __all__ = ['mobilenetb_w1', 'mobilenetb_w3d4', 'mobilenetb_wd2', 'mobilenetb_wd4'] from .mobilenet import get_mobilenet def mobilenetb_w1(**kwargs): """ 1.0 MobileNet(B)-224 model with simplified depthwise separable convolution block 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=1.0, dws_simplified=True, model_name="mobilenetb_w1", **kwargs) def mobilenetb_w3d4(**kwargs): """ 0.75 MobileNet(B)-224 model with simplified depthwise separable convolution block 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=0.75, dws_simplified=True, model_name="mobilenetb_w3d4", **kwargs) def mobilenetb_wd2(**kwargs): """ 0.5 MobileNet(B)-224 model with simplified depthwise separable convolution block 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=0.5, dws_simplified=True, model_name="mobilenetb_wd2", **kwargs) def mobilenetb_wd4(**kwargs): """ 0.25 MobileNet(B)-224 model with simplified depthwise separable convolution block 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenet(width_scale=0.25, dws_simplified=True, model_name="mobilenetb_wd4", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ mobilenetb_w1, mobilenetb_w3d4, mobilenetb_wd2, mobilenetb_wd4, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mobilenetb_w1 or weight_count == 4222056) assert (model != mobilenetb_w3d4 or weight_count == 2578120) assert (model != mobilenetb_wd2 or weight_count == 1326632) assert (model != mobilenetb_wd4 or weight_count == 467592) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
3,794
32.289474
113
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/shakedropresnet_cifar.py
""" ShakeDrop-ResNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'ShakeDrop Regularization for Deep Residual Learning,' https://arxiv.org/abs/1802.02375. """ __all__ = ['CIFARShakeDropResNet', 'shakedropresnet20_cifar10', 'shakedropresnet20_cifar100', 'shakedropresnet20_svhn'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block from .resnet import ResBlock, ResBottleneck class ShakeDrop(torch.autograd.Function): """ ShakeDrop function. """ @staticmethod def forward(ctx, x, b, alpha): y = (b + alpha - b * alpha) * x ctx.save_for_backward(b) return y @staticmethod def backward(ctx, dy): beta = torch.rand(dy.size(0), dtype=dy.dtype, device=dy.device).view(-1, 1, 1, 1) b, = ctx.saved_tensors return (b + beta - b * beta) * dy, None, None class ShakeDropResUnit(nn.Module): """ ShakeDrop-ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. life_prob : float Residual branch life probability. """ def __init__(self, in_channels, out_channels, stride, bottleneck, life_prob): super(ShakeDropResUnit, self).__init__() self.life_prob = life_prob self.resize_identity = (in_channels != out_channels) or (stride != 1) body_class = ResBottleneck if bottleneck else ResBlock self.body = body_class( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) self.shake_drop = ShakeDrop.apply def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) if self.training: b = torch.bernoulli(torch.full((1,), self.life_prob, dtype=x.dtype, device=x.device)) alpha = torch.empty(x.size(0), dtype=x.dtype, device=x.device).view(-1, 1, 1, 1).uniform_(-1.0, 1.0) x = self.shake_drop(x, b, alpha) else: x = self.life_prob * x x = x + identity x = self.activ(x) return x class CIFARShakeDropResNet(nn.Module): """ ShakeDrop-ResNet model for CIFAR from 'ShakeDrop Regularization for Deep Residual Learning,' https://arxiv.org/abs/1802.02375. 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. life_probs : list of float Residual branch life probability for each unit. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, life_probs, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARShakeDropResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels k = 0 for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ShakeDropResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, life_prob=life_probs[k])) in_channels = out_channels k += 1 self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_shakedropresnet_cifar(classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create ShakeDrop-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. root : str, default '~/.torch/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 init_block_channels = 16 channels_per_layers = [16, 32, 64] 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] total_layers = sum(layers) final_death_prob = 0.5 life_probs = [1.0 - float(i + 1) / float(total_layers) * final_death_prob for i in range(total_layers)] net = CIFARShakeDropResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, life_probs=life_probs, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def shakedropresnet20_cifar10(classes=10, **kwargs): """ ShakeDrop-ResNet-20 model for CIFAR-10 from 'ShakeDrop Regularization for Deep Residual Learning,' https://arxiv.org/abs/1802.02375. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakedropresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="shakedropresnet20_cifar10", **kwargs) def shakedropresnet20_cifar100(classes=100, **kwargs): """ ShakeDrop-ResNet-20 model for CIFAR-100 from 'ShakeDrop Regularization for Deep Residual Learning,' https://arxiv.org/abs/1802.02375. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakedropresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="shakedropresnet20_cifar100", **kwargs) def shakedropresnet20_svhn(classes=10, **kwargs): """ ShakeDrop-ResNet-20 model for SVHN from 'ShakeDrop Regularization for Deep Residual Learning,' https://arxiv.org/abs/1802.02375. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakedropresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="shakedropresnet20_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (shakedropresnet20_cifar10, 10), (shakedropresnet20_cifar100, 100), (shakedropresnet20_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != shakedropresnet20_cifar10 or weight_count == 272474) assert (model != shakedropresnet20_cifar100 or weight_count == 278324) assert (model != shakedropresnet20_svhn or weight_count == 272474) x = torch.randn(14, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (14, num_classes)) if __name__ == "__main__": _test()
10,750
31.677812
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/inceptionresnetv1.py
""" InceptionResNetV1 for ImageNet-1K, implemented in PyTorch. Original paper: 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. """ __all__ = ['InceptionResNetV1', 'inceptionresnetv1', 'InceptionAUnit', 'InceptionBUnit', 'InceptionCUnit', 'ReductionAUnit', 'ReductionBUnit'] import os import torch.nn as nn from .common import conv1x1, conv1x1_block, conv3x3_block, Concurrent from .inceptionv3 import MaxPoolBranch, Conv1x1Branch, ConvSeqBranch class InceptionAUnit(nn.Module): """ InceptionResNetV1 type Inception-A unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int List for numbers of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels_list, bn_eps): super(InceptionAUnit, self).__init__() self.scale = 0.17 self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=out_channels_list[0], bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[1:3], kernel_size_list=(1, 3), strides_list=(1, 1), padding_list=(0, 1), bn_eps=bn_eps)) self.branches.add_module("branch3", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[3:6], kernel_size_list=(1, 3, 3), strides_list=(1, 1, 1), padding_list=(0, 1, 1), bn_eps=bn_eps)) conv_in_channels = out_channels_list[0] + out_channels_list[2] + out_channels_list[5] self.conv = conv1x1( in_channels=conv_in_channels, out_channels=in_channels, bias=True) self.activ = nn.ReLU(inplace=True) def forward(self, x): identity = x x = self.branches(x) x = self.conv(x) x = self.scale * x + identity x = self.activ(x) return x class InceptionBUnit(nn.Module): """ InceptionResNetV1 type Inception-B unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int List for numbers of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels_list, bn_eps): super(InceptionBUnit, self).__init__() self.scale = 0.10 self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=out_channels_list[0], bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[1:4], kernel_size_list=(1, (1, 7), (7, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 3), (3, 0)), bn_eps=bn_eps)) conv_in_channels = out_channels_list[0] + out_channels_list[3] self.conv = conv1x1( in_channels=conv_in_channels, out_channels=in_channels, bias=True) self.activ = nn.ReLU(inplace=True) def forward(self, x): identity = x x = self.branches(x) x = self.conv(x) x = self.scale * x + identity x = self.activ(x) return x class InceptionCUnit(nn.Module): """ InceptionResNetV1 type Inception-C unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int List for numbers of output channels. bn_eps : float Small float added to variance in Batch norm. scale : float, default 0.2 Scale value for residual branch. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels_list, bn_eps, scale=0.2, activate=True): super(InceptionCUnit, self).__init__() self.activate = activate self.scale = scale self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=out_channels_list[0], bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[1:4], kernel_size_list=(1, (1, 3), (3, 1)), strides_list=(1, 1, 1), padding_list=(0, (0, 1), (1, 0)), bn_eps=bn_eps)) conv_in_channels = out_channels_list[0] + out_channels_list[3] self.conv = conv1x1( in_channels=conv_in_channels, out_channels=in_channels, bias=True) if self.activate: self.activ = nn.ReLU(inplace=True) def forward(self, x): identity = x x = self.branches(x) x = self.conv(x) x = self.scale * x + identity if self.activate: x = self.activ(x) return x class ReductionAUnit(nn.Module): """ InceptionResNetV1 type Reduction-A unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int List for numbers of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels_list, bn_eps): super(ReductionAUnit, self).__init__() self.branches = Concurrent() self.branches.add_module("branch1", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[0:1], kernel_size_list=(3,), strides_list=(2,), padding_list=(0,), bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[1:4], kernel_size_list=(1, 3, 3), strides_list=(1, 1, 2), padding_list=(0, 1, 0), bn_eps=bn_eps)) self.branches.add_module("branch3", MaxPoolBranch()) def forward(self, x): x = self.branches(x) return x class ReductionBUnit(nn.Module): """ InceptionResNetV1 type Reduction-B unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : list of int List for numbers of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels_list, bn_eps): super(ReductionBUnit, self).__init__() self.branches = Concurrent() self.branches.add_module("branch1", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[0:2], kernel_size_list=(1, 3), strides_list=(1, 2), padding_list=(0, 0), bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[2:4], kernel_size_list=(1, 3), strides_list=(1, 2), padding_list=(0, 0), bn_eps=bn_eps)) self.branches.add_module("branch3", ConvSeqBranch( in_channels=in_channels, out_channels_list=out_channels_list[4:7], kernel_size_list=(1, 3, 3), strides_list=(1, 1, 2), padding_list=(0, 1, 0), bn_eps=bn_eps)) self.branches.add_module("branch4", MaxPoolBranch()) def forward(self, x): x = self.branches(x) return x class InceptInitBlock(nn.Module): """ InceptionResNetV1 specific initial block. Parameters: ---------- in_channels : int Number of input channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, bn_eps): super(InceptInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=32, stride=2, padding=0, bn_eps=bn_eps) self.conv2 = conv3x3_block( in_channels=32, out_channels=32, stride=1, padding=0, bn_eps=bn_eps) self.conv3 = conv3x3_block( in_channels=32, out_channels=64, stride=1, padding=1, bn_eps=bn_eps) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=0) self.conv4 = conv1x1_block( in_channels=64, out_channels=80, stride=1, padding=0, bn_eps=bn_eps) self.conv5 = conv3x3_block( in_channels=80, out_channels=192, stride=1, padding=0, bn_eps=bn_eps) self.conv6 = conv3x3_block( in_channels=192, out_channels=256, stride=2, padding=0, bn_eps=bn_eps) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.pool(x) x = self.conv4(x) x = self.conv5(x) x = self.conv6(x) return x class InceptHead(nn.Module): """ InceptionResNetV1 specific classification block. Parameters: ---------- in_channels : int Number of input channels. bn_eps : float Small float added to variance in Batch norm. dropout_rate : float Fraction of the input units to drop. Must be a number between 0 and 1. num_classes : int Number of classification classes. """ def __init__(self, in_channels, bn_eps, dropout_rate, num_classes): super(InceptHead, self).__init__() self.use_dropout = (dropout_rate != 0.0) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) self.fc1 = nn.Linear( in_features=in_channels, out_features=512, bias=False) self.bn = nn.BatchNorm1d( num_features=512, eps=bn_eps) self.fc2 = nn.Linear( in_features=512, out_features=num_classes) def forward(self, x): if self.use_dropout: x = self.dropout(x) x = self.fc1(x) x = self.bn(x) x = self.fc2(x) return x class InceptionResNetV1(nn.Module): """ InceptionResNetV1 model from 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. Parameters: ---------- dropout_rate : float, default 0.0 Fraction of the input units to drop. Must be a number between 0 and 1. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, dropout_prob=0.6, bn_eps=1e-5, in_channels=3, in_size=(299, 299), num_classes=1000): super(InceptionResNetV1, self).__init__() self.in_size = in_size self.num_classes = num_classes layers = [5, 11, 7] in_channels_list = [256, 896, 1792] normal_out_channels_list = [[32, 32, 32, 32, 32, 32], [128, 128, 128, 128], [192, 192, 192, 192]] reduction_out_channels_list = [[384, 192, 192, 256], [256, 384, 256, 256, 256, 256, 256]] normal_units = [InceptionAUnit, InceptionBUnit, InceptionCUnit] reduction_units = [ReductionAUnit, ReductionBUnit] self.features = nn.Sequential() self.features.add_module("init_block", InceptInitBlock( in_channels=in_channels, bn_eps=bn_eps)) in_channels = in_channels_list[0] for i, layers_per_stage in enumerate(layers): stage = nn.Sequential() for j in range(layers_per_stage): if (j == 0) and (i != 0): unit = reduction_units[i - 1] out_channels_list_per_stage = reduction_out_channels_list[i - 1] else: unit = normal_units[i] out_channels_list_per_stage = normal_out_channels_list[i] if (i == len(layers) - 1) and (j == layers_per_stage - 1): unit_kwargs = {"scale": 1.0, "activate": False} else: unit_kwargs = {} stage.add_module("unit{}".format(j + 1), unit( in_channels=in_channels, out_channels_list=out_channels_list_per_stage, bn_eps=bn_eps, **unit_kwargs)) if (j == 0) and (i != 0): in_channels = in_channels_list[i] self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = InceptHead( in_channels=in_channels, bn_eps=bn_eps, dropout_rate=dropout_prob, num_classes=num_classes) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_inceptionresnetv1(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create InceptionResNetV1 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = InceptionResNetV1(**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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def inceptionresnetv1(**kwargs): """ InceptionResNetV1 model from 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_inceptionresnetv1(model_name="inceptionresnetv1", bn_eps=1e-3, **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ inceptionresnetv1, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != inceptionresnetv1 or weight_count == 23995624) x = torch.randn(1, 3, 299, 299) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
16,987
30.285451
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/scnet.py
""" SCNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Improving Convolutional Networks with Self-Calibrated Convolutions,' http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf. """ __all__ = ['SCNet', 'scnet50', 'scnet101', 'scneta50', 'scneta101'] import os import torch import torch.nn as nn from .common import conv1x1_block, conv3x3_block, InterpolationBlock from .resnet import ResInitBlock from .senet import SEInitBlock from .resnesta import ResNeStADownBlock class ScDownBlock(nn.Module): """ SCNet specific convolutional downscale block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. pool_size: int or list/tuple of 2 ints, default 2 Size of the average pooling windows. """ def __init__(self, in_channels, out_channels, pool_size=2): super(ScDownBlock, self).__init__() self.pool = nn.AvgPool2d( kernel_size=pool_size, stride=pool_size) self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.pool(x) x = self.conv(x) return x class ScConv(nn.Module): """ Self-calibrated convolutional block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. scale_factor : int Scale factor. """ def __init__(self, in_channels, out_channels, stride, scale_factor): super(ScConv, self).__init__() self.down = ScDownBlock( in_channels=in_channels, out_channels=out_channels, pool_size=scale_factor) self.up = InterpolationBlock( scale_factor=scale_factor, mode="nearest", align_corners=None) self.sigmoid = nn.Sigmoid() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, activation=None) self.conv2 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride) def forward(self, x): w = self.sigmoid(x + self.up(self.down(x), size=x.shape[2:])) x = self.conv1(x) * w x = self.conv2(x) return x class ScBottleneck(nn.Module): """ SCNet specific bottleneck block for residual path in SCNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck_factor : int, default 4 Bottleneck factor. scale_factor : int, default 4 Scale factor. avg_downsample : bool, default False Whether to use average downsampling. """ def __init__(self, in_channels, out_channels, stride, bottleneck_factor=4, scale_factor=4, avg_downsample=False): super(ScBottleneck, self).__init__() self.avg_resize = (stride > 1) and avg_downsample mid_channels = out_channels // bottleneck_factor // 2 self.conv1a = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2a = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=(1 if self.avg_resize else stride)) self.conv1b = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2b = ScConv( in_channels=mid_channels, out_channels=mid_channels, stride=(1 if self.avg_resize else stride), scale_factor=scale_factor) if self.avg_resize: self.pool = nn.AvgPool2d( kernel_size=3, stride=stride, padding=1) self.conv3 = conv1x1_block( in_channels=(2 * mid_channels), out_channels=out_channels, activation=None) def forward(self, x): y = self.conv1a(x) y = self.conv2a(y) z = self.conv1b(x) z = self.conv2b(z) if self.avg_resize: y = self.pool(y) z = self.pool(z) x = torch.cat((y, z), dim=1) x = self.conv3(x) return x class ScUnit(nn.Module): """ SCNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. avg_downsample : bool, default False Whether to use average downsampling. """ def __init__(self, in_channels, out_channels, stride, avg_downsample=False): super(ScUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = ScBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, avg_downsample=avg_downsample) if self.resize_identity: if avg_downsample: self.identity_block = ResNeStADownBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) else: self.identity_block = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 SCNet(nn.Module): """ SCNet model from 'Improving Convolutional Networks with Self-Calibrated Convolutions,' http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf. 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. se_init_block : bool, default False SENet-like initial block. avg_downsample : bool, default False Whether to use average downsampling. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, se_init_block=False, avg_downsample=False, in_channels=3, in_size=(224, 224), num_classes=1000): super(SCNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() init_block_class = SEInitBlock if se_init_block else ResInitBlock self.features.add_module("init_block", init_block_class( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ScUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, avg_downsample=avg_downsample)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_scnet(blocks, width_scale=1.0, se_init_block=False, avg_downsample=False, init_block_channels_scale=1, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SCNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. width_scale : float, default 1.0 Scale factor for width of layers. se_init_block : bool, default False SENet-like initial block. avg_downsample : bool, default False Whether to use average downsampling. init_block_channels_scale : int, default 1 Scale factor for number of output channels in the initial unit. 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. root : str, default '~/.torch/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] elif blocks == 152: layers = [3, 8, 36, 3] elif blocks == 200: layers = [3, 24, 36, 3] else: raise ValueError("Unsupported SCNet with number of blocks: {}".format(blocks)) assert (sum(layers) * 3 + 2 == blocks) init_block_channels = 64 channels_per_layers = [64, 128, 256, 512] init_block_channels *= init_block_channels_scale 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 = SCNet( channels=channels, init_block_channels=init_block_channels, se_init_block=se_init_block, avg_downsample=avg_downsample, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def scnet50(**kwargs): """ SCNet-50 model from 'Improving Convolutional Networks with Self-Calibrated Convolutions,' http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_scnet(blocks=50, model_name="scnet50", **kwargs) def scnet101(**kwargs): """ SCNet-101 model from 'Improving Convolutional Networks with Self-Calibrated Convolutions,' http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_scnet(blocks=101, model_name="scnet101", **kwargs) def scneta50(**kwargs): """ SCNet(A)-50 with average downsampling model from 'Improving Convolutional Networks with Self-Calibrated Convolutions,' http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_scnet(blocks=50, se_init_block=True, avg_downsample=True, model_name="scneta50", **kwargs) def scneta101(**kwargs): """ SCNet(A)-101 with average downsampling model from 'Improving Convolutional Networks with Self-Calibrated Convolutions,' http://mftp.mmcheng.net/Papers/20cvprSCNet.pdf. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_scnet(blocks=101, se_init_block=True, avg_downsample=True, init_block_channels_scale=2, model_name="scneta101", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ scnet50, scnet101, scneta50, scneta101, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != scnet50 or weight_count == 25564584) assert (model != scnet101 or weight_count == 44565416) assert (model != scneta50 or weight_count == 25583816) assert (model != scneta101 or weight_count == 44689192) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
14,943
29.876033
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/igcv3.py
""" IGCV3 for ImageNet-1K, implemented in PyTorch. Original paper: 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,' https://arxiv.org/abs/1806.00178. """ __all__ = ['IGCV3', 'igcv3_w1', 'igcv3_w3d4', 'igcv3_wd2', 'igcv3_wd4'] import os import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block, dwconv3x3_block, ChannelShuffle class InvResUnit(nn.Module): """ So-called 'Inverted Residual Unit' layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the second convolution layer. expansion : bool Whether do expansion of channels. """ def __init__(self, in_channels, out_channels, stride, expansion): super(InvResUnit, self).__init__() self.residual = (in_channels == out_channels) and (stride == 1) mid_channels = in_channels * 6 if expansion else in_channels groups = 2 self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, groups=groups, activation=None) self.c_shuffle = ChannelShuffle( channels=mid_channels, groups=groups) self.conv2 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, activation="relu6") self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, groups=groups, activation=None) def forward(self, x): if self.residual: identity = x x = self.conv1(x) x = self.c_shuffle(x) x = self.conv2(x) x = self.conv3(x) if self.residual: x = x + identity return x class IGCV3(nn.Module): """ IGCV3 model from 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,' https://arxiv.org/abs/1806.00178. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, in_channels=3, in_size=(224, 224), num_classes=1000): super(IGCV3, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2, activation="relu6")) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 expansion = (i != 0) or (j != 0) stage.add_module("unit{}".format(j + 1), InvResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, expansion=expansion)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, activation="relu6")) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_igcv3(width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create IGCV3-D 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 32 final_block_channels = 1280 layers = [1, 4, 6, 8, 6, 6, 1] downsample = [0, 1, 1, 1, 0, 1, 0] channels_per_layers = [16, 24, 32, 64, 96, 160, 320] from functools import reduce channels = reduce( lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(channels_per_layers, layers, downsample), [[]]) if width_scale != 1.0: def make_even(x): return x if (x % 2 == 0) else x + 1 channels = [[make_even(int(cij * width_scale)) for cij in ci] for ci in channels] init_block_channels = make_even(int(init_block_channels * width_scale)) if width_scale > 1.0: final_block_channels = make_even(int(final_block_channels * width_scale)) net = IGCV3( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def igcv3_w1(**kwargs): """ IGCV3-D 1.0x model from 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,' https://arxiv.org/abs/1806.00178. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_igcv3(width_scale=1.0, model_name="igcv3_w1", **kwargs) def igcv3_w3d4(**kwargs): """ IGCV3-D 0.75x model from 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,' https://arxiv.org/abs/1806.00178. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_igcv3(width_scale=0.75, model_name="igcv3_w3d4", **kwargs) def igcv3_wd2(**kwargs): """ IGCV3-D 0.5x model from 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,' https://arxiv.org/abs/1806.00178. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_igcv3(width_scale=0.5, model_name="igcv3_wd2", **kwargs) def igcv3_wd4(**kwargs): """ IGCV3-D 0.25x model from 'IGCV3: Interleaved Low-Rank Group Convolutions for Efficient Deep Neural Networks,' https://arxiv.org/abs/1806.00178. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_igcv3(width_scale=0.25, model_name="igcv3_wd4", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ igcv3_w1, igcv3_w3d4, igcv3_wd2, igcv3_wd4, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != igcv3_w1 or weight_count == 3491688) assert (model != igcv3_w3d4 or weight_count == 2638084) assert (model != igcv3_wd2 or weight_count == 1985528) assert (model != igcv3_wd4 or weight_count == 1534020) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
9,829
30.709677
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/seresnet_cifar.py
""" SE-ResNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['CIFARSEResNet', 'seresnet20_cifar10', 'seresnet20_cifar100', 'seresnet20_svhn', 'seresnet56_cifar10', 'seresnet56_cifar100', 'seresnet56_svhn', 'seresnet110_cifar10', 'seresnet110_cifar100', 'seresnet110_svhn', 'seresnet164bn_cifar10', 'seresnet164bn_cifar100', 'seresnet164bn_svhn', 'seresnet272bn_cifar10', 'seresnet272bn_cifar100', 'seresnet272bn_svhn', 'seresnet542bn_cifar10', 'seresnet542bn_cifar100', 'seresnet542bn_svhn', 'seresnet1001_cifar10', 'seresnet1001_cifar100', 'seresnet1001_svhn', 'seresnet1202_cifar10', 'seresnet1202_cifar100', 'seresnet1202_svhn'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3_block from .seresnet import SEResUnit class CIFARSEResNet(nn.Module): """ SE-ResNet model for CIFAR 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. 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. num_classes : int, default 10 Number of classification num_classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARSEResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), SEResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=False)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_seresnet_cifar(num_classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SE-ResNet model for CIFAR with specific parameters. Parameters: ---------- num_classes : int Number of classification num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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 = CIFARSEResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def seresnet20_cifar10(num_classes=10, **kwargs): """ SE-ResNet-20 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="seresnet20_cifar10", **kwargs) def seresnet20_cifar100(num_classes=100, **kwargs): """ SE-ResNet-20 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="seresnet20_cifar100", **kwargs) def seresnet20_svhn(num_classes=10, **kwargs): """ SE-ResNet-20 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="seresnet20_svhn", **kwargs) def seresnet56_cifar10(num_classes=10, **kwargs): """ SE-ResNet-56 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="seresnet56_cifar10", **kwargs) def seresnet56_cifar100(num_classes=100, **kwargs): """ SE-ResNet-56 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="seresnet56_cifar100", **kwargs) def seresnet56_svhn(num_classes=10, **kwargs): """ SE-ResNet-56 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="seresnet56_svhn", **kwargs) def seresnet110_cifar10(num_classes=10, **kwargs): """ SE-ResNet-110 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="seresnet110_cifar10", **kwargs) def seresnet110_cifar100(num_classes=100, **kwargs): """ SE-ResNet-110 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="seresnet110_cifar100", **kwargs) def seresnet110_svhn(num_classes=10, **kwargs): """ SE-ResNet-110 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="seresnet110_svhn", **kwargs) def seresnet164bn_cifar10(num_classes=10, **kwargs): """ SE-ResNet-164(BN) model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="seresnet164bn_cifar10", **kwargs) def seresnet164bn_cifar100(num_classes=100, **kwargs): """ SE-ResNet-164(BN) model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="seresnet164bn_cifar100", **kwargs) def seresnet164bn_svhn(num_classes=10, **kwargs): """ SE-ResNet-164(BN) model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="seresnet164bn_svhn", **kwargs) def seresnet272bn_cifar10(num_classes=10, **kwargs): """ SE-ResNet-272(BN) model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="seresnet272bn_cifar10", **kwargs) def seresnet272bn_cifar100(num_classes=100, **kwargs): """ SE-ResNet-272(BN) model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="seresnet272bn_cifar100", **kwargs) def seresnet272bn_svhn(num_classes=10, **kwargs): """ SE-ResNet-272(BN) model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="seresnet272bn_svhn", **kwargs) def seresnet542bn_cifar10(num_classes=10, **kwargs): """ SE-ResNet-542(BN) model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="seresnet542bn_cifar10", **kwargs) def seresnet542bn_cifar100(num_classes=100, **kwargs): """ SE-ResNet-542(BN) model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="seresnet542bn_cifar100", **kwargs) def seresnet542bn_svhn(num_classes=10, **kwargs): """ SE-ResNet-542(BN) model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="seresnet542bn_svhn", **kwargs) def seresnet1001_cifar10(num_classes=10, **kwargs): """ SE-ResNet-1001 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="seresnet1001_cifar10", **kwargs) def seresnet1001_cifar100(num_classes=100, **kwargs): """ SE-ResNet-1001 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="seresnet1001_cifar100", **kwargs) def seresnet1001_svhn(num_classes=10, **kwargs): """ SE-ResNet-1001 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="seresnet1001_svhn", **kwargs) def seresnet1202_cifar10(num_classes=10, **kwargs): """ SE-ResNet-1202 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="seresnet1202_cifar10", **kwargs) def seresnet1202_cifar100(num_classes=100, **kwargs): """ SE-ResNet-1202 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="seresnet1202_cifar100", **kwargs) def seresnet1202_svhn(num_classes=10, **kwargs): """ SE-ResNet-1202 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="seresnet1202_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (seresnet20_cifar10, 10), (seresnet20_cifar100, 100), (seresnet20_svhn, 10), (seresnet56_cifar10, 10), (seresnet56_cifar100, 100), (seresnet56_svhn, 10), (seresnet110_cifar10, 10), (seresnet110_cifar100, 100), (seresnet110_svhn, 10), (seresnet164bn_cifar10, 10), (seresnet164bn_cifar100, 100), (seresnet164bn_svhn, 10), (seresnet272bn_cifar10, 10), (seresnet272bn_cifar100, 100), (seresnet272bn_svhn, 10), (seresnet542bn_cifar10, 10), (seresnet542bn_cifar100, 100), (seresnet542bn_svhn, 10), (seresnet1001_cifar10, 10), (seresnet1001_cifar100, 100), (seresnet1001_svhn, 10), (seresnet1202_cifar10, 10), (seresnet1202_cifar100, 100), (seresnet1202_svhn, 10), ] for model, num_num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != seresnet20_cifar10 or weight_count == 274847) assert (model != seresnet20_cifar100 or weight_count == 280697) assert (model != seresnet20_svhn or weight_count == 274847) assert (model != seresnet56_cifar10 or weight_count == 862889) assert (model != seresnet56_cifar100 or weight_count == 868739) assert (model != seresnet56_svhn or weight_count == 862889) assert (model != seresnet110_cifar10 or weight_count == 1744952) assert (model != seresnet110_cifar100 or weight_count == 1750802) assert (model != seresnet110_svhn or weight_count == 1744952) assert (model != seresnet164bn_cifar10 or weight_count == 1906258) assert (model != seresnet164bn_cifar100 or weight_count == 1929388) assert (model != seresnet164bn_svhn or weight_count == 1906258) assert (model != seresnet272bn_cifar10 or weight_count == 3153826) assert (model != seresnet272bn_cifar100 or weight_count == 3176956) assert (model != seresnet272bn_svhn or weight_count == 3153826) assert (model != seresnet542bn_cifar10 or weight_count == 6272746) assert (model != seresnet542bn_cifar100 or weight_count == 6295876) assert (model != seresnet542bn_svhn or weight_count == 6272746) assert (model != seresnet1001_cifar10 or weight_count == 11574910) assert (model != seresnet1001_cifar100 or weight_count == 11598040) assert (model != seresnet1001_svhn or weight_count == 11574910) assert (model != seresnet1202_cifar10 or weight_count == 19582226) assert (model != seresnet1202_cifar100 or weight_count == 19588076) assert (model != seresnet1202_svhn or weight_count == 19582226) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_num_classes)) if __name__ == "__main__": _test()
24,036
36.324534
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resnetd.py
""" ResNet(D) with dilation for ImageNet-1K, implemented in PyTorch. Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. """ __all__ = ['ResNetD', 'resnetd50b', 'resnetd101b', 'resnetd152b'] import os import torch.nn as nn import torch.nn.init as init from .common import MultiOutputSequential from .resnet import ResUnit, ResInitBlock from .senet import SEInitBlock class ResNetD(nn.Module): """ ResNet(D) with dilation 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. ordinary_init : bool, default False Whether to use original initial block or SENet one. bends : tuple of int, default None Numbers of bends for multiple output. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, ordinary_init=False, bends=None, in_channels=3, in_size=(224, 224), num_classes=1000): super(ResNetD, self).__init__() self.in_size = in_size self.num_classes = num_classes self.multi_output = (bends is not None) self.features = MultiOutputSequential() if ordinary_init: self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) else: init_block_channels = 2 * init_block_channels self.features.add_module("init_block", SEInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if ((j == 0) and (i != 0) and (i < 2)) else 1 dilation = (2 ** max(0, i - 1 - int(j == 0))) stage.add_module("unit{}".format(j + 1), ResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, padding=dilation, dilation=dilation, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels if self.multi_output and ((i + 1) in bends): stage.do_output = True self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): outs = self.features(x) x = outs[0] x = x.view(x.size(0), -1) x = self.output(x) if self.multi_output: return [x] + outs[1:] else: return x def get_resnetd(blocks, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create ResNet(D) with dilation 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. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14: layers = [2, 2, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif 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 ResNet(D) 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)] 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 = ResNetD( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def resnetd50b(**kwargs): """ ResNet(D)-50 with dilation 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnetd(blocks=50, conv1_stride=False, model_name="resnetd50b", **kwargs) def resnetd101b(**kwargs): """ ResNet(D)-101 with dilation 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnetd(blocks=101, conv1_stride=False, model_name="resnetd101b", **kwargs) def resnetd152b(**kwargs): """ ResNet(D)-152 with dilation 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnetd(blocks=152, conv1_stride=False, model_name="resnetd152b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch ordinary_init = False bends = None pretrained = False models = [ resnetd50b, resnetd101b, resnetd152b, ] for model in models: net = model( pretrained=pretrained, ordinary_init=ordinary_init, bends=bends) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) if ordinary_init: assert (model != resnetd50b or weight_count == 25557032) assert (model != resnetd101b or weight_count == 44549160) assert (model != resnetd152b or weight_count == 60192808) else: assert (model != resnetd50b or weight_count == 25680808) assert (model != resnetd101b or weight_count == 44672936) assert (model != resnetd152b or weight_count == 60316584) x = torch.randn(1, 3, 224, 224) y = net(x) if bends is not None: y = y[0] y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
9,674
32.362069
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/quartznet.py
""" QuartzNet for ASR, implemented in PyTorch. Original paper: 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. """ __all__ = ['quartznet5x5_en_ls', 'quartznet15x5_en', 'quartznet15x5_en_nr', 'quartznet15x5_fr', 'quartznet15x5_de', 'quartznet15x5_it', 'quartznet15x5_es', 'quartznet15x5_ca', 'quartznet15x5_pl', 'quartznet15x5_ru', 'quartznet15x5_ru34'] from .jasper import get_jasper def quartznet5x5_en_ls(num_classes=29, **kwargs): """ QuartzNet 5x5 model for English language (trained on LibriSpeech dataset) from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 29 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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', "'"] return get_jasper(num_classes=num_classes, version=("quartznet", "5x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet5x5_en_ls", **kwargs) def quartznet15x5_en(num_classes=29, **kwargs): """ QuartzNet 15x5 model for English language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 29 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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', "'"] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_en", **kwargs) def quartznet15x5_en_nr(num_classes=29, **kwargs): """ QuartzNet 15x5 model for English language (with presence of noise) from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 29 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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', "'"] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_en_nr", **kwargs) def quartznet15x5_fr(num_classes=43, **kwargs): """ QuartzNet 15x5 model for French language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 43 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_fr", **kwargs) def quartznet15x5_de(num_classes=32, **kwargs): """ QuartzNet 15x5 model for German language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 32 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_de", **kwargs) def quartznet15x5_it(num_classes=39, **kwargs): """ QuartzNet 15x5 model for Italian language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 39 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_it", **kwargs) def quartznet15x5_es(num_classes=36, **kwargs): """ QuartzNet 15x5 model for Spanish language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 36 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_es", **kwargs) def quartznet15x5_ca(num_classes=39, **kwargs): """ QuartzNet 15x5 model for Spanish language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 39 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_ca", **kwargs) def quartznet15x5_pl(num_classes=34, **kwargs): """ QuartzNet 15x5 model for Spanish language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 34 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_pl", **kwargs) def quartznet15x5_ru(num_classes=35, **kwargs): """ QuartzNet 15x5 model for Russian language from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 35 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_ru", **kwargs) def quartznet15x5_ru34(num_classes=34, **kwargs): """ QuartzNet 15x5 model for Russian language (32 graphemes) from 'QuartzNet: Deep Automatic Speech Recognition with 1D Time-Channel Separable Convolutions,' https://arxiv.org/abs/1910.10261. Parameters: ---------- num_classes : int, default 34 Number of classification classes (number of graphemes). pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ vocabulary = [' ', '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'] return get_jasper(num_classes=num_classes, version=("quartznet", "15x5"), use_dw=True, vocabulary=vocabulary, model_name="quartznet15x5_ru34", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import numpy as np import torch pretrained = False from_audio = False audio_features = 64 use_cuda = True models = [ quartznet5x5_en_ls, quartznet15x5_en, quartznet15x5_en_nr, quartznet15x5_fr, quartznet15x5_de, quartznet15x5_it, quartznet15x5_es, quartznet15x5_ca, quartznet15x5_pl, quartznet15x5_ru, quartznet15x5_ru34, ] for model in models: net = model( in_channels=audio_features, from_audio=from_audio, pretrained=pretrained) if use_cuda: net = net.cuda() # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != quartznet5x5_en_ls or weight_count == 6713181) assert (model != quartznet15x5_en or weight_count == 18924381) assert (model != quartznet15x5_en_nr or weight_count == 18924381) assert (model != quartznet15x5_fr or weight_count == 18938731) assert (model != quartznet15x5_de or weight_count == 18927456) assert (model != quartznet15x5_it or weight_count == 18934631) assert (model != quartznet15x5_es or weight_count == 18931556) assert (model != quartznet15x5_ca or weight_count == 18934631) assert (model != quartznet15x5_pl or weight_count == 18929506) assert (model != quartznet15x5_ru or weight_count == 18930531) assert (model != quartznet15x5_ru34 or weight_count == 18929506) 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 = torch.randn(x_shape) x_len = torch.tensor(seq_len, dtype=torch.long, device=x.device) if use_cuda: x = x.cuda() x_len = x_len.cuda() y, y_len = net(x, x_len) # y.sum().backward() assert (tuple(y.size())[:2] == (batch, net.num_classes)) if from_audio: assert (y.size()[2] in range(seq_len_max // aud_scale * 2, seq_len_max // aud_scale * 2 + 9)) else: assert (y.size()[2] in [seq_len_max // 2, seq_len_max // 2 + 1]) if __name__ == "__main__": _test()
13,675
42.141956
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/preresnet.py
""" PreResNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. """ __all__ = ['PreResNet', 'preresnet10', 'preresnet12', 'preresnet14', 'preresnetbc14b', 'preresnet16', 'preresnet18_wd4', 'preresnet18_wd2', 'preresnet18_w3d4', 'preresnet18', 'preresnet26', 'preresnetbc26b', 'preresnet34', 'preresnetbc38b', 'preresnet50', 'preresnet50b', 'preresnet101', 'preresnet101b', 'preresnet152', 'preresnet152b', 'preresnet200', 'preresnet200b', 'preresnet269b', 'PreResBlock', 'PreResBottleneck', 'PreResUnit', 'PreResInitBlock', 'PreResActivation'] import os import torch.nn as nn import torch.nn.init as init from .common import pre_conv1x1_block, pre_conv3x3_block, conv1x1 class PreResBlock(nn.Module): """ Simple PreResNet block for residual path in PreResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. """ def __init__(self, in_channels, out_channels, stride, bias=False, use_bn=True): super(PreResBlock, self).__init__() self.conv1 = pre_conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias, use_bn=use_bn, return_preact=True) self.conv2 = pre_conv3x3_block( in_channels=out_channels, out_channels=out_channels, bias=bias, use_bn=use_bn) def forward(self, x): x, x_pre_activ = self.conv1(x) x = self.conv2(x) return x, x_pre_activ class PreResBottleneck(nn.Module): """ PreResNet bottleneck block for residual path in PreResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. 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, stride, conv1_stride): super(PreResBottleneck, self).__init__() mid_channels = out_channels // 4 self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, stride=(stride if conv1_stride else 1), return_preact=True) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=(1 if conv1_stride else stride)) self.conv3 = pre_conv1x1_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x, x_pre_activ = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x, x_pre_activ class PreResUnit(nn.Module): """ PreResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. 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, stride, bias=False, use_bn=True, bottleneck=True, conv1_stride=False): super(PreResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = PreResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, conv1_stride=conv1_stride) else: self.body = PreResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias, use_bn=use_bn) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias) def forward(self, 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 PreResInitBlock(nn.Module): """ PreResNet 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): super(PreResInitBlock, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=2, padding=3, bias=False) self.bn = nn.BatchNorm2d(num_features=out_channels) self.activ = nn.ReLU(inplace=True) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) x = self.pool(x) return x class PreResActivation(nn.Module): """ PreResNet pure pre-activation block without convolution layer. It's used by itself as the final block. Parameters: ---------- in_channels : int Number of input channels. """ def __init__(self, in_channels): super(PreResActivation, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class PreResNet(nn.Module): """ PreResNet model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, in_channels=3, in_size=(224, 224), num_classes=1000): super(PreResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 1 if (i == 0) or (j != 0) else 2 stage.add_module("unit{}".format(j + 1), PreResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_preresnet(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create 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. root : str, default '~/.torch/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 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 = PreResNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def preresnet10(**kwargs): """ PreResNet-10 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=10, model_name="preresnet10", **kwargs) def preresnet12(**kwargs): """ PreResNet-12 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=12, model_name="preresnet12", **kwargs) def preresnet14(**kwargs): """ PreResNet-14 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=14, model_name="preresnet14", **kwargs) def preresnetbc14b(**kwargs): """ PreResNet-BC-14b model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=14, bottleneck=True, conv1_stride=False, model_name="preresnetbc14b", **kwargs) def preresnet16(**kwargs): """ PreResNet-16 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=16, model_name="preresnet16", **kwargs) def preresnet18_wd4(**kwargs): """ PreResNet-18 model with 0.25 width scale from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=18, width_scale=0.25, model_name="preresnet18_wd4", **kwargs) def preresnet18_wd2(**kwargs): """ PreResNet-18 model with 0.5 width scale from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=18, width_scale=0.5, model_name="preresnet18_wd2", **kwargs) def preresnet18_w3d4(**kwargs): """ PreResNet-18 model with 0.75 width scale from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=18, width_scale=0.75, model_name="preresnet18_w3d4", **kwargs) def preresnet18(**kwargs): """ PreResNet-18 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=18, model_name="preresnet18", **kwargs) def preresnet26(**kwargs): """ PreResNet-26 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=26, bottleneck=False, model_name="preresnet26", **kwargs) def preresnetbc26b(**kwargs): """ PreResNet-BC-26b model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="preresnetbc26b", **kwargs) def preresnet34(**kwargs): """ PreResNet-34 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=34, model_name="preresnet34", **kwargs) def preresnetbc38b(**kwargs): """ PreResNet-BC-38b model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=38, bottleneck=True, conv1_stride=False, model_name="preresnetbc38b", **kwargs) def preresnet50(**kwargs): """ PreResNet-50 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=50, model_name="preresnet50", **kwargs) def preresnet50b(**kwargs): """ PreResNet-50 model with stride at the second convolution in bottleneck block from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=50, conv1_stride=False, model_name="preresnet50b", **kwargs) def preresnet101(**kwargs): """ PreResNet-101 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=101, model_name="preresnet101", **kwargs) def preresnet101b(**kwargs): """ PreResNet-101 model with stride at the second convolution in bottleneck block from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=101, conv1_stride=False, model_name="preresnet101b", **kwargs) def preresnet152(**kwargs): """ PreResNet-152 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=152, model_name="preresnet152", **kwargs) def preresnet152b(**kwargs): """ PreResNet-152 model with stride at the second convolution in bottleneck block from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=152, conv1_stride=False, model_name="preresnet152b", **kwargs) def preresnet200(**kwargs): """ PreResNet-200 model from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=200, model_name="preresnet200", **kwargs) def preresnet200b(**kwargs): """ PreResNet-200 model with stride at the second convolution in bottleneck block from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=200, conv1_stride=False, model_name="preresnet200b", **kwargs) def preresnet269b(**kwargs): """ PreResNet-269 model with stride at the second convolution in bottleneck block from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet(blocks=269, conv1_stride=False, model_name="preresnet269b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ preresnet10, preresnet12, preresnet14, preresnetbc14b, preresnet16, preresnet18_wd4, preresnet18_wd2, preresnet18_w3d4, preresnet18, preresnet26, preresnetbc26b, preresnet34, preresnetbc38b, preresnet50, preresnet50b, preresnet101, preresnet101b, preresnet152, preresnet152b, preresnet200, preresnet200b, preresnet269b, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != preresnet10 or weight_count == 5417128) assert (model != preresnet12 or weight_count == 5491112) assert (model != preresnet14 or weight_count == 5786536) assert (model != preresnetbc14b or weight_count == 10057384) assert (model != preresnet16 or weight_count == 6967208) assert (model != preresnet18_wd4 or weight_count == 3935960) assert (model != preresnet18_wd2 or weight_count == 5802440) assert (model != preresnet18_w3d4 or weight_count == 8473784) assert (model != preresnet18 or weight_count == 11687848) assert (model != preresnet26 or weight_count == 17958568) assert (model != preresnetbc26b or weight_count == 15987624) assert (model != preresnet34 or weight_count == 21796008) assert (model != preresnetbc38b or weight_count == 21917864) assert (model != preresnet50 or weight_count == 25549480) assert (model != preresnet50b or weight_count == 25549480) assert (model != preresnet101 or weight_count == 44541608) assert (model != preresnet101b or weight_count == 44541608) assert (model != preresnet152 or weight_count == 60185256) assert (model != preresnet152b or weight_count == 60185256) assert (model != preresnet200 or weight_count == 64666280) assert (model != preresnet200b or weight_count == 64666280) assert (model != preresnet269b or weight_count == 102065832) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
26,501
32.044888
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/lednet.py
""" LEDNet for image segmentation, implemented in PyTorch. Original paper: 'LEDNet: A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation,' https://arxiv.org/abs/1905.02423. """ __all__ = ['LEDNet', 'lednet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1_block, conv3x3_block, conv5x5_block, conv7x7_block, asym_conv3x3_block, ChannelShuffle,\ InterpolationBlock, Hourglass, BreakBlock from .enet import ENetMixDownBlock class LEDBranch(nn.Module): """ LEDNet encoder branch. Parameters: ---------- channels : int Number of input/output channels. dilation : int Dilation value for convolution layer. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, channels, dilation, dropout_rate, bn_eps): super(LEDBranch, self).__init__() self.use_dropout = (dropout_rate != 0.0) self.conv1 = asym_conv3x3_block( channels=channels, bias=True, lw_use_bn=False, bn_eps=bn_eps) self.conv2 = asym_conv3x3_block( channels=channels, padding=dilation, dilation=dilation, bias=True, lw_use_bn=False, bn_eps=bn_eps, rw_activation=None) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): x = self.conv1(x) x = self.conv2(x) if self.use_dropout: x = self.dropout(x) return x class LEDUnit(nn.Module): """ LEDNet encoder unit (Split-Shuffle-non-bottleneck). Parameters: ---------- channels : int Number of input/output channels. dilation : int Dilation value for convolution layer. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, channels, dilation, dropout_rate, bn_eps): super(LEDUnit, self).__init__() mid_channels = channels // 2 self.left_branch = LEDBranch( channels=mid_channels, dilation=dilation, dropout_rate=dropout_rate, bn_eps=bn_eps) self.right_branch = LEDBranch( channels=mid_channels, dilation=dilation, dropout_rate=dropout_rate, bn_eps=bn_eps) self.activ = nn.ReLU(inplace=True) self.shuffle = ChannelShuffle( channels=channels, groups=2) def forward(self, x): identity = x x1, x2 = torch.chunk(x, chunks=2, dim=1) x1 = self.left_branch(x1) x2 = self.right_branch(x2) x = torch.cat((x1, x2), dim=1) x = x + identity x = self.activ(x) x = self.shuffle(x) return x class PoolingBranch(nn.Module): """ Pooling branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. bn_eps : float Small float added to variance in Batch norm. in_size : tuple of 2 int or None Spatial size of input image. down_size : int Spatial size of downscaled image. """ def __init__(self, in_channels, out_channels, bias, bn_eps, in_size, down_size): super(PoolingBranch, self).__init__() self.in_size = in_size self.pool = nn.AdaptiveAvgPool2d(output_size=down_size) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=bias, bn_eps=bn_eps) self.up = InterpolationBlock( scale_factor=None, out_size=in_size) def forward(self, x): in_size = self.in_size if self.in_size is not None else x.shape[2:] x = self.pool(x) x = self.conv(x) x = self.up(x, in_size) return x class APN(nn.Module): """ Attention pyramid network block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. in_size : tuple of 2 int or None Spatial size of input image. """ def __init__(self, in_channels, out_channels, bn_eps, in_size): super(APN, self).__init__() self.in_size = in_size att_out_channels = 1 self.pool_branch = PoolingBranch( in_channels=in_channels, out_channels=out_channels, bias=True, bn_eps=bn_eps, in_size=in_size, down_size=1) self.body = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=True, bn_eps=bn_eps) down_seq = nn.Sequential() down_seq.add_module("down1", conv7x7_block( in_channels=in_channels, out_channels=att_out_channels, stride=2, bias=True, bn_eps=bn_eps)) down_seq.add_module("down2", conv5x5_block( in_channels=att_out_channels, out_channels=att_out_channels, stride=2, bias=True, bn_eps=bn_eps)) down3_subseq = nn.Sequential() down3_subseq.add_module("conv1", conv3x3_block( in_channels=att_out_channels, out_channels=att_out_channels, stride=2, bias=True, bn_eps=bn_eps)) down3_subseq.add_module("conv2", conv3x3_block( in_channels=att_out_channels, out_channels=att_out_channels, bias=True, bn_eps=bn_eps)) down_seq.add_module("down3", down3_subseq) up_seq = nn.Sequential() up = InterpolationBlock(scale_factor=2) up_seq.add_module("up1", up) up_seq.add_module("up2", up) up_seq.add_module("up3", up) skip_seq = nn.Sequential() skip_seq.add_module("skip1", BreakBlock()) skip_seq.add_module("skip2", conv7x7_block( in_channels=att_out_channels, out_channels=att_out_channels, bias=True, bn_eps=bn_eps)) skip_seq.add_module("skip3", conv5x5_block( in_channels=att_out_channels, out_channels=att_out_channels, bias=True, bn_eps=bn_eps)) self.hg = Hourglass( down_seq=down_seq, up_seq=up_seq, skip_seq=skip_seq) def forward(self, x): y = self.pool_branch(x) w = self.hg(x) x = self.body(x) x = x * w x = x + y return x class LEDNet(nn.Module): """ LEDNet model from 'LEDNet: A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation,' https://arxiv.org/abs/1905.02423. Parameters: ---------- channels : list of int Number of output channels for each unit. dilations : list of int Dilations for units. dropout_rates : list of list of int Dropout rates for each unit in encoder. correct_size_mistmatch : bool Whether to correct downscaled sizes of images in encoder. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, channels, dilations, dropout_rates, correct_size_mismatch=False, bn_eps=1e-5, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(LEDNet, self).__init__() 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.num_classes = num_classes self.fixed_size = fixed_size self.encoder = nn.Sequential() for i, dilations_per_stage in enumerate(dilations): out_channels = channels[i] dropout_rate = dropout_rates[i] stage = nn.Sequential() for j, dilation in enumerate(dilations_per_stage): if j == 0: stage.add_module("unit{}".format(j + 1), ENetMixDownBlock( in_channels=in_channels, out_channels=out_channels, bias=True, bn_eps=bn_eps, correct_size_mismatch=correct_size_mismatch)) in_channels = out_channels else: stage.add_module("unit{}".format(j + 1), LEDUnit( channels=in_channels, dilation=dilation, dropout_rate=dropout_rate, bn_eps=bn_eps)) self.encoder.add_module("stage{}".format(i + 1), stage) self.apn = APN( in_channels=in_channels, out_channels=num_classes, bn_eps=bn_eps, in_size=(in_size[0] // 8, in_size[1] // 8) if fixed_size else None) self.up = InterpolationBlock( scale_factor=8, align_corners=True) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.encoder(x) x = self.apn(x) x = self.up(x) return x def get_lednet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create LEDNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [32, 64, 128] dilations = [[0, 1, 1, 1], [0, 1, 1], [0, 1, 2, 5, 9, 2, 5, 9, 17]] dropout_rates = [0.03, 0.03, 0.3] bn_eps = 1e-3 net = LEDNet( channels=channels, dilations=dilations, dropout_rates=dropout_rates, bn_eps=bn_eps, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def lednet_cityscapes(num_classes=19, **kwargs): """ LEDNet model for Cityscapes from 'LEDNet: A Lightweight Encoder-Decoder Network for Real-Time Semantic Segmentation,' https://arxiv.org/abs/1905.02423. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_lednet(num_classes=num_classes, model_name="lednet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True correct_size_mismatch = False in_size = (1024, 2048) classes = 19 models = [ lednet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size, correct_size_mismatch=correct_size_mismatch) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != lednet_cityscapes or weight_count == 922821) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
13,638
29.241685
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/superpointnet.py
""" SuperPointNet for HPatches (image matching), implemented in PyTorch. Original paper: 'SuperPoint: Self-Supervised Interest Point Detection and Description,' https://arxiv.org/abs/1712.07629. """ __all__ = ['SuperPointNet', 'superpointnet'] import os import torch import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F from .common import conv1x1, conv3x3_block class SPHead(nn.Module): """ SuperPointNet head block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, mid_channels, out_channels): super(SPHead, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, bias=True, use_bn=False) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class SPDetector(nn.Module): """ SuperPointNet detector. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. conf_thresh : float, default 0.015 Confidence threshold. nms_dist : int, default 4 NMS distance. border_size : int, default 4 Image border size to remove points. reduction : int, default 8 Feature reduction factor. """ def __init__(self, in_channels, mid_channels, conf_thresh=0.015, nms_dist=4, border_size=4, reduction=8): super(SPDetector, self).__init__() self.conf_thresh = conf_thresh self.nms_dist = nms_dist self.border_size = border_size self.reduction = reduction num_classes = reduction * reduction + 1 self.detector = SPHead( in_channels=in_channels, mid_channels=mid_channels, out_channels=num_classes) def forward(self, x): batch = x.size(0) x_height, x_width = x.size()[-2:] img_height = x_height * self.reduction img_width = x_width * self.reduction semi = self.detector(x) dense = semi.softmax(dim=1) nodust = dense[:, :-1, :, :] heatmap = nodust.permute(0, 2, 3, 1) heatmap = heatmap.reshape((-1, x_height, x_width, self.reduction, self.reduction)) heatmap = heatmap.permute(0, 1, 3, 2, 4) heatmap = heatmap.reshape((-1, 1, x_height * self.reduction, x_width * self.reduction)) heatmap_mask = (heatmap >= self.conf_thresh) pad = self.nms_dist bord = self.border_size + pad heatmap_mask2 = F.pad(heatmap_mask, pad=(pad, pad, pad, pad)) pts_list = [] confs_list = [] for i in range(batch): heatmap_i = heatmap[i, 0] heatmap_mask_i = heatmap_mask[i, 0] heatmap_mask2_i = heatmap_mask2[i, 0] src_pts = torch.nonzero(heatmap_mask_i) src_confs = torch.masked_select(heatmap_i, heatmap_mask_i) src_inds = torch.argsort(src_confs, descending=True) dst_inds = torch.zeros_like(src_inds) dst_pts_count = 0 for ind_j in src_inds: pt = src_pts[ind_j] + pad assert (pad <= pt[0] < heatmap_mask2_i.shape[0] - pad) assert (pad <= pt[1] < heatmap_mask2_i.shape[1] - pad) assert (0 <= pt[0] - pad < img_height) assert (0 <= pt[1] - pad < img_width) if heatmap_mask2_i[pt[0], pt[1]] == 1: heatmap_mask2_i[(pt[0] - pad):(pt[0] + pad + 1), (pt[1] - pad):(pt[1] + pad + 1)] = 0 if (bord < pt[0] - pad <= img_height - bord) and (bord < pt[1] - pad <= img_width - bord): dst_inds[dst_pts_count] = ind_j dst_pts_count += 1 dst_inds = dst_inds[:dst_pts_count] dst_pts = torch.index_select(src_pts, dim=0, index=dst_inds) dst_confs = torch.index_select(src_confs, dim=0, index=dst_inds) pts_list.append(dst_pts) confs_list.append(dst_confs) return pts_list, confs_list class SPDescriptor(nn.Module): """ SuperPointNet descriptor generator. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. descriptor_length : int, default 256 Descriptor length. transpose_descriptors : bool, default True Whether transpose descriptors with respect to points. reduction : int, default 8 Feature reduction factor. """ def __init__(self, in_channels, mid_channels, descriptor_length=256, transpose_descriptors=True, reduction=8): super(SPDescriptor, self).__init__() self.desc_length = descriptor_length self.transpose_descriptors = transpose_descriptors self.reduction = reduction self.head = SPHead( in_channels=in_channels, mid_channels=mid_channels, out_channels=descriptor_length) def forward(self, x, pts_list): x_height, x_width = x.size()[-2:] coarse_desc_map = self.head(x) coarse_desc_map = F.normalize(coarse_desc_map) descriptors_list = [] for i, pts in enumerate(pts_list): pts = pts.float() pts[:, 0] = pts[:, 0] / (0.5 * x_height * self.reduction) - 1.0 pts[:, 1] = pts[:, 1] / (0.5 * x_width * self.reduction) - 1.0 if self.transpose_descriptors: pts = torch.index_select(pts, dim=1, index=torch.tensor([1, 0], device=pts.device)) pts = pts.unsqueeze(0).unsqueeze(0) descriptors = F.grid_sample(coarse_desc_map[i:(i + 1)], pts) descriptors = descriptors.squeeze(0).squeeze(1) descriptors = descriptors.transpose(0, 1) descriptors = F.normalize(descriptors) descriptors_list.append(descriptors) return descriptors_list class SuperPointNet(nn.Module): """ SuperPointNet model from 'SuperPoint: Self-Supervised Interest Point Detection and Description,' https://arxiv.org/abs/1712.07629. Parameters: ---------- channels : list of list of int Number of output channels for each unit. final_block_channels : int Number of output channels for the final units. transpose_descriptors : bool, default True Whether transpose descriptors with respect to points. in_channels : int, default 1 Number of input channels. """ def __init__(self, channels, final_block_channels, transpose_descriptors=True, in_channels=1): super(SuperPointNet, self).__init__() self.features = nn.Sequential() for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): if (j == 0) and (i != 0): stage.add_module("reduce{}".format(i + 1), nn.MaxPool2d( kernel_size=2, stride=2)) stage.add_module("unit{}".format(j + 1), conv3x3_block( in_channels=in_channels, out_channels=out_channels, bias=True, use_bn=False)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.detector = SPDetector( in_channels=in_channels, mid_channels=final_block_channels) self.descriptor = SPDescriptor( in_channels=in_channels, mid_channels=final_block_channels, transpose_descriptors=transpose_descriptors) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): assert (x.size(1) == 1) x = self.features(x) pts_list, confs_list = self.detector(x) descriptors_list = self.descriptor(x, pts_list) return pts_list, confs_list, descriptors_list def get_superpointnet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SuperPointNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels_per_layers = [64, 64, 128, 128] layers = [2, 2, 2, 2] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] final_block_channels = 256 net = SuperPointNet( channels=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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def superpointnet(**kwargs): """ SuperPointNet model from 'SuperPoint: Self-Supervised Interest Point Detection and Description,' https://arxiv.org/abs/1712.07629. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_superpointnet(model_name="superpointnet", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ superpointnet, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != superpointnet or weight_count == 1300865) # x = torch.randn(1, 1, 224, 224) x = torch.randn(1, 1, 1000, 2000) y = net(x) # y.sum().backward() assert (len(y) == 3) if __name__ == "__main__": _test()
11,418
31.719198
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/ibndensenet.py
""" IBN-DenseNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Two at Once: Enhancing Learning and Generalization Capacities via IBN-Net,' https://arxiv.org/abs/1807.09441. """ __all__ = ['IBNDenseNet', 'ibn_densenet121', 'ibn_densenet161', 'ibn_densenet169', 'ibn_densenet201'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import pre_conv3x3_block, IBN from .preresnet import PreResInitBlock, PreResActivation from .densenet import TransitionBlock class IBNPreConvBlock(nn.Module): """ IBN-Net specific convolution block with BN/IBN 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. stride : 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_ibn : bool, default False Whether use Instance-Batch Normalization. return_preact : bool, default False Whether return pre-activation. It's used by PreResNet. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, use_ibn=False, return_preact=False): super(IBNPreConvBlock, self).__init__() self.use_ibn = use_ibn self.return_preact = return_preact if self.use_ibn: self.ibn = IBN( channels=in_channels, first_fraction=0.6, inst_first=False) else: self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=False) def forward(self, x): if self.use_ibn: x = self.ibn(x) else: 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 ibn_pre_conv1x1_block(in_channels, out_channels, stride=1, use_ibn=False, return_preact=False): """ 1x1 version of the IBN-Net specific pre-activated convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. use_ibn : bool, default False Whether use Instance-Batch Normalization. return_preact : bool, default False Whether return pre-activation. """ return IBNPreConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, use_ibn=use_ibn, return_preact=return_preact) class IBNDenseUnit(nn.Module): """ IBN-DenseNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. conv1_ibn : bool Whether to use IBN normalization in the first convolution layer of the block. """ def __init__(self, in_channels, out_channels, dropout_rate, conv1_ibn): super(IBNDenseUnit, self).__init__() self.use_dropout = (dropout_rate != 0.0) bn_size = 4 inc_channels = out_channels - in_channels mid_channels = inc_channels * bn_size self.conv1 = ibn_pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_ibn=conv1_ibn) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=inc_channels) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) if self.use_dropout: x = self.dropout(x) x = torch.cat((identity, x), dim=1) return x class IBNDenseNet(nn.Module): """ IBN-DenseNet 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. 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 (224, 224) Spatial size of the expected input image. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, dropout_rate=0.0, in_channels=3, in_size=(224, 224), num_classes=1000): super(IBNDenseNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() if i != 0: stage.add_module("trans{}".format(i + 1), TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2))) in_channels = in_channels // 2 for j, out_channels in enumerate(channels_per_stage): conv1_ibn = (i < 3) and (j % 3 == 0) stage.add_module("unit{}".format(j + 1), IBNDenseUnit( in_channels=in_channels, out_channels=out_channels, dropout_rate=dropout_rate, conv1_ibn=conv1_ibn)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_ibndensenet(num_layers, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create IBN-DenseNet model with specific parameters. Parameters: ---------- num_layers : int Number 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if num_layers == 121: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 24, 16] elif num_layers == 161: init_block_channels = 96 growth_rate = 48 layers = [6, 12, 36, 24] elif num_layers == 169: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 32, 32] elif num_layers == 201: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 48, 32] else: raise ValueError("Unsupported IBN-DenseNet version with number of layers {}".format(num_layers)) 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 = IBNDenseNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def ibn_densenet121(**kwargs): """ IBN-DenseNet-121 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibndensenet(num_layers=121, model_name="ibn_densenet121", **kwargs) def ibn_densenet161(**kwargs): """ IBN-DenseNet-161 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibndensenet(num_layers=161, model_name="ibn_densenet161", **kwargs) def ibn_densenet169(**kwargs): """ IBN-DenseNet-169 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibndensenet(num_layers=169, model_name="ibn_densenet169", **kwargs) def ibn_densenet201(**kwargs): """ IBN-DenseNet-201 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibndensenet(num_layers=201, model_name="ibn_densenet201", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ ibn_densenet121, ibn_densenet161, ibn_densenet169, ibn_densenet201, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ibn_densenet121 or weight_count == 7978856) assert (model != ibn_densenet161 or weight_count == 28681000) assert (model != ibn_densenet169 or weight_count == 14149480) assert (model != ibn_densenet201 or weight_count == 20013928) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,647
30.384615
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/hardnet.py
""" HarDNet for ImageNet-1K, implemented in PyTorch. Original paper: 'HarDNet: A Low Memory Traffic Network,' https://arxiv.org/abs/1909.00948. """ __all__ = ['HarDNet', 'hardnet39ds', 'hardnet68ds', 'hardnet68', 'hardnet85'] import os import torch import torch.nn as nn from .common import conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv_block class InvDwsConvBlock(nn.Module): """ Inverse 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. pw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the pointwise convolution block. dw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the depthwise convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, pw_activation=(lambda: nn.ReLU(inplace=True)), dw_activation=(lambda: nn.ReLU(inplace=True))): super(InvDwsConvBlock, self).__init__() self.pw_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=pw_activation) self.dw_conv = dwconv_block( in_channels=out_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=dw_activation) def forward(self, x): x = self.pw_conv(x) x = self.dw_conv(x) return x def invdwsconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, bn_eps=1e-5, pw_activation=(lambda: nn.ReLU(inplace=True)), dw_activation=(lambda: nn.ReLU(inplace=True))): """ 3x3 inverse depthwise separable version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. pw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the pointwise convolution block. dw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the depthwise convolution block. """ return InvDwsConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, bn_eps=bn_eps, pw_activation=pw_activation, dw_activation=dw_activation) class HarDUnit(nn.Module): """ HarDNet unit. Parameters: ---------- in_channels_list : list of int Number of input channels for each block. out_channels_list : list of int Number of output channels for each block. links_list : list of list of int List of indices for each layer. use_deptwise : bool Whether to use depthwise downsampling. use_dropout : bool Whether to use dropout module. downsampling : bool Whether to downsample input. activation : str Name of activation function. """ def __init__(self, in_channels_list, out_channels_list, links_list, use_deptwise, use_dropout, downsampling, activation): super(HarDUnit, self).__init__() self.links_list = links_list self.use_dropout = use_dropout self.downsampling = downsampling self.blocks = nn.Sequential() for i in range(len(links_list)): in_channels = in_channels_list[i] out_channels = out_channels_list[i] if use_deptwise: unit = invdwsconv3x3_block( in_channels=in_channels, out_channels=out_channels, pw_activation=activation, dw_activation=None) else: unit = conv3x3_block( in_channels=in_channels, out_channels=out_channels) self.blocks.add_module("block{}".format(i + 1), unit) if self.use_dropout: self.dropout = nn.Dropout(p=0.1) self.conv = conv1x1_block( in_channels=in_channels_list[-1], out_channels=out_channels_list[-1], activation=activation) if self.downsampling: if use_deptwise: self.downsample = dwconv3x3_block( in_channels=out_channels_list[-1], out_channels=out_channels_list[-1], stride=2, activation=None) else: self.downsample = nn.MaxPool2d( kernel_size=2, stride=2) def forward(self, x): layer_outs = [x] for links_i, layer_i in zip(self.links_list, self.blocks._modules.values()): layer_in = [] for idx_ij in links_i: layer_in.append(layer_outs[idx_ij]) if len(layer_in) > 1: x = torch.cat(layer_in, dim=1) else: x = layer_in[0] out = layer_i(x) layer_outs.append(out) outs = [] for i, layer_out_i in enumerate(layer_outs): if (i == len(layer_outs) - 1) or (i % 2 == 1): outs.append(layer_out_i) x = torch.cat(outs, dim=1) if self.use_dropout: x = self.dropout(x) x = self.conv(x) if self.downsampling: x = self.downsample(x) return x class HarDInitBlock(nn.Module): """ HarDNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. use_deptwise : bool Whether to use depthwise downsampling. activation : str Name of activation function. """ def __init__(self, in_channels, out_channels, use_deptwise, activation): super(HarDInitBlock, self).__init__() mid_channels = out_channels // 2 self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2, activation=activation) conv2_block_class = conv1x1_block if use_deptwise else conv3x3_block self.conv2 = conv2_block_class( in_channels=mid_channels, out_channels=out_channels, activation=activation) if use_deptwise: self.downsample = dwconv3x3_block( in_channels=out_channels, out_channels=out_channels, stride=2, activation=None) else: self.downsample = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.downsample(x) return x class HarDNet(nn.Module): """ HarDNet model from 'HarDNet: A Low Memory Traffic Network,' https://arxiv.org/abs/1909.00948. Parameters: ---------- init_block_channels : int Number of output channels for the initial unit. unit_in_channels : list of list of list of int Number of input channels for each layer in each stage. unit_out_channels : list list of of list of int Number of output channels for each layer in each stage. unit_links : list of list of list of int List of indices for each layer in each stage. use_deptwise : bool Whether to use depthwise downsampling. use_last_dropout : bool Whether to use dropouts in the last unit. output_dropout_rate : float Parameter of Dropout layer before classifier. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, init_block_channels, unit_in_channels, unit_out_channels, unit_links, use_deptwise, use_last_dropout, output_dropout_rate, in_channels=3, in_size=(224, 224), num_classes=1000): super(HarDNet, self).__init__() self.in_size = in_size self.num_classes = num_classes activation = "relu6" self.features = nn.Sequential() self.features.add_module("init_block", HarDInitBlock( in_channels=in_channels, out_channels=init_block_channels, use_deptwise=use_deptwise, activation=activation)) for i, (in_channels_list_i, out_channels_list_i) in enumerate(zip(unit_in_channels, unit_out_channels)): stage = nn.Sequential() for j, (in_channels_list_ij, out_channels_list_ij) in enumerate(zip(in_channels_list_i, out_channels_list_i)): use_dropout = ((j == len(in_channels_list_i) - 1) and (i == len(unit_in_channels) - 1) and use_last_dropout) downsampling = ((j == len(in_channels_list_i) - 1) and (i != len(unit_in_channels) - 1)) stage.add_module("unit{}".format(j + 1), HarDUnit( in_channels_list=in_channels_list_ij, out_channels_list=out_channels_list_ij, links_list=unit_links[i][j], use_deptwise=use_deptwise, use_dropout=use_dropout, downsampling=downsampling, activation=activation)) self.features.add_module("stage{}".format(i + 1), stage) in_channels = unit_out_channels[-1][-1][-1] self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Sequential() self.output.add_module("dropout", nn.Dropout(p=output_dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight, mode="fan_out", nonlinearity="relu") if module.bias is not None: nn.init.constant_(module.bias, 0) elif isinstance(module, nn.BatchNorm2d): nn.init.constant_(module.weight, 1) nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_hardnet(blocks, use_deptwise=True, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create HarDNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. use_deepwise : bool, default True Whether to use depthwise separable version of the 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 39: init_block_channels = 48 growth_factor = 1.6 dropout_rate = 0.05 if use_deptwise else 0.1 layers = [4, 16, 8, 4] channels_per_layers = [96, 320, 640, 1024] growth_rates = [16, 20, 64, 160] downsamples = [1, 1, 1, 0] use_dropout = False elif blocks == 68: init_block_channels = 64 growth_factor = 1.7 dropout_rate = 0.05 if use_deptwise else 0.1 layers = [8, 16, 16, 16, 4] channels_per_layers = [128, 256, 320, 640, 1024] growth_rates = [14, 16, 20, 40, 160] downsamples = [1, 0, 1, 1, 0] use_dropout = False elif blocks == 85: init_block_channels = 96 growth_factor = 1.7 dropout_rate = 0.05 if use_deptwise else 0.2 layers = [8, 16, 16, 16, 16, 4] channels_per_layers = [192, 256, 320, 480, 720, 1280] growth_rates = [24, 24, 28, 36, 48, 256] downsamples = [1, 0, 1, 0, 1, 0] use_dropout = True else: raise ValueError("Unsupported HarDNet version with number of layers {}".format(blocks)) assert (downsamples[-1] == 0) def calc_stage_params(): def calc_unit_params(): def calc_blocks_params(layer_idx, base_channels, growth_rate): if layer_idx == 0: return base_channels, 0, [] out_channels_ij = growth_rate links_ij = [] for k in range(10): dv = 2 ** k if layer_idx % dv == 0: t = layer_idx - dv links_ij.append(t) if k > 0: out_channels_ij *= growth_factor out_channels_ij = int(int(out_channels_ij + 1) / 2) * 2 in_channels_ij = 0 for t in links_ij: out_channels_ik, _, _ = calc_blocks_params( layer_idx=t, base_channels=base_channels, growth_rate=growth_rate) in_channels_ij += out_channels_ik return out_channels_ij, in_channels_ij, links_ij unit_out_channels = [] unit_in_channels = [] unit_links = [] for num_layers, growth_rate, base_channels, channels_per_layers_i in zip( layers, growth_rates, [init_block_channels] + channels_per_layers[:-1], channels_per_layers): stage_out_channels_i = 0 unit_out_channels_i = [] unit_in_channels_i = [] unit_links_i = [] for j in range(num_layers): out_channels_ij, in_channels_ij, links_ij = calc_blocks_params( layer_idx=(j + 1), base_channels=base_channels, growth_rate=growth_rate) unit_out_channels_i.append(out_channels_ij) unit_in_channels_i.append(in_channels_ij) unit_links_i.append(links_ij) if (j % 2 == 0) or (j == num_layers - 1): stage_out_channels_i += out_channels_ij unit_in_channels_i.append(stage_out_channels_i) unit_out_channels_i.append(channels_per_layers_i) unit_out_channels.append(unit_out_channels_i) unit_in_channels.append(unit_in_channels_i) unit_links.append(unit_links_i) return unit_out_channels, unit_in_channels, unit_links unit_out_channels, unit_in_channels, unit_links = calc_unit_params() stage_out_channels = [] stage_in_channels = [] stage_links = [] stage_out_channels_k = None for i in range(len(layers)): if stage_out_channels_k is None: stage_out_channels_k = [] stage_in_channels_k = [] stage_links_k = [] stage_out_channels_k.append(unit_out_channels[i]) stage_in_channels_k.append(unit_in_channels[i]) stage_links_k.append(unit_links[i]) if (downsamples[i] == 1) or (i == len(layers) - 1): stage_out_channels.append(stage_out_channels_k) stage_in_channels.append(stage_in_channels_k) stage_links.append(stage_links_k) stage_out_channels_k = None return stage_out_channels, stage_in_channels, stage_links stage_out_channels, stage_in_channels, stage_links = calc_stage_params() net = HarDNet( init_block_channels=init_block_channels, unit_in_channels=stage_in_channels, unit_out_channels=stage_out_channels, unit_links=stage_links, use_deptwise=use_deptwise, use_last_dropout=use_dropout, output_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def hardnet39ds(**kwargs): """ HarDNet-39DS (Depthwise Separable) model from 'HarDNet: A Low Memory Traffic Network,' https://arxiv.org/abs/1909.00948. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hardnet(blocks=39, use_deptwise=True, model_name="hardnet39ds", **kwargs) def hardnet68ds(**kwargs): """ HarDNet-68DS (Depthwise Separable) model from 'HarDNet: A Low Memory Traffic Network,' https://arxiv.org/abs/1909.00948. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hardnet(blocks=68, use_deptwise=True, model_name="hardnet68ds", **kwargs) def hardnet68(**kwargs): """ HarDNet-68 model from 'HarDNet: A Low Memory Traffic Network,' https://arxiv.org/abs/1909.00948. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hardnet(blocks=68, use_deptwise=False, model_name="hardnet68", **kwargs) def hardnet85(**kwargs): """ HarDNet-85 model from 'HarDNet: A Low Memory Traffic Network,' https://arxiv.org/abs/1909.00948. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_hardnet(blocks=85, use_deptwise=False, model_name="hardnet85", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ hardnet39ds, hardnet68ds, hardnet68, hardnet85, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != hardnet39ds or weight_count == 3488228) assert (model != hardnet68ds or weight_count == 4180602) assert (model != hardnet68 or weight_count == 17565348) assert (model != hardnet85 or weight_count == 36670212) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
21,984
34.176
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/sinet.py
""" SINet for image segmentation, implemented in PyTorch. Original paper: 'SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder,' https://arxiv.org/abs/1911.09099. """ __all__ = ['SINet', 'sinet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1, get_activation_layer, conv1x1_block, conv3x3_block, round_channels, dwconv_block,\ Concurrent, InterpolationBlock, ChannelShuffle class SEBlock(nn.Module): """ SINet version of 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. round_mid : bool, default False Whether to round middle channel number (make divisible by 8). activation : function, or str, or nn.Module, default 'relu' Activation function after the first convolution. out_activation : function, or str, or nn.Module, default 'sigmoid' Activation function after the last convolution. """ def __init__(self, channels, reduction=16, round_mid=False, mid_activation=(lambda: nn.ReLU(inplace=True)), out_activation=(lambda: nn.Sigmoid())): super(SEBlock, self).__init__() self.use_conv2 = (reduction > 1) mid_channels = channels // reduction if not round_mid else round_channels(float(channels) / reduction) self.pool = nn.AdaptiveAvgPool2d(output_size=1) self.fc1 = nn.Linear( in_features=channels, out_features=mid_channels) if self.use_conv2: self.activ = get_activation_layer(mid_activation) self.fc2 = nn.Linear( in_features=mid_channels, out_features=channels) self.sigmoid = get_activation_layer(out_activation) def forward(self, x): w = self.pool(x) w = w.squeeze(dim=-1).squeeze(dim=-1) w = self.fc1(w) if self.use_conv2: w = self.activ(w) w = self.fc2(w) w = self.sigmoid(w) w = w.unsqueeze(dim=-1).unsqueeze(dim=-1) x = x * w return x class DwsConvBlock(nn.Module): """ SINet version of 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. stride : 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. 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_eps : float, default 1e-5 Small float added to variance in Batch norm. dw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the depthwise convolution block. pw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the pointwise convolution block. se_reduction : int, default 0 Squeeze reduction value (0 means no-se). """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, bias=False, dw_use_bn=True, pw_use_bn=True, bn_eps=1e-5, dw_activation=(lambda: nn.ReLU(inplace=True)), pw_activation=(lambda: nn.ReLU(inplace=True)), se_reduction=0): super(DwsConvBlock, self).__init__() self.use_se = (se_reduction > 0) self.dw_conv = dwconv_block( in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=dw_use_bn, bn_eps=bn_eps, activation=dw_activation) if self.use_se: self.se = SEBlock( channels=in_channels, reduction=se_reduction, round_mid=False, mid_activation=(lambda: nn.PReLU(in_channels // se_reduction)), out_activation=(lambda: nn.PReLU(in_channels))) self.pw_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=pw_use_bn, bn_eps=bn_eps, activation=pw_activation) def forward(self, x): x = self.dw_conv(x) if self.use_se: x = self.se(x) x = self.pw_conv(x) return x def dwsconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, dw_use_bn=True, pw_use_bn=True, bn_eps=1e-5, dw_activation=(lambda: nn.ReLU(inplace=True)), pw_activation=(lambda: nn.ReLU(inplace=True)), se_reduction=0): """ 3x3 depthwise separable version of the standard convolution block (SINet version). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. 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_eps : float, default 1e-5 Small float added to variance in Batch norm. dw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the depthwise convolution block. pw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the pointwise convolution block. se_reduction : int, default 0 Squeeze reduction value (0 means no-se). """ return DwsConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, dw_use_bn=dw_use_bn, pw_use_bn=pw_use_bn, bn_eps=bn_eps, dw_activation=dw_activation, pw_activation=pw_activation, se_reduction=se_reduction) def dwconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 3x3 depthwise version of the standard convolution block (SINet version). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, bn_eps=bn_eps, activation=activation) class FDWConvBlock(nn.Module): """ Factorized 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 Convolution window size. stride : int or tuple/list of 2 int Strides of the convolution. padding : int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the each convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(FDWConvBlock, self).__init__() assert use_bn self.activate = (activation is not None) self.v_conv = dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=(kernel_size, 1), stride=stride, padding=(padding, 0), dilation=dilation, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=None) self.h_conv = dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=(1, kernel_size), stride=stride, padding=(0, padding), dilation=dilation, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=None) if self.activate: self.act = get_activation_layer(activation) def forward(self, x): x = self.v_conv(x) + self.h_conv(x) if self.activate: x = self.act(x) return x def fdwconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 3x3 factorized depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return FDWConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def fdwconv5x5_block(in_channels, out_channels, stride=1, padding=2, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 5x5 factorized depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return FDWConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=5, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) class SBBlock(nn.Module): """ SB-block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int Convolution window size for a factorized depthwise separable convolution block. scale_factor : int Scale factor. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, kernel_size, scale_factor, bn_eps): super(SBBlock, self).__init__() self.use_scale = (scale_factor > 1) if self.use_scale: self.down_scale = nn.AvgPool2d( kernel_size=scale_factor, stride=scale_factor) self.up_scale = InterpolationBlock(scale_factor=scale_factor) use_fdw = (scale_factor > 0) if use_fdw: fdwconv3x3_class = fdwconv3x3_block if kernel_size == 3 else fdwconv5x5_block self.conv1 = fdwconv3x3_class( in_channels=in_channels, out_channels=in_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(in_channels))) else: self.conv1 = dwconv3x3_block( in_channels=in_channels, out_channels=in_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(in_channels))) self.conv2 = conv1x1( in_channels=in_channels, out_channels=out_channels) self.bn = nn.BatchNorm2d( num_features=out_channels, eps=bn_eps) def forward(self, x): if self.use_scale: x = self.down_scale(x) x = self.conv1(x) x = self.conv2(x) if self.use_scale: x = self.up_scale(x) x = self.bn(x) return x class PreActivation(nn.Module): """ PreResNet like pure pre-activation block without convolution layer. Parameters: ---------- in_channels : int Number of input channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, bn_eps): super(PreActivation, self).__init__() self.bn = nn.BatchNorm2d( num_features=in_channels, eps=bn_eps) self.activ = nn.PReLU(num_parameters=in_channels) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class ESPBlock(nn.Module): """ ESP block, which is based on the following principle: Reduce ---> Split ---> Transform --> Merge. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_sizes : list of int Convolution window size for branches. scale_factors : list of int Scale factor for branches. use_residual : bool Whether to use residual connection. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, kernel_sizes, scale_factors, use_residual, bn_eps): super(ESPBlock, self).__init__() self.use_residual = use_residual groups = len(kernel_sizes) mid_channels = int(out_channels / groups) res_channels = out_channels - groups * mid_channels self.conv = conv1x1( in_channels=in_channels, out_channels=mid_channels, groups=groups) self.c_shuffle = ChannelShuffle( channels=mid_channels, groups=groups) self.branches = Concurrent() for i in range(groups): out_channels_i = (mid_channels + res_channels) if i == 0 else mid_channels self.branches.add_module("branch{}".format(i + 1), SBBlock( in_channels=mid_channels, out_channels=out_channels_i, kernel_size=kernel_sizes[i], scale_factor=scale_factors[i], bn_eps=bn_eps)) self.preactiv = PreActivation( in_channels=out_channels, bn_eps=bn_eps) def forward(self, x): if self.use_residual: identity = x x = self.conv(x) x = self.c_shuffle(x) x = self.branches(x) if self.use_residual: x = identity + x x = self.preactiv(x) return x class SBStage(nn.Module): """ SB stage. Parameters: ---------- in_channels : int Number of input channels. down_channels : int Number of output channels for a downscale block. channels_list : list of int Number of output channels for all residual block. kernel_sizes_list : list of int Convolution window size for branches. scale_factors_list : list of int Scale factor for branches. use_residual_list : list of int List of flags for using residual in each ESP-block. se_reduction : int Squeeze reduction value (0 means no-se). bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, down_channels, channels_list, kernel_sizes_list, scale_factors_list, use_residual_list, se_reduction, bn_eps): super(SBStage, self).__init__() self.down_conv = dwsconv3x3_block( in_channels=in_channels, out_channels=down_channels, stride=2, dw_use_bn=False, bn_eps=bn_eps, dw_activation=None, pw_activation=(lambda: nn.PReLU(down_channels)), se_reduction=se_reduction) in_channels = down_channels self.main_branch = nn.Sequential() for i, out_channels in enumerate(channels_list): use_residual = (use_residual_list[i] == 1) kernel_sizes = kernel_sizes_list[i] scale_factors = scale_factors_list[i] self.main_branch.add_module("block{}".format(i + 1), ESPBlock( in_channels=in_channels, out_channels=out_channels, kernel_sizes=kernel_sizes, scale_factors=scale_factors, use_residual=use_residual, bn_eps=bn_eps)) in_channels = out_channels self.preactiv = PreActivation( in_channels=(down_channels + in_channels), bn_eps=bn_eps) def forward(self, x): x = self.down_conv(x) y = self.main_branch(x) x = torch.cat((x, y), dim=1) x = self.preactiv(x) return x, y class SBEncoderInitBlock(nn.Module): """ SB encoder specific initial block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, mid_channels, out_channels, bn_eps): super(SBEncoderInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid_channels))) self.conv2 = dwsconv3x3_block( in_channels=mid_channels, out_channels=out_channels, stride=2, dw_use_bn=False, bn_eps=bn_eps, dw_activation=None, pw_activation=(lambda: nn.PReLU(out_channels)), se_reduction=1) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class SBEncoder(nn.Module): """ SB encoder for SINet. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of input channels. init_block_channels : list int Number of output channels for convolutions in the initial block. down_channels_list : list of int Number of downsample channels for each residual block. channels_list : list of list of int Number of output channels for all residual block. kernel_sizes_list : list of list of int Convolution window size for each residual block. scale_factors_list : list of list of int Scale factor for each residual block. use_residual_list : list of list of int List of flags for using residual in each residual block. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, init_block_channels, down_channels_list, channels_list, kernel_sizes_list, scale_factors_list, use_residual_list, bn_eps): super(SBEncoder, self).__init__() self.init_block = SBEncoderInitBlock( in_channels=in_channels, mid_channels=init_block_channels[0], out_channels=init_block_channels[1], bn_eps=bn_eps) in_channels = init_block_channels[1] self.stage1 = SBStage( in_channels=in_channels, down_channels=down_channels_list[0], channels_list=channels_list[0], kernel_sizes_list=kernel_sizes_list[0], scale_factors_list=scale_factors_list[0], use_residual_list=use_residual_list[0], se_reduction=1, bn_eps=bn_eps) in_channels = down_channels_list[0] + channels_list[0][-1] self.stage2 = SBStage( in_channels=in_channels, down_channels=down_channels_list[1], channels_list=channels_list[1], kernel_sizes_list=kernel_sizes_list[1], scale_factors_list=scale_factors_list[1], use_residual_list=use_residual_list[1], se_reduction=2, bn_eps=bn_eps) in_channels = down_channels_list[1] + channels_list[1][-1] self.output = conv1x1( in_channels=in_channels, out_channels=out_channels) def forward(self, x): y1 = self.init_block(x) x, y2 = self.stage1(y1) x, _ = self.stage2(x) x = self.output(x) return x, y2, y1 class SBDecodeBlock(nn.Module): """ SB decoder block for SINet. Parameters: ---------- channels : int Number of output classes. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, channels, bn_eps): super(SBDecodeBlock, self).__init__() self.up = InterpolationBlock( scale_factor=2, align_corners=False) self.bn = nn.BatchNorm2d( num_features=channels, eps=bn_eps) self.conf = nn.Softmax2d() def forward(self, x, y): x = self.up(x) x = self.bn(x) w_conf = self.conf(x) w_max = (torch.max(w_conf, dim=1)[0]).unsqueeze(1).expand_as(x) x = y * (1 - w_max) + x return x class SBDecoder(nn.Module): """ SB decoder for SINet. Parameters: ---------- dim2 : int Size of dimension #2. num_classes : int Number of segmentation classes. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, dim2, num_classes, bn_eps): super(SBDecoder, self).__init__() self.decode1 = SBDecodeBlock( channels=num_classes, bn_eps=bn_eps) self.decode2 = SBDecodeBlock( channels=num_classes, bn_eps=bn_eps) self.conv3c = conv1x1_block( in_channels=dim2, out_channels=num_classes, bn_eps=bn_eps, activation=(lambda: nn.PReLU(num_classes))) self.output = nn.ConvTranspose2d( in_channels=num_classes, out_channels=num_classes, kernel_size=2, stride=2, padding=0, output_padding=0, bias=False) self.up = InterpolationBlock(scale_factor=2) def forward(self, y3, y2, y1): y2 = self.conv3c(y2) x = self.decode1(y3, y2) x = self.decode2(x, y1) x = self.output(x) x = self.up(x) return x class SINet(nn.Module): """ SINet model from 'SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder,' https://arxiv.org/abs/1911.09099. Parameters: ---------- down_channels_list : list of int Number of downsample channels for each residual block. channels_list : list of list of int Number of output channels for all residual block. kernel_sizes_list : list of list of int Convolution window size for each residual block. scale_factors_list : list of list of int Scale factor for each residual block. use_residual_list : list of list of int List of flags for using residual in each residual block. dim2 : int Size of dimension #2. bn_eps : float Small float added to variance in Batch norm. 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. num_classes : int, default 21 Number of segmentation classes. """ def __init__(self, down_channels_list, channels_list, kernel_sizes_list, scale_factors_list, use_residual_list, dim2, bn_eps, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=21): super(SINet, self).__init__() assert (fixed_size is not None) assert (in_channels > 0) assert ((in_size[0] % 64 == 0) and (in_size[1] % 64 == 0)) self.in_size = in_size self.num_classes = num_classes self.aux = aux init_block_channels = [16, num_classes] out_channels = num_classes self.encoder = SBEncoder( in_channels=in_channels, out_channels=out_channels, init_block_channels=init_block_channels, down_channels_list=down_channels_list, channels_list=channels_list, kernel_sizes_list=kernel_sizes_list, scale_factors_list=scale_factors_list, use_residual_list=use_residual_list, bn_eps=bn_eps) self.decoder = SBDecoder( dim2=dim2, num_classes=num_classes, bn_eps=bn_eps) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): y3, y2, y1 = self.encoder(x) x = self.decoder(y3, y2, y1) if self.aux: return x, y3 else: return x def get_sinet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SINet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ kernel_sizes_list = [ [[3, 5], [3, 3], [3, 3]], [[3, 5], [3, 3], [5, 5], [3, 5], [3, 5], [3, 5], [3, 3], [5, 5], [3, 5], [3, 5]]] scale_factors_list = [ [[1, 1], [0, 1], [0, 1]], [[1, 1], [0, 1], [1, 4], [2, 8], [1, 1], [1, 1], [0, 1], [1, 8], [2, 4], [0, 2]]] chnn = 4 dims = [24] + [24 * (i + 2) + 4 * (chnn - 1) for i in range(3)] dim1 = dims[0] dim2 = dims[1] dim3 = dims[2] dim4 = dims[3] p = len(kernel_sizes_list[0]) q = len(kernel_sizes_list[1]) channels_list = [[dim2] * p, ([dim3] * (q // 2)) + ([dim4] * (q - q // 2))] use_residual_list = [[0] + ([1] * (p - 1)), [0] + ([1] * (q // 2 - 1)) + [0] + ([1] * (q - q // 2 - 1))] down_channels_list = [dim1, dim2] net = SINet( down_channels_list=down_channels_list, channels_list=channels_list, kernel_sizes_list=kernel_sizes_list, scale_factors_list=scale_factors_list, use_residual_list=use_residual_list, dim2=dims[1], **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def sinet_cityscapes(num_classes=19, **kwargs): """ SINet model for Cityscapes from 'SINet: Extreme Lightweight Portrait Segmentation Networks with Spatial Squeeze Modules and Information Blocking Decoder,' https://arxiv.org/abs/1911.09099. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sinet(num_classes=num_classes, bn_eps=1e-3, model_name="sinet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch in_size = (1024, 2048) aux = False fixed_size = True pretrained = False models = [ sinet_cityscapes, ] for model in models: net = model(pretrained=pretrained, aux=aux, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sinet_cityscapes or weight_count == 119418) batch = 14 x = torch.randn(batch, 3, in_size[0], in_size[1]) ys = net(x) y = ys[0] if aux else ys # y.sum().backward() assert (tuple(y.size()) == (batch, 19, in_size[0], in_size[1])) if __name__ == "__main__": _test()
33,876
30.929312
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/shufflenetv2b.py
""" ShuffleNet V2 for ImageNet-1K, implemented in PyTorch. The alternative version. Original paper: 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,' https://arxiv.org/abs/1807.11164. """ __all__ = ['ShuffleNetV2b', 'shufflenetv2b_wd2', 'shufflenetv2b_w1', 'shufflenetv2b_w3d2', 'shufflenetv2b_w2'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block, dwconv3x3_block, ChannelShuffle, ChannelShuffle2, SEBlock class ShuffleUnit(nn.Module): """ ShuffleNetV2(b) unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. downsample : bool Whether do downsample. use_se : bool Whether to use SE block. use_residual : bool Whether to use residual connection. shuffle_group_first : bool Whether to use channel shuffle in group first mode. """ def __init__(self, in_channels, out_channels, downsample, use_se, use_residual, shuffle_group_first): super(ShuffleUnit, self).__init__() self.downsample = downsample self.use_se = use_se self.use_residual = use_residual mid_channels = out_channels // 2 in_channels2 = in_channels // 2 assert (in_channels % 2 == 0) y2_in_channels = (in_channels if downsample else in_channels2) y2_out_channels = out_channels - y2_in_channels self.conv1 = conv1x1_block( in_channels=y2_in_channels, out_channels=mid_channels) self.dconv = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=(2 if self.downsample else 1), activation=None) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=y2_out_channels) if self.use_se: self.se = SEBlock(channels=y2_out_channels) if downsample: self.shortcut_dconv = dwconv3x3_block( in_channels=in_channels, out_channels=in_channels, stride=2, activation=None) self.shortcut_conv = conv1x1_block( in_channels=in_channels, out_channels=in_channels) if shuffle_group_first: self.c_shuffle = ChannelShuffle( channels=out_channels, groups=2) else: self.c_shuffle = ChannelShuffle2( channels=out_channels, groups=2) def forward(self, x): if self.downsample: y1 = self.shortcut_dconv(x) y1 = self.shortcut_conv(y1) x2 = x else: y1, x2 = torch.chunk(x, chunks=2, dim=1) y2 = self.conv1(x2) y2 = self.dconv(y2) y2 = self.conv2(y2) if self.use_se: y2 = self.se(y2) if self.use_residual and not self.downsample: y2 = y2 + x2 x = torch.cat((y1, y2), dim=1) x = self.c_shuffle(x) return x class ShuffleInitBlock(nn.Module): """ ShuffleNetV2(b) 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): super(ShuffleInitBlock, self).__init__() self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1, ceil_mode=False) def forward(self, x): x = self.conv(x) x = self.pool(x) return x class ShuffleNetV2b(nn.Module): """ ShuffleNetV2(b) model from 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,' https://arxiv.org/abs/1807.11164. 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. use_se : bool, default False Whether to use SE block. use_residual : bool, default False Whether to use residual connections. shuffle_group_first : bool, default True Whether to use channel shuffle in group first 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, use_se=False, use_residual=False, shuffle_group_first=True, in_channels=3, in_size=(224, 224), num_classes=1000): super(ShuffleNetV2b, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", 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.Sequential() for j, out_channels in enumerate(channels_per_stage): downsample = (j == 0) stage.add_module("unit{}".format(j + 1), ShuffleUnit( in_channels=in_channels, out_channels=out_channels, downsample=downsample, use_se=use_se, use_residual=use_residual, shuffle_group_first=shuffle_group_first)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_shufflenetv2b(width_scale, shuffle_group_first=True, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create ShuffleNetV2(b) model with specific parameters. Parameters: ---------- width_scale : float Scale factor for width of layers. shuffle_group_first : bool, default True Whether to use channel shuffle in group first mode. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 24 final_block_channels = 1024 layers = [4, 8, 4] channels_per_layers = [116, 232, 464] 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] if width_scale > 1.5: final_block_channels = int(final_block_channels * width_scale) net = ShuffleNetV2b( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, shuffle_group_first=shuffle_group_first, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def shufflenetv2b_wd2(**kwargs): """ ShuffleNetV2(b) 0.5x model from 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,' https://arxiv.org/abs/1807.11164. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shufflenetv2b( width_scale=(12.0 / 29.0), shuffle_group_first=True, model_name="shufflenetv2b_wd2", **kwargs) def shufflenetv2b_w1(**kwargs): """ ShuffleNetV2(b) 1x model from 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,' https://arxiv.org/abs/1807.11164. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shufflenetv2b( width_scale=1.0, shuffle_group_first=True, model_name="shufflenetv2b_w1", **kwargs) def shufflenetv2b_w3d2(**kwargs): """ ShuffleNetV2(b) 1.5x model from 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,' https://arxiv.org/abs/1807.11164. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shufflenetv2b( width_scale=(44.0 / 29.0), shuffle_group_first=True, model_name="shufflenetv2b_w3d2", **kwargs) def shufflenetv2b_w2(**kwargs): """ ShuffleNetV2(b) 2x model from 'ShuffleNet V2: Practical Guidelines for Efficient CNN Architecture Design,' https://arxiv.org/abs/1807.11164. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shufflenetv2b( width_scale=(61.0 / 29.0), shuffle_group_first=True, model_name="shufflenetv2b_w2", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ shufflenetv2b_wd2, shufflenetv2b_w1, shufflenetv2b_w3d2, shufflenetv2b_w2, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != shufflenetv2b_wd2 or weight_count == 1366792) assert (model != shufflenetv2b_w1 or weight_count == 2279760) assert (model != shufflenetv2b_w3d2 or weight_count == 4410194) assert (model != shufflenetv2b_w2 or weight_count == 7611290) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,431
30.553299
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/sparsenet.py
""" SparseNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. """ __all__ = ['SparseNet', 'sparsenet121', 'sparsenet161', 'sparsenet169', 'sparsenet201', 'sparsenet264'] import os import math import torch import torch.nn as nn import torch.nn.init as init from .common import pre_conv1x1_block, pre_conv3x3_block from .preresnet import PreResInitBlock, PreResActivation from .densenet import TransitionBlock def sparsenet_exponential_fetch(lst): """ SparseNet's specific exponential fetch. Parameters: ---------- lst : list List of something. Returns: ------- list Filtered list. """ return [lst[len(lst) - 2**i] for i in range(1 + math.floor(math.log(len(lst), 2)))] class SparseBlock(nn.Module): """ SparseNet block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, dropout_rate): super(SparseBlock, self).__init__() self.use_dropout = (dropout_rate != 0.0) bn_size = 4 mid_channels = out_channels * bn_size self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=out_channels) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): x = self.conv1(x) x = self.conv2(x) if self.use_dropout: x = self.dropout(x) return x class SparseStage(nn.Module): """ SparseNet stage. Parameters: ---------- in_channels : int Number of input channels. channels_per_stage : list of int Number of output channels for each unit in stage. growth_rate : int Growth rate for blocks. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. do_transition : bool Whether use transition block. """ def __init__(self, in_channels, channels_per_stage, growth_rate, dropout_rate, do_transition): super(SparseStage, self).__init__() self.do_transition = do_transition if self.do_transition: self.trans = TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2)) in_channels = in_channels // 2 self.blocks = nn.Sequential() for i, out_channels in enumerate(channels_per_stage): self.blocks.add_module("block{}".format(i + 1), SparseBlock( in_channels=in_channels, out_channels=growth_rate, dropout_rate=dropout_rate)) in_channels = out_channels def forward(self, x): if self.do_transition: x = self.trans(x) outs = [x] for block in self.blocks._modules.values(): y = block(x) outs.append(y) flt_outs = sparsenet_exponential_fetch(outs) x = torch.cat(tuple(flt_outs), dim=1) return x class SparseNet(nn.Module): """ SparseNet model from 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. 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. growth_rate : int Growth rate for blocks. 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 (224, 224) Spatial size of the expected input image. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, growth_rate, dropout_rate=0.0, in_channels=3, in_size=(224, 224), num_classes=1000): super(SparseNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = SparseStage( in_channels=in_channels, channels_per_stage=channels_per_stage, growth_rate=growth_rate, dropout_rate=dropout_rate, do_transition=(i != 0)) in_channels = channels_per_stage[-1] self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_sparsenet(num_layers, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SparseNet model with specific parameters. Parameters: ---------- num_layers : int Number 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if num_layers == 121: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 24, 16] elif num_layers == 161: init_block_channels = 96 growth_rate = 48 layers = [6, 12, 36, 24] elif num_layers == 169: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 32, 32] elif num_layers == 201: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 48, 32] elif num_layers == 264: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 64, 48] else: raise ValueError("Unsupported SparseNet version with number of layers {}".format(num_layers)) from functools import reduce channels = reduce( lambda xi, yi: xi + [reduce( lambda xj, yj: xj + [sum(sparsenet_exponential_fetch([xj[0]] + [yj[0]] * (yj[1] + 1)))], zip([growth_rate] * yi, range(yi)), [xi[-1][-1] // 2])[1:]], layers, [[init_block_channels * 2]])[1:] net = SparseNet( channels=channels, init_block_channels=init_block_channels, growth_rate=growth_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def sparsenet121(**kwargs): """ SparseNet-121 model from 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sparsenet(num_layers=121, model_name="sparsenet121", **kwargs) def sparsenet161(**kwargs): """ SparseNet-161 model from 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sparsenet(num_layers=161, model_name="sparsenet161", **kwargs) def sparsenet169(**kwargs): """ SparseNet-169 model from 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sparsenet(num_layers=169, model_name="sparsenet169", **kwargs) def sparsenet201(**kwargs): """ SparseNet-201 model from 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sparsenet(num_layers=201, model_name="sparsenet201", **kwargs) def sparsenet264(**kwargs): """ SparseNet-264 model from 'Sparsely Aggregated Convolutional Networks,' https://arxiv.org/abs/1801.05895. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sparsenet(num_layers=264, model_name="sparsenet264", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ sparsenet121, sparsenet161, sparsenet169, sparsenet201, sparsenet264, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sparsenet121 or weight_count == 3250824) assert (model != sparsenet161 or weight_count == 9853288) assert (model != sparsenet169 or weight_count == 4709864) assert (model != sparsenet201 or weight_count == 5703144) assert (model != sparsenet264 or weight_count == 7717224) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
11,646
29.569554
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/menet.py
""" MENet for ImageNet-1K, implemented in PyTorch. Original paper: 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. """ __all__ = ['MENet', 'menet108_8x1_g3', 'menet128_8x1_g4', 'menet160_8x1_g8', 'menet228_12x1_g3', 'menet256_12x1_g4', 'menet348_12x1_g3', 'menet352_12x1_g8', 'menet456_24x1_g3'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv3x3, depthwise_conv3x3, ChannelShuffle class MEUnit(nn.Module): """ MENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. side_channels : int Number of side 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, side_channels, groups, downsample, ignore_group): super(MEUnit, self).__init__() self.downsample = downsample mid_channels = out_channels // 4 if downsample: out_channels -= in_channels # residual branch self.compress_conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels, groups=(1 if ignore_group else groups)) self.compress_bn1 = nn.BatchNorm2d(num_features=mid_channels) self.c_shuffle = ChannelShuffle( channels=mid_channels, groups=groups) self.dw_conv2 = depthwise_conv3x3( channels=mid_channels, stride=(2 if self.downsample else 1)) self.dw_bn2 = nn.BatchNorm2d(num_features=mid_channels) self.expand_conv3 = conv1x1( in_channels=mid_channels, out_channels=out_channels, groups=groups) self.expand_bn3 = nn.BatchNorm2d(num_features=out_channels) if downsample: self.avgpool = nn.AvgPool2d(kernel_size=3, stride=2, padding=1) self.activ = nn.ReLU(inplace=True) # fusion branch self.s_merge_conv = conv1x1( in_channels=mid_channels, out_channels=side_channels) self.s_merge_bn = nn.BatchNorm2d(num_features=side_channels) self.s_conv = conv3x3( in_channels=side_channels, out_channels=side_channels, stride=(2 if self.downsample else 1)) self.s_conv_bn = nn.BatchNorm2d(num_features=side_channels) self.s_evolve_conv = conv1x1( in_channels=side_channels, out_channels=mid_channels) self.s_evolve_bn = nn.BatchNorm2d(num_features=mid_channels) def forward(self, x): identity = x # pointwise group convolution 1 x = self.compress_conv1(x) x = self.compress_bn1(x) x = self.activ(x) x = self.c_shuffle(x) # merging y = self.s_merge_conv(x) y = self.s_merge_bn(y) y = self.activ(y) # depthwise convolution (bottleneck) x = self.dw_conv2(x) x = self.dw_bn2(x) # evolution y = self.s_conv(y) y = self.s_conv_bn(y) y = self.activ(y) y = self.s_evolve_conv(y) y = self.s_evolve_bn(y) y = torch.sigmoid(y) x = x * y # pointwise group convolution 2 x = self.expand_conv3(x) x = self.expand_bn3(x) # identity branch if self.downsample: identity = self.avgpool(identity) x = torch.cat((x, identity), dim=1) else: x = x + identity x = self.activ(x) return x class MEInitBlock(nn.Module): """ MENet 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): super(MEInitBlock, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=2, padding=1, bias=False) self.bn = nn.BatchNorm2d(num_features=out_channels) self.activ = nn.ReLU(inplace=True) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) x = self.pool(x) return x class MENet(nn.Module): """ MENet model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. 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. side_channels : int Number of side channels in a ME-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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, side_channels, groups, in_channels=3, in_size=(224, 224), num_classes=1000): super(MENet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", MEInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): downsample = (j == 0) ignore_group = (i == 0) and (j == 0) stage.add_module("unit{}".format(j + 1), MEUnit( in_channels=in_channels, out_channels=out_channels, side_channels=side_channels, groups=groups, downsample=downsample, ignore_group=ignore_group)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_menet(first_stage_channels, side_channels, groups, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create MENet model with specific parameters. Parameters: ---------- first_stage_channels : int Number of output channels at the first stage. side_channels : int Number of side channels in a ME-unit. groups : int Number of groups in convolution 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ layers = [4, 8, 4] if first_stage_channels == 108: init_block_channels = 12 channels_per_layers = [108, 216, 432] elif first_stage_channels == 128: init_block_channels = 12 channels_per_layers = [128, 256, 512] elif first_stage_channels == 160: init_block_channels = 16 channels_per_layers = [160, 320, 640] elif first_stage_channels == 228: init_block_channels = 24 channels_per_layers = [228, 456, 912] elif first_stage_channels == 256: init_block_channels = 24 channels_per_layers = [256, 512, 1024] elif first_stage_channels == 348: init_block_channels = 24 channels_per_layers = [348, 696, 1392] elif first_stage_channels == 352: init_block_channels = 24 channels_per_layers = [352, 704, 1408] elif first_stage_channels == 456: init_block_channels = 48 channels_per_layers = [456, 912, 1824] else: raise ValueError("The {} of `first_stage_channels` is not supported".format(first_stage_channels)) channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = MENet( channels=channels, init_block_channels=init_block_channels, side_channels=side_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def menet108_8x1_g3(**kwargs): """ 108-MENet-8x1 (g=3) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=108, side_channels=8, groups=3, model_name="menet108_8x1_g3", **kwargs) def menet128_8x1_g4(**kwargs): """ 128-MENet-8x1 (g=4) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=128, side_channels=8, groups=4, model_name="menet128_8x1_g4", **kwargs) def menet160_8x1_g8(**kwargs): """ 160-MENet-8x1 (g=8) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=160, side_channels=8, groups=8, model_name="menet160_8x1_g8", **kwargs) def menet228_12x1_g3(**kwargs): """ 228-MENet-12x1 (g=3) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=228, side_channels=12, groups=3, model_name="menet228_12x1_g3", **kwargs) def menet256_12x1_g4(**kwargs): """ 256-MENet-12x1 (g=4) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=256, side_channels=12, groups=4, model_name="menet256_12x1_g4", **kwargs) def menet348_12x1_g3(**kwargs): """ 348-MENet-12x1 (g=3) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=348, side_channels=12, groups=3, model_name="menet348_12x1_g3", **kwargs) def menet352_12x1_g8(**kwargs): """ 352-MENet-12x1 (g=8) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=352, side_channels=12, groups=8, model_name="menet352_12x1_g8", **kwargs) def menet456_24x1_g3(**kwargs): """ 456-MENet-24x1 (g=3) model from 'Merging and Evolution: Improving Convolutional Neural Networks for Mobile Applications,' https://arxiv.org/abs/1803.09127. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_menet(first_stage_channels=456, side_channels=24, groups=3, model_name="menet456_24x1_g3", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ menet108_8x1_g3, menet128_8x1_g4, menet160_8x1_g8, menet228_12x1_g3, menet256_12x1_g4, menet348_12x1_g3, menet352_12x1_g8, menet456_24x1_g3, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != menet108_8x1_g3 or weight_count == 654516) assert (model != menet128_8x1_g4 or weight_count == 750796) assert (model != menet160_8x1_g8 or weight_count == 850120) assert (model != menet228_12x1_g3 or weight_count == 1806568) assert (model != menet256_12x1_g4 or weight_count == 1888240) assert (model != menet348_12x1_g3 or weight_count == 3368128) assert (model != menet352_12x1_g8 or weight_count == 2272872) assert (model != menet456_24x1_g3 or weight_count == 5304784) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
15,917
31.956522
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/voca.py
""" VOCA for speech-driven facial animation, implemented in PyTorch. Original paper: 'Capture, Learning, and Synthesis of 3D Speaking Styles,' https://arxiv.org/abs/1905.03079. """ __all__ = ['VOCA', 'voca8flame'] import os import torch import torch.nn as nn import torch.nn.functional as F from .common import ConvBlock class VocaEncoder(nn.Module): """ VOCA encoder. Parameters: ---------- audio_features : int Number of audio features (characters/sounds). audio_window_size : int Size of audio window (for time related audio features). base_persons : int Number of base persons (subjects). encoder_features : int Number of encoder features. """ def __init__(self, audio_features, audio_window_size, base_persons, encoder_features): super(VocaEncoder, self).__init__() self.audio_window_size = audio_window_size channels = (32, 32, 64, 64) fc1_channels = 128 self.bn = nn.BatchNorm2d(num_features=1) in_channels = audio_features + base_persons self.branch = nn.Sequential() for i, out_channels in enumerate(channels): self.branch.add_module("conv{}".format(i + 1), ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=(3, 1), stride=(2, 1), padding=(1, 0), bias=True, use_bn=False)) in_channels = out_channels in_channels += base_persons self.fc1 = nn.Linear( in_features=in_channels, out_features=fc1_channels) self.fc2 = nn.Linear( in_features=fc1_channels, out_features=encoder_features) def forward(self, x, pid): x = self.bn(x) x = x.transpose(1, 3).contiguous() y = pid.unsqueeze(-1).unsqueeze(-1) y = y.repeat(1, 1, self.audio_window_size, 1) x = torch.cat((x, y), dim=1) x = self.branch(x) x = x.view(x.size(0), -1) x = torch.cat((x, pid), dim=1) x = self.fc1(x) x = x.tanh() x = self.fc2(x) return x class VOCA(nn.Module): """ VOCA model from 'Capture, Learning, and Synthesis of 3D Speaking Styles,' https://arxiv.org/abs/1905.03079. 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). base_persons : int, default 8 Number of base persons (subjects). encoder_features : int, default 50 Number of encoder features. vertices : int, default 5023 Number of 3D geometry vertices. """ def __init__(self, audio_features=29, audio_window_size=16, base_persons=8, encoder_features=50, vertices=5023): super(VOCA, self).__init__() self.base_persons = base_persons self.encoder = VocaEncoder( audio_features=audio_features, audio_window_size=audio_window_size, base_persons=base_persons, encoder_features=encoder_features) self.decoder = nn.Linear( in_features=encoder_features, out_features=(3 * vertices)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x, pid): pid = F.one_hot(pid.long(), num_classes=self.base_persons).type_as(pid) x = self.encoder(x, pid) x = self.decoder(x) x = x.view(x.size(0), 1, -1, 3) return x def get_voca(base_persons, vertices, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create VOCA model with specific parameters. Parameters: ---------- base_persons : int Number of base persons (subjects). vertices : int Number of 3D geometry vertices. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = VOCA( base_persons=base_persons, vertices=vertices, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def voca8flame(**kwargs): """ VOCA-8-FLAME model for 8 base persons and FLAME topology from 'Capture, Learning, and Synthesis of 3D Speaking Styles,' https://arxiv.org/abs/1905.03079. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_voca(base_persons=8, vertices=5023, model_name="voca8flame", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ voca8flame, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != voca8flame or weight_count == 809563) batch = 14 audio_features = 29 audio_window_size = 16 vertices = 5023 x = torch.randn(batch, 1, audio_window_size, audio_features) pid = torch.full(size=(batch,), fill_value=3) y = net(x, pid) # y.sum().backward() assert (y.shape == (batch, 1, vertices, 3)) if __name__ == "__main__": _test()
6,683
28.575221
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/shakeshakeresnet_cifar.py
""" Shake-Shake-ResNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. """ __all__ = ['CIFARShakeShakeResNet', 'shakeshakeresnet20_2x16d_cifar10', 'shakeshakeresnet20_2x16d_cifar100', 'shakeshakeresnet20_2x16d_svhn', 'shakeshakeresnet26_2x32d_cifar10', 'shakeshakeresnet26_2x32d_cifar100', 'shakeshakeresnet26_2x32d_svhn'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv3x3_block from .resnet import ResBlock, ResBottleneck class ShakeShake(torch.autograd.Function): """ Shake-Shake function. """ @staticmethod def forward(ctx, x1, x2, alpha): y = alpha * x1 + (1 - alpha) * x2 return y @staticmethod def backward(ctx, dy): beta = torch.rand(dy.size(0), dtype=dy.dtype, device=dy.device).view(-1, 1, 1, 1) return beta * dy, (1 - beta) * dy, None class ShakeShakeShortcut(nn.Module): """ Shake-Shake-ResNet shortcut. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(ShakeShakeShortcut, self).__init__() assert (out_channels % 2 == 0) mid_channels = out_channels // 2 self.pool = nn.AvgPool2d( kernel_size=1, stride=stride) self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels) self.conv2 = conv1x1( in_channels=in_channels, out_channels=mid_channels) self.bn = nn.BatchNorm2d(num_features=out_channels) self.pad = nn.ZeroPad2d(padding=(1, 0, 1, 0)) def forward(self, x): x1 = self.pool(x) x1 = self.conv1(x1) x2 = x[:, :, :-1, :-1].contiguous() x2 = self.pad(x2) x2 = self.pool(x2) x2 = self.conv2(x2) x = torch.cat((x1, x2), dim=1) x = self.bn(x) return x class ShakeShakeResUnit(nn.Module): """ Shake-Shake-ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, stride, bottleneck): super(ShakeShakeResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) branch_class = ResBottleneck if bottleneck else ResBlock self.branch1 = branch_class( in_channels=in_channels, out_channels=out_channels, stride=stride) self.branch2 = branch_class( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_branch = ShakeShakeShortcut( in_channels=in_channels, out_channels=out_channels, stride=stride) self.activ = nn.ReLU(inplace=True) self.shake_shake = ShakeShake.apply def forward(self, x): if self.resize_identity: identity = self.identity_branch(x) else: identity = x x1 = self.branch1(x) x2 = self.branch2(x) if self.training: alpha = torch.rand(x1.size(0), dtype=x1.dtype, device=x1.device).view(-1, 1, 1, 1) x = self.shake_shake(x1, x2, alpha) else: x = 0.5 * (x1 + x2) x = x + identity x = self.activ(x) return x class CIFARShakeShakeResNet(nn.Module): """ Shake-Shake-ResNet model for CIFAR from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARShakeShakeResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ShakeShakeResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_shakeshakeresnet_cifar(classes, blocks, bottleneck, first_stage_channels=16, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create Shake-Shake-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. first_stage_channels : int, default 16 Number of output channels for the first stage. 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. root : str, default '~/.torch/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 init_block_channels = 16 from functools import reduce channels_per_layers = reduce(lambda x, y: x + [x[-1] * 2], range(2), [first_stage_channels]) 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 = CIFARShakeShakeResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def shakeshakeresnet20_2x16d_cifar10(classes=10, **kwargs): """ Shake-Shake-ResNet-20-2x16d model for CIFAR-10 from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakeshakeresnet_cifar(classes=classes, blocks=20, bottleneck=False, first_stage_channels=16, model_name="shakeshakeresnet20_2x16d_cifar10", **kwargs) def shakeshakeresnet20_2x16d_cifar100(classes=100, **kwargs): """ Shake-Shake-ResNet-20-2x16d model for CIFAR-100 from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakeshakeresnet_cifar(classes=classes, blocks=20, bottleneck=False, first_stage_channels=16, model_name="shakeshakeresnet20_2x16d_cifar100", **kwargs) def shakeshakeresnet20_2x16d_svhn(classes=10, **kwargs): """ Shake-Shake-ResNet-20-2x16d model for SVHN from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakeshakeresnet_cifar(classes=classes, blocks=20, bottleneck=False, first_stage_channels=16, model_name="shakeshakeresnet20_2x16d_svhn", **kwargs) def shakeshakeresnet26_2x32d_cifar10(classes=10, **kwargs): """ Shake-Shake-ResNet-26-2x32d model for CIFAR-10 from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakeshakeresnet_cifar(classes=classes, blocks=26, bottleneck=False, first_stage_channels=32, model_name="shakeshakeresnet26_2x32d_cifar10", **kwargs) def shakeshakeresnet26_2x32d_cifar100(classes=100, **kwargs): """ Shake-Shake-ResNet-26-2x32d model for CIFAR-100 from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakeshakeresnet_cifar(classes=classes, blocks=26, bottleneck=False, first_stage_channels=32, model_name="shakeshakeresnet26_2x32d_cifar100", **kwargs) def shakeshakeresnet26_2x32d_svhn(classes=10, **kwargs): """ Shake-Shake-ResNet-26-2x32d model for SVHN from 'Shake-Shake regularization,' https://arxiv.org/abs/1705.07485. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shakeshakeresnet_cifar(classes=classes, blocks=26, bottleneck=False, first_stage_channels=32, model_name="shakeshakeresnet26_2x32d_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (shakeshakeresnet20_2x16d_cifar10, 10), (shakeshakeresnet20_2x16d_cifar100, 100), (shakeshakeresnet20_2x16d_svhn, 10), (shakeshakeresnet26_2x32d_cifar10, 10), (shakeshakeresnet26_2x32d_cifar100, 100), (shakeshakeresnet26_2x32d_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != shakeshakeresnet20_2x16d_cifar10 or weight_count == 541082) assert (model != shakeshakeresnet20_2x16d_cifar100 or weight_count == 546932) assert (model != shakeshakeresnet20_2x16d_svhn or weight_count == 541082) assert (model != shakeshakeresnet26_2x32d_cifar10 or weight_count == 2923162) assert (model != shakeshakeresnet26_2x32d_cifar100 or weight_count == 2934772) assert (model != shakeshakeresnet26_2x32d_svhn or weight_count == 2923162) x = torch.randn(14, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (14, num_classes)) if __name__ == "__main__": _test()
14,392
33.269048
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/sqnet.py
""" SQNet for image segmentation, implemented in PyTorch. Original paper: 'Speeding up Semantic Segmentation for Autonomous Driving,' https://openreview.net/pdf?id=S1uHiFyyg. """ __all__ = ['SQNet', 'sqnet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1_block, conv3x3_block, deconv3x3_block, Concurrent, Hourglass class FireBlock(nn.Module): """ SQNet specific encoder block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. 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, bias, use_bn, activation): super(FireBlock, self).__init__() squeeze_channels = out_channels // 8 expand_channels = out_channels // 2 self.conv = conv1x1_block( in_channels=in_channels, out_channels=squeeze_channels, bias=bias, use_bn=use_bn, activation=activation) self.branches = Concurrent(merge_type="cat") self.branches.add_module("branch1", conv1x1_block( in_channels=squeeze_channels, out_channels=expand_channels, bias=bias, use_bn=use_bn, activation=None)) self.branches.add_module("branch2", conv3x3_block( in_channels=squeeze_channels, out_channels=expand_channels, bias=bias, use_bn=use_bn, activation=None)) self.activ = nn.ELU(inplace=True) def forward(self, x): x = self.conv(x) x = self.branches(x) x = self.activ(x) return x class ParallelDilatedConv(nn.Module): """ SQNet specific decoder block (parallel dilated convolution). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. 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, bias, use_bn, activation): super(ParallelDilatedConv, self).__init__() dilations = [1, 2, 3, 4] self.branches = Concurrent(merge_type="sum") for i, dilation in enumerate(dilations): self.branches.add_module("branch{}".format(i + 1), conv3x3_block( in_channels=in_channels, out_channels=out_channels, padding=dilation, dilation=dilation, bias=bias, use_bn=use_bn, activation=activation)) def forward(self, x): x = self.branches(x) return x class SQNetUpStage(nn.Module): """ SQNet upscale stage. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. use_bn : bool Whether to use BatchNorm layer. activation : function or str or None Activation function or name of activation function. use_parallel_conv : bool Whether to use parallel dilated convolution. """ def __init__(self, in_channels, out_channels, bias, use_bn, activation, use_parallel_conv): super(SQNetUpStage, self).__init__() if use_parallel_conv: self.conv = ParallelDilatedConv( in_channels=in_channels, out_channels=in_channels, bias=bias, use_bn=use_bn, activation=activation) else: self.conv = conv3x3_block( in_channels=in_channels, out_channels=in_channels, bias=bias, use_bn=use_bn, activation=activation) self.deconv = deconv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, bias=bias, use_bn=use_bn, activation=activation) def forward(self, x): x = self.conv(x) x = self.deconv(x) return x class SQNet(nn.Module): """ SQNet model from 'Speeding up Semantic Segmentation for Autonomous Driving,' https://openreview.net/pdf?id=S1uHiFyyg. Parameters: ---------- channels : list of list of int Number of output channels for each stage in encoder and decoder. init_block_channels : int Number of output channels for the initial unit. layers : list of int Number of layers for each stage in encoder. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, channels, init_block_channels, layers, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(SQNet, self).__init__() 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.num_classes = num_classes self.fixed_size = fixed_size bias = True use_bn = False activation = (lambda: nn.ELU(inplace=True)) self.stem = conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2, bias=bias, use_bn=use_bn, activation=activation) in_channels = init_block_channels down_seq = nn.Sequential() skip_seq = nn.Sequential() for i, out_channels in enumerate(channels[0]): skip_seq.add_module("skip{}".format(i + 1), conv3x3_block( in_channels=in_channels, out_channels=in_channels, bias=bias, use_bn=use_bn, activation=activation)) stage = nn.Sequential() stage.add_module("unit1", nn.MaxPool2d( kernel_size=2, stride=2)) for j in range(layers[i]): stage.add_module("unit{}".format(j + 2), FireBlock( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=use_bn, activation=activation)) in_channels = out_channels down_seq.add_module("down{}".format(i + 1), stage) in_channels = in_channels // 2 up_seq = nn.Sequential() for i, out_channels in enumerate(channels[1]): use_parallel_conv = True if i == 0 else False up_seq.add_module("up{}".format(i + 1), SQNetUpStage( in_channels=(2 * in_channels), out_channels=out_channels, bias=bias, use_bn=use_bn, activation=activation, use_parallel_conv=use_parallel_conv)) in_channels = out_channels up_seq = up_seq[::-1] self.hg = Hourglass( down_seq=down_seq, up_seq=up_seq, skip_seq=skip_seq, merge_type="cat") self.head = SQNetUpStage( in_channels=(2 * in_channels), out_channels=num_classes, bias=bias, use_bn=use_bn, activation=activation, use_parallel_conv=False) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.stem(x) x = self.hg(x) x = self.head(x) return x def get_sqnet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SQNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [[128, 256, 512], [256, 128, 96]] init_block_channels = 96 layers = [2, 2, 3] net = SQNet( channels=channels, init_block_channels=init_block_channels, layers=layers, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def sqnet_cityscapes(num_classes=19, **kwargs): """ SQNet model for Cityscapes from 'Speeding up Semantic Segmentation for Autonomous Driving,' https://openreview.net/pdf?id=S1uHiFyyg. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sqnet(num_classes=num_classes, model_name="sqnet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ sqnet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sqnet_cityscapes or weight_count == 16262771) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
11,602
29.374346
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/wrn_cifar.py
""" WRN for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. """ __all__ = ['CIFARWRN', 'wrn16_10_cifar10', 'wrn16_10_cifar100', 'wrn16_10_svhn', 'wrn28_10_cifar10', 'wrn28_10_cifar100', 'wrn28_10_svhn', 'wrn40_8_cifar10', 'wrn40_8_cifar100', 'wrn40_8_svhn'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3 from .preresnet import PreResUnit, PreResActivation class CIFARWRN(nn.Module): """ WRN model for CIFAR 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARWRN, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", 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.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), PreResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=False, conv1_stride=False)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_wrn_cifar(num_classes, blocks, width_factor, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create WRN model for CIFAR with specific parameters. Parameters: ---------- num_classes : int Number of classification classes. blocks : int Number of blocks. width_factor : int 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert ((blocks - 4) % 6 == 0) layers = [(blocks - 4) // 6] * 3 channels_per_layers = [16, 32, 64] init_block_channels = 16 channels = [[ci * width_factor] * li for (ci, li) in zip(channels_per_layers, layers)] net = CIFARWRN( channels=channels, init_block_channels=init_block_channels, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def wrn16_10_cifar10(num_classes=10, **kwargs): """ WRN-16-10 model for CIFAR-10 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=16, width_factor=10, model_name="wrn16_10_cifar10", **kwargs) def wrn16_10_cifar100(num_classes=100, **kwargs): """ WRN-16-10 model for CIFAR-100 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=16, width_factor=10, model_name="wrn16_10_cifar100", **kwargs) def wrn16_10_svhn(num_classes=10, **kwargs): """ WRN-16-10 model for SVHN from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=16, width_factor=10, model_name="wrn16_10_svhn", **kwargs) def wrn28_10_cifar10(num_classes=10, **kwargs): """ WRN-28-10 model for CIFAR-10 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=28, width_factor=10, model_name="wrn28_10_cifar10", **kwargs) def wrn28_10_cifar100(num_classes=100, **kwargs): """ WRN-28-10 model for CIFAR-100 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=28, width_factor=10, model_name="wrn28_10_cifar100", **kwargs) def wrn28_10_svhn(num_classes=10, **kwargs): """ WRN-28-10 model for SVHN from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=28, width_factor=10, model_name="wrn28_10_svhn", **kwargs) def wrn40_8_cifar10(num_classes=10, **kwargs): """ WRN-40-8 model for CIFAR-10 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=40, width_factor=8, model_name="wrn40_8_cifar10", **kwargs) def wrn40_8_cifar100(num_classes=100, **kwargs): """ WRN-40-8 model for CIFAR-100 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=40, width_factor=8, model_name="wrn40_8_cifar100", **kwargs) def wrn40_8_svhn(num_classes=10, **kwargs): """ WRN-40-8 model for SVHN from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn_cifar(num_classes=num_classes, blocks=40, width_factor=8, model_name="wrn40_8_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (wrn16_10_cifar10, 10), (wrn16_10_cifar100, 100), (wrn16_10_svhn, 10), (wrn28_10_cifar10, 10), (wrn28_10_cifar100, 100), (wrn28_10_svhn, 10), (wrn40_8_cifar10, 10), (wrn40_8_cifar100, 100), (wrn40_8_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != wrn16_10_cifar10 or weight_count == 17116634) assert (model != wrn16_10_cifar100 or weight_count == 17174324) assert (model != wrn16_10_svhn or weight_count == 17116634) assert (model != wrn28_10_cifar10 or weight_count == 36479194) assert (model != wrn28_10_cifar100 or weight_count == 36536884) assert (model != wrn28_10_svhn or weight_count == 36479194) assert (model != wrn40_8_cifar10 or weight_count == 35748314) assert (model != wrn40_8_cifar100 or weight_count == 35794484) assert (model != wrn40_8_svhn or weight_count == 35748314) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
11,329
33.126506
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/inceptionresnetv2.py
""" InceptionResNetV2 for ImageNet-1K, implemented in PyTorch. Original paper: 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. """ __all__ = ['InceptionResNetV2', 'inceptionresnetv2'] import os import torch.nn as nn from .common import conv1x1_block, conv3x3_block, Concurrent from .inceptionv3 import AvgPoolBranch, Conv1x1Branch, ConvSeqBranch from .inceptionresnetv1 import InceptionAUnit, InceptionBUnit, InceptionCUnit, ReductionAUnit, ReductionBUnit class InceptBlock5b(nn.Module): """ InceptionResNetV2 type Mixed-5b block. Parameters: ---------- bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, bn_eps): super(InceptBlock5b, self).__init__() in_channels = 192 self.branches = Concurrent() self.branches.add_module("branch1", Conv1x1Branch( in_channels=in_channels, out_channels=96, bn_eps=bn_eps)) self.branches.add_module("branch2", ConvSeqBranch( in_channels=in_channels, out_channels_list=(48, 64), kernel_size_list=(1, 5), strides_list=(1, 1), padding_list=(0, 2), bn_eps=bn_eps)) self.branches.add_module("branch3", 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_eps=bn_eps)) self.branches.add_module("branch4", AvgPoolBranch( in_channels=in_channels, out_channels=64, bn_eps=bn_eps, count_include_pad=False)) def forward(self, x): x = self.branches(x) return x class InceptInitBlock(nn.Module): """ InceptionResNetV2 specific initial block. Parameters: ---------- in_channels : int Number of input channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, bn_eps): super(InceptInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=32, stride=2, padding=0, bn_eps=bn_eps) self.conv2 = conv3x3_block( in_channels=32, out_channels=32, stride=1, padding=0, bn_eps=bn_eps) self.conv3 = conv3x3_block( in_channels=32, out_channels=64, stride=1, padding=1, bn_eps=bn_eps) self.pool1 = nn.MaxPool2d( kernel_size=3, stride=2, padding=0) self.conv4 = conv1x1_block( in_channels=64, out_channels=80, stride=1, padding=0, bn_eps=bn_eps) self.conv5 = conv3x3_block( in_channels=80, out_channels=192, stride=1, padding=0, bn_eps=bn_eps) self.pool2 = nn.MaxPool2d( kernel_size=3, stride=2, padding=0) self.block = InceptBlock5b(bn_eps=bn_eps) def forward(self, 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) x = self.block(x) return x class InceptionResNetV2(nn.Module): """ InceptionResNetV2 model from 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. Parameters: ---------- dropout_rate : float, default 0.0 Fraction of the input units to drop. Must be a number between 0 and 1. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, dropout_rate=0.0, bn_eps=1e-5, in_channels=3, in_size=(299, 299), num_classes=1000): super(InceptionResNetV2, self).__init__() self.in_size = in_size self.num_classes = num_classes layers = [10, 21, 11] in_channels_list = [320, 1088, 2080] normal_out_channels_list = [[32, 32, 32, 32, 48, 64], [192, 128, 160, 192], [192, 192, 224, 256]] reduction_out_channels_list = [[384, 256, 256, 384], [256, 384, 256, 288, 256, 288, 320]] normal_units = [InceptionAUnit, InceptionBUnit, InceptionCUnit] reduction_units = [ReductionAUnit, ReductionBUnit] self.features = nn.Sequential() self.features.add_module("init_block", InceptInitBlock( in_channels=in_channels, bn_eps=bn_eps)) in_channels = in_channels_list[0] for i, layers_per_stage in enumerate(layers): stage = nn.Sequential() for j in range(layers_per_stage): if (j == 0) and (i != 0): unit = reduction_units[i - 1] out_channels_list_per_stage = reduction_out_channels_list[i - 1] else: unit = normal_units[i] out_channels_list_per_stage = normal_out_channels_list[i] if (i == len(layers) - 1) and (j == layers_per_stage - 1): unit_kwargs = {"scale": 1.0, "activate": False} else: unit_kwargs = {} stage.add_module("unit{}".format(j + 1), unit( in_channels=in_channels, out_channels_list=out_channels_list_per_stage, bn_eps=bn_eps, **unit_kwargs)) if (j == 0) and (i != 0): in_channels = in_channels_list[i] self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_conv", conv1x1_block( in_channels=in_channels, out_channels=1536, bn_eps=bn_eps)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Sequential() if dropout_rate > 0.0: self.output.add_module("dropout", nn.Dropout(p=dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=1536, out_features=num_classes)) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_inceptionresnetv2(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create InceptionResNetV2 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = InceptionResNetV2(**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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def inceptionresnetv2(**kwargs): """ InceptionResNetV2 model from 'Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning,' https://arxiv.org/abs/1602.07261. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_inceptionresnetv2(model_name="inceptionresnetv2", bn_eps=1e-3, **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ inceptionresnetv2, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != inceptionresnetv2 or weight_count == 55843464) x = torch.randn(1, 3, 299, 299) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
9,577
30.926667
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/ghostnet.py
""" GhostNet for ImageNet-1K, implemented in PyTorch. Original paper: 'GhostNet: More Features from Cheap Operations,' https://arxiv.org/abs/1911.11907. """ __all__ = ['GhostNet', 'ghostnet'] import os import math import torch import torch.nn as nn from .common import round_channels, conv1x1, conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block,\ dwsconv3x3_block, SEBlock class GhostHSigmoid(nn.Module): """ Approximated sigmoid function, specific for GhostNet. """ def forward(self, x): return torch.clamp(x, min=0.0, max=1.0) class GhostConvBlock(nn.Module): """ GhostNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, activation=(lambda: nn.ReLU(inplace=True))): super(GhostConvBlock, self).__init__() main_out_channels = math.ceil(0.5 * out_channels) cheap_out_channels = out_channels - main_out_channels self.main_conv = conv1x1_block( in_channels=in_channels, out_channels=main_out_channels, activation=activation) self.cheap_conv = dwconv3x3_block( in_channels=main_out_channels, out_channels=cheap_out_channels, activation=activation) def forward(self, x): x = self.main_conv(x) y = self.cheap_conv(x) return torch.cat((x, y), dim=1) class GhostExpBlock(nn.Module): """ GhostNet expansion block for residual path in GhostNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. use_kernel3 : bool Whether to use 3x3 (instead of 5x5) kernel. exp_factor : float Expansion factor. use_se : bool Whether to use SE-module. """ def __init__(self, in_channels, out_channels, stride, use_kernel3, exp_factor, use_se): super(GhostExpBlock, self).__init__() self.use_dw_conv = (stride != 1) self.use_se = use_se mid_channels = int(math.ceil(exp_factor * in_channels)) self.exp_conv = GhostConvBlock( in_channels=in_channels, out_channels=mid_channels) if self.use_dw_conv: dw_conv_class = dwconv3x3_block if use_kernel3 else dwconv5x5_block self.dw_conv = dw_conv_class( in_channels=mid_channels, out_channels=mid_channels, stride=stride, activation=None) if self.use_se: self.se = SEBlock( channels=mid_channels, reduction=4, out_activation=GhostHSigmoid()) self.pw_conv = GhostConvBlock( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.exp_conv(x) if self.use_dw_conv: x = self.dw_conv(x) if self.use_se: x = self.se(x) x = self.pw_conv(x) return x class GhostUnit(nn.Module): """ GhostNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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 : float Expansion factor. use_se : bool Whether to use SE-module. """ def __init__(self, in_channels, out_channels, stride, use_kernel3, exp_factor, use_se): super(GhostUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = GhostExpBlock( in_channels=in_channels, out_channels=out_channels, stride=stride, use_kernel3=use_kernel3, exp_factor=exp_factor, use_se=use_se) if self.resize_identity: self.identity_conv = dwsconv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride, pw_activation=None) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = x + identity return x class GhostClassifier(nn.Module): """ GhostNet classifier. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. mid_channels : int Number of middle channels. """ def __init__(self, in_channels, out_channels, mid_channels): super(GhostClassifier, self).__init__() self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class GhostNet(nn.Module): """ GhostNet model from 'GhostNet: More Features from Cheap Operations,' https://arxiv.org/abs/1911.11907. 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. 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. exp_factors : list of list of int Expansion factor 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, classifier_mid_channels, kernels3, exp_factors, use_se, first_stride, in_channels=3, in_size=(224, 224), num_classes=1000): super(GhostNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and ((i != 0) or first_stride) else 1 use_kernel3 = kernels3[i][j] == 1 exp_factor = exp_factors[i][j] use_se_flag = use_se[i][j] == 1 stage.add_module("unit{}".format(j + 1), GhostUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, use_kernel3=use_kernel3, exp_factor=exp_factor, use_se=use_se_flag)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = GhostClassifier( in_channels=in_channels, out_channels=num_classes, mid_channels=classifier_mid_channels) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_normal_(module.weight, mode="fan_out", nonlinearity="relu") if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_ghostnet(width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create GhostNet model with specific parameters. Parameters: ---------- 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 16 channels = [[16], [24, 24], [40, 40], [80, 80, 80, 80, 112, 112], [160, 160, 160, 160, 160]] kernels3 = [[1], [1, 1], [0, 0], [1, 1, 1, 1, 1, 1], [0, 0, 0, 0, 0]] exp_factors = [[1], [3, 3], [3, 3], [6, 2.5, 2.3, 2.3, 6, 6], [6, 6, 6, 6, 6]] use_se = [[0], [0, 0], [1, 1], [0, 0, 0, 0, 1, 1], [1, 0, 1, 0, 1]] final_block_channels = 960 classifier_mid_channels = 1280 first_stride = False if width_scale != 1.0: channels = [[round_channels(cij * width_scale, divisor=4) for cij in ci] for ci in channels] init_block_channels = round_channels(init_block_channels * width_scale, divisor=4) if width_scale > 1.0: final_block_channels = round_channels(final_block_channels * width_scale, divisor=4) net = GhostNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, classifier_mid_channels=classifier_mid_channels, kernels3=kernels3, exp_factors=exp_factors, use_se=use_se, first_stride=first_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def ghostnet(**kwargs): """ GhostNet model from 'GhostNet: More Features from Cheap Operations,' https://arxiv.org/abs/1911.11907. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ghostnet(model_name="ghostnet", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ ghostnet, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ghostnet or weight_count == 5180840) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,819
30.268293
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/efficientnet.py
""" EfficientNet for ImageNet-1K, implemented in PyTorch. Original papers: - 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946, - 'Adversarial Examples Improve Image Recognition,' https://arxiv.org/abs/1911.09665. """ __all__ = ['EfficientNet', 'calc_tf_padding', 'EffiInvResUnit', 'EffiInitBlock', 'efficientnet_b0', 'efficientnet_b1', 'efficientnet_b2', 'efficientnet_b3', 'efficientnet_b4', 'efficientnet_b5', 'efficientnet_b6', 'efficientnet_b7', 'efficientnet_b8', 'efficientnet_b0b', 'efficientnet_b1b', 'efficientnet_b2b', 'efficientnet_b3b', 'efficientnet_b4b', 'efficientnet_b5b', 'efficientnet_b6b', 'efficientnet_b7b', 'efficientnet_b0c', 'efficientnet_b1c', 'efficientnet_b2c', 'efficientnet_b3c', 'efficientnet_b4c', 'efficientnet_b5c', 'efficientnet_b6c', 'efficientnet_b7c', 'efficientnet_b8c'] import os import math import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from .common import round_channels, conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block, SEBlock def calc_tf_padding(x, kernel_size, stride=1, dilation=1): """ Calculate TF-same like padding size. Parameters: ---------- x : tensor Input tensor. kernel_size : int Convolution window size. stride : int, default 1 Strides of the convolution. dilation : int, default 1 Dilation value for convolution layer. Returns: ------- tuple of 4 int The size of the padding. """ height, width = x.size()[2:] oh = math.ceil(float(height) / stride) ow = math.ceil(float(width) / stride) pad_h = max((oh - 1) * stride + (kernel_size - 1) * dilation + 1 - height, 0) pad_w = max((ow - 1) * stride + (kernel_size - 1) * dilation + 1 - width, 0) return pad_h // 2, pad_h - pad_h // 2, pad_w // 2, pad_w - pad_w // 2 class EffiDwsConvUnit(nn.Module): """ EfficientNet specific depthwise separable convolution block/unit with BatchNorms and activations at each convolution layers. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the second convolution layer. bn_eps : float Small float added to variance in Batch norm. activation : str Name of activation function. tf_mode : bool Whether to use TF-like mode. """ def __init__(self, in_channels, out_channels, stride, bn_eps, activation, tf_mode): super(EffiDwsConvUnit, self).__init__() self.tf_mode = tf_mode self.residual = (in_channels == out_channels) and (stride == 1) self.dw_conv = dwconv3x3_block( in_channels=in_channels, out_channels=in_channels, padding=(0 if tf_mode else 1), bn_eps=bn_eps, activation=activation) self.se = SEBlock( channels=in_channels, reduction=4, mid_activation=activation) self.pw_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bn_eps=bn_eps, activation=None) def forward(self, x): if self.residual: identity = x if self.tf_mode: x = F.pad(x, pad=calc_tf_padding(x, kernel_size=3)) x = self.dw_conv(x) x = self.se(x) x = self.pw_conv(x) if self.residual: x = x + identity return x class EffiInvResUnit(nn.Module): """ EfficientNet inverted residual unit. 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. stride : int or tuple/list of 2 int Strides of the second convolution layer. exp_factor : int Factor for expansion of channels. se_factor : int SE reduction factor for each unit. bn_eps : float Small float added to variance in Batch norm. activation : str Name of activation function. tf_mode : bool Whether to use TF-like mode. """ def __init__(self, in_channels, out_channels, kernel_size, stride, exp_factor, se_factor, bn_eps, activation, tf_mode): super(EffiInvResUnit, self).__init__() self.kernel_size = kernel_size self.stride = stride self.tf_mode = tf_mode self.residual = (in_channels == out_channels) and (stride == 1) self.use_se = se_factor > 0 mid_channels = in_channels * exp_factor dwconv_block_fn = dwconv3x3_block if kernel_size == 3 else (dwconv5x5_block if kernel_size == 5 else None) self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_eps=bn_eps, activation=activation) self.conv2 = dwconv_block_fn( in_channels=mid_channels, out_channels=mid_channels, stride=stride, padding=(0 if tf_mode else (kernel_size // 2)), bn_eps=bn_eps, activation=activation) if self.use_se: self.se = SEBlock( channels=mid_channels, reduction=(exp_factor * se_factor), mid_activation=activation) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_eps=bn_eps, activation=None) def forward(self, x): if self.residual: identity = x x = self.conv1(x) if self.tf_mode: x = F.pad(x, pad=calc_tf_padding(x, kernel_size=self.kernel_size, stride=self.stride)) x = self.conv2(x) if self.use_se: x = self.se(x) x = self.conv3(x) if self.residual: x = x + identity return x class EffiInitBlock(nn.Module): """ EfficientNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. activation : str Name of activation function. tf_mode : bool Whether to use TF-like mode. """ def __init__(self, in_channels, out_channels, bn_eps, activation, tf_mode): super(EffiInitBlock, self).__init__() self.tf_mode = tf_mode self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, padding=(0 if tf_mode else 1), bn_eps=bn_eps, activation=activation) def forward(self, x): if self.tf_mode: x = F.pad(x, pad=calc_tf_padding(x, kernel_size=3, stride=2)) x = self.conv(x) return x class EfficientNet(nn.Module): """ EfficientNet model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for initial unit. final_block_channels : int Number of output channels for the final block of the feature extractor. kernel_sizes : list of list of int Number of kernel sizes for each unit. strides_per_stage : list int Stride value for the first unit of each stage. expansion_factors : list of list of int Number of expansion factors for each unit. dropout_rate : float, default 0.2 Fraction of the input units to drop. Must be a number between 0 and 1. tf_mode : bool, default False Whether to use TF-like mode. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, kernel_sizes, strides_per_stage, expansion_factors, dropout_rate=0.2, tf_mode=False, bn_eps=1e-5, in_channels=3, in_size=(224, 224), num_classes=1000): super(EfficientNet, self).__init__() self.in_size = in_size self.num_classes = num_classes activation = "swish" self.features = nn.Sequential() self.features.add_module("init_block", EffiInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_eps=bn_eps, activation=activation, tf_mode=tf_mode)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): kernel_sizes_per_stage = kernel_sizes[i] expansion_factors_per_stage = expansion_factors[i] stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): kernel_size = kernel_sizes_per_stage[j] expansion_factor = expansion_factors_per_stage[j] stride = strides_per_stage[i] if (j == 0) else 1 if i == 0: stage.add_module("unit{}".format(j + 1), EffiDwsConvUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bn_eps=bn_eps, activation=activation, tf_mode=tf_mode)) else: stage.add_module("unit{}".format(j + 1), EffiInvResUnit( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, exp_factor=expansion_factor, se_factor=4, bn_eps=bn_eps, activation=activation, tf_mode=tf_mode)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, bn_eps=bn_eps, activation=activation)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Sequential() if dropout_rate > 0.0: self.output.add_module("dropout", nn.Dropout(p=dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_efficientnet(version, in_size, tf_mode=False, bn_eps=1e-5, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create EfficientNet model with specific parameters. Parameters: ---------- version : str Version of EfficientNet ('b0'...'b8'). in_size : tuple of two ints Spatial size of the expected input image. tf_mode : bool, default False Whether to use TF-like mode. bn_eps : 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if version == "b0": assert (in_size == (224, 224)) depth_factor = 1.0 width_factor = 1.0 dropout_rate = 0.2 elif version == "b1": assert (in_size == (240, 240)) depth_factor = 1.1 width_factor = 1.0 dropout_rate = 0.2 elif version == "b2": assert (in_size == (260, 260)) depth_factor = 1.2 width_factor = 1.1 dropout_rate = 0.3 elif version == "b3": assert (in_size == (300, 300)) depth_factor = 1.4 width_factor = 1.2 dropout_rate = 0.3 elif version == "b4": assert (in_size == (380, 380)) depth_factor = 1.8 width_factor = 1.4 dropout_rate = 0.4 elif version == "b5": assert (in_size == (456, 456)) depth_factor = 2.2 width_factor = 1.6 dropout_rate = 0.4 elif version == "b6": assert (in_size == (528, 528)) depth_factor = 2.6 width_factor = 1.8 dropout_rate = 0.5 elif version == "b7": assert (in_size == (600, 600)) depth_factor = 3.1 width_factor = 2.0 dropout_rate = 0.5 elif version == "b8": assert (in_size == (672, 672)) depth_factor = 3.6 width_factor = 2.2 dropout_rate = 0.5 else: raise ValueError("Unsupported EfficientNet version {}".format(version)) init_block_channels = 32 layers = [1, 2, 2, 3, 3, 4, 1] downsample = [1, 1, 1, 1, 0, 1, 0] channels_per_layers = [16, 24, 40, 80, 112, 192, 320] expansion_factors_per_layers = [1, 6, 6, 6, 6, 6, 6] kernel_sizes_per_layers = [3, 3, 5, 3, 5, 5, 3] strides_per_stage = [1, 2, 2, 2, 1, 2, 1] final_block_channels = 1280 layers = [int(math.ceil(li * depth_factor)) for li in layers] channels_per_layers = [round_channels(ci * width_factor) for ci in channels_per_layers] from functools import reduce channels = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(channels_per_layers, layers, downsample), []) kernel_sizes = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(kernel_sizes_per_layers, layers, downsample), []) expansion_factors = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(expansion_factors_per_layers, layers, downsample), []) strides_per_stage = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(strides_per_stage, layers, downsample), []) strides_per_stage = [si[0] for si in strides_per_stage] init_block_channels = round_channels(init_block_channels * width_factor) if width_factor > 1.0: assert (int(final_block_channels * width_factor) == round_channels(final_block_channels * width_factor)) final_block_channels = round_channels(final_block_channels * width_factor) net = EfficientNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, kernel_sizes=kernel_sizes, strides_per_stage=strides_per_stage, expansion_factors=expansion_factors, dropout_rate=dropout_rate, tf_mode=tf_mode, bn_eps=bn_eps, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def efficientnet_b0(in_size=(224, 224), **kwargs): """ EfficientNet-B0 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b0", in_size=in_size, model_name="efficientnet_b0", **kwargs) def efficientnet_b1(in_size=(240, 240), **kwargs): """ EfficientNet-B1 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (240, 240) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b1", in_size=in_size, model_name="efficientnet_b1", **kwargs) def efficientnet_b2(in_size=(260, 260), **kwargs): """ EfficientNet-B2 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (260, 260) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b2", in_size=in_size, model_name="efficientnet_b2", **kwargs) def efficientnet_b3(in_size=(300, 300), **kwargs): """ EfficientNet-B3 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (300, 300) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b3", in_size=in_size, model_name="efficientnet_b3", **kwargs) def efficientnet_b4(in_size=(380, 380), **kwargs): """ EfficientNet-B4 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (380, 380) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b4", in_size=in_size, model_name="efficientnet_b4", **kwargs) def efficientnet_b5(in_size=(456, 456), **kwargs): """ EfficientNet-B5 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (456, 456) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b5", in_size=in_size, model_name="efficientnet_b5", **kwargs) def efficientnet_b6(in_size=(528, 528), **kwargs): """ EfficientNet-B6 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (528, 528) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b6", in_size=in_size, model_name="efficientnet_b6", **kwargs) def efficientnet_b7(in_size=(600, 600), **kwargs): """ EfficientNet-B7 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (600, 600) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b7", in_size=in_size, model_name="efficientnet_b7", **kwargs) def efficientnet_b8(in_size=(672, 672), **kwargs): """ EfficientNet-B8 model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (672, 672) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b8", in_size=in_size, model_name="efficientnet_b8", **kwargs) def efficientnet_b0b(in_size=(224, 224), **kwargs): """ EfficientNet-B0-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b0", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b0b", **kwargs) def efficientnet_b1b(in_size=(240, 240), **kwargs): """ EfficientNet-B1-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (240, 240) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b1", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b1b", **kwargs) def efficientnet_b2b(in_size=(260, 260), **kwargs): """ EfficientNet-B2-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (260, 260) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b2", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b2b", **kwargs) def efficientnet_b3b(in_size=(300, 300), **kwargs): """ EfficientNet-B3-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (300, 300) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b3", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b3b", **kwargs) def efficientnet_b4b(in_size=(380, 380), **kwargs): """ EfficientNet-B4-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (380, 380) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b4", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b4b", **kwargs) def efficientnet_b5b(in_size=(456, 456), **kwargs): """ EfficientNet-B5-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (456, 456) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b5", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b5b", **kwargs) def efficientnet_b6b(in_size=(528, 528), **kwargs): """ EfficientNet-B6-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (528, 528) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b6", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b6b", **kwargs) def efficientnet_b7b(in_size=(600, 600), **kwargs): """ EfficientNet-B7-b (like TF-implementation) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (600, 600) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b7", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b7b", **kwargs) def efficientnet_b0c(in_size=(224, 224), **kwargs): """ EfficientNet-B0-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b0", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b0c", **kwargs) def efficientnet_b1c(in_size=(240, 240), **kwargs): """ EfficientNet-B1-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (240, 240) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b1", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b1c", **kwargs) def efficientnet_b2c(in_size=(260, 260), **kwargs): """ EfficientNet-B2-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (260, 260) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b2", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b2c", **kwargs) def efficientnet_b3c(in_size=(300, 300), **kwargs): """ EfficientNet-B3-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (300, 300) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b3", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b3c", **kwargs) def efficientnet_b4c(in_size=(380, 380), **kwargs): """ EfficientNet-B4-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (380, 380) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b4", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b4c", **kwargs) def efficientnet_b5c(in_size=(456, 456), **kwargs): """ EfficientNet-B5-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (456, 456) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b5", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b5c", **kwargs) def efficientnet_b6c(in_size=(528, 528), **kwargs): """ EfficientNet-B6-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (528, 528) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b6", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b6c", **kwargs) def efficientnet_b7c(in_size=(600, 600), **kwargs): """ EfficientNet-B7-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (600, 600) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b7", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b7c", **kwargs) def efficientnet_b8c(in_size=(672, 672), **kwargs): """ EfficientNet-B8-c (like TF-implementation, trained with AdvProp) model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (672, 672) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet(version="b8", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_b8c", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ efficientnet_b0, efficientnet_b1, efficientnet_b2, efficientnet_b3, efficientnet_b4, efficientnet_b5, efficientnet_b6, efficientnet_b7, efficientnet_b8, efficientnet_b0b, efficientnet_b1b, efficientnet_b2b, efficientnet_b3b, efficientnet_b4b, efficientnet_b5b, efficientnet_b6b, efficientnet_b7b, efficientnet_b0c, efficientnet_b1c, efficientnet_b2c, efficientnet_b3c, efficientnet_b4c, efficientnet_b5c, efficientnet_b6c, efficientnet_b7c, efficientnet_b8c, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != efficientnet_b0 or weight_count == 5288548) assert (model != efficientnet_b1 or weight_count == 7794184) assert (model != efficientnet_b2 or weight_count == 9109994) assert (model != efficientnet_b3 or weight_count == 12233232) assert (model != efficientnet_b4 or weight_count == 19341616) assert (model != efficientnet_b5 or weight_count == 30389784) assert (model != efficientnet_b6 or weight_count == 43040704) assert (model != efficientnet_b7 or weight_count == 66347960) assert (model != efficientnet_b8 or weight_count == 87413142) assert (model != efficientnet_b0b or weight_count == 5288548) assert (model != efficientnet_b1b or weight_count == 7794184) assert (model != efficientnet_b2b or weight_count == 9109994) assert (model != efficientnet_b3b or weight_count == 12233232) assert (model != efficientnet_b4b or weight_count == 19341616) assert (model != efficientnet_b5b or weight_count == 30389784) assert (model != efficientnet_b6b or weight_count == 43040704) assert (model != efficientnet_b7b or weight_count == 66347960) x = torch.randn(1, 3, net.in_size[0], net.in_size[1]) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
37,745
35.933464
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/edanet.py
""" EDANet for image segmentation, implemented in PyTorch. Original paper: 'Efficient Dense Modules of Asymmetric Convolution for Real-Time Semantic Segmentation,' https://arxiv.org/abs/1809.06323. """ __all__ = ['EDANet', 'edanet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1, conv3x3, conv1x1_block, asym_conv3x3_block, NormActivation, InterpolationBlock class DownBlock(nn.Module): """ EDANet specific downsample block for the main branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, bn_eps): super(DownBlock, self).__init__() self.expand = (in_channels < out_channels) mid_channels = out_channels - in_channels if self.expand else out_channels self.conv = conv3x3( in_channels=in_channels, out_channels=mid_channels, bias=True, stride=2) if self.expand: self.pool = nn.MaxPool2d( kernel_size=2, stride=2) self.norm_activ = NormActivation( in_channels=out_channels, bn_eps=bn_eps) def forward(self, x): y = self.conv(x) if self.expand: z = self.pool(x) y = torch.cat((y, z), dim=1) y = self.norm_activ(y) return y class EDABlock(nn.Module): """ EDANet base block. Parameters: ---------- channels : int Number of input/output channels. dilation : int Dilation value for convolution layer. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, channels, dilation, dropout_rate, bn_eps): super(EDABlock, self).__init__() self.use_dropout = (dropout_rate != 0.0) self.conv1 = asym_conv3x3_block( channels=channels, bias=True, lw_use_bn=False, bn_eps=bn_eps, lw_activation=None) self.conv2 = asym_conv3x3_block( channels=channels, padding=dilation, dilation=dilation, bias=True, lw_use_bn=False, bn_eps=bn_eps, rw_activation=None) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): x = self.conv1(x) x = self.conv2(x) if self.use_dropout: x = self.dropout(x) return x class EDAUnit(nn.Module): """ EDANet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. dilation : int Dilation value for convolution layer. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, dilation, dropout_rate, bn_eps): super(EDAUnit, self).__init__() self.use_dropout = (dropout_rate != 0.0) mid_channels = out_channels - in_channels self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bias=True) self.conv2 = EDABlock( channels=mid_channels, dilation=dilation, dropout_rate=dropout_rate, bn_eps=bn_eps) self.activ = nn.ReLU(inplace=True) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) x = torch.cat((x, identity), dim=1) x = self.activ(x) return x class EDANet(nn.Module): """ EDANet model from 'Efficient Dense Modules of Asymmetric Convolution for Real-Time Semantic Segmentation,' https://arxiv.org/abs/1809.06323. Parameters: ---------- channels : list of int Number of output channels for the first unit of each stage. dilations : list of list of int Dilations for blocks. growth_rate : int Growth rate for numbers of output channels for each non-first unit. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, channels, dilations, growth_rate, bn_eps=1e-5, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(EDANet, self).__init__() 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.num_classes = num_classes self.fixed_size = fixed_size dropout_rate = 0.02 self.features = nn.Sequential() for i, dilations_per_stage in enumerate(dilations): out_channels = channels[i] stage = nn.Sequential() for j, dilation in enumerate(dilations_per_stage): if j == 0: stage.add_module("unit{}".format(j + 1), DownBlock( in_channels=in_channels, out_channels=out_channels, bn_eps=bn_eps)) else: out_channels += growth_rate stage.add_module("unit{}".format(j + 1), EDAUnit( in_channels=in_channels, out_channels=out_channels, dilation=dilation, dropout_rate=dropout_rate, bn_eps=bn_eps)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.head = conv1x1( in_channels=in_channels, out_channels=num_classes, bias=True) self.up = InterpolationBlock( scale_factor=8, align_corners=True) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.head(x) x = self.up(x) return x def get_edanet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create EDANet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [15, 60, 130, 450] dilations = [[0], [0, 1, 1, 1, 2, 2], [0, 2, 2, 4, 4, 8, 8, 16, 16]] growth_rate = 40 bn_eps = 1e-3 net = EDANet( channels=channels, dilations=dilations, growth_rate=growth_rate, bn_eps=bn_eps, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def edanet_cityscapes(num_classes=19, **kwargs): """ EDANet model for Cityscapes from 'Efficient Dense Modules of Asymmetric Convolution for Real-Time Semantic Segmentation,' https://arxiv.org/abs/1809.06323. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_edanet(num_classes=num_classes, model_name="edanet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ edanet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != edanet_cityscapes or weight_count == 689485) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
10,158
28.618076
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/channelnet.py
""" ChannelNet for ImageNet-1K, implemented in PyTorch. Original paper: 'ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions,' https://arxiv.org/abs/1809.01330. """ __all__ = ['ChannelNet', 'channelnet'] import os import torch import torch.nn as nn import torch.nn.init as init def dwconv3x3(in_channels, out_channels, stride, bias=False): """ 3x3 depthwise version of the standard convolution layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=1, groups=out_channels, bias=bias) class ChannetConv(nn.Module): """ ChannelNet specific convolution block with Batch normalization and 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. dropout_rate : float, default 0.0 Dropout rate. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False, dropout_rate=0.0, activate=True): super(ChannetConv, self).__init__() self.use_dropout = (dropout_rate > 0.0) self.activate = activate self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) self.bn = nn.BatchNorm2d(num_features=out_channels) if self.activate: self.activ = nn.ReLU6(inplace=True) def forward(self, x): x = self.conv(x) if self.use_dropout: x = self.dropout(x) x = self.bn(x) if self.activate: x = self.activ(x) return x def channet_conv1x1(in_channels, out_channels, stride=1, groups=1, bias=False, dropout_rate=0.0, activate=True): """ 1x1 version of ChannelNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. dropout_rate : float, default 0.0 Dropout rate. activate : bool, default True Whether activate the convolution block. """ return ChannetConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, groups=groups, bias=bias, dropout_rate=dropout_rate, activate=activate) def channet_conv3x3(in_channels, out_channels, stride, padding=1, dilation=1, groups=1, bias=False, dropout_rate=0.0, activate=True): """ 3x3 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. dropout_rate : float, default 0.0 Dropout rate. activate : bool, default True Whether activate the convolution block. """ return ChannetConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, dropout_rate=dropout_rate, activate=activate) class ChannetDwsConvBlock(nn.Module): """ ChannelNet specific depthwise separable convolution block with BatchNorms and activations at last convolution layers. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. groups : int, default 1 Number of groups. dropout_rate : float, default 0.0 Dropout rate. """ def __init__(self, in_channels, out_channels, stride, groups=1, dropout_rate=0.0): super(ChannetDwsConvBlock, self).__init__() self.dw_conv = dwconv3x3( in_channels=in_channels, out_channels=in_channels, stride=stride) self.pw_conv = channet_conv1x1( in_channels=in_channels, out_channels=out_channels, groups=groups, dropout_rate=dropout_rate) def forward(self, x): x = self.dw_conv(x) x = self.pw_conv(x) return x class SimpleGroupBlock(nn.Module): """ ChannelNet specific block with a sequence of depthwise separable group convolution layers. Parameters: ---------- channels : int Number of input/output channels. multi_blocks : int Number of DWS layers in the sequence. groups : int Number of groups. dropout_rate : float Dropout rate. """ def __init__(self, channels, multi_blocks, groups, dropout_rate): super(SimpleGroupBlock, self).__init__() self.blocks = nn.Sequential() for i in range(multi_blocks): self.blocks.add_module("block{}".format(i + 1), ChannetDwsConvBlock( in_channels=channels, out_channels=channels, stride=1, groups=groups, dropout_rate=dropout_rate)) def forward(self, x): x = self.blocks(x) return x class ChannelwiseConv2d(nn.Module): """ ChannelNet specific block with channel-wise convolution. Parameters: ---------- groups : int Number of groups. dropout_rate : float Dropout rate. """ def __init__(self, groups, dropout_rate): super(ChannelwiseConv2d, self).__init__() self.use_dropout = (dropout_rate > 0.0) self.conv = nn.Conv3d( in_channels=1, out_channels=groups, kernel_size=(4 * groups, 1, 1), stride=(groups, 1, 1), padding=(2 * groups - 1, 0, 0), bias=False) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): batch, channels, height, width = x.size() x = x.unsqueeze(dim=1) x = self.conv(x) if self.use_dropout: x = self.dropout(x) x = x.view(batch, channels, height, width) return x class ConvGroupBlock(nn.Module): """ ChannelNet specific block with a combination of channel-wise convolution, depthwise separable group convolutions. Parameters: ---------- channels : int Number of input/output channels. multi_blocks : int Number of DWS layers in the sequence. groups : int Number of groups. dropout_rate : float Dropout rate. """ def __init__(self, channels, multi_blocks, groups, dropout_rate): super(ConvGroupBlock, self).__init__() self.conv = ChannelwiseConv2d( groups=groups, dropout_rate=dropout_rate) self.block = SimpleGroupBlock( channels=channels, multi_blocks=multi_blocks, groups=groups, dropout_rate=dropout_rate) def forward(self, x): x = self.conv(x) x = self.block(x) return x class ChannetUnit(nn.Module): """ ChannelNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels_list : tuple/list of 2 int Number of output channels for each sub-block. strides : int or tuple/list of 2 int Strides of the convolution. multi_blocks : int Number of DWS layers in the sequence. groups : int Number of groups. dropout_rate : float Dropout rate. block_names : tuple/list of 2 str Sub-block names. merge_type : str Type of sub-block output merging. """ def __init__(self, in_channels, out_channels_list, strides, multi_blocks, groups, dropout_rate, block_names, merge_type): super(ChannetUnit, self).__init__() assert (len(block_names) == 2) assert (merge_type in ["seq", "add", "cat"]) self.merge_type = merge_type self.blocks = nn.Sequential() for i, (out_channels, block_name) in enumerate(zip(out_channels_list, block_names)): stride_i = (strides if i == 0 else 1) if block_name == "channet_conv3x3": self.blocks.add_module("block{}".format(i + 1), channet_conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride_i, dropout_rate=dropout_rate, activate=False)) elif block_name == "channet_dws_conv_block": self.blocks.add_module("block{}".format(i + 1), ChannetDwsConvBlock( in_channels=in_channels, out_channels=out_channels, stride=stride_i, dropout_rate=dropout_rate)) elif block_name == "simple_group_block": self.blocks.add_module("block{}".format(i + 1), SimpleGroupBlock( channels=in_channels, multi_blocks=multi_blocks, groups=groups, dropout_rate=dropout_rate)) elif block_name == "conv_group_block": self.blocks.add_module("block{}".format(i + 1), ConvGroupBlock( channels=in_channels, multi_blocks=multi_blocks, groups=groups, dropout_rate=dropout_rate)) else: raise NotImplementedError() in_channels = out_channels def forward(self, x): x_outs = [] for block in self.blocks._modules.values(): x = block(x) x_outs.append(x) if self.merge_type == "add": for i in range(len(x_outs) - 1): x = x + x_outs[i] elif self.merge_type == "cat": x = torch.cat(tuple(x_outs), dim=1) return x class ChannelNet(nn.Module): """ ChannelNet model from 'ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions,' https://arxiv.org/abs/1809.01330. Parameters: ---------- channels : list of list of list of int Number of output channels for each unit. block_names : list of list of list of str Names of blocks for each unit. block_names : list of list of str Merge types for each unit. dropout_rate : float, default 0.0001 Dropout rate. multi_blocks : int, default 2 Block count architectural parameter. groups : int, default 2 Group count architectural parameter. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, block_names, merge_types, dropout_rate=0.0001, multi_blocks=2, groups=2, in_channels=3, in_size=(224, 224), num_classes=1000): super(ChannelNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): strides = 2 if (j == 0) else 1 stage.add_module("unit{}".format(j + 1), ChannetUnit( in_channels=in_channels, out_channels_list=out_channels, strides=strides, multi_blocks=multi_blocks, groups=groups, dropout_rate=dropout_rate, block_names=block_names[i][j], merge_type=merge_types[i][j])) if merge_types[i][j] == "cat": in_channels = sum(out_channels) else: in_channels = out_channels[-1] self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_channelnet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create ChannelNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [[[32, 64]], [[128, 128]], [[256, 256]], [[512, 512], [512, 512]], [[1024, 1024]]] block_names = [[["channet_conv3x3", "channet_dws_conv_block"]], [["channet_dws_conv_block", "channet_dws_conv_block"]], [["channet_dws_conv_block", "channet_dws_conv_block"]], [["channet_dws_conv_block", "simple_group_block"], ["conv_group_block", "conv_group_block"]], [["channet_dws_conv_block", "channet_dws_conv_block"]]] merge_types = [["cat"], ["cat"], ["cat"], ["add", "add"], ["seq"]] net = ChannelNet( channels=channels, block_names=block_names, merge_types=merge_types, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def channelnet(**kwargs): """ ChannelNet model from 'ChannelNets: Compact and Efficient Convolutional Neural Networks via Channel-Wise Convolutions,' https://arxiv.org/abs/1809.01330. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_channelnet(model_name="channelnet", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ channelnet, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != channelnet or weight_count == 3875112) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
18,471
29.633499
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/pnasnet.py
""" PNASNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Progressive Neural Architecture Search,' https://arxiv.org/abs/1712.00559. """ __all__ = ['PNASNet', 'pnasnet5large'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1 from .nasnet import nasnet_dual_path_sequential, nasnet_batch_norm, NasConv, NasDwsConv, NasPathBlock, NASNetInitBlock class PnasMaxPoolBlock(nn.Module): """ PNASNet specific Max pooling layer with extra padding. Parameters: ---------- stride : int or tuple/list of 2 int, default 2 Strides of the convolution. extra_padding : bool, default False Whether to use extra padding. """ def __init__(self, stride=2, extra_padding=False): super(PnasMaxPoolBlock, self).__init__() self.extra_padding = extra_padding self.pool = nn.MaxPool2d( kernel_size=3, stride=stride, padding=1) if self.extra_padding: self.pad = nn.ZeroPad2d(padding=(1, 0, 1, 0)) def forward(self, x): if self.extra_padding: x = self.pad(x) x = self.pool(x) if self.extra_padding: x = x[:, :, 1:, 1:].contiguous() return x def pnas_conv1x1(in_channels, out_channels, stride=1): """ 1x1 version of the PNASNet specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. """ return NasConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, groups=1) class DwsBranch(nn.Module): """ PNASNet 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. stride : int or tuple/list of 2 int Strides of the convolution. 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, stride, extra_padding=False, stem=False): super(DwsBranch, self).__init__() assert (not stem) or (not extra_padding) mid_channels = out_channels if stem else in_channels padding = kernel_size // 2 self.conv1 = NasDwsConv( in_channels=in_channels, out_channels=mid_channels, kernel_size=kernel_size, stride=stride, padding=padding, extra_padding=extra_padding) self.conv2 = NasDwsConv( in_channels=mid_channels, out_channels=out_channels, kernel_size=kernel_size, stride=1, padding=padding) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x def dws_branch_k3(in_channels, out_channels, stride=2, extra_padding=False, stem=False): """ 3x3 version of the PNASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 2 Strides of the convolution. 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=3, stride=stride, extra_padding=extra_padding, stem=stem) def dws_branch_k5(in_channels, out_channels, stride=2, extra_padding=False, stem=False): """ 5x5 version of the PNASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 2 Strides of the convolution. 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, stride=stride, extra_padding=extra_padding, stem=stem) def dws_branch_k7(in_channels, out_channels, stride=2, extra_padding=False): """ 7x7 version of the PNASNet specific depthwise separable convolution branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 2 Strides of the convolution. extra_padding : bool, default False Whether to use extra padding. """ return DwsBranch( in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=stride, extra_padding=extra_padding, stem=False) class PnasMaxPathBlock(nn.Module): """ PNASNet specific `max path` auxiliary block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(PnasMaxPathBlock, self).__init__() self.maxpool = PnasMaxPoolBlock() self.conv = conv1x1( in_channels=in_channels, out_channels=out_channels) self.bn = nasnet_batch_norm(channels=out_channels) def forward(self, x): x = self.maxpool(x) x = self.conv(x) x = self.bn(x) return x class PnasBaseUnit(nn.Module): """ PNASNet base unit. """ def __init__(self): super(PnasBaseUnit, self).__init__() def cell_forward(self, x, x_prev): assert (hasattr(self, 'comb0_left')) x_left = x_prev x_right = x x0 = self.comb0_left(x_left) + self.comb0_right(x_left) x1 = self.comb1_left(x_right) + self.comb1_right(x_right) x2 = self.comb2_left(x_right) + self.comb2_right(x_right) x3 = self.comb3_left(x2) + self.comb3_right(x_right) x4 = self.comb4_left(x_left) + (self.comb4_right(x_right) if self.comb4_right else x_right) x_out = torch.cat((x0, x1, x2, x3, x4), dim=1) return x_out class Stem1Unit(PnasBaseUnit): """ PNASNet Stem1 unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(Stem1Unit, self).__init__() mid_channels = out_channels // 5 self.conv_1x1 = pnas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5( in_channels=in_channels, out_channels=mid_channels, stem=True) self.comb0_right = PnasMaxPathBlock( in_channels=in_channels, out_channels=mid_channels) self.comb1_left = dws_branch_k7( in_channels=mid_channels, out_channels=mid_channels) self.comb1_right = PnasMaxPoolBlock() self.comb2_left = dws_branch_k5( in_channels=mid_channels, out_channels=mid_channels) self.comb2_right = dws_branch_k3( in_channels=mid_channels, out_channels=mid_channels) self.comb3_left = dws_branch_k3( in_channels=mid_channels, out_channels=mid_channels, stride=1) self.comb3_right = PnasMaxPoolBlock() self.comb4_left = dws_branch_k3( in_channels=in_channels, out_channels=mid_channels, stem=True) self.comb4_right = pnas_conv1x1( in_channels=mid_channels, out_channels=mid_channels, stride=2) def forward(self, x): x_prev = x x = self.conv_1x1(x) x_out = self.cell_forward(x, x_prev) return x_out class PnasUnit(PnasBaseUnit): """ PNASNet ordinary 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. reduction : bool, default False Whether to use reduction. extra_padding : bool, default False Whether to use extra padding. match_prev_layer_dimensions : bool, default False Whether to match previous layer dimensions. """ def __init__(self, in_channels, prev_in_channels, out_channels, reduction=False, extra_padding=False, match_prev_layer_dimensions=False): super(PnasUnit, self).__init__() mid_channels = out_channels // 5 stride = 2 if reduction else 1 if match_prev_layer_dimensions: self.conv_prev_1x1 = NasPathBlock( in_channels=prev_in_channels, out_channels=mid_channels) else: self.conv_prev_1x1 = pnas_conv1x1( in_channels=prev_in_channels, out_channels=mid_channels) self.conv_1x1 = pnas_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.comb0_left = dws_branch_k5( in_channels=mid_channels, out_channels=mid_channels, stride=stride, extra_padding=extra_padding) self.comb0_right = PnasMaxPoolBlock( stride=stride, extra_padding=extra_padding) self.comb1_left = dws_branch_k7( in_channels=mid_channels, out_channels=mid_channels, stride=stride, extra_padding=extra_padding) self.comb1_right = PnasMaxPoolBlock( stride=stride, extra_padding=extra_padding) self.comb2_left = dws_branch_k5( in_channels=mid_channels, out_channels=mid_channels, stride=stride, extra_padding=extra_padding) self.comb2_right = dws_branch_k3( in_channels=mid_channels, out_channels=mid_channels, stride=stride, extra_padding=extra_padding) self.comb3_left = dws_branch_k3( in_channels=mid_channels, out_channels=mid_channels, stride=1) self.comb3_right = PnasMaxPoolBlock( stride=stride, extra_padding=extra_padding) self.comb4_left = dws_branch_k3( in_channels=mid_channels, out_channels=mid_channels, stride=stride, extra_padding=extra_padding) if reduction: self.comb4_right = pnas_conv1x1( in_channels=mid_channels, out_channels=mid_channels, stride=stride) else: self.comb4_right = None def forward(self, x, x_prev): # print("x.shape={}, x_prev.shape={}".format(x.shape, x_prev.shape)) x_prev = self.conv_prev_1x1(x_prev) x = self.conv_1x1(x) x_out = self.cell_forward(x, x_prev) return x_out class PNASNet(nn.Module): """ PNASNet model from 'Progressive Neural Architecture Search,' https://arxiv.org/abs/1712.00559. 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. stem1_blocks_channels : list of 2 int Number of output channels for the Stem1 unit. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, stem1_blocks_channels, in_channels=3, in_size=(331, 331), num_classes=1000): super(PNASNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nasnet_dual_path_sequential( return_two=False, first_ordinals=2, last_ordinals=2) self.features.add_module("init_block", NASNetInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels self.features.add_module("stem1_unit", Stem1Unit( in_channels=in_channels, out_channels=stem1_blocks_channels)) prev_in_channels = in_channels in_channels = stem1_blocks_channels for i, channels_per_stage in enumerate(channels): stage = nasnet_dual_path_sequential() for j, out_channels in enumerate(channels_per_stage): reduction = (j == 0) extra_padding = (j == 0) and (i not in [0, 2]) match_prev_layer_dimensions = (j == 1) or ((j == 0) and (i == 0)) stage.add_module("unit{}".format(j + 1), PnasUnit( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels, reduction=reduction, extra_padding=extra_padding, match_prev_layer_dimensions=match_prev_layer_dimensions)) prev_in_channels = in_channels in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("activ", nn.ReLU()) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=11, stride=1)) self.output = nn.Sequential() self.output.add_module("dropout", nn.Dropout(p=0.5)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_pnasnet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create PNASNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ repeat = 4 init_block_channels = 96 stem_blocks_channels = [270, 540] norm_channels = [1080, 2160, 4320] channels = [[ci] * repeat for ci in norm_channels] stem1_blocks_channels = stem_blocks_channels[0] channels[0] = [stem_blocks_channels[1]] + channels[0] net = PNASNet( channels=channels, init_block_channels=init_block_channels, stem1_blocks_channels=stem1_blocks_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def pnasnet5large(**kwargs): """ PNASNet-5-Large model from 'Progressive Neural Architecture Search,' https://arxiv.org/abs/1712.00559. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pnasnet(model_name="pnasnet5large", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ pnasnet5large, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != pnasnet5large or weight_count == 86057668) x = torch.randn(1, 3, 331, 331) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
18,176
28.945634
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/efficientnetedge.py
""" EfficientNet-Edge for ImageNet-1K, implemented in PyTorch. Original paper: 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. """ __all__ = ['EfficientNetEdge', 'efficientnet_edge_small_b', 'efficientnet_edge_medium_b', 'efficientnet_edge_large_b'] import os import math import torch.nn as nn import torch.nn.init as init from .common import round_channels, conv1x1_block, conv3x3_block, SEBlock from .efficientnet import EffiInvResUnit, EffiInitBlock class EffiEdgeResUnit(nn.Module): """ EfficientNet-Edge edge residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the second convolution layer. exp_factor : int Factor for expansion of channels. se_factor : int SE reduction factor for each unit. mid_from_in : bool Whether to use input channel count for middle channel count calculation. use_skip : bool Whether to use skip connection. bn_eps : float Small float added to variance in Batch norm. activation : str Name of activation function. """ def __init__(self, in_channels, out_channels, stride, exp_factor, se_factor, mid_from_in, use_skip, bn_eps, activation): super(EffiEdgeResUnit, self).__init__() self.residual = (in_channels == out_channels) and (stride == 1) and use_skip self.use_se = se_factor > 0 mid_channels = in_channels * exp_factor if mid_from_in else out_channels * exp_factor self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, bn_eps=bn_eps, activation=activation) if self.use_se: self.se = SEBlock( channels=mid_channels, reduction=(exp_factor * se_factor), mid_activation=activation) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, stride=stride, bn_eps=bn_eps, activation=None) def forward(self, x): if self.residual: identity = 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 EfficientNetEdge(nn.Module): """ EfficientNet-Edge model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- channels : list of list of int Number of output channels for each unit. init_block_channels : int Number of output channels for initial unit. final_block_channels : int Number of output channels for the final block of the feature extractor. kernel_sizes : list of list of int Number of kernel sizes for each unit. strides_per_stage : list int Stride value for the first unit of each stage. expansion_factors : list of list of int Number of expansion factors for each unit. dropout_rate : float, default 0.2 Fraction of the input units to drop. Must be a number between 0 and 1. tf_mode : bool, default False Whether to use TF-like mode. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, kernel_sizes, strides_per_stage, expansion_factors, dropout_rate=0.2, tf_mode=False, bn_eps=1e-5, in_channels=3, in_size=(224, 224), num_classes=1000): super(EfficientNetEdge, self).__init__() self.in_size = in_size self.num_classes = num_classes activation = "relu" self.features = nn.Sequential() self.features.add_module("init_block", EffiInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_eps=bn_eps, activation=activation, tf_mode=tf_mode)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): kernel_sizes_per_stage = kernel_sizes[i] expansion_factors_per_stage = expansion_factors[i] mid_from_in = (i != 0) use_skip = (i != 0) stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): kernel_size = kernel_sizes_per_stage[j] expansion_factor = expansion_factors_per_stage[j] stride = strides_per_stage[i] if (j == 0) else 1 if i < 3: stage.add_module("unit{}".format(j + 1), EffiEdgeResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, exp_factor=expansion_factor, se_factor=0, mid_from_in=mid_from_in, use_skip=use_skip, bn_eps=bn_eps, activation=activation)) else: stage.add_module("unit{}".format(j + 1), EffiInvResUnit( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, exp_factor=expansion_factor, se_factor=0, bn_eps=bn_eps, activation=activation, tf_mode=tf_mode)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, bn_eps=bn_eps, activation=activation)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Sequential() if dropout_rate > 0.0: self.output.add_module("dropout", nn.Dropout(p=dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_efficientnet_edge(version, in_size, tf_mode=False, bn_eps=1e-5, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create EfficientNet-Edge model with specific parameters. Parameters: ---------- version : str Version of EfficientNet ('small', 'medium', 'large'). in_size : tuple of two ints Spatial size of the expected input image. tf_mode : bool, default False Whether to use TF-like mode. bn_eps : 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ dropout_rate = 0.0 if version == "small": assert (in_size == (224, 224)) depth_factor = 1.0 width_factor = 1.0 # dropout_rate = 0.2 elif version == "medium": assert (in_size == (240, 240)) depth_factor = 1.1 width_factor = 1.0 # dropout_rate = 0.2 elif version == "large": assert (in_size == (300, 300)) depth_factor = 1.4 width_factor = 1.2 # dropout_rate = 0.3 else: raise ValueError("Unsupported EfficientNet-Edge version {}".format(version)) init_block_channels = 32 layers = [1, 2, 4, 5, 4, 2] downsample = [1, 1, 1, 1, 0, 1] channels_per_layers = [24, 32, 48, 96, 144, 192] expansion_factors_per_layers = [4, 8, 8, 8, 8, 8] kernel_sizes_per_layers = [3, 3, 3, 5, 5, 5] strides_per_stage = [1, 2, 2, 2, 1, 2] final_block_channels = 1280 layers = [int(math.ceil(li * depth_factor)) for li in layers] channels_per_layers = [round_channels(ci * width_factor) for ci in channels_per_layers] from functools import reduce channels = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(channels_per_layers, layers, downsample), []) kernel_sizes = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(kernel_sizes_per_layers, layers, downsample), []) expansion_factors = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(expansion_factors_per_layers, layers, downsample), []) strides_per_stage = reduce(lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(strides_per_stage, layers, downsample), []) strides_per_stage = [si[0] for si in strides_per_stage] init_block_channels = round_channels(init_block_channels * width_factor) if width_factor > 1.0: assert (int(final_block_channels * width_factor) == round_channels(final_block_channels * width_factor)) final_block_channels = round_channels(final_block_channels * width_factor) net = EfficientNetEdge( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, kernel_sizes=kernel_sizes, strides_per_stage=strides_per_stage, expansion_factors=expansion_factors, dropout_rate=dropout_rate, tf_mode=tf_mode, bn_eps=bn_eps, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def efficientnet_edge_small_b(in_size=(224, 224), **kwargs): """ EfficientNet-Edge-Small-b model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (224, 224) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet_edge(version="small", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_edge_small_b", **kwargs) def efficientnet_edge_medium_b(in_size=(240, 240), **kwargs): """ EfficientNet-Edge-Medium-b model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (240, 240) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet_edge(version="medium", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_edge_medium_b", **kwargs) def efficientnet_edge_large_b(in_size=(300, 300), **kwargs): """ EfficientNet-Edge-Large-b model from 'EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks,' https://arxiv.org/abs/1905.11946. Parameters: ---------- in_size : tuple of two ints, default (300, 300) Spatial size of the expected input image. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_efficientnet_edge(version="large", in_size=in_size, tf_mode=True, bn_eps=1e-3, model_name="efficientnet_edge_large_b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ efficientnet_edge_small_b, efficientnet_edge_medium_b, efficientnet_edge_large_b, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != efficientnet_edge_small_b or weight_count == 5438392) assert (model != efficientnet_edge_medium_b or weight_count == 6899496) assert (model != efficientnet_edge_large_b or weight_count == 10589712) x = torch.randn(1, 3, net.in_size[0], net.in_size[1]) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
14,866
35.799505
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/ibnresnext.py
""" IBN-ResNeXt for ImageNet-1K, implemented in PyTorch. Original paper: 'Aggregated Residual Transformations for Deep Neural Networks,' http://arxiv.org/abs/1611.05431. """ __all__ = ['IBNResNeXt', 'ibn_resnext50_32x4d', 'ibn_resnext101_32x4d', 'ibn_resnext101_64x4d'] import os import math import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block from .resnet import ResInitBlock from .ibnresnet import ibn_conv1x1_block class IBNResNeXtBottleneck(nn.Module): """ IBN-ResNeXt bottleneck block for residual path in IBN-ResNeXt unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. conv1_ibn : bool Whether to use IBN normalization in the first convolution layer of the block. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width, conv1_ibn): super(IBNResNeXtBottleneck, self).__init__() mid_channels = out_channels // 4 D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D self.conv1 = ibn_conv1x1_block( in_channels=in_channels, out_channels=group_width, use_ibn=conv1_ibn) self.conv2 = conv3x3_block( in_channels=group_width, out_channels=group_width, stride=stride, groups=cardinality) self.conv3 = conv1x1_block( in_channels=group_width, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class IBNResNeXtUnit(nn.Module): """ IBN-ResNeXt unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. conv1_ibn : bool Whether to use IBN normalization in the first convolution layer of the block. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width, conv1_ibn): super(IBNResNeXtUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = IBNResNeXtBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width, conv1_ibn=conv1_ibn) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 IBNResNeXt(nn.Module): """ IBN-ResNeXt 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. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, in_channels=3, in_size=(224, 224), num_classes=1000): super(IBNResNeXt, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 conv1_ibn = (out_channels < 2048) stage.add_module("unit{}".format(j + 1), IBNResNeXtUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width, conv1_ibn=conv1_ibn)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_ibnresnext(blocks, cardinality, bottleneck_width, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create IBN-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] else: raise ValueError("Unsupported IBN-ResNeXt 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 = IBNResNeXt( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def ibn_resnext50_32x4d(**kwargs): """ IBN-ResNeXt-50 (32x4d) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibnresnext(blocks=50, cardinality=32, bottleneck_width=4, model_name="ibn_resnext50_32x4d", **kwargs) def ibn_resnext101_32x4d(**kwargs): """ IBN-ResNeXt-101 (32x4d) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibnresnext(blocks=101, cardinality=32, bottleneck_width=4, model_name="ibn_resnext101_32x4d", **kwargs) def ibn_resnext101_64x4d(**kwargs): """ IBN-ResNeXt-101 (64x4d) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibnresnext(blocks=101, cardinality=64, bottleneck_width=4, model_name="ibn_resnext101_64x4d", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ ibn_resnext50_32x4d, ibn_resnext101_32x4d, ibn_resnext101_64x4d, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != ibn_resnext50_32x4d or weight_count == 25028904) assert (model != ibn_resnext101_32x4d or weight_count == 44177704) assert (model != ibn_resnext101_64x4d or weight_count == 83455272) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
10,749
30.341108
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/squeezenext.py
""" SqueezeNext for ImageNet-1K, implemented in PyTorch. Original paper: 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. """ __all__ = ['SqueezeNext', 'sqnxt23_w1', 'sqnxt23_w3d2', 'sqnxt23_w2', 'sqnxt23v5_w1', 'sqnxt23v5_w3d2', 'sqnxt23v5_w2'] import os import torch.nn as nn import torch.nn.init as init from .common import ConvBlock, conv1x1_block, conv7x7_block class SqnxtUnit(nn.Module): """ SqueezeNext unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(SqnxtUnit, self).__init__() if stride == 2: reduction_den = 1 self.resize_identity = True elif in_channels > out_channels: reduction_den = 4 self.resize_identity = True else: reduction_den = 2 self.resize_identity = False self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=(in_channels // reduction_den), stride=stride, bias=True) self.conv2 = conv1x1_block( in_channels=(in_channels // reduction_den), out_channels=(in_channels // (2 * reduction_den)), bias=True) self.conv3 = ConvBlock( in_channels=(in_channels // (2 * reduction_den)), out_channels=(in_channels // reduction_den), kernel_size=(1, 3), stride=1, padding=(0, 1), bias=True) self.conv4 = ConvBlock( in_channels=(in_channels // reduction_den), out_channels=(in_channels // reduction_den), kernel_size=(3, 1), stride=1, padding=(1, 0), bias=True) self.conv5 = conv1x1_block( in_channels=(in_channels // reduction_den), out_channels=out_channels, bias=True) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=True) self.activ = nn.ReLU(inplace=True) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.conv4(x) x = self.conv5(x) x = x + identity x = self.activ(x) return x class SqnxtInitBlock(nn.Module): """ SqueezeNext 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): super(SqnxtInitBlock, self).__init__() self.conv = conv7x7_block( in_channels=in_channels, out_channels=out_channels, stride=2, padding=1, bias=True) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, ceil_mode=True) def forward(self, x): x = self.conv(x) x = self.pool(x) return x class SqueezeNext(nn.Module): """ SqueezeNext model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, in_channels=3, in_size=(224, 224), num_classes=1000): super(SqueezeNext, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", SqnxtInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), SqnxtUnit( in_channels=in_channels, out_channels=out_channels, stride=stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, bias=True)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_squeezenext(version, width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SqueezeNext model with specific parameters. Parameters: ---------- version : str Version of SqueezeNet ('23' or '23v5'). 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 64 final_block_channels = 128 channels_per_layers = [32, 64, 128, 256] if version == '23': layers = [6, 6, 8, 1] elif version == '23v5': layers = [2, 4, 14, 1] else: raise ValueError("Unsupported SqueezeNet version {}".format(version)) channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] if width_scale != 1: channels = [[int(cij * width_scale) for cij in ci] for ci in channels] init_block_channels = int(init_block_channels * width_scale) final_block_channels = int(final_block_channels * width_scale) net = SqueezeNext( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def sqnxt23_w1(**kwargs): """ 1.0-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenext(version="23", width_scale=1.0, model_name="sqnxt23_w1", **kwargs) def sqnxt23_w3d2(**kwargs): """ 1.5-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenext(version="23", width_scale=1.5, model_name="sqnxt23_w3d2", **kwargs) def sqnxt23_w2(**kwargs): """ 2.0-SqNxt-23 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenext(version="23", width_scale=2.0, model_name="sqnxt23_w2", **kwargs) def sqnxt23v5_w1(**kwargs): """ 1.0-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenext(version="23v5", width_scale=1.0, model_name="sqnxt23v5_w1", **kwargs) def sqnxt23v5_w3d2(**kwargs): """ 1.5-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenext(version="23v5", width_scale=1.5, model_name="sqnxt23v5_w3d2", **kwargs) def sqnxt23v5_w2(**kwargs): """ 2.0-SqNxt-23v5 model from 'SqueezeNext: Hardware-Aware Neural Network Design,' https://arxiv.org/abs/1803.10615. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenext(version="23v5", width_scale=2.0, model_name="sqnxt23v5_w2", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ sqnxt23_w1, sqnxt23_w3d2, sqnxt23_w2, sqnxt23v5_w1, sqnxt23v5_w3d2, sqnxt23v5_w2, ] for model in models: net = model(pretrained=pretrained) # net.eval() net.train() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sqnxt23_w1 or weight_count == 724056) assert (model != sqnxt23_w3d2 or weight_count == 1511824) assert (model != sqnxt23_w2 or weight_count == 2583752) assert (model != sqnxt23v5_w1 or weight_count == 921816) assert (model != sqnxt23v5_w3d2 or weight_count == 1953616) assert (model != sqnxt23v5_w2 or weight_count == 3366344) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,238
30.543814
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/xdensenet.py
""" X-DenseNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. """ __all__ = ['XDenseNet', 'xdensenet121_2', 'xdensenet161_2', 'xdensenet169_2', 'xdensenet201_2', 'pre_xconv3x3_block', 'XDenseUnit'] import os import torch import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from .preresnet import PreResInitBlock, PreResActivation from .densenet import TransitionBlock class XConv2d(nn.Conv2d): """ X-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. groups : int, default 1 Number of groups. expand_ratio : int, default 2 Ratio of expansion. """ def __init__(self, in_channels, out_channels, kernel_size, groups=1, expand_ratio=2, **kwargs): super(XConv2d, self).__init__( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, groups=groups, **kwargs) self.expand_ratio = expand_ratio if isinstance(kernel_size, int): kernel_size = (kernel_size, kernel_size) grouped_in_channels = in_channels // groups self.mask = torch.nn.Parameter( data=torch.Tensor(out_channels, grouped_in_channels, *kernel_size), requires_grad=False) self.init_parameters() def init_parameters(self): shape = self.mask.shape expand_size = max(shape[1] // self.expand_ratio, 1) self.mask[:] = 0 for i in range(shape[0]): jj = torch.randperm(shape[1], device=self.mask.device)[:expand_size] self.mask[i, jj, :, :] = 1 def forward(self, input): masked_weight = self.weight.mul(self.mask) return F.conv2d( input=input, weight=masked_weight, bias=self.bias, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups) class PreXConvBlock(nn.Module): """ X-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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. return_preact : bool, default False Whether return pre-activation. It's used by PreResNet. activate : bool, default True Whether activate the convolution block. expand_ratio : int, default 2 Ratio of expansion. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, bias=False, return_preact=False, activate=True, expand_ratio=2): super(PreXConvBlock, self).__init__() self.return_preact = return_preact self.activate = activate self.bn = nn.BatchNorm2d(num_features=in_channels) if self.activate: self.activ = nn.ReLU(inplace=True) self.conv = XConv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, expand_ratio=expand_ratio) def forward(self, x): 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_xconv1x1_block(in_channels, out_channels, stride=1, bias=False, return_preact=False, activate=True, expand_ratio=2): """ 1x1 version of the pre-activated x-convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. return_preact : bool, default False Whether return pre-activation. activate : bool, default True Whether activate the convolution block. expand_ratio : int, default 2 Ratio of expansion. """ return PreXConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, bias=bias, return_preact=return_preact, activate=activate, expand_ratio=expand_ratio) def pre_xconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, return_preact=False, activate=True, expand_ratio=2): """ 3x3 version of the pre-activated x-convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. return_preact : bool, default False Whether return pre-activation. activate : bool, default True Whether activate the convolution block. expand_ratio : int, default 2 Ratio of expansion. """ return PreXConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, return_preact=return_preact, activate=activate, expand_ratio=expand_ratio) class XDenseUnit(nn.Module): """ X-DenseNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. 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, dropout_rate, expand_ratio): super(XDenseUnit, self).__init__() self.use_dropout = (dropout_rate != 0.0) bn_size = 4 inc_channels = out_channels - in_channels mid_channels = inc_channels * bn_size self.conv1 = pre_xconv1x1_block( in_channels=in_channels, out_channels=mid_channels, expand_ratio=expand_ratio) self.conv2 = pre_xconv3x3_block( in_channels=mid_channels, out_channels=inc_channels, expand_ratio=expand_ratio) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) if self.use_dropout: x = self.dropout(x) x = torch.cat((identity, x), dim=1) return x class XDenseNet(nn.Module): """ X-DenseNet model 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. 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 (224, 224) Spatial size of the expected input image. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, dropout_rate=0.0, expand_ratio=2, in_channels=3, in_size=(224, 224), num_classes=1000): super(XDenseNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() if i != 0: stage.add_module("trans{}".format(i + 1), TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2))) in_channels = in_channels // 2 for j, out_channels in enumerate(channels_per_stage): stage.add_module("unit{}".format(j + 1), XDenseUnit( in_channels=in_channels, out_channels=out_channels, dropout_rate=dropout_rate, expand_ratio=expand_ratio)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_xdensenet(blocks, expand_ratio=2, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create X-DenseNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 121: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 24, 16] elif blocks == 161: init_block_channels = 96 growth_rate = 48 layers = [6, 12, 36, 24] elif blocks == 169: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 32, 32] elif blocks == 201: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 48, 32] else: raise ValueError("Unsupported X-DenseNet version with number of layers {}".format(blocks)) 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 = XDenseNet( channels=channels, init_block_channels=init_block_channels, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def xdensenet121_2(**kwargs): """ X-DenseNet-121-2 model from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet(blocks=121, model_name="xdensenet121_2", **kwargs) def xdensenet161_2(**kwargs): """ X-DenseNet-161-2 model from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet(blocks=161, model_name="xdensenet161_2", **kwargs) def xdensenet169_2(**kwargs): """ X-DenseNet-169-2 model from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet(blocks=169, model_name="xdensenet169_2", **kwargs) def xdensenet201_2(**kwargs): """ X-DenseNet-201-2 model from 'Deep Expander Networks: Efficient Deep Networks from Graph Theory,' https://arxiv.org/abs/1711.08757. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet(blocks=201, model_name="xdensenet201_2", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ xdensenet121_2, xdensenet161_2, xdensenet169_2, xdensenet201_2, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != xdensenet121_2 or weight_count == 7978856) assert (model != xdensenet161_2 or weight_count == 28681000) assert (model != xdensenet169_2 or weight_count == 14149480) assert (model != xdensenet201_2 or weight_count == 20013928) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
16,251
30.015267
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/linknet.py
""" LinkNet for image segmentation, implemented in PyTorch. Original paper: 'LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation,' https://arxiv.org/abs/1707.03718. """ __all__ = ['LinkNet', 'linknet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1_block, conv3x3_block, deconv3x3_block, Hourglass, Identity from .resnet import resnet18 class DecoderStage(nn.Module): """ LinkNet specific decoder stage. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the deconvolution. out_padding : int or tuple/list of 2 int Output padding value for deconvolution layer. bias : bool, default False Whether the layer uses a bias vector. """ def __init__(self, in_channels, out_channels, stride, output_padding, bias): super(DecoderStage, self).__init__() mid_channels = in_channels // 4 self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bias=bias) self.conv2 = deconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, out_padding=output_padding, bias=bias) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bias=bias) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class LinkNetHead(nn.Module): """ LinkNet head block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(LinkNetHead, self).__init__() mid_channels = in_channels // 2 self.conv1 = deconv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2, padding=1, out_padding=1, bias=True) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, bias=True) self.conv3 = nn.ConvTranspose2d( in_channels=mid_channels, out_channels=out_channels, kernel_size=2, stride=2, padding=0) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class LinkNet(nn.Module): """ LinkNet model from 'LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation,' https://arxiv.org/abs/1707.03718. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int Number of output channels form feature extractor. channels : list of int Number of output channels for the first unit of each stage. dilations : list of list of int Dilation values for each unit. dropout_rates : list of float Parameter of dropout layer for each stage. downs : list of int Whether to downscale or upscale in each stage. correct_size_mistmatch : bool Whether to correct downscaled sizes of images in encoder. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, backbone, backbone_out_channels, channels, strides, output_paddings, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(LinkNet, self).__init__() assert (in_channels == 3) 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.num_classes = num_classes self.fixed_size = fixed_size bias = False self.stem = backbone.init_block in_channels = backbone_out_channels down_seq = nn.Sequential() down_seq.add_module("down1", backbone.stage1) down_seq.add_module("down2", backbone.stage2) down_seq.add_module("down3", backbone.stage3) down_seq.add_module("down4", backbone.stage4) up_seq = nn.Sequential() skip_seq = nn.Sequential() for i, out_channels in enumerate(channels): up_seq.add_module("up{}".format(i + 1), DecoderStage( in_channels=in_channels, out_channels=out_channels, stride=strides[i], output_padding=output_paddings[i], bias=bias)) in_channels = out_channels skip_seq.add_module("skip{}".format(i + 1), Identity()) up_seq = up_seq[::-1] self.hg = Hourglass( down_seq=down_seq, up_seq=up_seq, skip_seq=skip_seq) self.head = LinkNetHead( in_channels=in_channels, out_channels=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.stem(x) x = self.hg(x) x = self.head(x) return x def get_linknet(backbone, backbone_out_channels, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create LinkNet model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. backbone_out_channels : int Number of output channels form feature extractor. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [256, 128, 64, 64] strides = [2, 2, 2, 1] output_paddings = [1, 1, 1, 0] net = LinkNet( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, strides=strides, output_paddings=output_paddings, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def linknet_cityscapes(pretrained_backbone=False, num_classes=19, **kwargs): """ LinkNet model for Cityscapes from 'LinkNet: Exploiting Encoder Representations for Efficient Semantic Segmentation,' https://arxiv.org/abs/1707.03718. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features del backbone[-1] backbone_out_channels = 512 return get_linknet(backbone=backbone, backbone_out_channels=backbone_out_channels, num_classes=num_classes, model_name="linknet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ linknet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != linknet_cityscapes or weight_count == 11535699) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
9,565
29.5623
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/diaresnet_cifar.py
""" DIA-ResNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. """ __all__ = ['CIFARDIAResNet', 'diaresnet20_cifar10', 'diaresnet20_cifar100', 'diaresnet20_svhn', 'diaresnet56_cifar10', 'diaresnet56_cifar100', 'diaresnet56_svhn', 'diaresnet110_cifar10', 'diaresnet110_cifar100', 'diaresnet110_svhn', 'diaresnet164bn_cifar10', 'diaresnet164bn_cifar100', 'diaresnet164bn_svhn', 'diaresnet1001_cifar10', 'diaresnet1001_cifar100', 'diaresnet1001_svhn', 'diaresnet1202_cifar10', 'diaresnet1202_cifar100', 'diaresnet1202_svhn'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3_block, DualPathSequential from .diaresnet import DIAAttention, DIAResUnit class CIFARDIAResNet(nn.Module): """ DIA-ResNet 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARDIAResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( 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) attention = DIAAttention( in_x_features=channels_per_stage[0], in_h_features=channels_per_stage[0]) for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), DIAResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=False, attention=attention)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_diaresnet_cifar(num_classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DIA-ResNet model for CIFAR with specific parameters. Parameters: ---------- num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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 = CIFARDIAResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def diaresnet20_cifar10(num_classes=10, **kwargs): """ DIA-ResNet-20 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="diaresnet20_cifar10", **kwargs) def diaresnet20_cifar100(num_classes=100, **kwargs): """ DIA-ResNet-20 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="diaresnet20_cifar100", **kwargs) def diaresnet20_svhn(num_classes=10, **kwargs): """ DIA-ResNet-20 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="diaresnet20_svhn", **kwargs) def diaresnet56_cifar10(num_classes=10, **kwargs): """ DIA-ResNet-56 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="diaresnet56_cifar10", **kwargs) def diaresnet56_cifar100(num_classes=100, **kwargs): """ DIA-ResNet-56 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="diaresnet56_cifar100", **kwargs) def diaresnet56_svhn(num_classes=10, **kwargs): """ DIA-ResNet-56 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="diaresnet56_svhn", **kwargs) def diaresnet110_cifar10(num_classes=10, **kwargs): """ DIA-ResNet-110 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="diaresnet110_cifar10", **kwargs) def diaresnet110_cifar100(num_classes=100, **kwargs): """ DIA-ResNet-110 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="diaresnet110_cifar100", **kwargs) def diaresnet110_svhn(num_classes=10, **kwargs): """ DIA-ResNet-110 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="diaresnet110_svhn", **kwargs) def diaresnet164bn_cifar10(num_classes=10, **kwargs): """ DIA-ResNet-164(BN) model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="diaresnet164bn_cifar10", **kwargs) def diaresnet164bn_cifar100(num_classes=100, **kwargs): """ DIA-ResNet-164(BN) model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="diaresnet164bn_cifar100", **kwargs) def diaresnet164bn_svhn(num_classes=10, **kwargs): """ DIA-ResNet-164(BN) model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="diaresnet164bn_svhn", **kwargs) def diaresnet1001_cifar10(num_classes=10, **kwargs): """ DIA-ResNet-1001 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="diaresnet1001_cifar10", **kwargs) def diaresnet1001_cifar100(num_classes=100, **kwargs): """ DIA-ResNet-1001 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="diaresnet1001_cifar100", **kwargs) def diaresnet1001_svhn(num_classes=10, **kwargs): """ DIA-ResNet-1001 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="diaresnet1001_svhn", **kwargs) def diaresnet1202_cifar10(num_classes=10, **kwargs): """ DIA-ResNet-1202 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="diaresnet1202_cifar10", **kwargs) def diaresnet1202_cifar100(num_classes=100, **kwargs): """ DIA-ResNet-1202 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="diaresnet1202_cifar100", **kwargs) def diaresnet1202_svhn(num_classes=10, **kwargs): """ DIA-ResNet-1202 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="diaresnet1202_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (diaresnet20_cifar10, 10), (diaresnet20_cifar100, 100), (diaresnet20_svhn, 10), (diaresnet56_cifar10, 10), (diaresnet56_cifar100, 100), (diaresnet56_svhn, 10), (diaresnet110_cifar10, 10), (diaresnet110_cifar100, 100), (diaresnet110_svhn, 10), (diaresnet164bn_cifar10, 10), (diaresnet164bn_cifar100, 100), (diaresnet164bn_svhn, 10), (diaresnet1001_cifar10, 10), (diaresnet1001_cifar100, 100), (diaresnet1001_svhn, 10), (diaresnet1202_cifar10, 10), (diaresnet1202_cifar100, 100), (diaresnet1202_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != diaresnet20_cifar10 or weight_count == 286866) assert (model != diaresnet20_cifar100 or weight_count == 292716) assert (model != diaresnet20_svhn or weight_count == 286866) assert (model != diaresnet56_cifar10 or weight_count == 870162) assert (model != diaresnet56_cifar100 or weight_count == 876012) assert (model != diaresnet56_svhn or weight_count == 870162) assert (model != diaresnet110_cifar10 or weight_count == 1745106) assert (model != diaresnet110_cifar100 or weight_count == 1750956) assert (model != diaresnet110_svhn or weight_count == 1745106) assert (model != diaresnet164bn_cifar10 or weight_count == 1923002) assert (model != diaresnet164bn_cifar100 or weight_count == 1946132) assert (model != diaresnet164bn_svhn or weight_count == 1923002) assert (model != diaresnet1001_cifar10 or weight_count == 10547450) assert (model != diaresnet1001_cifar100 or weight_count == 10570580) assert (model != diaresnet1001_svhn or weight_count == 10547450) assert (model != diaresnet1202_cifar10 or weight_count == 19438418) assert (model != diaresnet1202_cifar100 or weight_count == 19444268) assert (model != diaresnet1202_svhn or weight_count == 19438418) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
19,959
35.489945
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resdropresnet_cifar.py
""" ResDrop-ResNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Deep Networks with Stochastic Depth,' https://arxiv.org/abs/1603.09382. """ __all__ = ['CIFARResDropResNet', 'resdropresnet20_cifar10', 'resdropresnet20_cifar100', 'resdropresnet20_svhn'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block from .resnet import ResBlock, ResBottleneck class ResDropResUnit(nn.Module): """ ResDrop-ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. life_prob : float Residual branch life probability. """ def __init__(self, in_channels, out_channels, stride, bottleneck, life_prob): super(ResDropResUnit, self).__init__() self.life_prob = life_prob self.resize_identity = (in_channels != out_channels) or (stride != 1) body_class = ResBottleneck if bottleneck else ResBlock self.body = body_class( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) if self.training: b = torch.bernoulli(torch.full((1,), self.life_prob, dtype=x.dtype, device=x.device)) x = float(b) / self.life_prob * x x = x + identity x = self.activ(x) return x class CIFARResDropResNet(nn.Module): """ ResDrop-ResNet model for CIFAR from 'Deep Networks with Stochastic Depth,' https://arxiv.org/abs/1603.09382. 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. life_probs : list of float Residual branch life probability for each unit. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, life_probs, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARResDropResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels k = 0 for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ResDropResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, life_prob=life_probs[k])) in_channels = out_channels k += 1 self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resdropresnet_cifar(classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create ResDrop-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. root : str, default '~/.torch/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 init_block_channels = 16 channels_per_layers = [16, 32, 64] 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] total_layers = sum(layers) final_death_prob = 0.5 life_probs = [1.0 - float(i + 1) / float(total_layers) * final_death_prob for i in range(total_layers)] net = CIFARResDropResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, life_probs=life_probs, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def resdropresnet20_cifar10(classes=10, **kwargs): """ ResDrop-ResNet-20 model for CIFAR-10 from 'Deep Networks with Stochastic Depth,' https://arxiv.org/abs/1603.09382. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resdropresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="resdropresnet20_cifar10", **kwargs) def resdropresnet20_cifar100(classes=100, **kwargs): """ ResDrop-ResNet-20 model for CIFAR-100 from 'Deep Networks with Stochastic Depth,' https://arxiv.org/abs/1603.09382. Parameters: ---------- classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resdropresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="resdropresnet20_cifar100", **kwargs) def resdropresnet20_svhn(classes=10, **kwargs): """ ResDrop-ResNet-20 model for SVHN from 'Deep Networks with Stochastic Depth,' https://arxiv.org/abs/1603.09382. Parameters: ---------- classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resdropresnet_cifar(classes=classes, blocks=20, bottleneck=False, model_name="resdropresnet20_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (resdropresnet20_cifar10, 10), (resdropresnet20_cifar100, 100), (resdropresnet20_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resdropresnet20_cifar10 or weight_count == 272474) assert (model != resdropresnet20_cifar100 or weight_count == 278324) assert (model != resdropresnet20_svhn or weight_count == 272474) x = torch.randn(14, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (14, num_classes)) if __name__ == "__main__": _test()
9,918
31.735974
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/bisenet.py
""" BiSeNet for CelebAMask-HQ, implemented in PyTorch. Original paper: 'BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1808.00897. """ __all__ = ['BiSeNet', 'bisenet_resnet18_celebamaskhq'] import os import torch import torch.nn as nn from .common import conv1x1, conv1x1_block, conv3x3_block, InterpolationBlock, MultiOutputSequential from .resnet import resnet18 class PyramidPoolingZeroBranch(nn.Module): """ Pyramid pooling zero branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. in_size : tuple of 2 int Spatial size of output image for the upsampling operation. """ def __init__(self, in_channels, out_channels, in_size): super(PyramidPoolingZeroBranch, self).__init__() self.in_size = in_size self.pool = nn.AdaptiveAvgPool2d(1) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels) self.up = InterpolationBlock( scale_factor=None, mode="nearest", align_corners=None) def forward(self, x): in_size = self.in_size if self.in_size is not None else x.shape[2:] x = self.pool(x) x = self.conv(x) x = self.up(x, size=in_size) return x class AttentionRefinementBlock(nn.Module): """ Attention refinement block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(AttentionRefinementBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels) self.pool = nn.AdaptiveAvgPool2d(1) self.conv2 = conv1x1_block( in_channels=out_channels, out_channels=out_channels, activation=(lambda: nn.Sigmoid())) def forward(self, x): x = self.conv1(x) w = self.pool(x) w = self.conv2(w) x = x * w return x class PyramidPoolingMainBranch(nn.Module): """ Pyramid pooling main branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. scale_factor : float Multiplier for spatial size. """ def __init__(self, in_channels, out_channels, scale_factor): super(PyramidPoolingMainBranch, self).__init__() self.att = AttentionRefinementBlock( in_channels=in_channels, out_channels=out_channels) self.up = InterpolationBlock( scale_factor=scale_factor, mode="nearest", align_corners=None) self.conv = conv3x3_block( in_channels=out_channels, out_channels=out_channels) def forward(self, x, y): x = self.att(x) x = x + y x = self.up(x) x = self.conv(x) return x class FeatureFusion(nn.Module): """ Feature fusion 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): super(FeatureFusion, self).__init__() mid_channels = out_channels // reduction self.conv_merge = conv1x1_block( in_channels=in_channels, out_channels=out_channels) self.pool = nn.AdaptiveAvgPool2d(1) self.conv1 = conv1x1( in_channels=out_channels, out_channels=mid_channels) self.activ = nn.ReLU(inplace=True) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels) self.sigmoid = nn.Sigmoid() def forward(self, x, y): x = torch.cat((x, y), dim=1) x = self.conv_merge(x) w = self.pool(x) w = self.conv1(w) w = self.activ(w) w = self.conv2(w) w = self.sigmoid(w) x_att = x * w x = x + x_att return x class PyramidPooling(nn.Module): """ Pyramid Pooling module. Parameters: ---------- x16_in_channels : int Number of input channels for x16. x32_in_channels : int Number of input channels for x32. y_out_channels : int Number of output channels for y-outputs. y32_out_size : tuple of 2 int Spatial size of the y32 tensor. """ def __init__(self, x16_in_channels, x32_in_channels, y_out_channels, y32_out_size): super(PyramidPooling, self).__init__() z_out_channels = 2 * y_out_channels self.pool32 = PyramidPoolingZeroBranch( in_channels=x32_in_channels, out_channels=y_out_channels, in_size=y32_out_size) self.pool16 = PyramidPoolingMainBranch( in_channels=x32_in_channels, out_channels=y_out_channels, scale_factor=2) self.pool8 = PyramidPoolingMainBranch( in_channels=x16_in_channels, out_channels=y_out_channels, scale_factor=2) self.fusion = FeatureFusion( in_channels=z_out_channels, out_channels=z_out_channels) def forward(self, x8, x16, x32): y32 = self.pool32(x32) y16 = self.pool16(x32, y32) y8 = self.pool8(x16, y16) z8 = self.fusion(x8, y8) return z8, y8, y16 class BiSeHead(nn.Module): """ BiSeNet head (final) block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, mid_channels, out_channels): super(BiSeHead, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class BiSeNet(nn.Module): """ BiSeNet model from 'BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1808.00897. Parameters: ---------- backbone : func -> nn.Sequential Feature extractor. aux : bool, default True Whether to output an auxiliary results. 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 (640, 480) Spatial size of the expected input image. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, backbone, aux=True, fixed_size=True, in_channels=3, in_size=(640, 480), num_classes=19): super(BiSeNet, self).__init__() assert (in_channels == 3) self.in_size = in_size self.num_classes = num_classes self.aux = aux self.fixed_size = fixed_size self.backbone, backbone_out_channels = backbone() y_out_channels = backbone_out_channels[0] z_out_channels = 2 * y_out_channels y32_out_size = (self.in_size[0] // 32, self.in_size[1] // 32) if fixed_size else None self.pool = PyramidPooling( x16_in_channels=backbone_out_channels[1], x32_in_channels=backbone_out_channels[2], y_out_channels=y_out_channels, y32_out_size=y32_out_size) self.head_z8 = BiSeHead( in_channels=z_out_channels, mid_channels=z_out_channels, out_channels=num_classes) self.up8 = InterpolationBlock(scale_factor=(8 if fixed_size else None)) if self.aux: mid_channels = y_out_channels // 2 self.head_y8 = BiSeHead( in_channels=y_out_channels, mid_channels=mid_channels, out_channels=num_classes) self.head_y16 = BiSeHead( in_channels=y_out_channels, mid_channels=mid_channels, out_channels=num_classes) self.up16 = InterpolationBlock(scale_factor=(16 if fixed_size else None)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_normal_(module.weight, a=1) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): assert (x.shape[2] % 32 == 0) and (x.shape[3] % 32 == 0) x8, x16, x32 = self.backbone(x) z8, y8, y16 = self.pool(x8, x16, x32) z8 = self.head_z8(z8) z8 = self.up8(z8) if self.aux: y8 = self.head_y8(y8) y16 = self.head_y16(y16) y8 = self.up8(y8) y16 = self.up16(y16) return z8, y8, y16 else: return z8 def get_bisenet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create BiSeNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = BiSeNet( **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def bisenet_resnet18_celebamaskhq(pretrained_backbone=False, num_classes=19, **kwargs): """ BiSeNet model on the base of ResNet-18 for face segmentation on CelebAMask-HQ from 'BiSeNet: Bilateral Segmentation Network for Real-time Semantic Segmentation,' https://arxiv.org/abs/1808.00897. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_classes : int, default 19 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ def backbone(): features_raw = resnet18(pretrained=pretrained_backbone).features del features_raw[-1] features = MultiOutputSequential(return_last=False) features.add_module("init_block", features_raw[0]) for i, stage in enumerate(features_raw[1:]): if i != 0: stage.do_output = True features.add_module("stage{}".format(i + 1), stage) out_channels = [128, 256, 512] return features, out_channels return get_bisenet(backbone=backbone, num_classes=num_classes, model_name="bisenet_resnet18_celebamaskhq", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch in_size = (640, 480) aux = True pretrained = False models = [ bisenet_resnet18_celebamaskhq, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, aux=aux) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) if aux: assert (model != bisenet_resnet18_celebamaskhq or weight_count == 13300416) else: assert (model != bisenet_resnet18_celebamaskhq or weight_count == 13150272) batch = 1 x = torch.randn(batch, 3, in_size[0], in_size[1]) ys = net(x) y = ys[0] if aux else ys # y.sum().backward() assert (tuple(y.size()) == (batch, 19, in_size[0], in_size[1])) if __name__ == "__main__": _test()
13,181
28.959091
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resnet.py
""" ResNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. """ __all__ = ['ResNet', 'resnet10', 'resnet12', 'resnet14', 'resnetbc14b', 'resnet16', 'resnet18_wd4', 'resnet18_wd2', 'resnet18_w3d4', 'resnet18', 'resnet26', 'resnetbc26b', 'resnet34', 'resnetbc38b', 'resnet50', 'resnet50b', 'resnet101', 'resnet101b', 'resnet152', 'resnet152b', 'resnet200', 'resnet200b', 'ResBlock', 'ResBottleneck', 'ResUnit', 'ResInitBlock'] import os import torch.nn as nn from .common import conv1x1_block, conv3x3_block, conv7x7_block class ResBlock(nn.Module): """ Simple ResNet block for residual path in ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. """ def __init__(self, in_channels, out_channels, stride, bias=False, use_bn=True): super(ResBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias, use_bn=use_bn) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bias=bias, use_bn=use_bn, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class ResBottleneck(nn.Module): """ ResNet bottleneck block for residual path in ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. 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, stride, padding=1, dilation=1, conv1_stride=False, bottleneck_factor=4): super(ResBottleneck, self).__init__() mid_channels = out_channels // bottleneck_factor self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, stride=(stride if conv1_stride else 1)) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=(1 if conv1_stride else stride), padding=padding, dilation=dilation) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class ResUnit(nn.Module): """ ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. 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, stride, padding=1, dilation=1, bias=False, use_bn=True, bottleneck=True, conv1_stride=False): super(ResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, padding=padding, dilation=dilation, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias, use_bn=use_bn) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias, use_bn=use_bn, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 ResInitBlock(nn.Module): """ 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): super(ResInitBlock, self).__init__() self.conv = conv7x7_block( in_channels=in_channels, out_channels=out_channels, stride=2) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.pool(x) return x class ResNet(nn.Module): """ ResNet 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, in_channels=3, in_size=(224, 224), num_classes=1000): super(ResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resnet(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create 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. root : str, default '~/.torch/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 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 = ResNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def resnet10(**kwargs): """ ResNet-10 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=10, model_name="resnet10", **kwargs) def resnet12(**kwargs): """ ResNet-12 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=12, model_name="resnet12", **kwargs) def resnet14(**kwargs): """ ResNet-14 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=14, model_name="resnet14", **kwargs) def resnetbc14b(**kwargs): """ ResNet-BC-14b 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=14, bottleneck=True, conv1_stride=False, model_name="resnetbc14b", **kwargs) def resnet16(**kwargs): """ ResNet-16 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=16, model_name="resnet16", **kwargs) def resnet18_wd4(**kwargs): """ ResNet-18 model with 0.25 width scale from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=18, width_scale=0.25, model_name="resnet18_wd4", **kwargs) def resnet18_wd2(**kwargs): """ ResNet-18 model with 0.5 width scale from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=18, width_scale=0.5, model_name="resnet18_wd2", **kwargs) def resnet18_w3d4(**kwargs): """ ResNet-18 model with 0.75 width scale from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=18, width_scale=0.75, model_name="resnet18_w3d4", **kwargs) def resnet18(**kwargs): """ ResNet-18 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=18, model_name="resnet18", **kwargs) def resnet26(**kwargs): """ ResNet-26 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=26, bottleneck=False, model_name="resnet26", **kwargs) def resnetbc26b(**kwargs): """ ResNet-BC-26b 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="resnetbc26b", **kwargs) def resnet34(**kwargs): """ ResNet-34 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=34, model_name="resnet34", **kwargs) def resnetbc38b(**kwargs): """ ResNet-BC-38b 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=38, bottleneck=True, conv1_stride=False, model_name="resnetbc38b", **kwargs) def resnet50(**kwargs): """ ResNet-50 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=50, model_name="resnet50", **kwargs) def resnet50b(**kwargs): """ ResNet-50 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=50, conv1_stride=False, model_name="resnet50b", **kwargs) def resnet101(**kwargs): """ ResNet-101 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=101, model_name="resnet101", **kwargs) def resnet101b(**kwargs): """ ResNet-101 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=101, conv1_stride=False, model_name="resnet101b", **kwargs) def resnet152(**kwargs): """ ResNet-152 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=152, model_name="resnet152", **kwargs) def resnet152b(**kwargs): """ ResNet-152 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=152, conv1_stride=False, model_name="resnet152b", **kwargs) def resnet200(**kwargs): """ ResNet-200 model from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=200, model_name="resnet200", **kwargs) def resnet200b(**kwargs): """ ResNet-200 model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=200, conv1_stride=False, model_name="resnet200b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ resnet10, resnet12, resnet14, resnetbc14b, resnet16, resnet18_wd4, resnet18_wd2, resnet18_w3d4, resnet18, resnet26, resnetbc26b, resnet34, resnetbc38b, resnet50, resnet50b, resnet101, resnet101b, resnet152, resnet152b, resnet200, resnet200b, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resnet10 or weight_count == 5418792) assert (model != resnet12 or weight_count == 5492776) assert (model != resnet14 or weight_count == 5788200) assert (model != resnetbc14b or weight_count == 10064936) assert (model != resnet16 or weight_count == 6968872) assert (model != resnet18_wd4 or weight_count == 3937400) assert (model != resnet18_wd2 or weight_count == 5804296) assert (model != resnet18_w3d4 or weight_count == 8476056) assert (model != resnet18 or weight_count == 11689512) assert (model != resnet26 or weight_count == 17960232) assert (model != resnetbc26b or weight_count == 15995176) assert (model != resnet34 or weight_count == 21797672) assert (model != resnetbc38b or weight_count == 21925416) assert (model != resnet50 or weight_count == 25557032) assert (model != resnet50b or weight_count == 25557032) assert (model != resnet101 or weight_count == 44549160) assert (model != resnet101b or weight_count == 44549160) assert (model != resnet152 or weight_count == 60192808) assert (model != resnet152b or weight_count == 60192808) assert (model != resnet200 or weight_count == 64673832) assert (model != resnet200b or weight_count == 64673832) batch = 4 x = torch.randn(batch, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (batch, 1000)) if __name__ == "__main__": _test()
25,346
31.579692
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/simpleposemobile_coco.py
""" SimplePose(Mobile) for COCO Keypoint, implemented in PyTorch. Original paper: 'Simple Baselines for Human Pose Estimation and Tracking,' https://arxiv.org/abs/1804.06208. """ __all__ = ['SimplePoseMobile', 'simplepose_mobile_resnet18_coco', 'simplepose_mobile_resnet50b_coco', 'simplepose_mobile_mobilenet_w1_coco', 'simplepose_mobile_mobilenetv2b_w1_coco', 'simplepose_mobile_mobilenetv3_small_w1_coco', 'simplepose_mobile_mobilenetv3_large_w1_coco'] import os import torch import torch.nn as nn from .common import conv1x1, DucBlock, HeatmapMaxDetBlock from .resnet import resnet18, resnet50b from .mobilenet import mobilenet_w1 from .mobilenetv2 import mobilenetv2b_w1 from .mobilenetv3 import mobilenetv3_small_w1, mobilenetv3_large_w1 class SimplePoseMobile(nn.Module): """ SimplePose(Mobile) 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. decoder_init_block_channels : int Number of output channels for the initial unit of the decoder. return_heatmap : bool, default False Whether to return only heatmap. 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, decoder_init_block_channels, return_heatmap=False, in_channels=3, in_size=(256, 192), keypoints=17): super(SimplePoseMobile, self).__init__() assert (in_channels == 3) self.in_size = in_size self.keypoints = keypoints self.return_heatmap = return_heatmap self.backbone = backbone self.decoder = nn.Sequential() in_channels = backbone_out_channels self.decoder.add_module("init_block", conv1x1( in_channels=in_channels, out_channels=decoder_init_block_channels)) in_channels = decoder_init_block_channels for i, out_channels in enumerate(channels): self.decoder.add_module("unit{}".format(i + 1), DucBlock( in_channels=in_channels, out_channels=out_channels, scale_factor=2)) in_channels = out_channels self.decoder.add_module("final_block", conv1x1( in_channels=in_channels, out_channels=keypoints)) self.heatmap_max_det = HeatmapMaxDetBlock() self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, 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_simpleposemobile(backbone, backbone_out_channels, keypoints, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SimplePose(Mobile) 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. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [128, 64, 32] decoder_init_block_channels = 256 net = SimplePoseMobile( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, decoder_init_block_channels=decoder_init_block_channels, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def simplepose_mobile_resnet18_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ SimplePose(Mobile) 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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features del backbone[-1] return get_simpleposemobile(backbone=backbone, backbone_out_channels=512, keypoints=keypoints, model_name="simplepose_mobile_resnet18_coco", **kwargs) def simplepose_mobile_resnet50b_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ SimplePose(Mobile) 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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features del backbone[-1] return get_simpleposemobile(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_mobile_resnet50b_coco", **kwargs) def simplepose_mobile_mobilenet_w1_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ SimplePose(Mobile) model on the base of 1.0 MobileNet-224 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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = mobilenet_w1(pretrained=pretrained_backbone).features del backbone[-1] return get_simpleposemobile(backbone=backbone, backbone_out_channels=1024, keypoints=keypoints, model_name="simplepose_mobile_mobilenet_w1_coco", **kwargs) def simplepose_mobile_mobilenetv2b_w1_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ SimplePose(Mobile) model on the base of 1.0 MobileNetV2b-224 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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = mobilenetv2b_w1(pretrained=pretrained_backbone).features del backbone[-1] return get_simpleposemobile(backbone=backbone, backbone_out_channels=1280, keypoints=keypoints, model_name="simplepose_mobile_mobilenetv2b_w1_coco", **kwargs) def simplepose_mobile_mobilenetv3_small_w1_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ SimplePose(Mobile) model on the base of MobileNetV3 Small 224/1.0 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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = mobilenetv3_small_w1(pretrained=pretrained_backbone).features del backbone[-1] return get_simpleposemobile(backbone=backbone, backbone_out_channels=576, keypoints=keypoints, model_name="simplepose_mobile_mobilenetv3_small_w1_coco", **kwargs) def simplepose_mobile_mobilenetv3_large_w1_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ SimplePose(Mobile) model on the base of MobileNetV3 Large 224/1.0 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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = mobilenetv3_large_w1(pretrained=pretrained_backbone).features del backbone[-1] return get_simpleposemobile(backbone=backbone, backbone_out_channels=960, keypoints=keypoints, model_name="simplepose_mobile_mobilenetv3_large_w1_coco", **kwargs) def _calc_width(net): import numpy as np 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 _test(): in_size = (256, 192) keypoints = 17 return_heatmap = False pretrained = False models = [ simplepose_mobile_resnet18_coco, simplepose_mobile_resnet50b_coco, simplepose_mobile_mobilenet_w1_coco, simplepose_mobile_mobilenetv2b_w1_coco, simplepose_mobile_mobilenetv3_small_w1_coco, simplepose_mobile_mobilenetv3_large_w1_coco, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, return_heatmap=return_heatmap) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != simplepose_mobile_resnet18_coco or weight_count == 12858208) assert (model != simplepose_mobile_resnet50b_coco or weight_count == 25582944) assert (model != simplepose_mobile_mobilenet_w1_coco or weight_count == 5019744) assert (model != simplepose_mobile_mobilenetv2b_w1_coco or weight_count == 4102176) assert (model != simplepose_mobile_mobilenetv3_small_w1_coco or weight_count == 2625088) assert (model != simplepose_mobile_mobilenetv3_large_w1_coco or weight_count == 4768336) batch = 14 x = torch.randn(batch, 3, in_size[0], in_size[1]) 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()
12,743
37.735562
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/cbamresnet.py
""" CBAM-ResNet for ImageNet-1K, implemented in PyTorch. Original paper: 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. """ __all__ = ['CbamResNet', 'cbam_resnet18', 'cbam_resnet34', 'cbam_resnet50', 'cbam_resnet101', 'cbam_resnet152'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv7x7_block from .resnet import ResInitBlock, ResBlock, ResBottleneck class MLP(nn.Module): """ Multilayer perceptron block. Parameters: ---------- channels : int Number of input/output channels. reduction_ratio : int, default 16 Channel reduction ratio. """ def __init__(self, channels, reduction_ratio=16): super(MLP, self).__init__() mid_channels = channels // reduction_ratio self.fc1 = nn.Linear( in_features=channels, out_features=mid_channels) self.activ = nn.ReLU(inplace=True) self.fc2 = nn.Linear( in_features=mid_channels, out_features=channels) def forward(self, x): x = x.view(x.size(0), -1) x = self.fc1(x) x = self.activ(x) x = self.fc2(x) return x class ChannelGate(nn.Module): """ CBAM channel gate block. Parameters: ---------- channels : int Number of input/output channels. reduction_ratio : int, default 16 Channel reduction ratio. """ def __init__(self, channels, reduction_ratio=16): super(ChannelGate, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d(output_size=(1, 1)) self.max_pool = nn.AdaptiveMaxPool2d(output_size=(1, 1)) self.mlp = MLP( channels=channels, reduction_ratio=reduction_ratio) self.sigmoid = nn.Sigmoid() def forward(self, x): att1 = self.avg_pool(x) att1 = self.mlp(att1) att2 = self.max_pool(x) att2 = self.mlp(att2) att = att1 + att2 att = self.sigmoid(att) att = att.unsqueeze(2).unsqueeze(3).expand_as(x) x = x * att return x class SpatialGate(nn.Module): """ CBAM spatial gate block. """ def __init__(self): super(SpatialGate, self).__init__() self.conv = conv7x7_block( in_channels=2, out_channels=1, activation=None) self.sigmoid = nn.Sigmoid() def forward(self, x): att1 = x.max(dim=1)[0].unsqueeze(1) att2 = x.mean(dim=1).unsqueeze(1) att = torch.cat((att1, att2), dim=1) att = self.conv(att) att = self.sigmoid(att) x = x * att return x class CbamBlock(nn.Module): """ CBAM attention block for CBAM-ResNet. Parameters: ---------- channels : int Number of input/output channels. reduction_ratio : int, default 16 Channel reduction ratio. """ def __init__(self, channels, reduction_ratio=16): super(CbamBlock, self).__init__() self.ch_gate = ChannelGate( channels=channels, reduction_ratio=reduction_ratio) self.sp_gate = SpatialGate() def forward(self, x): x = self.ch_gate(x) x = self.sp_gate(x) return x class CbamResUnit(nn.Module): """ CBAM-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, stride, bottleneck): super(CbamResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, conv1_stride=False) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.cbam = CbamBlock(channels=out_channels) self.activ = nn.ReLU(inplace=True) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) x = self.cbam(x) x = x + identity x = self.activ(x) return x class CbamResNet(nn.Module): """ CBAM-ResNet model from 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(224, 224), num_classes=1000): super(CbamResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), CbamResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resnet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create CBAM-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. root : str, default '~/.torch/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 CBAM-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 = CbamResNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def cbam_resnet18(**kwargs): """ CBAM-ResNet-18 model from 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=18, model_name="cbam_resnet18", **kwargs) def cbam_resnet34(**kwargs): """ CBAM-ResNet-34 model from 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=34, model_name="cbam_resnet34", **kwargs) def cbam_resnet50(**kwargs): """ CBAM-ResNet-50 model from 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=50, model_name="cbam_resnet50", **kwargs) def cbam_resnet101(**kwargs): """ CBAM-ResNet-101 model from 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=101, model_name="cbam_resnet101", **kwargs) def cbam_resnet152(**kwargs): """ CBAM-ResNet-152 model from 'CBAM: Convolutional Block Attention Module,' https://arxiv.org/abs/1807.06521. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=152, model_name="cbam_resnet152", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ cbam_resnet18, cbam_resnet34, cbam_resnet50, cbam_resnet101, cbam_resnet152, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != cbam_resnet18 or weight_count == 11779392) assert (model != cbam_resnet34 or weight_count == 21960468) assert (model != cbam_resnet50 or weight_count == 28089624) assert (model != cbam_resnet101 or weight_count == 49330172) assert (model != cbam_resnet152 or weight_count == 66826848) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,908
28.405467
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/diracnetv2.py
""" DiracNetV2 for ImageNet-1K, implemented in PyTorch. Original paper: 'DiracNets: Training Very Deep Neural Networks Without Skip-Connections,' https://arxiv.org/abs/1706.00388. """ __all__ = ['DiracNetV2', 'diracnet18v2', 'diracnet34v2'] import os import torch.nn as nn import torch.nn.init as init class DiracConv(nn.Module): """ DiracNetV2 specific convolution block with 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. stride : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding): super(DiracConv, self).__init__() self.activ = nn.ReLU(inplace=True) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=True) def forward(self, x): x = self.activ(x) x = self.conv(x) return x def dirac_conv3x3(in_channels, out_channels): """ 3x3 version of the DiracNetV2 specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ return DiracConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1) class DiracInitBlock(nn.Module): """ DiracNetV2 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): super(DiracInitBlock, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=2, padding=3, bias=True) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.pool(x) return x class DiracNetV2(nn.Module): """ DiracNetV2 model from 'DiracNets: Training Very Deep Neural Networks Without Skip-Connections,' https://arxiv.org/abs/1706.00388. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, in_channels=3, in_size=(224, 224), num_classes=1000): super(DiracNetV2, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", DiracInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stage.add_module("unit{}".format(j + 1), dirac_conv3x3( in_channels=in_channels, out_channels=out_channels)) in_channels = out_channels if i != len(channels) - 1: stage.add_module("pool{}".format(i + 1), nn.MaxPool2d( kernel_size=2, stride=2, padding=0)) self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_activ", nn.ReLU(inplace=True)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_diracnetv2(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DiracNetV2 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 18: layers = [4, 4, 4, 4] elif blocks == 34: layers = [6, 8, 12, 6] else: raise ValueError("Unsupported DiracNetV2 with number of blocks: {}".format(blocks)) channels_per_layers = [64, 128, 256, 512] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] init_block_channels = 64 net = DiracNetV2( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def diracnet18v2(**kwargs): """ DiracNetV2 model from 'DiracNets: Training Very Deep Neural Networks Without Skip-Connections,' https://arxiv.org/abs/1706.00388. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diracnetv2(blocks=18, model_name="diracnet18v2", **kwargs) def diracnet34v2(**kwargs): """ DiracNetV2 model from 'DiracNets: Training Very Deep Neural Networks Without Skip-Connections,' https://arxiv.org/abs/1706.00388. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diracnetv2(blocks=34, model_name="diracnet34v2", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ diracnet18v2, diracnet34v2, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != diracnet18v2 or weight_count == 11511784) assert (model != diracnet34v2 or weight_count == 21616232) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
8,444
27.72449
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/sepreresnet_cifar.py
""" SE-PreResNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['CIFARSEPreResNet', 'sepreresnet20_cifar10', 'sepreresnet20_cifar100', 'sepreresnet20_svhn', 'sepreresnet56_cifar10', 'sepreresnet56_cifar100', 'sepreresnet56_svhn', 'sepreresnet110_cifar10', 'sepreresnet110_cifar100', 'sepreresnet110_svhn', 'sepreresnet164bn_cifar10', 'sepreresnet164bn_cifar100', 'sepreresnet164bn_svhn', 'sepreresnet272bn_cifar10', 'sepreresnet272bn_cifar100', 'sepreresnet272bn_svhn', 'sepreresnet542bn_cifar10', 'sepreresnet542bn_cifar100', 'sepreresnet542bn_svhn', 'sepreresnet1001_cifar10', 'sepreresnet1001_cifar100', 'sepreresnet1001_svhn', 'sepreresnet1202_cifar10', 'sepreresnet1202_cifar100', 'sepreresnet1202_svhn'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3_block from .sepreresnet import SEPreResUnit class CIFARSEPreResNet(nn.Module): """ SE-PreResNet model for CIFAR 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. 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. num_classes : int, default 10 Number of classification num_classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARSEPreResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), SEPreResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=False)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_sepreresnet_cifar(num_classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SE-PreResNet model for CIFAR with specific parameters. Parameters: ---------- num_classes : int Number of classification num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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 = CIFARSEPreResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def sepreresnet20_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-20 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="sepreresnet20_cifar10", **kwargs) def sepreresnet20_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-20 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="sepreresnet20_cifar100", **kwargs) def sepreresnet20_svhn(num_classes=10, **kwargs): """ SE-PreResNet-20 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="sepreresnet20_svhn", **kwargs) def sepreresnet56_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-56 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="sepreresnet56_cifar10", **kwargs) def sepreresnet56_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-56 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="sepreresnet56_cifar100", **kwargs) def sepreresnet56_svhn(num_classes=10, **kwargs): """ SE-PreResNet-56 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="sepreresnet56_svhn", **kwargs) def sepreresnet110_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-110 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="sepreresnet110_cifar10", **kwargs) def sepreresnet110_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-110 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="sepreresnet110_cifar100", **kwargs) def sepreresnet110_svhn(num_classes=10, **kwargs): """ SE-PreResNet-110 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="sepreresnet110_svhn", **kwargs) def sepreresnet164bn_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-164(BN) model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="sepreresnet164bn_cifar10", **kwargs) def sepreresnet164bn_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-164(BN) model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="sepreresnet164bn_cifar100", **kwargs) def sepreresnet164bn_svhn(num_classes=10, **kwargs): """ SE-PreResNet-164(BN) model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="sepreresnet164bn_svhn", **kwargs) def sepreresnet272bn_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-272(BN) model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="sepreresnet272bn_cifar10", **kwargs) def sepreresnet272bn_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-272(BN) model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="sepreresnet272bn_cifar100", **kwargs) def sepreresnet272bn_svhn(num_classes=10, **kwargs): """ SE-PreResNet-272(BN) model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="sepreresnet272bn_svhn", **kwargs) def sepreresnet542bn_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-542(BN) model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="sepreresnet542bn_cifar10", **kwargs) def sepreresnet542bn_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-542(BN) model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="sepreresnet542bn_cifar100", **kwargs) def sepreresnet542bn_svhn(num_classes=10, **kwargs): """ SE-PreResNet-542(BN) model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="sepreresnet542bn_svhn", **kwargs) def sepreresnet1001_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-1001 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="sepreresnet1001_cifar10", **kwargs) def sepreresnet1001_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-1001 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="sepreresnet1001_cifar100", **kwargs) def sepreresnet1001_svhn(num_classes=10, **kwargs): """ SE-PreResNet-1001 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="sepreresnet1001_svhn", **kwargs) def sepreresnet1202_cifar10(num_classes=10, **kwargs): """ SE-PreResNet-1202 model for CIFAR-10 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="sepreresnet1202_cifar10", **kwargs) def sepreresnet1202_cifar100(num_classes=100, **kwargs): """ SE-PreResNet-1202 model for CIFAR-100 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 100 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="sepreresnet1202_cifar100", **kwargs) def sepreresnet1202_svhn(num_classes=10, **kwargs): """ SE-PreResNet-1202 model for SVHN from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 10 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="sepreresnet1202_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (sepreresnet20_cifar10, 10), (sepreresnet20_cifar100, 100), (sepreresnet20_svhn, 10), (sepreresnet56_cifar10, 10), (sepreresnet56_cifar100, 100), (sepreresnet56_svhn, 10), (sepreresnet110_cifar10, 10), (sepreresnet110_cifar100, 100), (sepreresnet110_svhn, 10), (sepreresnet164bn_cifar10, 10), (sepreresnet164bn_cifar100, 100), (sepreresnet164bn_svhn, 10), (sepreresnet272bn_cifar10, 10), (sepreresnet272bn_cifar100, 100), (sepreresnet272bn_svhn, 10), (sepreresnet542bn_cifar10, 10), (sepreresnet542bn_cifar100, 100), (sepreresnet542bn_svhn, 10), (sepreresnet1001_cifar10, 10), (sepreresnet1001_cifar100, 100), (sepreresnet1001_svhn, 10), (sepreresnet1202_cifar10, 10), (sepreresnet1202_cifar100, 100), (sepreresnet1202_svhn, 10), ] for model, num_num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != sepreresnet20_cifar10 or weight_count == 274559) assert (model != sepreresnet20_cifar100 or weight_count == 280409) assert (model != sepreresnet20_svhn or weight_count == 274559) assert (model != sepreresnet56_cifar10 or weight_count == 862601) assert (model != sepreresnet56_cifar100 or weight_count == 868451) assert (model != sepreresnet56_svhn or weight_count == 862601) assert (model != sepreresnet110_cifar10 or weight_count == 1744664) assert (model != sepreresnet110_cifar100 or weight_count == 1750514) assert (model != sepreresnet110_svhn or weight_count == 1744664) assert (model != sepreresnet164bn_cifar10 or weight_count == 1904882) assert (model != sepreresnet164bn_cifar100 or weight_count == 1928012) assert (model != sepreresnet164bn_svhn or weight_count == 1904882) assert (model != sepreresnet272bn_cifar10 or weight_count == 3152450) assert (model != sepreresnet272bn_cifar100 or weight_count == 3175580) assert (model != sepreresnet272bn_svhn or weight_count == 3152450) assert (model != sepreresnet542bn_cifar10 or weight_count == 6271370) assert (model != sepreresnet542bn_cifar100 or weight_count == 6294500) assert (model != sepreresnet542bn_svhn or weight_count == 6271370) assert (model != sepreresnet1001_cifar10 or weight_count == 11573534) assert (model != sepreresnet1001_cifar100 or weight_count == 11596664) assert (model != sepreresnet1001_svhn or weight_count == 11573534) assert (model != sepreresnet1202_cifar10 or weight_count == 19581938) assert (model != sepreresnet1202_cifar100 or weight_count == 19587788) assert (model != sepreresnet1202_svhn or weight_count == 19581938) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_num_classes)) if __name__ == "__main__": _test()
24,663
37.298137
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/danet.py
""" DANet for image segmentation, implemented in Gluon. Original paper: 'Dual Attention Network for Scene Segmentation,' https://arxiv.org/abs/1809.02983. """ __all__ = ['DANet', 'danet_resnetd50b_cityscapes', 'danet_resnetd101b_cityscapes', 'ScaleBlock'] import os import torch import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from torch.nn.parameter import Parameter from .common import conv1x1, conv3x3_block from .resnetd import resnetd50b, resnetd101b class ScaleBlock(nn.Module): """ Simple scale block. """ def __init__(self): super(ScaleBlock, self).__init__() self.alpha = Parameter(torch.Tensor((1,))) def forward(self, x): return self.alpha * x def __repr__(self): s = '{name}(alpha={alpha})' return s.format( name=self.__class__.__name__, gamma=self.alpha.shape[0]) def calc_flops(self, x): assert (x.shape[0] == 1) num_flops = x.numel() num_macs = 0 return num_flops, num_macs class PosAttBlock(nn.Module): """ Position attention block from 'Dual Attention Network for Scene Segmentation,' https://arxiv.org/abs/1809.02983. It captures long-range spatial contextual information. Parameters: ---------- channels : int Number of channels. reduction : int, default 8 Squeeze reduction value. """ def __init__(self, channels, reduction=8): super(PosAttBlock, self).__init__() mid_channels = channels // reduction self.query_conv = conv1x1( in_channels=channels, out_channels=mid_channels, bias=True) self.key_conv = conv1x1( in_channels=channels, out_channels=mid_channels, bias=True) self.value_conv = conv1x1( in_channels=channels, out_channels=channels, bias=True) self.scale = ScaleBlock() self.softmax = nn.Softmax(dim=-1) def forward(self, x): batch, channels, height, width = x.shape proj_query = self.query_conv(x).view((batch, -1, height * width)) proj_key = self.key_conv(x).view((batch, -1, height * width)) proj_value = self.value_conv(x).view((batch, -1, height * width)) energy = proj_query.transpose(1, 2).contiguous().bmm(proj_key) w = self.softmax(energy) y = proj_value.bmm(w.transpose(1, 2).contiguous()) y = y.reshape((batch, -1, height, width)) y = self.scale(y) + x return y class ChaAttBlock(nn.Module): """ Channel attention block from 'Dual Attention Network for Scene Segmentation,' https://arxiv.org/abs/1809.02983. It explicitly models interdependencies between channels. """ def __init__(self): super(ChaAttBlock, self).__init__() self.scale = ScaleBlock() self.softmax = nn.Softmax(dim=-1) def forward(self, x): batch, channels, height, width = x.shape proj_query = x.view((batch, -1, height * width)) proj_key = x.view((batch, -1, height * width)) proj_value = x.view((batch, -1, height * width)) energy = proj_query.bmm(proj_key.transpose(1, 2).contiguous()) energy_max, _ = energy.max(dim=-1, keepdims=True) energy_new = energy_max.expand_as(energy) - energy w = self.softmax(energy_new) y = w.bmm(proj_value) y = y.reshape((batch, -1, height, width)) y = self.scale(y) + x return y class DANetHeadBranch(nn.Module): """ DANet head branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. pose_att : bool, default True Whether to use position attention instead of channel one. """ def __init__(self, in_channels, out_channels, pose_att=True): super(DANetHeadBranch, self).__init__() mid_channels = in_channels // 4 dropout_rate = 0.1 self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels) if pose_att: self.att = PosAttBlock(mid_channels) else: self.att = ChaAttBlock() self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels) self.conv3 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) self.dropout = nn.Dropout(p=dropout_rate, inplace=False) def forward(self, x): x = self.conv1(x) x = self.att(x) y = self.conv2(x) x = self.conv3(y) x = self.dropout(x) return x, y class DANetHead(nn.Module): """ DANet head block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(DANetHead, self).__init__() mid_channels = in_channels // 4 dropout_rate = 0.1 self.branch_pa = DANetHeadBranch( in_channels=in_channels, out_channels=out_channels, pose_att=True) self.branch_ca = DANetHeadBranch( in_channels=in_channels, out_channels=out_channels, pose_att=False) self.conv = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) self.dropout = nn.Dropout(p=dropout_rate, inplace=False) def forward(self, x): pa_x, pa_y = self.branch_pa(x) ca_x, ca_y = self.branch_ca(x) y = pa_y + ca_y x = self.conv(y) x = self.dropout(x) return x, pa_x, ca_x class DANet(nn.Module): """ DANet model from 'Dual Attention Network for Scene Segmentation,' https://arxiv.org/abs/1809.02983. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, backbone, backbone_out_channels=2048, aux=False, fixed_size=True, in_channels=3, in_size=(480, 480), num_classes=19): super(DANet, self).__init__() assert (in_channels > 0) assert ((in_size[0] % 8 == 0) and (in_size[1] % 8 == 0)) self.in_size = in_size self.num_classes = num_classes self.aux = aux self.fixed_size = fixed_size self.backbone = backbone self.head = DANetHead( in_channels=backbone_out_channels, out_channels=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): in_size = self.in_size if self.fixed_size else x.shape[2:] x, _ = self.backbone(x) x, y, z = self.head(x) x = F.interpolate(x, size=in_size, mode="bilinear", align_corners=True) if self.aux: y = F.interpolate(y, size=in_size, mode="bilinear", align_corners=True) z = F.interpolate(z, size=in_size, mode="bilinear", align_corners=True) return x, y, z else: return x def get_danet(backbone, num_classes, aux=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DANet model with specific parameters. Parameters: ---------- backbone : nn.Sequential Feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ net = DANet( backbone=backbone, num_classes=num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def danet_resnetd50b_cityscapes(pretrained_backbone=False, num_classes=19, aux=True, **kwargs): """ DANet model on the base of ResNet(D)-50b for Cityscapes from 'Dual Attention Network for Scene Segmentation,' https://arxiv.org/abs/1809.02983. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd50b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_danet(backbone=backbone, num_classes=num_classes, aux=aux, model_name="danet_resnetd50b_cityscapes", **kwargs) def danet_resnetd101b_cityscapes(pretrained_backbone=False, num_classes=19, aux=True, **kwargs): """ DANet model on the base of ResNet(D)-101b for Cityscapes from 'Dual Attention Network for Scene Segmentation,' https://arxiv.org/abs/1809.02983. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnetd101b(pretrained=pretrained_backbone, ordinary_init=False, bends=(3,)).features del backbone[-1] return get_danet(backbone=backbone, num_classes=num_classes, aux=aux, model_name="danet_resnetd101b_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch in_size = (480, 480) aux = True pretrained = False models = [ danet_resnetd50b_cityscapes, danet_resnetd101b_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, aux=aux) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != danet_resnetd50b_cityscapes or weight_count == 47586427) assert (model != danet_resnetd101b_cityscapes or weight_count == 66578555) batch = 2 num_classes = 19 x = torch.randn(batch, 3, in_size[0], in_size[1]) ys = net(x) y = ys[0] if aux else ys y.sum().backward() assert ((y.size(0) == x.size(0)) and (y.size(1) == num_classes) and (y.size(2) == x.size(2)) and (y.size(3) == x.size(3))) if __name__ == "__main__": _test()
12,721
30.568238
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/mobilenetv2.py
""" MobileNetV2 for ImageNet-1K, implemented in PyTorch. Original paper: 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. """ __all__ = ['MobileNetV2', 'mobilenetv2_w1', 'mobilenetv2_w3d4', 'mobilenetv2_wd2', 'mobilenetv2_wd4', 'mobilenetv2b_w1', 'mobilenetv2b_w3d4', 'mobilenetv2b_wd2', 'mobilenetv2b_wd4'] import os import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv1x1_block, conv3x3_block, dwconv3x3_block class LinearBottleneck(nn.Module): """ So-called 'Linear Bottleneck' layer. It is used as a MobileNetV2 unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the second convolution layer. expansion : bool Whether do expansion of channels. remove_exp_conv : bool Whether to remove expansion convolution. """ def __init__(self, in_channels, out_channels, stride, expansion, remove_exp_conv): super(LinearBottleneck, self).__init__() self.residual = (in_channels == out_channels) and (stride == 1) mid_channels = in_channels * 6 if expansion else in_channels self.use_exp_conv = (expansion or (not remove_exp_conv)) if self.use_exp_conv: self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, activation="relu6") self.conv2 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, activation="relu6") self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): if self.residual: identity = x if self.use_exp_conv: x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) if self.residual: x = x + identity return x class MobileNetV2(nn.Module): """ MobileNetV2 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. 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. remove_exp_conv : bool Whether to remove expansion convolution. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, remove_exp_conv, in_channels=3, in_size=(224, 224), num_classes=1000): super(MobileNetV2, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2, activation="relu6")) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 expansion = (i != 0) or (j != 0) stage.add_module("unit{}".format(j + 1), LinearBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, expansion=expansion, remove_exp_conv=remove_exp_conv)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, activation="relu6")) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = conv1x1( in_channels=in_channels, out_channels=num_classes, bias=False) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_mobilenetv2(width_scale, remove_exp_conv=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create MobileNetV2 model with specific parameters. Parameters: ---------- width_scale : float Scale factor for width of layers. remove_exp_conv : bool, default False Whether to remove expansion convolution. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 32 final_block_channels = 1280 layers = [1, 2, 3, 4, 3, 3, 1] downsample = [0, 1, 1, 1, 0, 1, 0] channels_per_layers = [16, 24, 32, 64, 96, 160, 320] from functools import reduce channels = reduce( lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(channels_per_layers, layers, downsample), [[]]) 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) if width_scale > 1.0: final_block_channels = int(final_block_channels * width_scale) net = MobileNetV2( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, remove_exp_conv=remove_exp_conv, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def mobilenetv2_w1(**kwargs): """ 1.0 MobileNetV2-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=1.0, model_name="mobilenetv2_w1", **kwargs) def mobilenetv2_w3d4(**kwargs): """ 0.75 MobileNetV2-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=0.75, model_name="mobilenetv2_w3d4", **kwargs) def mobilenetv2_wd2(**kwargs): """ 0.5 MobileNetV2-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=0.5, model_name="mobilenetv2_wd2", **kwargs) def mobilenetv2_wd4(**kwargs): """ 0.25 MobileNetV2-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=0.25, model_name="mobilenetv2_wd4", **kwargs) def mobilenetv2b_w1(**kwargs): """ 1.0 MobileNetV2b-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=1.0, remove_exp_conv=True, model_name="mobilenetv2b_w1", **kwargs) def mobilenetv2b_w3d4(**kwargs): """ 0.75 MobileNetV2b-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=0.75, remove_exp_conv=True, model_name="mobilenetv2b_w3d4", **kwargs) def mobilenetv2b_wd2(**kwargs): """ 0.5 MobileNetV2b-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=0.5, remove_exp_conv=True, model_name="mobilenetv2b_wd2", **kwargs) def mobilenetv2b_wd4(**kwargs): """ 0.25 MobileNetV2b-224 model from 'MobileNetV2: Inverted Residuals and Linear Bottlenecks,' https://arxiv.org/abs/1801.04381. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv2(width_scale=0.25, remove_exp_conv=True, model_name="mobilenetv2b_wd4", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ mobilenetv2_w1, mobilenetv2_w3d4, mobilenetv2_wd2, mobilenetv2_wd4, mobilenetv2b_w1, mobilenetv2b_w3d4, mobilenetv2b_wd2, mobilenetv2b_wd4, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mobilenetv2_w1 or weight_count == 3504960) assert (model != mobilenetv2_w3d4 or weight_count == 2627592) assert (model != mobilenetv2_wd2 or weight_count == 1964736) assert (model != mobilenetv2_wd4 or weight_count == 1516392) assert (model != mobilenetv2b_w1 or weight_count == 3503872) assert (model != mobilenetv2b_w3d4 or weight_count == 2626968) assert (model != mobilenetv2b_wd2 or weight_count == 1964448) assert (model != mobilenetv2b_wd4 or weight_count == 1516312) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,761
32.321149
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/squeezenet.py
""" SqueezeNet for ImageNet-1K, implemented in PyTorch. Original paper: 'SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,' https://arxiv.org/abs/1602.07360. """ __all__ = ['SqueezeNet', 'squeezenet_v1_0', 'squeezenet_v1_1', 'squeezeresnet_v1_0', 'squeezeresnet_v1_1'] import os import torch import torch.nn as nn import torch.nn.init as init class FireConv(nn.Module): """ SqueezeNet 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. padding : int or tuple/list of 2 int Padding value for convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, padding): super(FireConv, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, padding=padding) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.conv(x) x = self.activ(x) return x class FireUnit(nn.Module): """ SqueezeNet unit, so-called 'Fire' unit. Parameters: ---------- in_channels : int Number of input channels. squeeze_channels : int Number of output channels for squeeze convolution blocks. expand1x1_channels : int Number of output channels for expand 1x1 convolution blocks. expand3x3_channels : int Number of output channels for expand 3x3 convolution blocks. residual : bool Whether use residual connection. """ def __init__(self, in_channels, squeeze_channels, expand1x1_channels, expand3x3_channels, residual): super(FireUnit, self).__init__() self.residual = residual self.squeeze = FireConv( in_channels=in_channels, out_channels=squeeze_channels, kernel_size=1, padding=0) self.expand1x1 = FireConv( in_channels=squeeze_channels, out_channels=expand1x1_channels, kernel_size=1, padding=0) self.expand3x3 = FireConv( in_channels=squeeze_channels, out_channels=expand3x3_channels, kernel_size=3, padding=1) def forward(self, x): if self.residual: identity = x x = self.squeeze(x) y1 = self.expand1x1(x) y2 = self.expand3x3(x) out = torch.cat((y1, y2), dim=1) if self.residual: out = out + identity return out class SqueezeInitBlock(nn.Module): """ SqueezeNet 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. """ def __init__(self, in_channels, out_channels, kernel_size): super(SqueezeInitBlock, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=2) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.conv(x) x = self.activ(x) return x class SqueezeNet(nn.Module): """ SqueezeNet model from 'SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,' https://arxiv.org/abs/1602.07360. Parameters: ---------- channels : list of list of int Number of output channels for each unit. residuals : bool Whether to use residual units. init_block_kernel_size : int or tuple/list of 2 int The dimensions of the convolution window for the initial unit. init_block_channels : int Number of output channels for the initial unit. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, residuals, init_block_kernel_size, init_block_channels, in_channels=3, in_size=(224, 224), num_classes=1000): super(SqueezeNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", SqueezeInitBlock( in_channels=in_channels, out_channels=init_block_channels, kernel_size=init_block_kernel_size)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() stage.add_module("pool{}".format(i + 1), nn.MaxPool2d( kernel_size=3, stride=2, ceil_mode=True)) for j, out_channels in enumerate(channels_per_stage): expand_channels = out_channels // 2 squeeze_channels = out_channels // 8 stage.add_module("unit{}".format(j + 1), FireUnit( in_channels=in_channels, squeeze_channels=squeeze_channels, expand1x1_channels=expand_channels, expand3x3_channels=expand_channels, residual=((residuals is not None) and (residuals[i][j] == 1)))) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("dropout", nn.Dropout(p=0.5)) self.output = nn.Sequential() self.output.add_module("final_conv", nn.Conv2d( in_channels=in_channels, out_channels=num_classes, kernel_size=1)) self.output.add_module("final_activ", nn.ReLU(inplace=True)) self.output.add_module("final_pool", nn.AvgPool2d( kernel_size=13, stride=1)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): if 'final_conv' in name: init.normal_(module.weight, mean=0.0, std=0.01) else: init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_squeezenet(version, residual=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SqueezeNet model with specific parameters. Parameters: ---------- version : str Version of SqueezeNet ('1.0' or '1.1'). residual : bool, default False Whether to use residual connections. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if version == '1.0': channels = [[128, 128, 256], [256, 384, 384, 512], [512]] residuals = [[0, 1, 0], [1, 0, 1, 0], [1]] init_block_kernel_size = 7 init_block_channels = 96 elif version == '1.1': channels = [[128, 128], [256, 256], [384, 384, 512, 512]] residuals = [[0, 1], [0, 1], [0, 1, 0, 1]] init_block_kernel_size = 3 init_block_channels = 64 else: raise ValueError("Unsupported SqueezeNet version {}".format(version)) if not residual: residuals = None net = SqueezeNet( channels=channels, residuals=residuals, init_block_kernel_size=init_block_kernel_size, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def squeezenet_v1_0(**kwargs): """ SqueezeNet 'vanilla' model from 'SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,' https://arxiv.org/abs/1602.07360. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenet(version="1.0", residual=False, model_name="squeezenet_v1_0", **kwargs) def squeezenet_v1_1(**kwargs): """ SqueezeNet v1.1 model from 'SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,' https://arxiv.org/abs/1602.07360. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenet(version="1.1", residual=False, model_name="squeezenet_v1_1", **kwargs) def squeezeresnet_v1_0(**kwargs): """ SqueezeNet model with residual connections from 'SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,' https://arxiv.org/abs/1602.07360. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenet(version="1.0", residual=True, model_name="squeezeresnet_v1_0", **kwargs) def squeezeresnet_v1_1(**kwargs): """ SqueezeNet v1.1 model with residual connections from 'SqueezeNet: AlexNet-level accuracy with 50x fewer parameters and <0.5MB model size,' https://arxiv.org/abs/1602.07360. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_squeezenet(version="1.1", residual=True, model_name="squeezeresnet_v1_1", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False models = [ squeezenet_v1_0, squeezenet_v1_1, squeezeresnet_v1_0, squeezeresnet_v1_1, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != squeezenet_v1_0 or weight_count == 1248424) assert (model != squeezenet_v1_1 or weight_count == 1235496) assert (model != squeezeresnet_v1_0 or weight_count == 1248424) assert (model != squeezeresnet_v1_1 or weight_count == 1235496) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,164
30.929134
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/nin_cifar.py
""" NIN for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Network In Network,' https://arxiv.org/abs/1312.4400. """ __all__ = ['CIFARNIN', 'nin_cifar10', 'nin_cifar100', 'nin_svhn'] import os import torch.nn as nn import torch.nn.init as init class NINConv(nn.Module): """ NIN 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. stride : 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, stride=1, padding=0): super(NINConv, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=True) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.conv(x) x = self.activ(x) return x class CIFARNIN(nn.Module): """ NIN model for CIFAR from 'Network In Network,' https://arxiv.org/abs/1312.4400. Parameters: ---------- channels : list of list of int Number of output channels for each unit. first_kernel_sizes : list of int Convolution window sizes for the first units in each stage. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, first_kernel_sizes, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARNIN, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): if (j == 0) and (i != 0): if i == 1: stage.add_module("pool{}".format(i + 1), nn.MaxPool2d( kernel_size=3, stride=2, padding=1)) else: stage.add_module("pool{}".format(i + 1), nn.AvgPool2d( kernel_size=3, stride=2, padding=1)) stage.add_module("dropout{}".format(i + 1), nn.Dropout(p=0.5)) kernel_size = first_kernel_sizes[i] if j == 0 else 1 padding = (kernel_size - 1) // 2 stage.add_module("unit{}".format(j + 1), NINConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, padding=padding)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.output = nn.Sequential() self.output.add_module("final_conv", NINConv( in_channels=in_channels, out_channels=num_classes, kernel_size=1)) self.output.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_nin_cifar(num_classes, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create NIN 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [[192, 160, 96], [192, 192, 192], [192, 192]] first_kernel_sizes = [5, 5, 3] net = CIFARNIN( channels=channels, first_kernel_sizes=first_kernel_sizes, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def nin_cifar10(num_classes=10, **kwargs): """ NIN model for CIFAR-10 from 'Network In Network,' https://arxiv.org/abs/1312.4400. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_nin_cifar(num_classes=num_classes, model_name="nin_cifar10", **kwargs) def nin_cifar100(num_classes=100, **kwargs): """ NIN model for CIFAR-100 from 'Network In Network,' https://arxiv.org/abs/1312.4400. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_nin_cifar(num_classes=num_classes, model_name="nin_cifar100", **kwargs) def nin_svhn(num_classes=10, **kwargs): """ NIN model for SVHN from 'Network In Network,' https://arxiv.org/abs/1312.4400. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_nin_cifar(num_classes=num_classes, model_name="nin_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (nin_cifar10, 10), (nin_cifar100, 100), (nin_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != nin_cifar10 or weight_count == 966986) assert (model != nin_cifar100 or weight_count == 984356) assert (model != nin_svhn or weight_count == 966986) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
8,048
29.957692
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/vgg.py
""" VGG for ImageNet-1K, implemented in PyTorch. Original paper: 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. """ __all__ = ['VGG', 'vgg11', 'vgg13', 'vgg16', 'vgg19', 'bn_vgg11', 'bn_vgg13', 'bn_vgg16', 'bn_vgg19', 'bn_vgg11b', 'bn_vgg13b', 'bn_vgg16b', 'bn_vgg19b'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3_block class VGGDense(nn.Module): """ VGG 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): super(VGGDense, self).__init__() self.fc = nn.Linear( in_features=in_channels, out_features=out_channels) self.activ = nn.ReLU(inplace=True) self.dropout = nn.Dropout(p=0.5) def forward(self, x): x = self.fc(x) x = self.activ(x) x = self.dropout(x) return x class VGGOutputBlock(nn.Module): """ VGG specific output block. Parameters: ---------- in_channels : int Number of input channels. classes : int Number of classification classes. """ def __init__(self, in_channels, classes): super(VGGOutputBlock, self).__init__() mid_channels = 4096 self.fc1 = VGGDense( in_channels=in_channels, out_channels=mid_channels) self.fc2 = VGGDense( in_channels=mid_channels, out_channels=mid_channels) self.fc3 = nn.Linear( in_features=mid_channels, out_features=classes) def forward(self, x): x = self.fc1(x) x = self.fc2(x) x = self.fc3(x) return x class VGG(nn.Module): """ VGG models from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- channels : list of list of int Number of output channels for each unit. bias : bool, default True Whether the convolution layer uses a bias vector. use_bn : bool, default False Whether to use 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, bias=True, use_bn=False, in_channels=3, in_size=(224, 224), num_classes=1000): super(VGG, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stage.add_module("unit{}".format(j + 1), conv3x3_block( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=use_bn)) in_channels = out_channels stage.add_module("pool{}".format(i + 1), nn.MaxPool2d( kernel_size=2, stride=2, padding=0)) self.features.add_module("stage{}".format(i + 1), stage) self.output = VGGOutputBlock( in_channels=(in_channels * 7 * 7), classes=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_vgg(blocks, bias=True, use_bn=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create VGG model with specific parameters. Parameters: ---------- blocks : int Number of blocks. bias : bool, default True Whether the convolution layer uses a bias vector. use_bn : bool, default False Whether to use BatchNorm 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 11: layers = [1, 1, 2, 2, 2] elif blocks == 13: layers = [2, 2, 2, 2, 2] elif blocks == 16: layers = [2, 2, 3, 3, 3] elif blocks == 19: layers = [2, 2, 4, 4, 4] else: raise ValueError("Unsupported VGG with number of blocks: {}".format(blocks)) channels_per_layers = [64, 128, 256, 512, 512] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = VGG( channels=channels, bias=bias, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def vgg11(**kwargs): """ VGG-11 model from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=11, model_name="vgg11", **kwargs) def vgg13(**kwargs): """ VGG-13 model from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=13, model_name="vgg13", **kwargs) def vgg16(**kwargs): """ VGG-16 model from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=16, model_name="vgg16", **kwargs) def vgg19(**kwargs): """ VGG-19 model from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=19, model_name="vgg19", **kwargs) def bn_vgg11(**kwargs): """ VGG-11 model with batch normalization from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=11, bias=False, use_bn=True, model_name="bn_vgg11", **kwargs) def bn_vgg13(**kwargs): """ VGG-13 model with batch normalization from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=13, bias=False, use_bn=True, model_name="bn_vgg13", **kwargs) def bn_vgg16(**kwargs): """ VGG-16 model with batch normalization from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=16, bias=False, use_bn=True, model_name="bn_vgg16", **kwargs) def bn_vgg19(**kwargs): """ VGG-19 model with batch normalization from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=19, bias=False, use_bn=True, model_name="bn_vgg19", **kwargs) def bn_vgg11b(**kwargs): """ VGG-11 model with batch normalization and biases in convolution layers from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=11, bias=True, use_bn=True, model_name="bn_vgg11b", **kwargs) def bn_vgg13b(**kwargs): """ VGG-13 model with batch normalization and biases in convolution layers from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=13, bias=True, use_bn=True, model_name="bn_vgg13b", **kwargs) def bn_vgg16b(**kwargs): """ VGG-16 model with batch normalization and biases in convolution layers from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=16, bias=True, use_bn=True, model_name="bn_vgg16b", **kwargs) def bn_vgg19b(**kwargs): """ VGG-19 model with batch normalization and biases in convolution layers from 'Very Deep Convolutional Networks for Large-Scale Image Recognition,' https://arxiv.org/abs/1409.1556. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vgg(blocks=19, bias=True, use_bn=True, model_name="bn_vgg19b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ vgg11, vgg13, vgg16, vgg19, bn_vgg11, bn_vgg13, bn_vgg16, bn_vgg19, bn_vgg11b, bn_vgg13b, bn_vgg16b, bn_vgg19b, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != vgg11 or weight_count == 132863336) assert (model != vgg13 or weight_count == 133047848) assert (model != vgg16 or weight_count == 138357544) assert (model != vgg19 or weight_count == 143667240) assert (model != bn_vgg11 or weight_count == 132866088) assert (model != bn_vgg13 or weight_count == 133050792) assert (model != bn_vgg16 or weight_count == 138361768) assert (model != bn_vgg19 or weight_count == 143672744) assert (model != bn_vgg11b or weight_count == 132868840) assert (model != bn_vgg13b or weight_count == 133053736) assert (model != bn_vgg16b or weight_count == 138365992) assert (model != bn_vgg19b or weight_count == 143678248) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
13,528
29.678005
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resnet_cub.py
""" ResNet for CUB-200-2011, implemented in PyTorch. Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. """ __all__ = ['resnet10_cub', 'resnet12_cub', 'resnet14_cub', 'resnetbc14b_cub', 'resnet16_cub', 'resnet18_cub', 'resnet26_cub', 'resnetbc26b_cub', 'resnet34_cub', 'resnetbc38b_cub', 'resnet50_cub', 'resnet50b_cub', 'resnet101_cub', 'resnet101b_cub', 'resnet152_cub', 'resnet152b_cub', 'resnet200_cub', 'resnet200b_cub'] from .resnet import get_resnet def resnet10_cub(num_classes=200, **kwargs): """ ResNet-10 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=10, model_name="resnet10_cub", **kwargs) def resnet12_cub(num_classes=200, **kwargs): """ ResNet-12 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=12, model_name="resnet12_cub", **kwargs) def resnet14_cub(num_classes=200, **kwargs): """ ResNet-14 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=14, model_name="resnet14_cub", **kwargs) def resnetbc14b_cub(num_classes=200, **kwargs): """ ResNet-BC-14b model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model (bottleneck compressed). Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=14, bottleneck=True, conv1_stride=False, model_name="resnetbc14b_cub", **kwargs) def resnet16_cub(num_classes=200, **kwargs): """ ResNet-16 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=16, model_name="resnet16_cub", **kwargs) def resnet18_cub(num_classes=200, **kwargs): """ ResNet-18 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=18, model_name="resnet18_cub", **kwargs) def resnet26_cub(num_classes=200, **kwargs): """ ResNet-26 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=26, bottleneck=False, model_name="resnet26_cub", **kwargs) def resnetbc26b_cub(num_classes=200, **kwargs): """ ResNet-BC-26b model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model (bottleneck compressed). Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=26, bottleneck=True, conv1_stride=False, model_name="resnetbc26b_cub", **kwargs) def resnet34_cub(num_classes=200, **kwargs): """ ResNet-34 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=34, model_name="resnet34_cub", **kwargs) def resnetbc38b_cub(num_classes=200, **kwargs): """ ResNet-BC-38b model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model (bottleneck compressed). Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=38, bottleneck=True, conv1_stride=False, model_name="resnetbc38b_cub", **kwargs) def resnet50_cub(num_classes=200, **kwargs): """ ResNet-50 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=50, model_name="resnet50_cub", **kwargs) def resnet50b_cub(num_classes=200, **kwargs): """ ResNet-50 model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=50, conv1_stride=False, model_name="resnet50b_cub", **kwargs) def resnet101_cub(num_classes=200, **kwargs): """ ResNet-101 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=101, model_name="resnet101_cub", **kwargs) def resnet101b_cub(num_classes=200, **kwargs): """ ResNet-101 model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=101, conv1_stride=False, model_name="resnet101b_cub", **kwargs) def resnet152_cub(num_classes=200, **kwargs): """ ResNet-152 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=152, model_name="resnet152_cub", **kwargs) def resnet152b_cub(num_classes=200, **kwargs): """ ResNet-152 model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=152, conv1_stride=False, model_name="resnet152b_cub", **kwargs) def resnet200_cub(num_classes=200, **kwargs): """ ResNet-200 model for CUB-200-2011 from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=200, model_name="resnet200_cub", **kwargs) def resnet200b_cub(num_classes=200, **kwargs): """ ResNet-200 model with stride at the second convolution in bottleneck block from 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. It's an experimental model. Parameters: ---------- classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(num_classes=num_classes, blocks=200, conv1_stride=False, model_name="resnet200b_cub", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ resnet10_cub, resnet12_cub, resnet14_cub, resnetbc14b_cub, resnet16_cub, resnet18_cub, resnet26_cub, resnetbc26b_cub, resnet34_cub, resnetbc38b_cub, resnet50_cub, resnet50b_cub, resnet101_cub, resnet101b_cub, resnet152_cub, resnet152b_cub, resnet200_cub, resnet200b_cub, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != resnet10_cub or weight_count == 5008392) assert (model != resnet12_cub or weight_count == 5082376) assert (model != resnet14_cub or weight_count == 5377800) assert (model != resnetbc14b_cub or weight_count == 8425736) assert (model != resnet16_cub or weight_count == 6558472) assert (model != resnet18_cub or weight_count == 11279112) assert (model != resnet26_cub or weight_count == 17549832) assert (model != resnetbc26b_cub or weight_count == 14355976) assert (model != resnet34_cub or weight_count == 21387272) assert (model != resnetbc38b_cub or weight_count == 20286216) assert (model != resnet50_cub or weight_count == 23917832) assert (model != resnet50b_cub or weight_count == 23917832) assert (model != resnet101_cub or weight_count == 42909960) assert (model != resnet101b_cub or weight_count == 42909960) assert (model != resnet152_cub or weight_count == 58553608) assert (model != resnet152b_cub or weight_count == 58553608) assert (model != resnet200_cub or weight_count == 63034632) assert (model != resnet200b_cub or weight_count == 63034632) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 200)) if __name__ == "__main__": _test()
14,148
35.094388
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/bagnet.py
""" BagNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,' https://openreview.net/pdf?id=SkfMWhAqYQ. """ __all__ = ['BagNet', 'bagnet9', 'bagnet17', 'bagnet33'] import os import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv1x1_block, conv3x3_block, ConvBlock class BagNetBottleneck(nn.Module): """ BagNet bottleneck block for residual path in BagNet unit. 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 of the second convolution. stride : int or tuple/list of 2 int Strides of the second convolution. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, kernel_size, stride, bottleneck_factor=4): super(BagNetBottleneck, self).__init__() mid_channels = out_channels // bottleneck_factor self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = ConvBlock( in_channels=mid_channels, out_channels=mid_channels, kernel_size=kernel_size, stride=stride, padding=0) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class BagNetUnit(nn.Module): """ BagNet unit. 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 of the second body convolution. stride : int or tuple/list of 2 int Strides of the second body convolution. """ def __init__(self, in_channels, out_channels, kernel_size, stride): super(BagNetUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = BagNetBottleneck( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) if x.size(-1) != identity.size(-1): diff = identity.size(-1) - x.size(-1) identity = identity[:, :, :-diff, :-diff] x = x + identity x = self.activ(x) return x class BagNetInitBlock(nn.Module): """ BagNet 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): super(BagNetInitBlock, self).__init__() self.conv1 = conv1x1( in_channels=in_channels, out_channels=out_channels) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, padding=0) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class BagNet(nn.Module): """ BagNet model from 'Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,' https://openreview.net/pdf?id=SkfMWhAqYQ. 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_pool_size : int Size of the pooling windows for final pool. normal_kernel_sizes : list of int Count of the first units with 3x3 convolution window size for each stage. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_pool_size, normal_kernel_sizes, in_channels=3, in_size=(224, 224), num_classes=1000): super(BagNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", BagNetInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != len(channels) - 1) else 1 kernel_size = 3 if j < normal_kernel_sizes[i] else 1 stage.add_module("unit{}".format(j + 1), BagNetUnit( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=final_pool_size, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_bagnet(field, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create BagNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ layers = [3, 4, 6, 3] if field == 9: normal_kernel_sizes = [1, 1, 0, 0] final_pool_size = 27 elif field == 17: normal_kernel_sizes = [1, 1, 1, 0] final_pool_size = 26 elif field == 33: normal_kernel_sizes = [1, 1, 1, 1] final_pool_size = 24 else: raise ValueError("Unsupported BagNet with field: {}".format(field)) init_block_channels = 64 channels_per_layers = [256, 512, 1024, 2048] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = BagNet( channels=channels, init_block_channels=init_block_channels, final_pool_size=final_pool_size, normal_kernel_sizes=normal_kernel_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def bagnet9(**kwargs): """ BagNet-9 model from 'Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,' https://openreview.net/pdf?id=SkfMWhAqYQ. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_bagnet(field=9, model_name="bagnet9", **kwargs) def bagnet17(**kwargs): """ BagNet-17 model from 'Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,' https://openreview.net/pdf?id=SkfMWhAqYQ. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_bagnet(field=17, model_name="bagnet17", **kwargs) def bagnet33(**kwargs): """ BagNet-33 model from 'Approximating CNNs with Bag-of-local-Features models works surprisingly well on ImageNet,' https://openreview.net/pdf?id=SkfMWhAqYQ. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_bagnet(field=33, model_name="bagnet33", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ bagnet9, bagnet17, bagnet33, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != bagnet9 or weight_count == 15688744) assert (model != bagnet17 or weight_count == 16213032) assert (model != bagnet33 or weight_count == 18310184) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
10,903
29.373259
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/airnet.py
""" AirNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Attention Inspiring Receptive-Fields Network for Learning Invariant Representations,' https://ieeexplore.ieee.org/document/8510896. """ __all__ = ['AirNet', 'airnet50_1x64d_r2', 'airnet50_1x64d_r16', 'airnet101_1x64d_r2', 'AirBlock', 'AirInitBlock'] import os import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from .common import conv1x1_block, conv3x3_block class AirBlock(nn.Module): """ AirNet attention block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. groups : int, default 1 Number of groups. ratio: int, default 2 Air compression ratio. """ def __init__(self, in_channels, out_channels, groups=1, ratio=2): super(AirBlock, self).__init__() assert (out_channels % ratio == 0) mid_channels = out_channels // ratio self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, groups=groups) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) self.sigmoid = nn.Sigmoid() def forward(self, x): x = self.conv1(x) x = self.pool(x) x = self.conv2(x) x = F.interpolate( input=x, scale_factor=2, mode="bilinear", align_corners=True) x = self.conv3(x) x = self.sigmoid(x) return x class AirBottleneck(nn.Module): """ AirNet bottleneck block for residual path in AirNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. ratio: int Air compression ratio. """ def __init__(self, in_channels, out_channels, stride, ratio): super(AirBottleneck, self).__init__() mid_channels = out_channels // 4 self.use_air_block = (stride == 1 and mid_channels < 512) self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) if self.use_air_block: self.air = AirBlock( in_channels=in_channels, out_channels=mid_channels, ratio=ratio) def forward(self, x): if self.use_air_block: att = self.air(x) x = self.conv1(x) x = self.conv2(x) if self.use_air_block: x = x * att x = self.conv3(x) return x class AirUnit(nn.Module): """ AirNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. ratio: int Air compression ratio. """ def __init__(self, in_channels, out_channels, stride, ratio): super(AirUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = AirBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, ratio=ratio) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 AirInitBlock(nn.Module): """ AirNet 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): super(AirInitBlock, self).__init__() mid_channels = out_channels // 2 self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels) self.conv3 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.pool(x) return x class AirNet(nn.Module): """ AirNet model from 'Attention Inspiring Receptive-Fields Network for Learning Invariant Representations,' https://ieeexplore.ieee.org/document/8510896. 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. ratio: int Air compression ratio. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, ratio, in_channels=3, in_size=(224, 224), num_classes=1000): super(AirNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", AirInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), AirUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, ratio=ratio)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_airnet(blocks, base_channels, ratio, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create AirNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. base_channels: int Base number of channels. ratio: int Air compression ratio. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] else: raise ValueError("Unsupported AirNet with number of blocks: {}".format(blocks)) bottleneck_expansion = 4 init_block_channels = base_channels channels_per_layers = [base_channels * (2 ** i) * bottleneck_expansion for i in range(len(layers))] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = AirNet( channels=channels, init_block_channels=init_block_channels, ratio=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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def airnet50_1x64d_r2(**kwargs): """ AirNet50-1x64d (r=2) model from 'Attention Inspiring Receptive-Fields Network for Learning Invariant Representations,' https://ieeexplore.ieee.org/document/8510896. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_airnet(blocks=50, base_channels=64, ratio=2, model_name="airnet50_1x64d_r2", **kwargs) def airnet50_1x64d_r16(**kwargs): """ AirNet50-1x64d (r=16) model from 'Attention Inspiring Receptive-Fields Network for Learning Invariant Representations,' https://ieeexplore.ieee.org/document/8510896. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_airnet(blocks=50, base_channels=64, ratio=16, model_name="airnet50_1x64d_r16", **kwargs) def airnet101_1x64d_r2(**kwargs): """ AirNet101-1x64d (r=2) model from 'Attention Inspiring Receptive-Fields Network for Learning Invariant Representations,' https://ieeexplore.ieee.org/document/8510896. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_airnet(blocks=101, base_channels=64, ratio=2, model_name="airnet101_1x64d_r2", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ airnet50_1x64d_r2, airnet50_1x64d_r16, airnet101_1x64d_r2, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != airnet50_1x64d_r2 or weight_count == 27425864) assert (model != airnet50_1x64d_r16 or weight_count == 25714952) assert (model != airnet101_1x64d_r2 or weight_count == 51727432) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,525
28.612293
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/mnasnet.py
""" MnasNet for ImageNet-1K, implemented in PyTorch. Original paper: 'MnasNet: Platform-Aware Neural Architecture Search for Mobile,' https://arxiv.org/abs/1807.11626. """ __all__ = ['MnasNet', 'mnasnet_b1', 'mnasnet_a1', 'mnasnet_small'] import os import torch.nn as nn import torch.nn.init as init from .common import round_channels, conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block, SEBlock class DwsExpSEResUnit(nn.Module): """ Depthwise separable expanded residual unit with SE-block. Here it used as MnasNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the second convolution layer. use_kernel3 : bool, default True Whether to use 3x3 (instead of 5x5) kernel. exp_factor : int, default 1 Expansion factor for each unit. se_factor : int, default 0 SE reduction factor for each unit. use_skip : bool, default True Whether to use skip connection. activation : str, default 'relu' Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, stride=1, use_kernel3=True, exp_factor=1, se_factor=0, use_skip=True, activation="relu"): super(DwsExpSEResUnit, self).__init__() assert (exp_factor >= 1) self.residual = (in_channels == out_channels) and (stride == 1) and use_skip self.use_exp_conv = exp_factor > 1 self.use_se = se_factor > 0 mid_channels = exp_factor * in_channels dwconv_block_fn = dwconv3x3_block if use_kernel3 else dwconv5x5_block if self.use_exp_conv: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, activation=activation) self.dw_conv = dwconv_block_fn( in_channels=mid_channels, out_channels=mid_channels, stride=stride, activation=activation) if self.use_se: self.se = SEBlock( channels=mid_channels, reduction=(exp_factor * se_factor), round_mid=False, mid_activation=activation) self.pw_conv = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): if self.residual: identity = x if self.use_exp_conv: x = self.exp_conv(x) x = self.dw_conv(x) if self.use_se: x = self.se(x) x = self.pw_conv(x) if self.residual: x = x + identity return x class MnasInitBlock(nn.Module): """ MnasNet 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. use_skip : bool Whether to use skip connection in the second block. """ def __init__(self, in_channels, out_channels, mid_channels, use_skip): super(MnasInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2) self.conv2 = DwsExpSEResUnit( in_channels=mid_channels, out_channels=out_channels, use_skip=use_skip) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class MnasFinalBlock(nn.Module): """ MnasNet 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. use_skip : bool Whether to use skip connection in the second block. """ def __init__(self, in_channels, out_channels, mid_channels, use_skip): super(MnasFinalBlock, self).__init__() self.conv1 = DwsExpSEResUnit( in_channels=in_channels, out_channels=mid_channels, exp_factor=6, use_skip=use_skip) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class MnasNet(nn.Module): """ MnasNet model from 'MnasNet: Platform-Aware Neural Architecture Search for Mobile,' https://arxiv.org/abs/1807.11626. 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. se_factors : list of list of int SE reduction factor for each unit. init_block_use_skip : bool Whether to use skip connection in the initial unit. final_block_use_skip : bool Whether to use skip connection in the final block of the feature extractor. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, kernels3, exp_factors, se_factors, init_block_use_skip, final_block_use_skip, in_channels=3, in_size=(224, 224), num_classes=1000): super(MnasNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", MnasInitBlock( in_channels=in_channels, out_channels=init_block_channels[1], mid_channels=init_block_channels[0], use_skip=init_block_use_skip)) in_channels = init_block_channels[1] for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) else 1 use_kernel3 = kernels3[i][j] == 1 exp_factor = exp_factors[i][j] se_factor = se_factors[i][j] stage.add_module("unit{}".format(j + 1), DwsExpSEResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, use_kernel3=use_kernel3, exp_factor=exp_factor, se_factor=se_factor)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", MnasFinalBlock( in_channels=in_channels, out_channels=final_block_channels[1], mid_channels=final_block_channels[0], use_skip=final_block_use_skip)) in_channels = final_block_channels[1] self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_mnasnet(version, width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create MnasNet model with specific parameters. Parameters: ---------- version : str Version of MobileNetV3 ('b1', 'a1' or 'small'). 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if version == "b1": init_block_channels = [32, 16] final_block_channels = [320, 1280] channels = [[24, 24, 24], [40, 40, 40], [80, 80, 80, 96, 96], [192, 192, 192, 192]] kernels3 = [[1, 1, 1], [0, 0, 0], [0, 0, 0, 1, 1], [0, 0, 0, 0]] exp_factors = [[3, 3, 3], [3, 3, 3], [6, 6, 6, 6, 6], [6, 6, 6, 6]] se_factors = [[0, 0, 0], [0, 0, 0], [0, 0, 0, 0, 0], [0, 0, 0, 0]] init_block_use_skip = False final_block_use_skip = False elif version == "a1": init_block_channels = [32, 16] final_block_channels = [320, 1280] channels = [[24, 24], [40, 40, 40], [80, 80, 80, 80, 112, 112], [160, 160, 160]] kernels3 = [[1, 1], [0, 0, 0], [1, 1, 1, 1, 1, 1], [0, 0, 0]] exp_factors = [[6, 6], [3, 3, 3], [6, 6, 6, 6, 6, 6], [6, 6, 6]] se_factors = [[0, 0], [4, 4, 4], [0, 0, 0, 0, 4, 4], [4, 4, 4]] init_block_use_skip = False final_block_use_skip = True elif version == "small": init_block_channels = [8, 8] final_block_channels = [144, 1280] channels = [[16], [16, 16], [32, 32, 32, 32, 32, 32, 32], [88, 88, 88]] kernels3 = [[1], [1, 1], [0, 0, 0, 0, 1, 1, 1], [0, 0, 0]] exp_factors = [[3], [6, 6], [6, 6, 6, 6, 6, 6, 6], [6, 6, 6]] se_factors = [[0], [0, 0], [4, 4, 4, 4, 4, 4, 4], [4, 4, 4]] init_block_use_skip = True final_block_use_skip = True else: raise ValueError("Unsupported MnasNet version {}".format(version)) if width_scale != 1.0: channels = [[round_channels(cij * width_scale) for cij in ci] for ci in channels] init_block_channels = round_channels(init_block_channels * width_scale) net = MnasNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, kernels3=kernels3, exp_factors=exp_factors, se_factors=se_factors, init_block_use_skip=init_block_use_skip, final_block_use_skip=final_block_use_skip, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def mnasnet_b1(**kwargs): """ MnasNet-B1 model from 'MnasNet: Platform-Aware Neural Architecture Search for Mobile,' https://arxiv.org/abs/1807.11626. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mnasnet(version="b1", width_scale=1.0, model_name="mnasnet_b1", **kwargs) def mnasnet_a1(**kwargs): """ MnasNet-A1 model from 'MnasNet: Platform-Aware Neural Architecture Search for Mobile,' https://arxiv.org/abs/1807.11626. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mnasnet(version="a1", width_scale=1.0, model_name="mnasnet_a1", **kwargs) def mnasnet_small(**kwargs): """ MnasNet-Small model from 'MnasNet: Platform-Aware Neural Architecture Search for Mobile,' https://arxiv.org/abs/1807.11626. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mnasnet(version="small", width_scale=1.0, model_name="mnasnet_small", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ mnasnet_b1, mnasnet_a1, mnasnet_small, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mnasnet_b1 or weight_count == 4383312) assert (model != mnasnet_a1 or weight_count == 3887038) assert (model != mnasnet_small or weight_count == 2030264) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
14,189
32.388235
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/pyramidnet_cifar.py
""" PyramidNet for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. """ __all__ = ['CIFARPyramidNet', 'pyramidnet110_a48_cifar10', 'pyramidnet110_a48_cifar100', 'pyramidnet110_a48_svhn', 'pyramidnet110_a84_cifar10', 'pyramidnet110_a84_cifar100', 'pyramidnet110_a84_svhn', 'pyramidnet110_a270_cifar10', 'pyramidnet110_a270_cifar100', 'pyramidnet110_a270_svhn', 'pyramidnet164_a270_bn_cifar10', 'pyramidnet164_a270_bn_cifar100', 'pyramidnet164_a270_bn_svhn', 'pyramidnet200_a240_bn_cifar10', 'pyramidnet200_a240_bn_cifar100', 'pyramidnet200_a240_bn_svhn', 'pyramidnet236_a220_bn_cifar10', 'pyramidnet236_a220_bn_cifar100', 'pyramidnet236_a220_bn_svhn', 'pyramidnet272_a200_bn_cifar10', 'pyramidnet272_a200_bn_cifar100', 'pyramidnet272_a200_bn_svhn'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3_block from .preresnet import PreResActivation from .pyramidnet import PyrUnit class CIFARPyramidNet(nn.Module): """ PyramidNet model for CIFAR from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARPyramidNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, activation=None)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 1 if (i == 0) or (j != 0) else 2 stage.add_module("unit{}".format(j + 1), PyrUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_pyramidnet_cifar(num_classes, blocks, alpha, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create PyramidNet for CIFAR model with specific parameters. Parameters: ---------- num_classes : int Number of classification classes. blocks : int Number of blocks. alpha : int PyramidNet's alpha value. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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 init_block_channels = 16 growth_add = float(alpha) / float(sum(layers)) from functools import reduce channels = reduce( lambda xi, yi: xi + [[(i + 1) * growth_add + xi[-1][-1] for i in list(range(yi))]], layers, [[init_block_channels]])[1:] channels = [[int(round(cij)) for cij in ci] for ci in channels] if bottleneck: channels = [[cij * 4 for cij in ci] for ci in channels] net = CIFARPyramidNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def pyramidnet110_a48_cifar10(num_classes=10, **kwargs): """ PyramidNet-110 (a=48) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=48, bottleneck=False, model_name="pyramidnet110_a48_cifar10", **kwargs) def pyramidnet110_a48_cifar100(num_classes=100, **kwargs): """ PyramidNet-110 (a=48) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=48, bottleneck=False, model_name="pyramidnet110_a48_cifar100", **kwargs) def pyramidnet110_a48_svhn(num_classes=10, **kwargs): """ PyramidNet-110 (a=48) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=48, bottleneck=False, model_name="pyramidnet110_a48_svhn", **kwargs) def pyramidnet110_a84_cifar10(num_classes=10, **kwargs): """ PyramidNet-110 (a=84) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=84, bottleneck=False, model_name="pyramidnet110_a84_cifar10", **kwargs) def pyramidnet110_a84_cifar100(num_classes=100, **kwargs): """ PyramidNet-110 (a=84) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=84, bottleneck=False, model_name="pyramidnet110_a84_cifar100", **kwargs) def pyramidnet110_a84_svhn(num_classes=10, **kwargs): """ PyramidNet-110 (a=84) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=84, bottleneck=False, model_name="pyramidnet110_a84_svhn", **kwargs) def pyramidnet110_a270_cifar10(num_classes=10, **kwargs): """ PyramidNet-110 (a=270) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=270, bottleneck=False, model_name="pyramidnet110_a270_cifar10", **kwargs) def pyramidnet110_a270_cifar100(num_classes=100, **kwargs): """ PyramidNet-110 (a=270) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=270, bottleneck=False, model_name="pyramidnet110_a270_cifar100", **kwargs) def pyramidnet110_a270_svhn(num_classes=10, **kwargs): """ PyramidNet-110 (a=270) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=110, alpha=270, bottleneck=False, model_name="pyramidnet110_a270_svhn", **kwargs) def pyramidnet164_a270_bn_cifar10(num_classes=10, **kwargs): """ PyramidNet-164 (a=270, bn) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=164, alpha=270, bottleneck=True, model_name="pyramidnet164_a270_bn_cifar10", **kwargs) def pyramidnet164_a270_bn_cifar100(num_classes=100, **kwargs): """ PyramidNet-164 (a=270, bn) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=164, alpha=270, bottleneck=True, model_name="pyramidnet164_a270_bn_cifar100", **kwargs) def pyramidnet164_a270_bn_svhn(num_classes=10, **kwargs): """ PyramidNet-164 (a=270, bn) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=164, alpha=270, bottleneck=True, model_name="pyramidnet164_a270_bn_svhn", **kwargs) def pyramidnet200_a240_bn_cifar10(num_classes=10, **kwargs): """ PyramidNet-200 (a=240, bn) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=200, alpha=240, bottleneck=True, model_name="pyramidnet200_a240_bn_cifar10", **kwargs) def pyramidnet200_a240_bn_cifar100(num_classes=100, **kwargs): """ PyramidNet-200 (a=240, bn) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=200, alpha=240, bottleneck=True, model_name="pyramidnet200_a240_bn_cifar100", **kwargs) def pyramidnet200_a240_bn_svhn(num_classes=10, **kwargs): """ PyramidNet-200 (a=240, bn) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=200, alpha=240, bottleneck=True, model_name="pyramidnet200_a240_bn_svhn", **kwargs) def pyramidnet236_a220_bn_cifar10(num_classes=10, **kwargs): """ PyramidNet-236 (a=220, bn) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=236, alpha=220, bottleneck=True, model_name="pyramidnet236_a220_bn_cifar10", **kwargs) def pyramidnet236_a220_bn_cifar100(num_classes=100, **kwargs): """ PyramidNet-236 (a=220, bn) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=236, alpha=220, bottleneck=True, model_name="pyramidnet236_a220_bn_cifar100", **kwargs) def pyramidnet236_a220_bn_svhn(num_classes=10, **kwargs): """ PyramidNet-236 (a=220, bn) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=236, alpha=220, bottleneck=True, model_name="pyramidnet236_a220_bn_svhn", **kwargs) def pyramidnet272_a200_bn_cifar10(num_classes=10, **kwargs): """ PyramidNet-272 (a=200, bn) model for CIFAR-10 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=272, alpha=200, bottleneck=True, model_name="pyramidnet272_a200_bn_cifar10", **kwargs) def pyramidnet272_a200_bn_cifar100(num_classes=100, **kwargs): """ PyramidNet-272 (a=200, bn) model for CIFAR-100 from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=272, alpha=200, bottleneck=True, model_name="pyramidnet272_a200_bn_cifar100", **kwargs) def pyramidnet272_a200_bn_svhn(num_classes=10, **kwargs): """ PyramidNet-272 (a=200, bn) model for SVHN from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet_cifar( num_classes=num_classes, blocks=272, alpha=200, bottleneck=True, model_name="pyramidnet272_a200_bn_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (pyramidnet110_a48_cifar10, 10), (pyramidnet110_a48_cifar100, 100), (pyramidnet110_a48_svhn, 10), (pyramidnet110_a84_cifar10, 10), (pyramidnet110_a84_cifar100, 100), (pyramidnet110_a84_svhn, 10), (pyramidnet110_a270_cifar10, 10), (pyramidnet110_a270_cifar100, 100), (pyramidnet110_a270_svhn, 10), (pyramidnet164_a270_bn_cifar10, 10), (pyramidnet164_a270_bn_cifar100, 100), (pyramidnet164_a270_bn_svhn, 10), (pyramidnet200_a240_bn_cifar10, 10), (pyramidnet200_a240_bn_cifar100, 100), (pyramidnet200_a240_bn_svhn, 10), (pyramidnet236_a220_bn_cifar10, 10), (pyramidnet236_a220_bn_cifar100, 100), (pyramidnet236_a220_bn_svhn, 10), (pyramidnet272_a200_bn_cifar10, 10), (pyramidnet272_a200_bn_cifar100, 100), (pyramidnet272_a200_bn_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained, num_classes=num_classes) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != pyramidnet110_a48_cifar10 or weight_count == 1772706) assert (model != pyramidnet110_a48_cifar100 or weight_count == 1778556) assert (model != pyramidnet110_a48_svhn or weight_count == 1772706) assert (model != pyramidnet110_a84_cifar10 or weight_count == 3904446) assert (model != pyramidnet110_a84_cifar100 or weight_count == 3913536) assert (model != pyramidnet110_a84_svhn or weight_count == 3904446) assert (model != pyramidnet110_a270_cifar10 or weight_count == 28485477) assert (model != pyramidnet110_a270_cifar100 or weight_count == 28511307) assert (model != pyramidnet110_a270_svhn or weight_count == 28485477) assert (model != pyramidnet164_a270_bn_cifar10 or weight_count == 27216021) assert (model != pyramidnet164_a270_bn_cifar100 or weight_count == 27319071) assert (model != pyramidnet164_a270_bn_svhn or weight_count == 27216021) assert (model != pyramidnet200_a240_bn_cifar10 or weight_count == 26752702) assert (model != pyramidnet200_a240_bn_cifar100 or weight_count == 26844952) assert (model != pyramidnet200_a240_bn_svhn or weight_count == 26752702) assert (model != pyramidnet236_a220_bn_cifar10 or weight_count == 26969046) assert (model != pyramidnet236_a220_bn_cifar100 or weight_count == 27054096) assert (model != pyramidnet236_a220_bn_svhn or weight_count == 26969046) assert (model != pyramidnet272_a200_bn_cifar10 or weight_count == 26210842) assert (model != pyramidnet272_a200_bn_cifar100 or weight_count == 26288692) assert (model != pyramidnet272_a200_bn_svhn or weight_count == 26210842) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
23,823
32.413745
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/preresnet_cifar.py
""" PreResNet for CIFAR/SVHN, implemented in PyTorch. Original papers: 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. """ __all__ = ['CIFARPreResNet', 'preresnet20_cifar10', 'preresnet20_cifar100', 'preresnet20_svhn', 'preresnet56_cifar10', 'preresnet56_cifar100', 'preresnet56_svhn', 'preresnet110_cifar10', 'preresnet110_cifar100', 'preresnet110_svhn', 'preresnet164bn_cifar10', 'preresnet164bn_cifar100', 'preresnet164bn_svhn', 'preresnet272bn_cifar10', 'preresnet272bn_cifar100', 'preresnet272bn_svhn', 'preresnet542bn_cifar10', 'preresnet542bn_cifar100', 'preresnet542bn_svhn', 'preresnet1001_cifar10', 'preresnet1001_cifar100', 'preresnet1001_svhn', 'preresnet1202_cifar10', 'preresnet1202_cifar100', 'preresnet1202_svhn'] import os import torch.nn as nn import torch.nn.init as init from .common import conv3x3 from .preresnet import PreResUnit, PreResActivation class CIFARPreResNet(nn.Module): """ PreResNet model for CIFAR from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARPreResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", 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.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), PreResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=False)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_preresnet_cifar(num_classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create PreResNet model for CIFAR with specific parameters. Parameters: ---------- num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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 = CIFARPreResNet( channels=channels, init_block_channels=init_block_channels, bottleneck=bottleneck, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def preresnet20_cifar10(num_classes=10, **kwargs): """ PreResNet-20 model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="preresnet20_cifar10", **kwargs) def preresnet20_cifar100(num_classes=100, **kwargs): """ PreResNet-20 model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="preresnet20_cifar100", **kwargs) def preresnet20_svhn(num_classes=10, **kwargs): """ PreResNet-20 model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="preresnet20_svhn", **kwargs) def preresnet56_cifar10(num_classes=10, **kwargs): """ PreResNet-56 model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="preresnet56_cifar10", **kwargs) def preresnet56_cifar100(num_classes=100, **kwargs): """ PreResNet-56 model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="preresnet56_cifar100", **kwargs) def preresnet56_svhn(num_classes=10, **kwargs): """ PreResNet-56 model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="preresnet56_svhn", **kwargs) def preresnet110_cifar10(num_classes=10, **kwargs): """ PreResNet-110 model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="preresnet110_cifar10", **kwargs) def preresnet110_cifar100(num_classes=100, **kwargs): """ PreResNet-110 model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="preresnet110_cifar100", **kwargs) def preresnet110_svhn(num_classes=10, **kwargs): """ PreResNet-110 model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="preresnet110_svhn", **kwargs) def preresnet164bn_cifar10(num_classes=10, **kwargs): """ PreResNet-164(BN) model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="preresnet164bn_cifar10", **kwargs) def preresnet164bn_cifar100(num_classes=100, **kwargs): """ PreResNet-164(BN) model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="preresnet164bn_cifar100", **kwargs) def preresnet164bn_svhn(num_classes=10, **kwargs): """ PreResNet-164(BN) model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="preresnet164bn_svhn", **kwargs) def preresnet272bn_cifar10(num_classes=10, **kwargs): """ PreResNet-272(BN) model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="preresnet272bn_cifar10", **kwargs) def preresnet272bn_cifar100(num_classes=100, **kwargs): """ PreResNet-272(BN) model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="preresnet272bn_cifar100", **kwargs) def preresnet272bn_svhn(num_classes=10, **kwargs): """ PreResNet-272(BN) model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=272, bottleneck=True, model_name="preresnet272bn_svhn", **kwargs) def preresnet542bn_cifar10(num_classes=10, **kwargs): """ PreResNet-542(BN) model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="preresnet542bn_cifar10", **kwargs) def preresnet542bn_cifar100(num_classes=100, **kwargs): """ PreResNet-542(BN) model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="preresnet542bn_cifar100", **kwargs) def preresnet542bn_svhn(num_classes=10, **kwargs): """ PreResNet-542(BN) model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=542, bottleneck=True, model_name="preresnet542bn_svhn", **kwargs) def preresnet1001_cifar10(num_classes=10, **kwargs): """ PreResNet-1001 model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="preresnet1001_cifar10", **kwargs) def preresnet1001_cifar100(num_classes=100, **kwargs): """ PreResNet-1001 model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="preresnet1001_cifar100", **kwargs) def preresnet1001_svhn(num_classes=10, **kwargs): """ PreResNet-1001 model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="preresnet1001_svhn", **kwargs) def preresnet1202_cifar10(num_classes=10, **kwargs): """ PreResNet-1202 model for CIFAR-10 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="preresnet1202_cifar10", **kwargs) def preresnet1202_cifar100(num_classes=100, **kwargs): """ PreResNet-1202 model for CIFAR-100 from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="preresnet1202_cifar100", **kwargs) def preresnet1202_svhn(num_classes=10, **kwargs): """ PreResNet-1202 model for SVHN from 'Identity Mappings in Deep Residual Networks,' https://arxiv.org/abs/1603.05027. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_preresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="preresnet1202_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (preresnet20_cifar10, 10), (preresnet20_cifar100, 100), (preresnet20_svhn, 10), (preresnet56_cifar10, 10), (preresnet56_cifar100, 100), (preresnet56_svhn, 10), (preresnet110_cifar10, 10), (preresnet110_cifar100, 100), (preresnet110_svhn, 10), (preresnet164bn_cifar10, 10), (preresnet164bn_cifar100, 100), (preresnet164bn_svhn, 10), (preresnet272bn_cifar10, 10), (preresnet272bn_cifar100, 100), (preresnet272bn_svhn, 10), (preresnet542bn_cifar10, 10), (preresnet542bn_cifar100, 100), (preresnet542bn_svhn, 10), (preresnet1001_cifar10, 10), (preresnet1001_cifar100, 100), (preresnet1001_svhn, 10), (preresnet1202_cifar10, 10), (preresnet1202_cifar100, 100), (preresnet1202_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != preresnet20_cifar10 or weight_count == 272282) assert (model != preresnet20_cifar100 or weight_count == 278132) assert (model != preresnet20_svhn or weight_count == 272282) assert (model != preresnet56_cifar10 or weight_count == 855578) assert (model != preresnet56_cifar100 or weight_count == 861428) assert (model != preresnet56_svhn or weight_count == 855578) assert (model != preresnet110_cifar10 or weight_count == 1730522) assert (model != preresnet110_cifar100 or weight_count == 1736372) assert (model != preresnet110_svhn or weight_count == 1730522) assert (model != preresnet164bn_cifar10 or weight_count == 1703258) assert (model != preresnet164bn_cifar100 or weight_count == 1726388) assert (model != preresnet164bn_svhn or weight_count == 1703258) assert (model != preresnet272bn_cifar10 or weight_count == 2816090) assert (model != preresnet272bn_cifar100 or weight_count == 2839220) assert (model != preresnet272bn_svhn or weight_count == 2816090) assert (model != preresnet542bn_cifar10 or weight_count == 5598170) assert (model != preresnet542bn_cifar100 or weight_count == 5621300) assert (model != preresnet542bn_svhn or weight_count == 5598170) assert (model != preresnet1001_cifar10 or weight_count == 10327706) assert (model != preresnet1001_cifar100 or weight_count == 10350836) assert (model != preresnet1001_svhn or weight_count == 10327706) assert (model != preresnet1202_cifar10 or weight_count == 19423834) assert (model != preresnet1202_cifar100 or weight_count == 19429684) assert (model != preresnet1202_svhn or weight_count == 19423834) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
24,611
35.789238
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/alphapose_coco.py
""" AlphaPose for COCO Keypoint, implemented in PyTorch. Original paper: 'RMPE: Regional Multi-person Pose Estimation,' https://arxiv.org/abs/1612.00137. """ __all__ = ['AlphaPose', 'alphapose_fastseresnet101b_coco'] import os import torch import torch.nn as nn from .common import conv3x3, DucBlock, HeatmapMaxDetBlock from .fastseresnet import fastseresnet101b class AlphaPose(nn.Module): """ AlphaPose model from 'RMPE: Regional Multi-person Pose Estimation,' https://arxiv.org/abs/1612.00137. 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. return_heatmap : bool, default False Whether to return only heatmap. 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, return_heatmap=False, in_channels=3, in_size=(256, 192), keypoints=17): super(AlphaPose, self).__init__() assert (in_channels == 3) self.in_size = in_size self.keypoints = keypoints self.return_heatmap = return_heatmap self.backbone = backbone self.decoder = nn.Sequential() self.decoder.add_module("init_block", nn.PixelShuffle(upscale_factor=2)) in_channels = backbone_out_channels // 4 for i, out_channels in enumerate(channels): self.decoder.add_module("unit{}".format(i + 1), DucBlock( in_channels=in_channels, out_channels=out_channels, scale_factor=2)) in_channels = out_channels self.decoder.add_module("final_block", conv3x3( in_channels=in_channels, out_channels=keypoints, bias=True)) self.heatmap_max_det = HeatmapMaxDetBlock() self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, 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_alphapose(backbone, backbone_out_channels, keypoints, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create AlphaPose 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. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [256, 128] net = AlphaPose( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def alphapose_fastseresnet101b_coco(pretrained_backbone=False, keypoints=17, **kwargs): """ AlphaPose model on the base of ResNet-101b for COCO Keypoint from 'RMPE: Regional Multi-person Pose Estimation,' https://arxiv.org/abs/1612.00137. Parameters: ---------- pretrained_backbone : bool, default False Whether to load the pretrained weights for feature extractor. keypoints : int, default 17 Number of keypoints. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = fastseresnet101b(pretrained=pretrained_backbone).features del backbone[-1] return get_alphapose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="alphapose_fastseresnet101b_coco", **kwargs) def _calc_width(net): import numpy as np 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 _test(): in_size = (256, 192) keypoints = 17 return_heatmap = False pretrained = False models = [ alphapose_fastseresnet101b_coco, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, return_heatmap=return_heatmap) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != alphapose_fastseresnet101b_coco or weight_count == 59569873) batch = 14 x = torch.randn(batch, 3, in_size[0], in_size[1]) 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()
6,247
30.877551
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/pyramidnet.py
""" PyramidNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. """ __all__ = ['PyramidNet', 'pyramidnet101_a360', 'PyrUnit'] import os import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F from .common import pre_conv1x1_block, pre_conv3x3_block from .preresnet import PreResActivation class PyrBlock(nn.Module): """ Simple PyramidNet block for residual path in PyramidNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(PyrBlock, self).__init__() self.conv1 = pre_conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activate=False) self.conv2 = pre_conv3x3_block( in_channels=out_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class PyrBottleneck(nn.Module): """ PyramidNet bottleneck block for residual path in PyramidNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(PyrBottleneck, self).__init__() mid_channels = out_channels // 4 self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, activate=False) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride) self.conv3 = pre_conv1x1_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class PyrUnit(nn.Module): """ PyramidNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, stride, bottleneck): super(PyrUnit, self).__init__() assert (out_channels >= in_channels) self.resize_identity = (stride != 1) self.identity_pad_width = (0, 0, 0, 0, 0, out_channels - in_channels) if bottleneck: self.body = PyrBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride) else: self.body = PyrBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) self.bn = nn.BatchNorm2d(num_features=out_channels) if self.resize_identity: self.identity_pool = nn.AvgPool2d( kernel_size=2, stride=stride, ceil_mode=True) def forward(self, x): identity = x x = self.body(x) x = self.bn(x) if self.resize_identity: identity = self.identity_pool(identity) identity = F.pad(identity, pad=self.identity_pad_width) x = x + identity return x class PyrInitBlock(nn.Module): """ PyramidNet 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): super(PyrInitBlock, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=2, padding=3, bias=False) self.bn = nn.BatchNorm2d(num_features=out_channels) self.activ = nn.ReLU(inplace=True) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) x = self.pool(x) return x class PyramidNet(nn.Module): """ PyramidNet model from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(224, 224), num_classes=1000): super(PyramidNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PyrInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 1 if (i == 0) or (j != 0) else 2 stage.add_module("unit{}".format(j + 1), PyrUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_pyramidnet(blocks, alpha, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create PyramidNet model with specific parameters. Parameters: ---------- blocks : int Number of blocks. alpha : int PyramidNet's alpha 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 10: layers = [1, 1, 1, 1] elif blocks == 12: layers = [2, 1, 1, 1] elif blocks == 14: layers = [2, 2, 1, 1] elif blocks == 16: layers = [2, 2, 2, 1] elif 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 ResNet with number of blocks: {}".format(blocks)) init_block_channels = 64 growth_add = float(alpha) / float(sum(layers)) from functools import reduce channels = reduce( lambda xi, yi: xi + [[(i + 1) * growth_add + xi[-1][-1] for i in list(range(yi))]], layers, [[init_block_channels]])[1:] channels = [[int(round(cij)) for cij in ci] for ci in channels] if blocks < 50: bottleneck = False else: bottleneck = True channels = [[cij * 4 for cij in ci] for ci in channels] net = PyramidNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def pyramidnet101_a360(**kwargs): """ PyramidNet-101 model from 'Deep Pyramidal Residual Networks,' https://arxiv.org/abs/1610.02915. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_pyramidnet(blocks=101, alpha=360, model_name="pyramidnet101_a360", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ pyramidnet101_a360, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != pyramidnet101_a360 or weight_count == 42455070) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
11,038
28.126649
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/seresnet.py
""" SE-ResNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['SEResNet', 'seresnet10', 'seresnet12', 'seresnet14', 'seresnet16', 'seresnet18', 'seresnet26', 'seresnetbc26b', 'seresnet34', 'seresnetbc38b', 'seresnet50', 'seresnet50b', 'seresnet101', 'seresnet101b', 'seresnet152', 'seresnet152b', 'seresnet200', 'seresnet200b', 'SEResUnit', 'get_seresnet'] import os import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, SEBlock from .resnet import ResBlock, ResBottleneck, ResInitBlock class SEResUnit(nn.Module): """ SE-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. 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, stride, bottleneck, conv1_stride): super(SEResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) self.se = SEBlock(channels=out_channels) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 SEResNet(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, in_channels=3, in_size=(224, 224), num_classes=1000): super(SEResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), SEResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_seresnet(blocks, bottleneck=None, conv1_stride=True, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create 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. root : str, default '~/.torch/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 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 = SEResNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def seresnet10(**kwargs): """ SE-ResNet-10 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=10, model_name="seresnet10", **kwargs) def seresnet12(**kwargs): """ SE-ResNet-12 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=12, model_name="seresnet12", **kwargs) def seresnet14(**kwargs): """ SE-ResNet-14 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=14, model_name="seresnet14", **kwargs) def seresnet16(**kwargs): """ SE-ResNet-16 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=16, model_name="seresnet16", **kwargs) def seresnet18(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=18, model_name="seresnet18", **kwargs) def seresnet26(**kwargs): """ SE-ResNet-26 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=26, bottleneck=False, model_name="seresnet26", **kwargs) def seresnetbc26b(**kwargs): """ SE-ResNet-BC-26b model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=26, bottleneck=True, conv1_stride=False, model_name="seresnetbc26b", **kwargs) def seresnet34(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=34, model_name="seresnet34", **kwargs) def seresnetbc38b(**kwargs): """ SE-ResNet-BC-38b model from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model (bottleneck compressed). Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=38, bottleneck=True, conv1_stride=False, model_name="seresnetbc38b", **kwargs) def seresnet50(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=50, model_name="seresnet50", **kwargs) def seresnet50b(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=50, conv1_stride=False, model_name="seresnet50b", **kwargs) def seresnet101(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=101, model_name="seresnet101", **kwargs) def seresnet101b(**kwargs): """ 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=101, conv1_stride=False, model_name="seresnet101b", **kwargs) def seresnet152(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=152, model_name="seresnet152", **kwargs) def seresnet152b(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=152, conv1_stride=False, model_name="seresnet152b", **kwargs) def seresnet200(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=200, model_name="seresnet200", **kwargs) def seresnet200b(**kwargs): """ SE-ResNet-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(blocks=200, conv1_stride=False, model_name="seresnet200b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ seresnet10, seresnet12, seresnet14, seresnet16, seresnet18, seresnet26, seresnetbc26b, seresnet34, seresnetbc38b, seresnet50, seresnet50b, seresnet101, seresnet101b, seresnet152, seresnet152b, seresnet200, seresnet200b, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != seresnet10 or weight_count == 5463332) assert (model != seresnet12 or weight_count == 5537896) assert (model != seresnet14 or weight_count == 5835504) assert (model != seresnet16 or weight_count == 7024640) assert (model != seresnet18 or weight_count == 11778592) assert (model != seresnet26 or weight_count == 18093852) assert (model != seresnetbc26b or weight_count == 17395976) assert (model != seresnet34 or weight_count == 21958868) assert (model != seresnetbc38b or weight_count == 24026616) assert (model != seresnet50 or weight_count == 28088024) assert (model != seresnet50b or weight_count == 28088024) assert (model != seresnet101 or weight_count == 49326872) assert (model != seresnet101b or weight_count == 49326872) assert (model != seresnet152 or weight_count == 66821848) assert (model != seresnet152b or weight_count == 66821848) assert (model != seresnet200 or weight_count == 71835864) assert (model != seresnet200b or weight_count == 71835864) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
18,211
31.579606
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/seresnet_cub.py
""" SE-ResNet for CUB-200-2011, implemented in PyTorch. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['seresnet10_cub', 'seresnet12_cub', 'seresnet14_cub', 'seresnetbc14b_cub', 'seresnet16_cub', 'seresnet18_cub', 'seresnet26_cub', 'seresnetbc26b_cub', 'seresnet34_cub', 'seresnetbc38b_cub', 'seresnet50_cub', 'seresnet50b_cub', 'seresnet101_cub', 'seresnet101b_cub', 'seresnet152_cub', 'seresnet152b_cub', 'seresnet200_cub', 'seresnet200b_cub'] from .seresnet import get_seresnet def seresnet10_cub(num_classes=200, **kwargs): """ SE-ResNet-10 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=10, model_name="seresnet10_cub", **kwargs) def seresnet12_cub(num_classes=200, **kwargs): """ SE-ResNet-12 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=12, model_name="seresnet12_cub", **kwargs) def seresnet14_cub(num_classes=200, **kwargs): """ SE-ResNet-14 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=14, model_name="seresnet14_cub", **kwargs) def seresnetbc14b_cub(num_classes=200, **kwargs): """ SE-ResNet-BC-14b model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model (bottleneck compressed). Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=14, bottleneck=True, conv1_stride=False, model_name="seresnetbc14b_cub", **kwargs) def seresnet16_cub(num_classes=200, **kwargs): """ SE-ResNet-16 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=16, model_name="seresnet16_cub", **kwargs) def seresnet18_cub(num_classes=200, **kwargs): """ SE-ResNet-18 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=18, model_name="seresnet18_cub", **kwargs) def seresnet26_cub(num_classes=200, **kwargs): """ SE-ResNet-26 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=26, bottleneck=False, model_name="seresnet26_cub", **kwargs) def seresnetbc26b_cub(num_classes=200, **kwargs): """ SE-ResNet-BC-26b model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model (bottleneck compressed). Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=26, bottleneck=True, conv1_stride=False, model_name="seresnetbc26b_cub", **kwargs) def seresnet34_cub(num_classes=200, **kwargs): """ SE-ResNet-34 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=34, model_name="seresnet34_cub", **kwargs) def seresnetbc38b_cub(num_classes=200, **kwargs): """ SE-ResNet-BC-38b model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model (bottleneck compressed). Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=38, bottleneck=True, conv1_stride=False, model_name="seresnetbc38b_cub", **kwargs) def seresnet50_cub(num_classes=200, **kwargs): """ SE-ResNet-50 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=50, model_name="seresnet50_cub", **kwargs) def seresnet50b_cub(num_classes=200, **kwargs): """ SE-ResNet-50 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=50, conv1_stride=False, model_name="seresnet50b_cub", **kwargs) def seresnet101_cub(num_classes=200, **kwargs): """ SE-ResNet-101 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=101, model_name="seresnet101_cub", **kwargs) def seresnet101b_cub(num_classes=200, **kwargs): """ 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: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=101, conv1_stride=False, model_name="seresnet101b_cub", **kwargs) def seresnet152_cub(num_classes=200, **kwargs): """ SE-ResNet-152 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=152, model_name="seresnet152_cub", **kwargs) def seresnet152b_cub(num_classes=200, **kwargs): """ SE-ResNet-152 model with stride at the second convolution in bottleneck block from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=152, conv1_stride=False, model_name="seresnet152b_cub", **kwargs) def seresnet200_cub(num_classes=200, **kwargs): """ SE-ResNet-200 model for CUB-200-2011 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. It's an experimental model. Parameters: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=200, model_name="seresnet200_cub", **kwargs) def seresnet200b_cub(num_classes=200, **kwargs): """ SE-ResNet-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: ---------- num_classes : int, default 200 Number of classification num_classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnet(num_classes=num_classes, blocks=200, conv1_stride=False, model_name="seresnet200b_cub", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ seresnet10_cub, seresnet12_cub, seresnet14_cub, seresnetbc14b_cub, seresnet16_cub, seresnet18_cub, seresnet26_cub, seresnetbc26b_cub, seresnet34_cub, seresnetbc38b_cub, seresnet50_cub, seresnet50b_cub, seresnet101_cub, seresnet101b_cub, seresnet152_cub, seresnet152b_cub, seresnet200_cub, seresnet200b_cub, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != seresnet10_cub or weight_count == 5052932) assert (model != seresnet12_cub or weight_count == 5127496) assert (model != seresnet14_cub or weight_count == 5425104) assert (model != seresnetbc14b_cub or weight_count == 9126136) assert (model != seresnet16_cub or weight_count == 6614240) assert (model != seresnet18_cub or weight_count == 11368192) assert (model != seresnet26_cub or weight_count == 17683452) assert (model != seresnetbc26b_cub or weight_count == 15756776) assert (model != seresnet34_cub or weight_count == 21548468) assert (model != seresnetbc38b_cub or weight_count == 22387416) assert (model != seresnet50_cub or weight_count == 26448824) assert (model != seresnet50b_cub or weight_count == 26448824) assert (model != seresnet101_cub or weight_count == 47687672) assert (model != seresnet101b_cub or weight_count == 47687672) assert (model != seresnet152_cub or weight_count == 65182648) assert (model != seresnet152b_cub or weight_count == 65182648) assert (model != seresnet200_cub or weight_count == 70196664) assert (model != seresnet200b_cub or weight_count == 70196664) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 200)) if __name__ == "__main__": _test()
14,391
35.808184
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/densenet.py
""" DenseNet for ImageNet-1K, implemented in PyTorch. Original paper: 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. """ __all__ = ['DenseNet', 'densenet121', 'densenet161', 'densenet169', 'densenet201', 'DenseUnit', 'TransitionBlock'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import pre_conv1x1_block, pre_conv3x3_block from .preresnet import PreResInitBlock, PreResActivation class DenseUnit(nn.Module): """ DenseNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. dropout_rate : float Parameter of Dropout layer. Faction of the input units to drop. """ def __init__(self, in_channels, out_channels, dropout_rate): super(DenseUnit, self).__init__() self.use_dropout = (dropout_rate != 0.0) bn_size = 4 inc_channels = out_channels - in_channels mid_channels = inc_channels * bn_size self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=inc_channels) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) if self.use_dropout: x = self.dropout(x) x = torch.cat((identity, x), dim=1) return x class TransitionBlock(nn.Module): """ DenseNet's auxiliary block, which can be treated as the initial part of the DenseNet unit, triggered only in the first unit of each stage. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(TransitionBlock, self).__init__() self.conv = pre_conv1x1_block( in_channels=in_channels, out_channels=out_channels) self.pool = nn.AvgPool2d( kernel_size=2, stride=2, padding=0) def forward(self, x): x = self.conv(x) x = self.pool(x) return x class DenseNet(nn.Module): """ DenseNet model 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. 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 (224, 224) Spatial size of the expected input image. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, dropout_rate=0.0, in_channels=3, in_size=(224, 224), num_classes=1000): super(DenseNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() if i != 0: stage.add_module("trans{}".format(i + 1), TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2))) in_channels = in_channels // 2 for j, out_channels in enumerate(channels_per_stage): stage.add_module("unit{}".format(j + 1), DenseUnit( in_channels=in_channels, out_channels=out_channels, dropout_rate=dropout_rate)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_densenet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DenseNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 121: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 24, 16] elif blocks == 161: init_block_channels = 96 growth_rate = 48 layers = [6, 12, 36, 24] elif blocks == 169: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 32, 32] elif blocks == 201: init_block_channels = 64 growth_rate = 32 layers = [6, 12, 48, 32] else: raise ValueError("Unsupported DenseNet version with number of layers {}".format(blocks)) 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 = DenseNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def densenet121(**kwargs): """ DenseNet-121 model from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_densenet(blocks=121, model_name="densenet121", **kwargs) def densenet161(**kwargs): """ DenseNet-161 model from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_densenet(blocks=161, model_name="densenet161", **kwargs) def densenet169(**kwargs): """ DenseNet-169 model from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_densenet(blocks=169, model_name="densenet169", **kwargs) def densenet201(**kwargs): """ DenseNet-201 model from 'Densely Connected Convolutional Networks,' https://arxiv.org/abs/1608.06993. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_densenet(blocks=201, model_name="densenet201", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ densenet121, densenet161, densenet169, densenet201, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != densenet121 or weight_count == 7978856) assert (model != densenet161 or weight_count == 28681000) assert (model != densenet169 or weight_count == 14149480) assert (model != densenet201 or weight_count == 20013928) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
9,930
29.556923
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/seresnext.py
""" SE-ResNeXt for ImageNet-1K, implemented in PyTorch. Original paper: 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. """ __all__ = ['SEResNeXt', 'seresnext50_32x4d', 'seresnext101_32x4d', 'seresnext101_64x4d'] import os import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, SEBlock from .resnet import ResInitBlock from .resnext import ResNeXtBottleneck class SEResNeXtUnit(nn.Module): """ SE-ResNeXt unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width): super(SEResNeXtUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = ResNeXtBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width) self.se = SEBlock(channels=out_channels) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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 SEResNeXt(nn.Module): """ SE-ResNeXt 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, in_channels=3, in_size=(224, 224), num_classes=1000): super(SEResNeXt, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), SEResNeXtUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_seresnext(blocks, cardinality, bottleneck_width, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create SE-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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 50: layers = [3, 4, 6, 3] elif blocks == 101: layers = [3, 4, 23, 3] else: raise ValueError("Unsupported SE-ResNeXt 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 = SEResNeXt( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def seresnext50_32x4d(**kwargs): """ SE-ResNeXt-50 (32x4d) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnext(blocks=50, cardinality=32, bottleneck_width=4, model_name="seresnext50_32x4d", **kwargs) def seresnext101_32x4d(**kwargs): """ SE-ResNeXt-101 (32x4d) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnext(blocks=101, cardinality=32, bottleneck_width=4, model_name="seresnext101_32x4d", **kwargs) def seresnext101_64x4d(**kwargs): """ SE-ResNeXt-101 (64x4d) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_seresnext(blocks=101, cardinality=64, bottleneck_width=4, model_name="seresnext101_64x4d", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ seresnext50_32x4d, seresnext101_32x4d, seresnext101_64x4d, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != seresnext50_32x4d or weight_count == 27559896) assert (model != seresnext101_32x4d or weight_count == 48955416) assert (model != seresnext101_64x4d or weight_count == 88232984) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
8,721
29.929078
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/darts.py
""" DARTS for ImageNet-1K, implemented in PyTorch. Original paper: 'DARTS: Differentiable Architecture Search,' https://arxiv.org/abs/1806.09055. """ __all__ = ['DARTS', 'darts'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1, Identity from .nasnet import nasnet_dual_path_sequential class DwsConv(nn.Module): """ Standard dilated depthwise separable convolution block with. 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. stride : 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 Dilation value for convolution layer. bias : bool, default False Whether the layers use a bias vector. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation, bias=False): super(DwsConv, self).__init__() self.dw_conv = nn.Conv2d( in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=in_channels, bias=bias) self.pw_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, bias=bias) def forward(self, x): x = self.dw_conv(x) x = self.pw_conv(x) return x class DartsConv(nn.Module): """ DARTS 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. stride : 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, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, activate=True): super(DartsConv, self).__init__() self.activate = activate if self.activate: self.activ = nn.ReLU(inplace=False) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=False) self.bn = nn.BatchNorm2d(num_features=out_channels) def forward(self, x): if self.activate: x = self.activ(x) x = self.conv(x) x = self.bn(x) return x def darts_conv1x1(in_channels, out_channels, activate=True): """ 1x1 version of the DARTS specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. activate : bool, default True Whether activate the convolution block. """ return DartsConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0, activate=activate) def darts_conv3x3_s2(in_channels, out_channels, activate=True): """ 3x3 version of the DARTS specific convolution block with stride 2. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. activate : bool, default True Whether activate the convolution block. """ return DartsConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=2, padding=1, activate=activate) class DartsDwsConv(nn.Module): """ DARTS specific dilated 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. stride : 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 Dilation value for convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation): super(DartsDwsConv, self).__init__() self.activ = nn.ReLU(inplace=False) self.conv = DwsConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=False) self.bn = nn.BatchNorm2d(num_features=out_channels) def forward(self, x): x = self.activ(x) x = self.conv(x) x = self.bn(x) return x class DartsDwsBranch(nn.Module): """ DARTS 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. stride : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding): super(DartsDwsBranch, self).__init__() mid_channels = in_channels self.conv1 = DartsDwsConv( in_channels=in_channels, out_channels=mid_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=1) self.conv2 = DartsDwsConv( in_channels=mid_channels, out_channels=out_channels, kernel_size=kernel_size, stride=1, padding=padding, dilation=1) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class DartsReduceBranch(nn.Module): """ DARTS specific factorized reduce block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 2 Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride=2): super(DartsReduceBranch, self).__init__() assert (out_channels % 2 == 0) mid_channels = out_channels // 2 self.activ = nn.ReLU(inplace=False) self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels, stride=stride) self.conv2 = conv1x1( in_channels=in_channels, out_channels=mid_channels, stride=stride) self.bn = nn.BatchNorm2d(num_features=out_channels) def forward(self, x): x = self.activ(x) x1 = self.conv1(x) x = x[:, :, 1:, 1:].contiguous() x2 = self.conv2(x) x = torch.cat((x1, x2), dim=1) x = self.bn(x) return x class Stem1Unit(nn.Module): """ DARTS Stem1 unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(Stem1Unit, self).__init__() mid_channels = out_channels // 2 self.conv1 = darts_conv3x3_s2( in_channels=in_channels, out_channels=mid_channels, activate=False) self.conv2 = darts_conv3x3_s2( in_channels=mid_channels, out_channels=out_channels, activate=True) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x def stem2_unit(in_channels, out_channels): """ DARTS Stem2 unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ return darts_conv3x3_s2( in_channels=in_channels, out_channels=out_channels, activate=True) def darts_maxpool3x3(channels, stride): """ DARTS specific 3x3 Max pooling layer. Parameters: ---------- channels : int Number of input/output channels. Unused parameter. stride : int or tuple/list of 2 int Strides of the convolution. """ assert (channels > 0) return nn.MaxPool2d( kernel_size=3, stride=stride, padding=1) def darts_skip_connection(channels, stride): """ DARTS specific skip connection layer. Parameters: ---------- channels : int Number of input/output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ assert (channels > 0) if stride == 1: return Identity() else: assert (stride == 2) return DartsReduceBranch( in_channels=channels, out_channels=channels, stride=stride) def darts_dws_conv3x3(channels, stride): """ 3x3 version of DARTS specific dilated convolution block. Parameters: ---------- channels : int Number of input/output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ return DartsDwsConv( in_channels=channels, out_channels=channels, kernel_size=3, stride=stride, padding=2, dilation=2) def darts_dws_branch3x3(channels, stride): """ 3x3 version of DARTS specific dilated convolution branch. Parameters: ---------- channels : int Number of input/output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ return DartsDwsBranch( in_channels=channels, out_channels=channels, kernel_size=3, stride=stride, padding=1) # Set of operations in genotype. GENOTYPE_OPS = { 'max_pool_3x3': darts_maxpool3x3, 'skip_connect': darts_skip_connection, 'dil_conv_3x3': darts_dws_conv3x3, 'sep_conv_3x3': darts_dws_branch3x3, } class DartsMainBlock(nn.Module): """ DARTS main block, described by genotype. Parameters: ---------- genotype : list of tuples (str, int) List of genotype elements (operations and linked indices). channels : int Number of input/output channels. reduction : bool Whether use reduction. """ def __init__(self, genotype, channels, reduction): super(DartsMainBlock, self).__init__() self.concat = [2, 3, 4, 5] op_names, indices = zip(*genotype) self.indices = indices self.steps = len(op_names) // 2 self.ops = nn.ModuleList() for name, index in zip(op_names, indices): stride = 2 if reduction and index < 2 else 1 op = GENOTYPE_OPS[name](channels, stride) self.ops += [op] def forward(self, x, x_prev): s0 = x_prev s1 = x states = [s0, s1] for i in range(self.steps): j1 = 2 * i j2 = 2 * i + 1 op1 = self.ops[j1] op2 = self.ops[j2] y1 = states[self.indices[j1]] y2 = states[self.indices[j2]] y1 = op1(y1) y2 = op2(y2) s = y1 + y2 states += [s] x_out = torch.cat([states[i] for i in self.concat], dim=1) return x_out class DartsUnit(nn.Module): """ DARTS 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. genotype : list of tuples (str, int) List of genotype elements (operations and linked indices). reduction : bool Whether use reduction. prev_reduction : bool Whether use previous reduction. """ def __init__(self, in_channels, prev_in_channels, out_channels, genotype, reduction, prev_reduction): super(DartsUnit, self).__init__() mid_channels = out_channels // 4 if prev_reduction: self.preprocess_prev = DartsReduceBranch( in_channels=prev_in_channels, out_channels=mid_channels) else: self.preprocess_prev = darts_conv1x1( in_channels=prev_in_channels, out_channels=mid_channels) self.preprocess = darts_conv1x1( in_channels=in_channels, out_channels=mid_channels) self.body = DartsMainBlock( genotype=genotype, channels=mid_channels, reduction=reduction) def forward(self, x, x_prev): x = self.preprocess(x) x_prev = self.preprocess_prev(x_prev) x_out = self.body(x, x_prev) return x_out class DARTS(nn.Module): """ DARTS model from 'DARTS: Differentiable Architecture Search,' https://arxiv.org/abs/1806.09055. Parameters: ---------- channels : list of list of int Number of output channels for each unit. stem_blocks_channels : int Number of output channels for the Stem units. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, stem_blocks_channels, normal_genotype, reduce_genotype, in_channels=3, in_size=(224, 224), num_classes=1000): super(DARTS, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nasnet_dual_path_sequential( return_two=False, first_ordinals=2, last_ordinals=1) self.features.add_module("stem1_unit", Stem1Unit( in_channels=in_channels, out_channels=stem_blocks_channels)) in_channels = stem_blocks_channels self.features.add_module("stem2_unit", stem2_unit( in_channels=in_channels, out_channels=stem_blocks_channels)) prev_in_channels = in_channels in_channels = stem_blocks_channels for i, channels_per_stage in enumerate(channels): stage = nasnet_dual_path_sequential() for j, out_channels in enumerate(channels_per_stage): reduction = (i != 0) and (j == 0) prev_reduction = ((i == 0) and (j == 0)) or ((i != 0) and (j == 1)) genotype = reduce_genotype if reduction else normal_genotype stage.add_module("unit{}".format(j + 1), DartsUnit( in_channels=in_channels, prev_in_channels=prev_in_channels, out_channels=out_channels, genotype=genotype, reduction=reduction, prev_reduction=prev_reduction)) prev_in_channels = in_channels in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_darts(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DARTS 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ stem_blocks_channels = 48 layers = [4, 5, 5] channels_per_layers = [192, 384, 768] channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] normal_genotype = [ ('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 0), ('sep_conv_3x3', 1), ('sep_conv_3x3', 1), ('skip_connect', 0), ('skip_connect', 0), ('dil_conv_3x3', 2)] reduce_genotype = [ ('max_pool_3x3', 0), ('max_pool_3x3', 1), ('skip_connect', 2), ('max_pool_3x3', 1), ('max_pool_3x3', 0), ('skip_connect', 2), ('skip_connect', 2), ('max_pool_3x3', 1)] net = DARTS( channels=channels, stem_blocks_channels=stem_blocks_channels, normal_genotype=normal_genotype, reduce_genotype=reduce_genotype, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def darts(**kwargs): """ DARTS model from 'DARTS: Differentiable Architecture Search,' https://arxiv.org/abs/1806.09055. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_darts(model_name="darts", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ darts, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != darts or weight_count == 4718752) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
20,291
26.683492
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/drn.py
""" DRN for ImageNet-1K, implemented in PyTorch. Original paper: 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. """ __all__ = ['DRN', 'drnc26', 'drnc42', 'drnc58', 'drnd22', 'drnd38', 'drnd54', 'drnd105'] import os import torch.nn as nn import torch.nn.init as init class DRNConv(nn.Module): """ DRN 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. stride : 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 Dilation value for convolution layer. activate : bool Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation, activate): super(DRNConv, self).__init__() self.activate = activate self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=False) self.bn = nn.BatchNorm2d(num_features=out_channels) if self.activate: self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.conv(x) x = self.bn(x) if self.activate: x = self.activ(x) return x def drn_conv1x1(in_channels, out_channels, stride, activate): """ 1x1 version of the DRN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. activate : bool Whether activate the convolution block. """ return DRNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, dilation=1, activate=activate) def drn_conv3x3(in_channels, out_channels, stride, dilation, activate): """ 3x3 version of the DRN specific convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. dilation : int or tuple/list of 2 int Padding/dilation value for convolution layer. activate : bool Whether activate the convolution block. """ return DRNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=dilation, dilation=dilation, activate=activate) class DRNBlock(nn.Module): """ Simple DRN block for residual path in DRN unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. dilation : int or tuple/list of 2 int Padding/dilation value for convolution layers. """ def __init__(self, in_channels, out_channels, stride, dilation): super(DRNBlock, self).__init__() self.conv1 = drn_conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilation, activate=True) self.conv2 = drn_conv3x3( in_channels=out_channels, out_channels=out_channels, stride=1, dilation=dilation, activate=False) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class DRNBottleneck(nn.Module): """ DRN bottleneck block for residual path in DRN unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. dilation : int or tuple/list of 2 int Padding/dilation value for 3x3 convolution layer. """ def __init__(self, in_channels, out_channels, stride, dilation): super(DRNBottleneck, self).__init__() mid_channels = out_channels // 4 self.conv1 = drn_conv1x1( in_channels=in_channels, out_channels=mid_channels, stride=1, activate=True) self.conv2 = drn_conv3x3( in_channels=mid_channels, out_channels=mid_channels, stride=stride, dilation=dilation, activate=True) self.conv3 = drn_conv1x1( in_channels=mid_channels, out_channels=out_channels, stride=1, activate=False) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class DRNUnit(nn.Module): """ DRN unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. dilation : int or tuple/list of 2 int Padding/dilation value for 3x3 convolution layers. bottleneck : bool Whether to use a bottleneck or simple block in units. simplified : bool Whether to use a simple or simplified block in units. residual : bool Whether do residual calculations. """ def __init__(self, in_channels, out_channels, stride, dilation, bottleneck, simplified, residual): super(DRNUnit, self).__init__() assert residual or (not bottleneck) assert (not (bottleneck and simplified)) assert (not (residual and simplified)) self.residual = residual self.resize_identity = ((in_channels != out_channels) or (stride != 1)) and self.residual and (not simplified) if bottleneck: self.body = DRNBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilation) elif simplified: self.body = drn_conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilation, activate=False) else: self.body = DRNBlock( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilation) if self.resize_identity: self.identity_conv = drn_conv1x1( in_channels=in_channels, out_channels=out_channels, stride=stride, activate=False) self.activ = nn.ReLU(inplace=True) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) else: identity = x x = self.body(x) if self.residual: x = x + identity x = self.activ(x) return x def drn_init_block(in_channels, out_channels): """ DRN specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ return DRNConv( in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=1, padding=3, dilation=1, activate=True) class DRN(nn.Module): """ DRN-C&D model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. 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. dilations : list of list of int Dilation values for 3x3 convolution layers for each unit. bottlenecks : list of list of int Whether to use a bottleneck or simple block in each unit. simplifieds : list of list of int Whether to use a simple or simplified block in each unit. residuals : list of list of int Whether to use residual block in each unit. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, dilations, bottlenecks, simplifieds, residuals, in_channels=3, in_size=(224, 224), num_classes=1000): super(DRN, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", drn_init_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), DRNUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilations[i][j], bottleneck=(bottlenecks[i][j] == 1), simplified=(simplifieds[i][j] == 1), residual=(residuals[i][j] == 1))) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=28, stride=1)) self.output = nn.Conv2d( in_channels=in_channels, out_channels=num_classes, kernel_size=1) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_drn(blocks, simplified=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DRN-C or DRN-D model with specific parameters. Parameters: ---------- blocks : int Number of blocks. simplified : bool, default False Whether to use simplified scheme (D architecture). 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 22: assert simplified layers = [1, 1, 2, 2, 2, 2, 1, 1] elif blocks == 26: layers = [1, 1, 2, 2, 2, 2, 1, 1] elif blocks == 38: assert simplified layers = [1, 1, 3, 4, 6, 3, 1, 1] elif blocks == 42: layers = [1, 1, 3, 4, 6, 3, 1, 1] elif blocks == 54: assert simplified layers = [1, 1, 3, 4, 6, 3, 1, 1] elif blocks == 58: layers = [1, 1, 3, 4, 6, 3, 1, 1] elif blocks == 105: assert simplified layers = [1, 1, 3, 4, 23, 3, 1, 1] else: raise ValueError("Unsupported DRN with number of blocks: {}".format(blocks)) if blocks < 50: channels_per_layers = [16, 32, 64, 128, 256, 512, 512, 512] bottlenecks_per_layers = [0, 0, 0, 0, 0, 0, 0, 0] else: channels_per_layers = [16, 32, 256, 512, 1024, 2048, 512, 512] bottlenecks_per_layers = [0, 0, 1, 1, 1, 1, 0, 0] if simplified: simplifieds_per_layers = [1, 1, 0, 0, 0, 0, 1, 1] residuals_per_layers = [0, 0, 1, 1, 1, 1, 0, 0] else: simplifieds_per_layers = [0, 0, 0, 0, 0, 0, 0, 0] residuals_per_layers = [1, 1, 1, 1, 1, 1, 0, 0] dilations_per_layers = [1, 1, 1, 1, 2, 4, 2, 1] downsample = [0, 1, 1, 1, 0, 0, 0, 0] def expand(property_per_layers): from functools import reduce return reduce( lambda x, y: x + [[y[0]] * y[1]] if y[2] != 0 else x[:-1] + [x[-1] + [y[0]] * y[1]], zip(property_per_layers, layers, downsample), [[]]) channels = expand(channels_per_layers) dilations = expand(dilations_per_layers) bottlenecks = expand(bottlenecks_per_layers) residuals = expand(residuals_per_layers) simplifieds = expand(simplifieds_per_layers) init_block_channels = channels_per_layers[0] net = DRN( channels=channels, init_block_channels=init_block_channels, dilations=dilations, bottlenecks=bottlenecks, simplifieds=simplifieds, residuals=residuals, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def drnc26(**kwargs): """ DRN-C-26 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=26, model_name="drnc26", **kwargs) def drnc42(**kwargs): """ DRN-C-42 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=42, model_name="drnc42", **kwargs) def drnc58(**kwargs): """ DRN-C-58 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=58, model_name="drnc58", **kwargs) def drnd22(**kwargs): """ DRN-D-58 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=22, simplified=True, model_name="drnd22", **kwargs) def drnd38(**kwargs): """ DRN-D-38 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=38, simplified=True, model_name="drnd38", **kwargs) def drnd54(**kwargs): """ DRN-D-54 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=54, simplified=True, model_name="drnd54", **kwargs) def drnd105(**kwargs): """ DRN-D-105 model from 'Dilated Residual Networks,' https://arxiv.org/abs/1705.09914. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_drn(blocks=105, simplified=True, model_name="drnd105", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ drnc26, drnc42, drnc58, drnd22, drnd38, drnd54, drnd105, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != drnc26 or weight_count == 21126584) assert (model != drnc42 or weight_count == 31234744) assert (model != drnc58 or weight_count == 40542008) # 41591608 assert (model != drnd22 or weight_count == 16393752) assert (model != drnd38 or weight_count == 26501912) assert (model != drnd54 or weight_count == 35809176) assert (model != drnd105 or weight_count == 54801304) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
18,826
28.695584
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/mixnet.py
""" MixNet for ImageNet-1K, implemented in PyTorch. Original paper: 'MixConv: Mixed Depthwise Convolutional Kernels,' https://arxiv.org/abs/1907.09595. """ __all__ = ['MixNet', 'mixnet_s', 'mixnet_m', 'mixnet_l'] import os import torch import torch.nn as nn import torch.nn.init as init from .common import round_channels, get_activation_layer, conv1x1_block, conv3x3_block, dwconv3x3_block, SEBlock class MixConv(nn.Module): """ Mixed convolution layer from 'MixConv: Mixed Depthwise Convolutional Kernels,' https://arxiv.org/abs/1907.09595. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_size : int or tuple/list of int, or tuple/list of tuple/list of 2 int Convolution window size. stride : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of int, or tuple/list of 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. bias : bool, default False Whether the layer uses a bias vector. axis : int, default 1 The axis on which to concatenate the outputs. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False, axis=1): super(MixConv, self).__init__() kernel_size = kernel_size if isinstance(kernel_size, list) else [kernel_size] padding = padding if isinstance(padding, list) else [padding] kernel_count = len(kernel_size) self.splitted_in_channels = self.split_channels(in_channels, kernel_count) splitted_out_channels = self.split_channels(out_channels, kernel_count) for i, kernel_size_i in enumerate(kernel_size): in_channels_i = self.splitted_in_channels[i] out_channels_i = splitted_out_channels[i] padding_i = padding[i] self.add_module( name=str(i), module=nn.Conv2d( in_channels=in_channels_i, out_channels=out_channels_i, kernel_size=kernel_size_i, stride=stride, padding=padding_i, dilation=dilation, groups=(out_channels_i if out_channels == groups else groups), bias=bias)) self.axis = axis def forward(self, x): xx = torch.split(x, self.splitted_in_channels, dim=self.axis) out = [conv_i(x_i) for x_i, conv_i in zip(xx, self._modules.values())] x = torch.cat(tuple(out), dim=self.axis) return x @staticmethod def split_channels(channels, kernel_count): splitted_channels = [channels // kernel_count] * kernel_count splitted_channels[0] += channels - sum(splitted_channels) return splitted_channels class MixConvBlock(nn.Module): """ Mixed 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 int, or tuple/list of tuple/list of 2 int Convolution window size. stride : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of int, or tuple/list of 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) 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, stride, padding, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(MixConvBlock, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.conv = MixConv( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) if self.use_bn: self.bn = nn.BatchNorm2d( num_features=out_channels, eps=bn_eps) if self.activate: self.activ = get_activation_layer(activation) def forward(self, x): x = self.conv(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x def mixconv1x1_block(in_channels, out_channels, kernel_count, stride=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 1x1 version of the mixed convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. kernel_count : int Kernel count. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str, or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return MixConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=([1] * kernel_count), stride=stride, padding=([0] * kernel_count), groups=groups, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) class MixUnit(nn.Module): """ MixNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. exp_channels : int Number of middle (expanded) channels. stride : int or tuple/list of 2 int Strides of the second convolution layer. exp_kernel_count : int Expansion convolution kernel count for each unit. conv1_kernel_count : int Conv1 kernel count for each unit. conv2_kernel_count : int Conv2 kernel count for each unit. exp_factor : int Expansion factor for each unit. se_factor : int SE reduction factor for each unit. activation : str Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, stride, exp_kernel_count, conv1_kernel_count, conv2_kernel_count, exp_factor, se_factor, activation): super(MixUnit, self).__init__() assert (exp_factor >= 1) assert (se_factor >= 0) self.residual = (in_channels == out_channels) and (stride == 1) self.use_se = se_factor > 0 mid_channels = exp_factor * in_channels self.use_exp_conv = exp_factor > 1 if self.use_exp_conv: if exp_kernel_count == 1: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, activation=activation) else: self.exp_conv = mixconv1x1_block( in_channels=in_channels, out_channels=mid_channels, kernel_count=exp_kernel_count, activation=activation) if conv1_kernel_count == 1: self.conv1 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, activation=activation) else: self.conv1 = MixConvBlock( in_channels=mid_channels, out_channels=mid_channels, kernel_size=[3 + 2 * i for i in range(conv1_kernel_count)], stride=stride, padding=[1 + i for i in range(conv1_kernel_count)], groups=mid_channels, activation=activation) if self.use_se: self.se = SEBlock( channels=mid_channels, reduction=(exp_factor * se_factor), round_mid=False, mid_activation=activation) if conv2_kernel_count == 1: self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) else: self.conv2 = mixconv1x1_block( in_channels=mid_channels, out_channels=out_channels, kernel_count=conv2_kernel_count, activation=None) def forward(self, 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 MixInitBlock(nn.Module): """ MixNet 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): super(MixInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2) self.conv2 = MixUnit( in_channels=out_channels, out_channels=out_channels, stride=1, exp_kernel_count=1, conv1_kernel_count=1, conv2_kernel_count=1, exp_factor=1, se_factor=0, activation="relu") def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class MixNet(nn.Module): """ MixNet model from 'MixConv: Mixed Depthwise Convolutional Kernels,' https://arxiv.org/abs/1907.09595. 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. exp_kernel_counts : list of list of int Expansion convolution kernel count for each unit. conv1_kernel_counts : list of list of int Conv1 kernel count for each unit. conv2_kernel_counts : list of list of int Conv2 kernel count for each unit. exp_factors : list of list of int Expansion factor for each unit. se_factors : list of list of int SE reduction factor for each unit. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, exp_kernel_counts, conv1_kernel_counts, conv2_kernel_counts, exp_factors, se_factors, in_channels=3, in_size=(224, 224), num_classes=1000): super(MixNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", MixInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if ((j == 0) and (i != 3)) or ((j == len(channels_per_stage) // 2) and (i == 3)) else 1 exp_kernel_count = exp_kernel_counts[i][j] conv1_kernel_count = conv1_kernel_counts[i][j] conv2_kernel_count = conv2_kernel_counts[i][j] exp_factor = exp_factors[i][j] se_factor = se_factors[i][j] activation = "relu" if i == 0 else "swish" stage.add_module("unit{}".format(j + 1), MixUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, exp_kernel_count=exp_kernel_count, conv1_kernel_count=conv1_kernel_count, conv2_kernel_count=conv2_kernel_count, exp_factor=exp_factor, se_factor=se_factor, activation=activation)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_mixnet(version, width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create MixNet model with specific parameters. Parameters: ---------- version : str Version of MobileNetV3 ('s' or 'm'). 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if version == "s": init_block_channels = 16 channels = [[24, 24], [40, 40, 40, 40], [80, 80, 80], [120, 120, 120, 200, 200, 200]] exp_kernel_counts = [[2, 2], [1, 2, 2, 2], [1, 1, 1], [2, 2, 2, 1, 1, 1]] conv1_kernel_counts = [[1, 1], [3, 2, 2, 2], [3, 2, 2], [3, 4, 4, 5, 4, 4]] conv2_kernel_counts = [[2, 2], [1, 2, 2, 2], [2, 2, 2], [2, 2, 2, 1, 2, 2]] exp_factors = [[6, 3], [6, 6, 6, 6], [6, 6, 6], [6, 3, 3, 6, 6, 6]] se_factors = [[0, 0], [2, 2, 2, 2], [4, 4, 4], [2, 2, 2, 2, 2, 2]] elif version == "m": init_block_channels = 24 channels = [[32, 32], [40, 40, 40, 40], [80, 80, 80, 80], [120, 120, 120, 120, 200, 200, 200, 200]] exp_kernel_counts = [[2, 2], [1, 2, 2, 2], [1, 2, 2, 2], [1, 2, 2, 2, 1, 1, 1, 1]] conv1_kernel_counts = [[3, 1], [4, 2, 2, 2], [3, 4, 4, 4], [1, 4, 4, 4, 4, 4, 4, 4]] conv2_kernel_counts = [[2, 2], [1, 2, 2, 2], [1, 2, 2, 2], [1, 2, 2, 2, 1, 2, 2, 2]] exp_factors = [[6, 3], [6, 6, 6, 6], [6, 6, 6, 6], [6, 3, 3, 3, 6, 6, 6, 6]] se_factors = [[0, 0], [2, 2, 2, 2], [4, 4, 4, 4], [2, 2, 2, 2, 2, 2, 2, 2]] else: raise ValueError("Unsupported MixNet version {}".format(version)) final_block_channels = 1536 if width_scale != 1.0: channels = [[round_channels(cij * width_scale) for cij in ci] for ci in channels] init_block_channels = round_channels(init_block_channels * width_scale) net = MixNet( channels=channels, init_block_channels=init_block_channels, final_block_channels=final_block_channels, exp_kernel_counts=exp_kernel_counts, conv1_kernel_counts=conv1_kernel_counts, conv2_kernel_counts=conv2_kernel_counts, exp_factors=exp_factors, se_factors=se_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def mixnet_s(**kwargs): """ MixNet-S model from 'MixConv: Mixed Depthwise Convolutional Kernels,' https://arxiv.org/abs/1907.09595. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mixnet(version="s", width_scale=1.0, model_name="mixnet_s", **kwargs) def mixnet_m(**kwargs): """ MixNet-M model from 'MixConv: Mixed Depthwise Convolutional Kernels,' https://arxiv.org/abs/1907.09595. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mixnet(version="m", width_scale=1.0, model_name="mixnet_m", **kwargs) def mixnet_l(**kwargs): """ MixNet-L model from 'MixConv: Mixed Depthwise Convolutional Kernels,' https://arxiv.org/abs/1907.09595. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mixnet(version="m", width_scale=1.3, model_name="mixnet_l", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ mixnet_s, mixnet_m, mixnet_l, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != mixnet_s or weight_count == 4134606) assert (model != mixnet_m or weight_count == 5014382) assert (model != mixnet_l or weight_count == 7329252) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
20,528
33.386935
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/dabnet.py
""" DABNet for image segmentation, implemented in PyTorch. Original paper: 'DABNet: Depth-wise Asymmetric Bottleneck for Real-time Semantic Segmentation,' https://arxiv.org/abs/1907.11357. """ __all__ = ['DABNet', 'dabnet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1, conv3x3, conv3x3_block, ConvBlock, NormActivation, Concurrent, InterpolationBlock,\ DualPathSequential class DwaConvBlock(nn.Module): """ Depthwise asymmetric separable convolution block. Parameters: ---------- channels : int Number of input/output channels. kernel_size : int Convolution window size. stride : int or tuple/list of 2 int Strides of the convolution. padding : int Padding value for convolution layer. dilation : int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, channels, kernel_size, stride, padding, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(DwaConvBlock, self).__init__() self.conv1 = ConvBlock( in_channels=channels, out_channels=channels, kernel_size=(kernel_size, 1), stride=stride, padding=(padding, 0), dilation=(dilation, 1), groups=channels, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) self.conv2 = ConvBlock( in_channels=channels, out_channels=channels, kernel_size=(1, kernel_size), stride=stride, padding=(0, padding), dilation=(1, dilation), groups=channels, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x def dwa_conv3x3_block(channels, stride=1, padding=1, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 3x3 version of the depthwise asymmetric separable convolution block. Parameters: ---------- channels : int Number of input/output channels. stride : int, default 1 Strides of the convolution. padding : int, default 1 Padding value for convolution layer. dilation : int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return DwaConvBlock( channels=channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) class DABBlock(nn.Module): """ DABNet specific base block. Parameters: ---------- channels : int Number of input/output channels. dilation : int Dilation value for a dilated branch in the unit. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, channels, dilation, bn_eps): super(DABBlock, self).__init__() mid_channels = channels // 2 self.norm_activ1 = NormActivation( in_channels=channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(channels))) self.conv1 = conv3x3_block( in_channels=channels, out_channels=mid_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid_channels))) self.branches = Concurrent(stack=True) self.branches.add_module("branches1", dwa_conv3x3_block( channels=mid_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid_channels)))) self.branches.add_module("branches2", dwa_conv3x3_block( channels=mid_channels, padding=dilation, dilation=dilation, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid_channels)))) self.norm_activ2 = NormActivation( in_channels=mid_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid_channels))) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=channels) def forward(self, x): identity = x x = self.norm_activ1(x) x = self.conv1(x) x = self.branches(x) x = x.sum(dim=1) x = self.norm_activ2(x) x = self.conv2(x) x = x + identity return x class DownBlock(nn.Module): """ DABNet specific downsample block for the main branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, bn_eps): super(DownBlock, self).__init__() self.expand = (in_channels < out_channels) mid_channels = out_channels - in_channels if self.expand else out_channels self.conv = conv3x3( in_channels=in_channels, out_channels=mid_channels, stride=2) if self.expand: self.pool = nn.MaxPool2d( kernel_size=2, stride=2) self.norm_activ = NormActivation( in_channels=out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) def forward(self, x): y = self.conv(x) if self.expand: z = self.pool(x) y = torch.cat((y, z), dim=1) y = self.norm_activ(y) return y class DABUnit(nn.Module): """ DABNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. dilations : list of int Dilations for blocks. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, dilations, bn_eps): super(DABUnit, self).__init__() mid_channels = out_channels // 2 self.down = DownBlock( in_channels=in_channels, out_channels=mid_channels, bn_eps=bn_eps) self.blocks = nn.Sequential() for i, dilation in enumerate(dilations): self.blocks.add_module("block{}".format(i + 1), DABBlock( channels=mid_channels, dilation=dilation, bn_eps=bn_eps)) def forward(self, x): x = self.down(x) y = self.blocks(x) x = torch.cat((y, x), dim=1) return x class DABStage(nn.Module): """ DABNet stage. Parameters: ---------- x_channels : int Number of input/output channels for x. y_in_channels : int Number of input channels for y. y_out_channels : int Number of output channels for y. dilations : list of int Dilations for blocks. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, x_channels, y_in_channels, y_out_channels, dilations, bn_eps): super(DABStage, self).__init__() self.use_unit = (len(dilations) > 0) self.x_down = nn.AvgPool2d( kernel_size=3, stride=2, padding=1) if self.use_unit: self.unit = DABUnit( in_channels=y_in_channels, out_channels=(y_out_channels - x_channels), dilations=dilations, bn_eps=bn_eps) self.norm_activ = NormActivation( in_channels=y_out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(y_out_channels))) def forward(self, y, x): x = self.x_down(x) if self.use_unit: y = self.unit(y) y = torch.cat((y, x), dim=1) y = self.norm_activ(y) return y, x class DABInitBlock(nn.Module): """ DABNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, bn_eps): super(DABInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) self.conv3 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class DABNet(nn.Module): """ DABNet model from 'DABNet: Depth-wise Asymmetric Bottleneck for Real-time Semantic Segmentation,' https://arxiv.org/abs/1907.11357. Parameters: ---------- channels : list of int Number of output channels for each unit (for y-branch). init_block_channels : int Number of output channels for the initial unit. dilations : list of list of int Dilations for blocks. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, channels, init_block_channels, dilations, bn_eps=1e-5, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(DABNet, self).__init__() 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.num_classes = num_classes self.fixed_size = fixed_size self.features = DualPathSequential( return_two=False, first_ordinals=1, last_ordinals=0) self.features.add_module("init_block", DABInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_eps=bn_eps)) y_in_channels = init_block_channels for i, (y_out_channels, dilations_i) in enumerate(zip(channels, dilations)): self.features.add_module("stage{}".format(i + 1), DABStage( x_channels=in_channels, y_in_channels=y_in_channels, y_out_channels=y_out_channels, dilations=dilations_i, bn_eps=bn_eps)) y_in_channels = y_out_channels self.classifier = conv1x1( in_channels=y_in_channels, out_channels=num_classes) self.up = InterpolationBlock( scale_factor=8, align_corners=False) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): in_size = self.in_size if self.fixed_size else x.shape[2:] y = self.features(x, x) y = self.classifier(y) y = self.up(y, size=in_size) return y def get_dabnet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DABNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 32 channels = [35, 131, 259] dilations = [[], [2, 2, 2], [4, 4, 8, 8, 16, 16]] bn_eps = 1e-3 net = DABNet( channels=channels, init_block_channels=init_block_channels, dilations=dilations, bn_eps=bn_eps, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def dabnet_cityscapes(num_classes=19, **kwargs): """ DABNet model for Cityscapes from 'DABNet: Depth-wise Asymmetric Bottleneck for Real-time Semantic Segmentation,' https://arxiv.org/abs/1907.11357. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_dabnet(num_classes=num_classes, model_name="dabnet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ dabnet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != dabnet_cityscapes or weight_count == 756643) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
16,345
28.505415
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/cgnet.py
""" CGNet for image segmentation, implemented in PyTorch. Original paper: 'CGNet: A Light-weight Context Guided Network for Semantic Segmentation,' https://arxiv.org/abs/1811.08201. """ __all__ = ['CGNet', 'cgnet_cityscapes'] import os import torch import torch.nn as nn from .common import NormActivation, conv1x1, conv1x1_block, conv3x3_block, depthwise_conv3x3, SEBlock, Concurrent,\ DualPathSequential, InterpolationBlock class CGBlock(nn.Module): """ CGNet block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. dilation : int Dilation value. se_reduction : int SE-block reduction value. down : bool Whether to downsample. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, dilation, se_reduction, down, bn_eps): super(CGBlock, self).__init__() self.down = down if self.down: mid1_channels = out_channels mid2_channels = 2 * out_channels else: mid1_channels = out_channels // 2 mid2_channels = out_channels if self.down: self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) else: self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid1_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid1_channels))) self.branches = Concurrent() self.branches.add_module("branches1", depthwise_conv3x3(channels=mid1_channels)) self.branches.add_module("branches2", depthwise_conv3x3( channels=mid1_channels, padding=dilation, dilation=dilation)) self.norm_activ = NormActivation( in_channels=mid2_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(mid2_channels))) if self.down: self.conv2 = conv1x1( in_channels=mid2_channels, out_channels=out_channels) self.se = SEBlock( channels=out_channels, reduction=se_reduction, use_conv=False) def forward(self, x): if not self.down: identity = x x = self.conv1(x) x = self.branches(x) x = self.norm_activ(x) if self.down: x = self.conv2(x) x = self.se(x) if not self.down: x += identity return x class CGUnit(nn.Module): """ CGNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. layers : int Number of layers. dilation : int Dilation value. se_reduction : int SE-block reduction value. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, layers, dilation, se_reduction, bn_eps): super(CGUnit, self).__init__() mid_channels = out_channels // 2 self.down = CGBlock( in_channels=in_channels, out_channels=mid_channels, dilation=dilation, se_reduction=se_reduction, down=True, bn_eps=bn_eps) self.blocks = nn.Sequential() for i in range(layers - 1): self.blocks.add_module("block{}".format(i + 1), CGBlock( in_channels=mid_channels, out_channels=mid_channels, dilation=dilation, se_reduction=se_reduction, down=False, bn_eps=bn_eps)) def forward(self, x): x = self.down(x) y = self.blocks(x) x = torch.cat((y, x), dim=1) # NB: This differs from the original implementation. return x class CGStage(nn.Module): """ CGNet stage. Parameters: ---------- x_channels : int Number of input/output channels for x. y_in_channels : int Number of input channels for y. y_out_channels : int Number of output channels for y. layers : int Number of layers in the unit. dilation : int Dilation for blocks. se_reduction : int SE-block reduction value for blocks. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, x_channels, y_in_channels, y_out_channels, layers, dilation, se_reduction, bn_eps): super(CGStage, self).__init__() self.use_x = (x_channels > 0) self.use_unit = (layers > 0) if self.use_x: self.x_down = nn.AvgPool2d( kernel_size=3, stride=2, padding=1) if self.use_unit: self.unit = CGUnit( in_channels=y_in_channels, out_channels=(y_out_channels - x_channels), layers=layers, dilation=dilation, se_reduction=se_reduction, bn_eps=bn_eps) self.norm_activ = NormActivation( in_channels=y_out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(y_out_channels))) def forward(self, y, x=None): if self.use_unit: y = self.unit(y) if self.use_x: x = self.x_down(x) y = torch.cat((y, x), dim=1) y = self.norm_activ(y) return y, x class CGInitBlock(nn.Module): """ CGNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, bn_eps): super(CGInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) self.conv3 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bn_eps=bn_eps, activation=(lambda: nn.PReLU(out_channels))) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class CGNet(nn.Module): """ CGNet model from 'CGNet: A Light-weight Context Guided Network for Semantic Segmentation,' https://arxiv.org/abs/1811.08201. Parameters: ---------- layers : list of int Number of layers for each unit. channels : list of int Number of output channels for each unit (for y-branch). init_block_channels : int Number of output channels for the initial unit. dilations : list of int Dilations for each unit. se_reductions : list of int SE-block reduction value for each unit. cut_x : list of int Whether to concatenate with x-branch for each unit. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, layers, channels, init_block_channels, dilations, se_reductions, cut_x, bn_eps=1e-5, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(CGNet, self).__init__() 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.num_classes = num_classes self.fixed_size = fixed_size self.features = DualPathSequential( return_two=False, first_ordinals=1, last_ordinals=0) self.features.add_module("init_block", CGInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_eps=bn_eps)) y_in_channels = init_block_channels for i, (layers_i, y_out_channels) in enumerate(zip(layers, channels)): self.features.add_module("stage{}".format(i + 1), CGStage( x_channels=in_channels if cut_x[i] == 1 else 0, y_in_channels=y_in_channels, y_out_channels=y_out_channels, layers=layers_i, dilation=dilations[i], se_reduction=se_reductions[i], bn_eps=bn_eps)) y_in_channels = y_out_channels self.classifier = conv1x1( in_channels=y_in_channels, out_channels=num_classes) self.up = InterpolationBlock( scale_factor=8, align_corners=False) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): in_size = self.in_size if self.fixed_size else x.shape[2:] y = self.features(x, x) y = self.classifier(y) y = self.up(y, size=in_size) return y def get_cgnet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create CGNet 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ init_block_channels = 32 layers = [0, 3, 21] channels = [35, 131, 256] dilations = [0, 2, 4] se_reductions = [0, 8, 16] cut_x = [1, 1, 0] bn_eps = 1e-3 net = CGNet( layers=layers, channels=channels, init_block_channels=init_block_channels, dilations=dilations, se_reductions=se_reductions, cut_x=cut_x, bn_eps=bn_eps, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def cgnet_cityscapes(num_classes=19, **kwargs): """ CGNet model for Cityscapes from 'CGNet: A Light-weight Context Guided Network for Semantic Segmentation,' https://arxiv.org/abs/1811.08201. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_cgnet(num_classes=num_classes, model_name="cgnet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ cgnet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != cgnet_cityscapes or weight_count == 496306) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
13,575
28.577342
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/wrn1bit_cifar.py
""" WRN-1bit for CIFAR/SVHN, implemented in PyTorch. Original paper: 'Training wide residual networks for deployment using a single bit for each weight,' https://arxiv.org/abs/1802.08530. """ __all__ = ['CIFARWRN1bit', 'wrn20_10_1bit_cifar10', 'wrn20_10_1bit_cifar100', 'wrn20_10_1bit_svhn', 'wrn20_10_32bit_cifar10', 'wrn20_10_32bit_cifar100', 'wrn20_10_32bit_svhn'] import os import math import torch import torch.nn as nn import torch.nn.init as init import torch.nn.functional as F class Binarize(torch.autograd.Function): """ Fake sign op for 1-bit weights. """ @staticmethod def forward(ctx, x): return math.sqrt(2.0 / (x.shape[1] * x.shape[2] * x.shape[3])) * x.sign() @staticmethod def backward(ctx, dy): return dy class Conv2d1bit(nn.Conv2d): """ Standard convolution block with binarization. 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. binarized : bool, default False Whether to use binarization. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding=1, dilation=1, groups=1, bias=False, binarized=False): super(Conv2d1bit, self).__init__( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) self.binarized = binarized def forward(self, input): weight = Binarize.apply(self.weight) if self.binarized else self.weight bias = Binarize.apply(self.bias) if self.bias is not None and self.binarized else self.bias return F.conv2d( input=input, weight=weight, bias=bias, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups) def conv1x1_1bit(in_channels, out_channels, stride=1, groups=1, bias=False, binarized=False): """ Convolution 1x1 layer with binarization. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. binarized : bool, default False Whether to use binarization. """ return Conv2d1bit( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, groups=groups, bias=bias, binarized=binarized) def conv3x3_1bit(in_channels, out_channels, stride=1, padding=1, dilation=1, groups=1, bias=False, binarized=False): """ Convolution 3x3 layer with binarization. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. binarized : bool, default False Whether to use binarization. """ return Conv2d1bit( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, binarized=binarized) class ConvBlock1bit(nn.Module): """ Standard convolution block with Batch normalization and ReLU activation, and binarization. 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. bn_affine : bool, default True Whether the BatchNorm layer learns affine parameters. activate : bool, default True Whether activate the convolution block. binarized : bool, default False Whether to use binarization. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False, bn_affine=True, activate=True, binarized=False): super(ConvBlock1bit, self).__init__() self.activate = activate self.conv = Conv2d1bit( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, binarized=binarized) self.bn = nn.BatchNorm2d( num_features=out_channels, affine=bn_affine) if self.activate: self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.conv(x) x = self.bn(x) if self.activate: x = self.activ(x) return x def conv1x1_block_1bit(in_channels, out_channels, stride=1, padding=0, groups=1, bias=False, bn_affine=True, activate=True, binarized=False): """ 1x1 version of the standard convolution block with binarization. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. bn_affine : bool, default True Whether the BatchNorm layer learns affine parameters. activate : bool, default True Whether activate the convolution block. binarized : bool, default False Whether to use binarization. """ return ConvBlock1bit( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=padding, groups=groups, bias=bias, bn_affine=bn_affine, activate=activate, binarized=binarized) class PreConvBlock1bit(nn.Module): """ Convolution block with Batch normalization and ReLU pre-activation, and binarization. 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. bn_affine : bool, default True Whether the BatchNorm layer learns affine parameters. return_preact : bool, default False Whether return pre-activation. It's used by PreResNet. activate : bool, default True Whether activate the convolution block. binarized : bool, default False Whether to use binarization. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, bias=False, bn_affine=True, return_preact=False, activate=True, binarized=False): super(PreConvBlock1bit, self).__init__() self.return_preact = return_preact self.activate = activate self.bn = nn.BatchNorm2d( num_features=in_channels, affine=bn_affine) if self.activate: self.activ = nn.ReLU(inplace=True) self.conv = Conv2d1bit( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, binarized=binarized) def forward(self, x): 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_conv3x3_block_1bit(in_channels, out_channels, stride=1, padding=1, dilation=1, bn_affine=True, return_preact=False, activate=True, binarized=False): """ 3x3 version of the pre-activated convolution block with binarization. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bn_affine : bool, default True Whether the BatchNorm layer learns affine parameters. return_preact : bool, default False Whether return pre-activation. activate : bool, default True Whether activate the convolution block. binarized : bool, default False Whether to use binarization. """ return PreConvBlock1bit( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bn_affine=bn_affine, return_preact=return_preact, activate=activate, binarized=binarized) class PreResBlock1bit(nn.Module): """ Simple PreResNet block for residual path in ResNet unit (with binarization). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. binarized : bool, default False Whether to use binarization. """ def __init__(self, in_channels, out_channels, stride, binarized=False): super(PreResBlock1bit, self).__init__() self.conv1 = pre_conv3x3_block_1bit( in_channels=in_channels, out_channels=out_channels, stride=stride, bn_affine=False, return_preact=False, binarized=binarized) self.conv2 = pre_conv3x3_block_1bit( in_channels=out_channels, out_channels=out_channels, bn_affine=False, binarized=binarized) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class PreResUnit1bit(nn.Module): """ PreResNet unit with residual connection (with binarization). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. binarized : bool, default False Whether to use binarization. """ def __init__(self, in_channels, out_channels, stride, binarized=False): super(PreResUnit1bit, self).__init__() self.resize_identity = (stride != 1) self.body = PreResBlock1bit( in_channels=in_channels, out_channels=out_channels, stride=stride, binarized=binarized) if self.resize_identity: self.identity_pool = nn.AvgPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): identity = x x = self.body(x) if self.resize_identity: identity = self.identity_pool(identity) identity = torch.cat((identity, torch.zeros_like(identity)), dim=1) x = x + identity return x class PreResActivation(nn.Module): """ PreResNet pure pre-activation block without convolution layer. It's used by itself as the final block. Parameters: ---------- in_channels : int Number of input channels. bn_affine : bool, default True Whether the BatchNorm layer learns affine parameters. """ def __init__(self, in_channels, bn_affine=True): super(PreResActivation, self).__init__() self.bn = nn.BatchNorm2d( num_features=in_channels, affine=bn_affine) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class CIFARWRN1bit(nn.Module): """ WRN-1bit model for CIFAR 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. binarized : bool, default True Whether to use binarization. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, binarized=True, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARWRN1bit, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_1bit( in_channels=in_channels, out_channels=init_block_channels, binarized=binarized)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), PreResUnit1bit( in_channels=in_channels, out_channels=out_channels, stride=stride, binarized=binarized)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation( in_channels=in_channels, bn_affine=False)) self.output = nn.Sequential() self.output.add_module("final_conv", conv1x1_block_1bit( in_channels=in_channels, out_channels=num_classes, activate=False, binarized=binarized)) self.output.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_wrn1bit_cifar(num_classes, blocks, width_factor, binarized=True, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create WRN-1bit model for CIFAR with specific parameters. Parameters: ---------- num_classes : int Number of classification classes. blocks : int Number of blocks. width_factor : int Wide scale factor for width of layers. binarized : bool, default True Whether to use binarization. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert ((blocks - 2) % 6 == 0) layers = [(blocks - 2) // 6] * 3 channels_per_layers = [16, 32, 64] init_block_channels = 16 channels = [[ci * width_factor] * li for (ci, li) in zip(channels_per_layers, layers)] init_block_channels *= width_factor net = CIFARWRN1bit( channels=channels, init_block_channels=init_block_channels, binarized=binarized, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def wrn20_10_1bit_cifar10(num_classes=10, **kwargs): """ WRN-20-10-1bit model for CIFAR-10 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn1bit_cifar(num_classes=num_classes, blocks=20, width_factor=10, binarized=True, model_name="wrn20_10_1bit_cifar10", **kwargs) def wrn20_10_1bit_cifar100(num_classes=100, **kwargs): """ WRN-20-10-1bit model for CIFAR-100 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn1bit_cifar(num_classes=num_classes, blocks=20, width_factor=10, binarized=True, model_name="wrn20_10_1bit_cifar100", **kwargs) def wrn20_10_1bit_svhn(num_classes=10, **kwargs): """ WRN-20-10-1bit model for SVHN from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn1bit_cifar(num_classes=num_classes, blocks=20, width_factor=10, binarized=True, model_name="wrn20_10_1bit_svhn", **kwargs) def wrn20_10_32bit_cifar10(num_classes=10, **kwargs): """ WRN-20-10-32bit model for CIFAR-10 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn1bit_cifar(num_classes=num_classes, blocks=20, width_factor=10, binarized=False, model_name="wrn20_10_32bit_cifar10", **kwargs) def wrn20_10_32bit_cifar100(num_classes=100, **kwargs): """ WRN-20-10-32bit model for CIFAR-100 from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn1bit_cifar(num_classes=num_classes, blocks=20, width_factor=10, binarized=False, model_name="wrn20_10_32bit_cifar100", **kwargs) def wrn20_10_32bit_svhn(num_classes=10, **kwargs): """ WRN-20-10-32bit model for SVHN from 'Wide Residual Networks,' https://arxiv.org/abs/1605.07146. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_wrn1bit_cifar(num_classes=num_classes, blocks=20, width_factor=10, binarized=False, model_name="wrn20_10_32bit_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (wrn20_10_1bit_cifar10, 10), (wrn20_10_1bit_cifar100, 100), (wrn20_10_1bit_svhn, 10), (wrn20_10_32bit_cifar10, 10), (wrn20_10_32bit_cifar100, 100), (wrn20_10_32bit_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != wrn20_10_1bit_cifar10 or weight_count == 26737140) assert (model != wrn20_10_1bit_cifar100 or weight_count == 26794920) assert (model != wrn20_10_1bit_svhn or weight_count == 26737140) assert (model != wrn20_10_32bit_cifar10 or weight_count == 26737140) assert (model != wrn20_10_32bit_cifar100 or weight_count == 26794920) assert (model != wrn20_10_32bit_svhn or weight_count == 26737140) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
24,899
30.558935
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/condensenet.py
""" CondenseNet for ImageNet-1K, implemented in PyTorch. Original paper: 'CondenseNet: An Efficient DenseNet using Learned Group Convolutions,' https://arxiv.org/abs/1711.09224. """ __all__ = ['CondenseNet', 'condensenet74_c4_g4', 'condensenet74_c8_g8'] import os import torch import torch.nn as nn import torch.nn.init as init from torch.autograd import Variable from .common import ChannelShuffle class CondenseSimpleConv(nn.Module): """ CondenseNet specific simple 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. stride : 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, stride, padding, groups): super(CondenseSimpleConv, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, bias=False) def forward(self, x): x = self.bn(x) x = self.activ(x) x = self.conv(x) return x def condense_simple_conv3x3(in_channels, out_channels, groups): """ 3x3 version of the CondenseNet specific simple convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. groups : int Number of groups. """ return CondenseSimpleConv( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1, groups=groups) class CondenseComplexConv(nn.Module): """ CondenseNet specific complex 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. stride : 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, stride, padding, groups): super(CondenseComplexConv, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, groups=groups, bias=False) self.c_shuffle = ChannelShuffle( channels=out_channels, groups=groups) self.register_buffer('index', torch.LongTensor(in_channels)) self.index.fill_(0) def forward(self, x): x = torch.index_select(x, dim=1, index=Variable(self.index)) x = self.bn(x) x = self.activ(x) x = self.conv(x) x = self.c_shuffle(x) return x def condense_complex_conv1x1(in_channels, out_channels, groups): """ 1x1 version of the CondenseNet specific complex convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. groups : int Number of groups. """ return CondenseComplexConv( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0, groups=groups) class CondenseUnit(nn.Module): """ CondenseNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. groups : int Number of groups. """ def __init__(self, in_channels, out_channels, groups): super(CondenseUnit, self).__init__() bottleneck_size = 4 inc_channels = out_channels - in_channels mid_channels = inc_channels * bottleneck_size self.conv1 = condense_complex_conv1x1( in_channels=in_channels, out_channels=mid_channels, groups=groups) self.conv2 = condense_simple_conv3x3( in_channels=mid_channels, out_channels=inc_channels, groups=groups) def forward(self, x): identity = x x = self.conv1(x) x = self.conv2(x) x = torch.cat((identity, x), dim=1) return x class TransitionBlock(nn.Module): """ CondenseNet's auxiliary block, which can be treated as the initial part of the DenseNet unit, triggered only in the first unit of each stage. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self): super(TransitionBlock, self).__init__() self.pool = nn.AvgPool2d( kernel_size=2, stride=2, padding=0) def forward(self, x): x = self.pool(x) return x class CondenseInitBlock(nn.Module): """ CondenseNet 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): super(CondenseInitBlock, self).__init__() self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=2, padding=1, bias=False) def forward(self, x): x = self.conv(x) return x class PostActivation(nn.Module): """ CondenseNet final block, which performs the same function of postactivation as in PreResNet. Parameters: ---------- in_channels : int Number of input channels. """ def __init__(self, in_channels): super(PostActivation, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class CondenseLinear(nn.Module): """ CondenseNet specific linear block. Parameters: ---------- in_features : int Number of input channels. out_features : int Number of output channels. drop_rate : float Fraction of input channels for drop. """ def __init__(self, in_features, out_features, drop_rate=0.5): super(CondenseLinear, self).__init__() drop_in_features = int(in_features * drop_rate) self.linear = nn.Linear( in_features=drop_in_features, out_features=out_features) self.register_buffer('index', torch.LongTensor(drop_in_features)) self.index.fill_(0) def forward(self, x): x = torch.index_select(x, dim=1, index=Variable(self.index)) x = self.linear(x) return x class CondenseNet(nn.Module): """ CondenseNet model (converted) from 'CondenseNet: An Efficient DenseNet using Learned Group Convolutions,' https://arxiv.org/abs/1711.09224. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, groups, in_channels=3, in_size=(224, 224), num_classes=1000): super(CondenseNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", CondenseInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() if i != 0: stage.add_module("trans{}".format(i + 1), TransitionBlock()) for j, out_channels in enumerate(channels_per_stage): stage.add_module("unit{}".format(j + 1), CondenseUnit( in_channels=in_channels, out_channels=out_channels, groups=groups)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PostActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = CondenseLinear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) elif isinstance(module, nn.BatchNorm2d): init.constant_(module.weight, 1) init.constant_(module.bias, 0) elif isinstance(module, nn.Linear): init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_condensenet(num_layers, groups=4, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create CondenseNet (converted) model with specific parameters. Parameters: ---------- num_layers : int Number of layers. groups : int Number of groups in convolution 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if num_layers == 74: init_block_channels = 16 layers = [4, 6, 8, 10, 8] growth_rates = [8, 16, 32, 64, 128] else: raise ValueError("Unsupported CondenseNet version with number of layers {}".format(num_layers)) from functools import reduce channels = reduce(lambda xi, yi: xi + [reduce(lambda xj, yj: xj + [xj[-1] + yj], [yi[1]] * yi[0], [xi[-1][-1]])[1:]], zip(layers, growth_rates), [[init_block_channels]])[1:] net = CondenseNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def condensenet74_c4_g4(**kwargs): """ CondenseNet-74 (C=G=4) model (converted) from 'CondenseNet: An Efficient DenseNet using Learned Group Convolutions,' https://arxiv.org/abs/1711.09224. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_condensenet(num_layers=74, groups=4, model_name="condensenet74_c4_g4", **kwargs) def condensenet74_c8_g8(**kwargs): """ CondenseNet-74 (C=G=8) model (converted) from 'CondenseNet: An Efficient DenseNet using Learned Group Convolutions,' https://arxiv.org/abs/1711.09224. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_condensenet(num_layers=74, groups=8, model_name="condensenet74_c8_g8", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ condensenet74_c4_g4, condensenet74_c8_g8, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != condensenet74_c4_g4 or weight_count == 4773944) assert (model != condensenet74_c8_g8 or weight_count == 2935416) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
14,732
28.059172
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/fbnet.py
""" FBNet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block class FBNetUnit(nn.Module): """ FBNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the second convolution layer. bn_eps : float Small float added to variance in Batch norm. 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, stride, bn_eps, use_kernel3, exp_factor, activation="relu"): super(FBNetUnit, self).__init__() assert (exp_factor >= 1) self.residual = (in_channels == out_channels) and (stride == 1) self.use_exp_conv = True mid_channels = exp_factor * in_channels if self.use_exp_conv: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bn_eps=bn_eps, activation=activation) if use_kernel3: self.conv1 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, bn_eps=bn_eps, activation=activation) else: self.conv1 = dwconv5x5_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, bn_eps=bn_eps, activation=activation) self.conv2 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bn_eps=bn_eps, activation=None) def forward(self, 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(nn.Module): """ FBNet specific initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bn_eps : float Small float added to variance in Batch norm. """ def __init__(self, in_channels, out_channels, bn_eps): super(FBNetInitBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, bn_eps=bn_eps) self.conv2 = FBNetUnit( in_channels=out_channels, out_channels=out_channels, stride=1, bn_eps=bn_eps, use_kernel3=True, exp_factor=1) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class FBNet(nn.Module): """ 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_eps : float, default 1e-5 Small float added to variance in Batch norm. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, kernels3, exp_factors, bn_eps=1e-5, in_channels=3, in_size=(224, 224), num_classes=1000): super(FBNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", FBNetInitBlock( in_channels=in_channels, out_channels=init_block_channels, bn_eps=bn_eps)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) else 1 use_kernel3 = kernels3[i][j] == 1 exp_factor = exp_factors[i][j] stage.add_module("unit{}".format(j + 1), FBNetUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bn_eps=bn_eps, use_kernel3=use_kernel3, exp_factor=exp_factor)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", conv1x1_block( in_channels=in_channels, out_channels=final_block_channels, bn_eps=bn_eps)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_fbnet(version, bn_eps=1e-5, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create FBNet model with specific parameters. Parameters: ---------- version : str Version of MobileNetV3 ('a', 'b' or 'c'). bn_eps : 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. root : str, default '~/.torch/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_eps=bn_eps, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def fbnet_cb(**kwargs): """ FBNet-Cb model (bn_eps=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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_fbnet(version="c", bn_eps=1e-3, model_name="fbnet_cb", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ fbnet_cb, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fbnet_cb or weight_count == 5572200) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
9,969
30.352201
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/visemenet.py
""" VisemeNet for speech-driven facial animation, implemented in PyTorch. Original paper: 'VisemeNet: Audio-Driven Animator-Centric Speech Animation,' https://arxiv.org/abs/1805.09488. """ __all__ = ['VisemeNet', 'visemenet20'] import os import torch import torch.nn as nn from .common import DenseBlock class VisemeDenseBranch(nn.Module): """ 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): super(VisemeDenseBranch, self).__init__() self.branch = nn.Sequential() for i, out_channels in enumerate(out_channels_list[:-1]): self.branch.add_module("block{}".format(i + 1), DenseBlock( in_features=in_channels, out_features=out_channels, bias=True, use_bn=True)) in_channels = out_channels self.final_fc = nn.Linear( in_features=in_channels, out_features=out_channels_list[-1]) def forward(self, x): x = self.branch(x) y = self.final_fc(x) return y, x class VisemeRnnBranch(nn.Module): """ 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): super(VisemeRnnBranch, self).__init__() self.rnn = nn.LSTM( input_size=in_channels, hidden_size=out_channels_list[0], num_layers=rnn_num_layers, dropout=dropout_rate) self.fc_branch = VisemeDenseBranch( in_channels=out_channels_list[0], out_channels_list=out_channels_list[1:]) def forward(self, x): x, _ = self.rnn(x) x = x[:, -1, :] y, _ = self.fc_branch(x) return y class VisemeNet(nn.Module): """ 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): super(VisemeNet, self).__init__() 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 self.stage1_rnn = nn.LSTM( input_size=audio_features, hidden_size=stage1_rnn_hidden_size, num_layers=3, dropout=dropout_rate) 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 forward(self, x, pid): y, _ = self.stage1_rnn(x) y = y[:, -1, :] y = torch.cat((y, pid), dim=1) lm, _ = self.lm_branch(y) lm += pid ph, ph1 = self.ph_branch(y) z = torch.cat((lm, ph1), dim=1) z2 = torch.cat((z, x[:, self.audio_window_size // 2, :]), dim=1) n_net2_input = z2.shape[1] z2 = torch.cat((torch.zeros((self.stage2_window_size // 2, n_net2_input)), z2), dim=0) z = torch.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)], dim=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, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_visemenet(model_name="visemenet20", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ visemenet20, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() 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 = torch.randn(batch, audio_window_size, audio_features) pid = torch.full(size=(batch, num_face_ids), fill_value=3) 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()
8,396
30.215613
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/fractalnet_cifar.py
""" FractalNet for CIFAR, implemented in PyTorch. 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 torch import torch.nn as nn import torch.nn.init as init from .common import ParametricSequential class DropConvBlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. padding : int or tuple/list of 2 int Padding value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. 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, stride, padding, bias=False, dropout_prob=0.0): super(DropConvBlock, self).__init__() self.use_dropout = (dropout_prob != 0.0) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, bias=bias) self.bn = nn.BatchNorm2d(num_features=out_channels) self.activ = nn.ReLU(inplace=True) if self.use_dropout: self.dropout = nn.Dropout2d(p=dropout_prob) def forward(self, 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, stride=1, padding=1, bias=False, dropout_prob=0.0): """ 3x3 version of the convolution block with dropout. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. 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, stride=stride, padding=padding, bias=bias, dropout_prob=dropout_prob) class FractalBlock(nn.Module): """ 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): super(FractalBlock, self).__init__() assert (num_columns >= 1) self.num_columns = num_columns self.loc_drop_prob = loc_drop_prob self.blocks = nn.Sequential() depth = 2 ** (num_columns - 1) for i in range(depth): level_block_i = nn.Sequential() 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_module("subblock{}".format(j + 1), drop_conv3x3_block( in_channels=in_channels_ij, out_channels=out_channels, dropout_prob=dropout_prob)) self.blocks.add_module("block{}".format(i + 1), 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: ------- Tensor 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 torch.from_numpy(drop_mask) @staticmethod def join_outs(raw_outs, glob_num_columns, num_columns, loc_drop_prob, training): """ Join outputs for current level of block. Parameters: ---------- 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: ------- Tensor Joined output. """ curr_num_columns = len(raw_outs) out = torch.stack(raw_outs, dim=0) assert (out.size(0) == curr_num_columns) if training: batch_size = out.size(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 = batch_mask.to(out.device) assert (batch_mask.size(0) == curr_num_columns) assert (batch_mask.size(1) == batch_size) batch_mask = batch_mask.unsqueeze(2).unsqueeze(3).unsqueeze(4) masked_out = out * batch_mask num_alive = batch_mask.sum(dim=0) num_alive[num_alive == 0.0] = 1.0 out = masked_out.sum(dim=0) / num_alive else: out = out.mean(dim=0) return out def forward(self, x, glob_num_columns): outs = [x] * self.num_columns for level_block_i in self.blocks._modules.values(): outs_i = [] for j, block_ij in enumerate(level_block_i._modules.values()): input_i = outs[j] outs_i.append(block_ij(input_i)) joined_out = FractalBlock.join_outs( raw_outs=outs_i[::-1], glob_num_columns=glob_num_columns, num_columns=self.num_columns, loc_drop_prob=self.loc_drop_prob, training=self.training) len_level_block_i = len(level_block_i._modules.values()) for j in range(len_level_block_i): outs[j] = joined_out return outs[0] class FractalUnit(nn.Module): """ 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): super(FractalUnit, self).__init__() 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( kernel_size=2, stride=2) def forward(self, x, glob_num_columns): x = self.block(x, glob_num_columns=glob_num_columns) x = self.pool(x) return x class CIFARFractalNet(nn.Module): """ 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. num_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), num_classes=10): super(CIFARFractalNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.glob_drop_ratio = glob_drop_ratio self.num_columns = num_columns self.features = ParametricSequential() for i, out_channels in enumerate(channels): dropout_prob = dropout_probs[i] self.features.add_module("unit{}".format(i + 1), 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.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): glob_batch_size = int(x.size(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=glob_num_columns) x = x.view(x.size(0), -1) x = self.output(x) return x def get_fractalnet_cifar(num_classes, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_fractalnet_cifar(num_classes=num_classes, model_name="fractalnet_cifar100", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (fractalnet_cifar10, 10), (fractalnet_cifar100, 100), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != fractalnet_cifar10 or weight_count == 33724618) assert (model != fractalnet_cifar100 or weight_count == 33770788) x = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
15,954
31.038153
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/mobilenetv3.py
""" MobileNetV3 for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import round_channels, conv1x1, conv1x1_block, conv3x3_block, dwconv3x3_block, dwconv5x5_block, SEBlock,\ HSwish class MobileNetV3Unit(nn.Module): """ 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. stride : 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. """ def __init__(self, in_channels, out_channels, exp_channels, stride, use_kernel3, activation, use_se): super(MobileNetV3Unit, self).__init__() assert (exp_channels >= out_channels) self.residual = (in_channels == out_channels) and (stride == 1) self.use_se = use_se self.use_exp_conv = exp_channels != out_channels mid_channels = exp_channels if self.use_exp_conv: self.exp_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, activation=activation) if use_kernel3: self.conv1 = dwconv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, activation=activation) else: self.conv1 = dwconv5x5_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, 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, activation=None) def forward(self, 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(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, use_se): super(MobileNetV3FinalBlock, self).__init__() self.use_se = use_se self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, activation="hswish") if self.use_se: self.se = SEBlock( channels=out_channels, reduction=4, round_mid=True, out_activation="hsigmoid") def forward(self, x): x = self.conv(x) if self.use_se: x = self.se(x) return x class MobileNetV3Classifier(nn.Module): """ 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): super(MobileNetV3Classifier, self).__init__() self.use_dropout = (dropout_rate != 0.0) self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels) self.activ = HSwish(inplace=True) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) def forward(self, 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(nn.Module): """ 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. 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. num_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, in_channels=3, in_size=(224, 224), num_classes=1000): super(MobileNetV3, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2, activation="hswish")) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): exp_channels_ij = exp_channels[i][j] stride = 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_module("unit{}".format(j + 1), MobileNetV3Unit( in_channels=in_channels, out_channels=out_channels, exp_channels=exp_channels_ij, use_kernel3=use_kernel3, stride=stride, activation=activation, use_se=use_se_flag)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", MobileNetV3FinalBlock( in_channels=in_channels, out_channels=final_block_channels, use_se=final_use_se)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = MobileNetV3Classifier( in_channels=in_channels, out_channels=num_classes, mid_channels=classifier_mid_channels, dropout_rate=0.2) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_mobilenetv3(version, width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_mobilenetv3(version="large", width_scale=1.25, model_name="mobilenetv3_large_w5d4", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
18,999
33.234234
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/diaresnet.py
""" DIA-ResNet for ImageNet-1K, implemented in PyTorch. 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 import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, DualPathSequential from .resnet import ResBlock, ResBottleneck, ResInitBlock class FirstLSTMAmp(nn.Module): """ First LSTM amplifier branch. Parameters: ---------- in_features : int Number of input channels. out_features : int Number of output channels. """ def __init__(self, in_features, out_features): super(FirstLSTMAmp, self).__init__() mid_features = in_features // 4 self.fc1 = nn.Linear( in_features=in_features, out_features=mid_features) self.activ = nn.ReLU(inplace=True) self.fc2 = nn.Linear( in_features=mid_features, out_features=out_features) def forward(self, x): x = self.fc1(x) x = self.activ(x) x = self.fc2(x) return x class DIALSTMCell(nn.Module): """ 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): super(DIALSTMCell, self).__init__() self.num_layers = num_layers out_features = 4 * in_h_features self.x_amps = nn.Sequential() self.h_amps = nn.Sequential() for i in range(num_layers): amp_class = FirstLSTMAmp if i == 0 else nn.Linear self.x_amps.add_module("amp{}".format(i + 1), amp_class( in_features=in_x_features, out_features=out_features)) self.h_amps.add_module("amp{}".format(i + 1), amp_class( in_features=in_h_features, out_features=out_features)) in_x_features = in_h_features self.dropout = nn.Dropout(p=dropout_rate) def forward(self, 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 = gates.chunk(chunks=4, dim=1) i_gate = torch.sigmoid(i_gate) f_gate = torch.sigmoid(f_gate) c_gate = torch.tanh(c_gate) o_gate = torch.sigmoid(o_gate) cy_i = (f_gate * cx_i) + (i_gate * c_gate) hy_i = o_gate * torch.sigmoid(cy_i) cy.append(cy_i) hy.append(hy_i) x = self.dropout(hy_i) return hy, cy class DIAAttention(nn.Module): """ 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): super(DIAAttention, self).__init__() self.num_layers = num_layers self.pool = nn.AdaptiveAvgPool2d(output_size=1) self.lstm = DIALSTMCell( in_x_features=in_x_features, in_h_features=in_h_features, num_layers=num_layers) def forward(self, x, hc=None): w = self.pool(x) w = w.view(w.size(0), -1) if hc is None: h = [torch.zeros_like(w)] * self.num_layers c = [torch.zeros_like(w)] * self.num_layers else: h, c = hc h, c = self.lstm(w, h, c) w = h[-1].unsqueeze(dim=-1).unsqueeze(dim=-1) x = x * w return x, (h, c) class DIAResUnit(nn.Module): """ DIA-ResNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. 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, stride, padding=1, dilation=1, bottleneck=True, conv1_stride=False, attention=None): super(DIAResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, padding=padding, dilation=dilation, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) self.attention = attention def forward(self, 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(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, in_channels=3, in_size=(224, 224), num_classes=1000): super(DIAResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( 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) attention = DIAAttention( in_x_features=channels_per_stage[0], in_h_features=channels_per_stage[0]) for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), DIAResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=conv1_stride, attention=attention)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_diaresnet(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diaresnet(blocks=200, conv1_stride=False, model_name="diaresnet200b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
24,132
32.058904
116
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/lffd.py
""" LFFD for face detection, implemented in PyTorch. 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 import torch.nn as nn from .common import conv3x3, conv1x1_block, conv3x3_block, Concurrent, MultiOutputSequential, ParallelConcurent from .resnet import ResUnit from .preresnet import PreResUnit class LffdDetectionBranch(nn.Module): """ LFFD specific detection branch. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. use_bn : bool Whether to use BatchNorm layer. """ def __init__(self, in_channels, out_channels, bias, use_bn): super(LffdDetectionBranch, self).__init__() self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=in_channels, bias=bias, use_bn=use_bn) self.conv2 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=use_bn, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class LffdDetectionBlock(nn.Module): """ LFFD specific detection block. Parameters: ---------- in_channels : int Number of input channels. mid_channels : int Number of middle channels. bias : bool Whether the layer uses a bias vector. use_bn : bool Whether to use BatchNorm layer. """ def __init__(self, in_channels, mid_channels, bias, use_bn): super(LffdDetectionBlock, self).__init__() self.conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bias=bias, use_bn=use_bn) self.branches = Concurrent() self.branches.add_module("bbox_branch", LffdDetectionBranch( in_channels=mid_channels, out_channels=4, bias=bias, use_bn=use_bn)) self.branches.add_module("score_branch", LffdDetectionBranch( in_channels=mid_channels, out_channels=2, bias=bias, use_bn=use_bn)) def forward(self, x): x = self.conv(x) x = self.branches(x) return x class LFFD(nn.Module): """ 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. 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, in_channels=3, in_size=(640, 640)): super(LFFD, self).__init__() self.in_size = in_size unit_class = PreResUnit if use_preresnet else ResUnit bias = True use_bn = False self.encoder = MultiOutputSequential(return_last=False) self.encoder.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels, stride=2, padding=0, bias=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(multi_output=False, dual_output=True) stage.add_module("trans{}".format(i + 1), conv3x3( in_channels=in_channels, out_channels=channels_per_stage, stride=2, padding=0, bias=bias)) for j in range(layers_per_stage): unit = unit_class( in_channels=channels_per_stage, out_channels=channels_per_stage, stride=1, bias=bias, use_bn=use_bn, bottleneck=False) if layers_per_stage - j <= int_bends_per_stage: unit.do_output = True stage.add_module("unit{}".format(j + 1), unit) final_activ = nn.ReLU(inplace=True) final_activ.do_output = True stage.add_module("final_activ", final_activ) stage.do_output2 = True in_channels = channels_per_stage self.encoder.add_module("stage{}".format(i + 1), 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_module("unit{}".format(k + 1), LffdDetectionBlock( in_channels=channels_per_stage, mid_channels=dec_channels, bias=bias, use_bn=use_bn)) k += 1 self.decoder.add_module("unit{}".format(k + 1), LffdDetectionBlock( in_channels=channels_per_stage, mid_channels=dec_channels, bias=bias, use_bn=use_bn)) k += 1 self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.encoder(x) x = self.decoder(x) return x def get_lffd(blocks, use_preresnet, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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] elif blocks == 25: layers = [4, 2, 1, 3] enc_channels = [64, 64, 128, 128] int_bends = [1, 1, 0, 2] 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, **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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_lffd(blocks=25, use_preresnet=False, model_name="lffd25x8s560v1_widerface", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch in_size = (640, 640) pretrained = False models = [ (lffd20x5s320v2_widerface, 5), (lffd25x8s560v1_widerface, 8), ] for model, num_outs in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) assert (len(y) == num_outs) if __name__ == "__main__": _test()
10,582
30.685629
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/sepreresnet.py
""" SE-PreResNet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv1x1, SEBlock from .preresnet import PreResBlock, PreResBottleneck, PreResInitBlock, PreResActivation class SEPreResUnit(nn.Module): """ SE-PreResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. 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, stride, bottleneck, conv1_stride): super(SEPreResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = PreResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, conv1_stride=conv1_stride) else: self.body = PreResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) self.se = SEBlock(channels=out_channels) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, stride=stride) def forward(self, 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(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, in_channels=3, in_size=(224, 224), num_classes=1000): super(SEPreResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", PreResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 1 if (i == 0) or (j != 0) else 2 stage.add_module("unit{}".format(j + 1), SEPreResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_sepreresnet(blocks, bottleneck=None, conv1_stride=True, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=18, model_name="sepreresnet18", **kwargs) def sepreresnet26(**kwargs): """ SE-PreResNet-26 from 'Squeeze-and-Excitation Networks,' https://arxiv.org/abs/1709.01507. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_sepreresnet(blocks=200, conv1_stride=False, model_name="sepreresnet200b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
18,420
32.371377
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resnext.py
""" ResNeXt for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block from .resnet import ResInitBlock class ResNeXtBottleneck(nn.Module): """ ResNeXt bottleneck block for residual path in ResNeXt unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width, bottleneck_factor=4): super(ResNeXtBottleneck, self).__init__() mid_channels = out_channels // bottleneck_factor D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=group_width) self.conv2 = conv3x3_block( in_channels=group_width, out_channels=group_width, stride=stride, groups=cardinality) self.conv3 = conv1x1_block( in_channels=group_width, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class ResNeXtUnit(nn.Module): """ ResNeXt unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width): super(ResNeXtUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = ResNeXtBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, in_channels=3, in_size=(224, 224), num_classes=1000): super(ResNeXt, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ResNeXtUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resnext(blocks, cardinality, bottleneck_width, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnext(blocks=101, cardinality=64, bottleneck_width=4, model_name="resnext101_64x4d", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
14,857
31.090713
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/jasper.py
""" Jasper/DR for ASR, implemented in PyTorch. 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 torch import torch.nn as nn import torch.nn.functional as F from .common import DualPathSequential, DualPathParallelConcurent def outmask_fill(x, x_len, value=0.0): """ Masked fill a tensor. Parameters: ---------- x : tensor Input tensor. x_len : tensor Tensor with lengths. value : float, default 0.0 Filled value. Returns: ------- tensor Resulted tensor. """ max_len = x.size(2) mask = torch.arange(max_len).to(x_len.device).expand(len(x_len), max_len) >= x_len.unsqueeze(1) mask = mask.unsqueeze(dim=1).to(device=x.device) x = x.masked_fill(mask=mask, value=value) return x def masked_normalize(x, x_len): """ Normalize a tensor with mask. Parameters: ---------- x : tensor Input tensor. x_len : tensor Tensor with lengths. Returns: ------- tensor Resulted tensor. """ x = outmask_fill(x, x_len) x_mean = x.sum(dim=2) / x_len.unsqueeze(dim=1) x_m0 = x - x_mean.unsqueeze(dim=2) x_m0 = outmask_fill(x_m0, x_len) x_std = x_m0.sum(dim=2) / x_len.unsqueeze(dim=1) x = x_m0 / x_std.unsqueeze(dim=2) return x def masked_normalize2(x, x_len): """ Normalize a tensor with mask (scheme #2). Parameters: ---------- x : tensor Input tensor. x_len : tensor Tensor with lengths. Returns: ------- tensor Resulted tensor. """ x = outmask_fill(x, x_len) x_mean = x.sum(dim=2) / x_len.unsqueeze(dim=1) x2_mean = x.square().sum(dim=2) / x_len.unsqueeze(dim=1) x_std = (x2_mean - x_mean.square()).sqrt() x = (x - x_mean.unsqueeze(dim=2)) / x_std.unsqueeze(dim=2) return x def masked_normalize3(x, x_len): """ Normalize a tensor with mask (scheme #3). Parameters: ---------- x : tensor Input tensor. x_len : tensor Tensor with lengths. Returns: ------- tensor Resulted tensor. """ x_eps = 1e-5 x_mean = torch.zeros(x.shape[:2], dtype=x.dtype, device=x.device) x_std = torch.zeros(x.shape[:2], dtype=x.dtype, device=x.device) for i in range(x.shape[0]): x_mean[i, :] = x[i, :, : x_len[i]].mean(dim=1) x_std[i, :] = x[i, :, : x_len[i]].std(dim=1) x_std += x_eps return (x - x_mean.unsqueeze(dim=2)) / x_std.unsqueeze(dim=2) 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(nn.Module): """ 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): super(NemoMelSpecExtractor, self).__init__() 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 window_tensor = torch.hann_window(win_length, periodic=False) self.register_buffer("window", window_tensor) self.stft = lambda x: torch.stft( x, n_fft=n_fft, hop_length=self.hop_length, win_length=win_length, window=self.window.to(dtype=torch.float), center=True) self.dither = dither self.preemph = preemph self.pad_align = 16 from librosa.filters import mel as librosa_mel filter_bank = librosa_mel( sr=sample_rate, n_fft=n_fft, n_mels=n_filters, fmin=0.0, fmax=(sample_rate / 2.0)) fb_tensor = torch.from_numpy(filter_bank).unsqueeze(0) self.register_buffer("fb", fb_tensor) def forward(self, x, x_len): """ Preprocess audio. Parameters: ---------- xs : list of np.array Audio data. Returns: ------- x : np.array Audio data. x_len : np.array Audio data lengths. """ x_len = torch.ceil(x_len.float() / self.hop_length).long() if self.dither > 0: x += self.dither * torch.randn_like(x) x = torch.cat((x[:, :1], x[:, 1:] - self.preemph * x[:, :-1]), dim=1) with torch.cuda.amp.autocast(enabled=False): x = self.stft(x) x = x.pow(2).sum(-1) x = torch.matmul(self.fb.to(x.dtype), x) x = torch.log(x + self.log_zero_guard_value) x = masked_normalize2(x, x_len) x = outmask_fill(x, x_len) x_len_max = x.size(-1) pad_rem = x_len_max % self.pad_align if pad_rem != 0: x = F.pad(x, pad=(0, self.pad_align - pad_rem)) return x, x_len def calc_flops(self, x): assert (x.shape[0] == 1) num_flops = x.numel() 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, stride=1, groups=1, bias=False): """ 1-dim kernel version of the 1D convolution layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv1d( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, groups=groups, bias=bias) 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. stride : 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. 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, stride, padding=0, dilation=1, groups=1, bias=False, use_mask=True): super(MaskConv1d, self).__init__( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) self.use_mask = use_mask def forward(self, x, x_len): if self.use_mask: x = outmask_fill(x, x_len) x_len = (x_len + 2 * self.padding[0] - self.dilation[0] * (self.kernel_size[0] - 1) - 1) // self.stride[0] + 1 x = F.conv1d( input=x, weight=self.weight, bias=self.bias, stride=self.stride, padding=self.padding, dilation=self.dilation, groups=self.groups) return x, x_len def mask_conv1d1(in_channels, out_channels, stride=1, groups=1, bias=False): """ 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. stride : int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. """ return MaskConv1d( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, groups=groups, bias=bias) class MaskConvBlock1d(nn.Module): """ 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) 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, stride, padding, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True)), dropout_rate=0.0): super(MaskConvBlock1d, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.use_dropout = (dropout_rate != 0.0) self.conv = MaskConv1d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) if self.use_bn: self.bn = nn.BatchNorm1d( num_features=out_channels, eps=bn_eps) if self.activate: self.activ = activation() if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, 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, stride=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. stride : 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, stride=stride, padding=padding, **kwargs) class ChannelShuffle1d(nn.Module): """ 1D version of the channel shuffle layer. Parameters: ---------- channels : int Number of channels. groups : int Number of groups. """ def __init__(self, channels, groups): super(ChannelShuffle1d, self).__init__() assert (channels % groups == 0) self.groups = groups def forward(self, x): batch, channels, seq_len = x.size() channels_per_group = channels // self.groups x = x.view(batch, self.groups, channels_per_group, seq_len) x = torch.transpose(x, 1, 2).contiguous() x = x.view(batch, channels, seq_len) return x def __repr__(self): s = "{name}(groups={groups})" return s.format( name=self.__class__.__name__, groups=self.groups) class DwsConvBlock1d(nn.Module): """ 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) 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, stride, padding, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True)), dropout_rate=0.0): super(DwsConvBlock1d, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.use_dropout = (dropout_rate != 0.0) self.use_channel_shuffle = (groups > 1) self.dw_conv = MaskConv1d( in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=in_channels, bias=bias) self.pw_conv = mask_conv1d1( in_channels=in_channels, out_channels=out_channels, groups=groups, bias=bias) if self.use_channel_shuffle: self.shuffle = ChannelShuffle1d( channels=out_channels, groups=groups) if self.use_bn: self.bn = nn.BatchNorm1d( num_features=out_channels, eps=bn_eps) if self.activate: self.activ = activation() if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, 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(nn.Module): """ 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_eps : 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. """ def __init__(self, in_channels, out_channels, kernel_size, bn_eps, dropout_rate, repeat, use_dw, use_dr): super(JasperUnit, self).__init__() self.use_dropout = (dropout_rate != 0.0) self.use_dr = use_dr block_class = DwsConvBlock1d if use_dw else MaskConvBlock1d if self.use_dr: self.identity_block = DualPathParallelConcurent() for i, dense_in_channels_i in enumerate(in_channels): self.identity_block.add_module("block{}".format(i + 1), mask_conv1d1_block( in_channels=dense_in_channels_i, out_channels=out_channels, bn_eps=bn_eps, 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_eps=bn_eps, dropout_rate=0.0, activation=None) self.body = DualPathSequential() for i in range(repeat): activation = (lambda: nn.ReLU(inplace=True)) if i < repeat - 1 else None dropout_rate_i = dropout_rate if i < repeat - 1 else 0.0 self.body.add_module("block{}".format(i + 1), block_class( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=1, padding=(kernel_size // 2), bn_eps=bn_eps, dropout_rate=dropout_rate_i, activation=activation)) in_channels = out_channels self.activ = nn.ReLU(inplace=True) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, 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 = torch.stack(tuple(identity), dim=1) identity = identity.sum(dim=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(nn.Module): """ 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_eps : 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. """ def __init__(self, in_channels, channels, kernel_sizes, bn_eps, dropout_rates, use_dw, use_dr): super(JasperFinalBlock, self).__init__() self.use_dr = use_dr conv1_class = DwsConvBlock1d if use_dw else MaskConvBlock1d self.conv1 = conv1_class( in_channels=in_channels, out_channels=channels[-2], kernel_size=kernel_sizes[-2], stride=1, padding=(2 * kernel_sizes[-2] // 2 - 1), dilation=2, bn_eps=bn_eps, dropout_rate=dropout_rates[-2]) self.conv2 = MaskConvBlock1d( in_channels=channels[-2], out_channels=channels[-1], kernel_size=kernel_sizes[-1], stride=1, padding=(kernel_sizes[-1] // 2), bn_eps=bn_eps, dropout_rate=dropout_rates[-1]) def forward(self, 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(nn.Module): """ 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_eps : 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. 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). num_classes : int, default 29 Number of classification classes (number of graphemes). """ def __init__(self, channels, kernel_sizes, bn_eps, dropout_rates, repeat, use_dw, use_dr, from_audio=True, dither=0.0, return_text=False, vocabulary=None, in_channels=64, num_classes=29): super(Jasper, self).__init__() self.in_size = in_channels self.num_classes = num_classes self.vocabulary = vocabulary self.from_audio = from_audio self.return_text = return_text if self.from_audio: self.preprocessor = NemoMelSpecExtractor(dither=dither) self.features = DualPathSequential() init_block_class = DwsConvBlock1d if use_dw else MaskConvBlock1d self.features.add_module("init_block", init_block_class( in_channels=in_channels, out_channels=channels[0], kernel_size=kernel_sizes[0], stride=2, padding=(kernel_sizes[0] // 2), bn_eps=bn_eps, 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_module("unit{}".format(i + 1), JasperUnit( in_channels=(in_channels_list if use_dr else in_channels), out_channels=out_channels, kernel_size=kernel_size, bn_eps=bn_eps, dropout_rate=dropout_rate, repeat=repeat, use_dw=use_dw, use_dr=use_dr)) in_channels = out_channels self.features.add_module("final_block", JasperFinalBlock( in_channels=in_channels, channels=channels, kernel_sizes=kernel_sizes, bn_eps=bn_eps, dropout_rates=dropout_rates, use_dw=use_dw, use_dr=use_dr)) in_channels = channels[-1] self.output = conv1d1( in_channels=in_channels, out_channels=num_classes, bias=True) if self.return_text: self.ctc_decoder = CtcDecoder(vocabulary=vocabulary) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, 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.transpose(1, 2).log_softmax(dim=-1).argmax(dim=-1, keepdim=False).cpu().numpy() return self.ctc_decoder(greedy_predictions) else: return x, x_len def get_jasper(version, use_dw=False, use_dr=False, bn_eps=1e-3, vocabulary=None, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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_eps : 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. root : str, default '~/.torch/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_eps=bn_eps, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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 = 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 _test(): import torch pretrained = False from_audio = True audio_features = 64 num_classes = 29 use_cuda = True models = [ jasper5x3, jasper10x4, jasper10x5, ] for model in models: net = model( in_channels=audio_features, num_classes=num_classes, from_audio=from_audio, pretrained=pretrained) if use_cuda: net = net.cuda() # net.train() net.eval() 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 = torch.randn(x_shape) x_len = torch.tensor(seq_len, dtype=torch.long, device=x.device) if use_cuda: x = x.cuda() x_len = x_len.cuda() y, y_len = net(x, x_len) # y.sum().backward() assert (tuple(y.size())[:2] == (batch, net.num_classes)) if from_audio: assert (y.size()[2] in range(seq_len_max // aud_scale * 2, seq_len_max // aud_scale * 2 + 9)) else: assert (y.size()[2] in [seq_len_max // 2, seq_len_max // 2 + 1]) if __name__ == "__main__": _test()
35,202
29.347414
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resneta.py
""" ResNet(A) with average downsampling for ImageNet-1K, implemented in PyTorch. Original paper: 'Deep Residual Learning for Image Recognition,' https://arxiv.org/abs/1512.03385. """ __all__ = ['ResNetA', 'resneta10', 'resnetabc14b', 'resneta18', 'resneta50b', 'resneta101b', 'resneta152b'] import os import torch.nn as nn from .common import conv1x1_block from .resnet import ResBlock, ResBottleneck from .senet import SEInitBlock class ResADownBlock(nn.Module): """ 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. stride : 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. """ def __init__(self, in_channels, out_channels, stride, dilation=1): super(ResADownBlock, self).__init__() self.pool = nn.AvgPool2d( kernel_size=(stride if dilation == 1 else 1), stride=(stride if dilation == 1 else 1), ceil_mode=True, count_include_pad=False) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.pool(x) x = self.conv(x) return x class ResAUnit(nn.Module): """ ResNet(A) unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. 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, stride, padding=1, dilation=1, bottleneck=True, conv1_stride=False): super(ResAUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = ResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, padding=padding, dilation=dilation, conv1_stride=conv1_stride) else: self.body = ResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_block = ResADownBlock( in_channels=in_channels, out_channels=out_channels, stride=stride, dilation=dilation) self.activ = nn.ReLU(inplace=True) def forward(self, 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(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, conv1_stride, dilated=False, in_channels=3, in_size=(224, 224), num_classes=1000): super(ResNetA, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", SEInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): if dilated: stride = 2 if ((j == 0) and (i != 0) and (i < 2)) else 1 dilation = (2 ** max(0, i - 1 - int(j == 0))) else: stride = 2 if (j == 0) and (i != 0) else 1 dilation = 1 stage.add_module("unit{}".format(j + 1), ResAUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, padding=dilation, dilation=dilation, bottleneck=bottleneck, conv1_stride=conv1_stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resneta(blocks, bottleneck=None, conv1_stride=True, width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resneta(blocks=152, conv1_stride=False, model_name="resneta152b", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ resneta10, resnetabc14b, resneta18, resneta50b, resneta101b, resneta152b, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
14,395
32.094253
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resnesta.py
""" ResNeSt(A) with average downsampling for ImageNet-1K, implemented in PyTorch. Original paper: 'ResNeSt: Split-Attention Networks,' https://arxiv.org/abs/2004.08955. """ __all__ = ['ResNeStA', 'resnestabc14', 'resnesta18', 'resnestabc26', 'resnesta50', 'resnesta101', 'resnesta152', 'resnesta200', 'resnesta269', 'ResNeStADownBlock'] import os import torch.nn as nn from .common import conv1x1_block, conv3x3_block, saconv3x3_block from .senet import SEInitBlock class ResNeStABlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. """ def __init__(self, in_channels, out_channels, stride, bias=False, use_bn=True): super(ResNeStABlock, self).__init__() self.resize = (stride > 1) self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=use_bn) if self.resize: self.pool = nn.AvgPool2d( kernel_size=3, stride=stride, padding=1) self.conv2 = saconv3x3_block( in_channels=out_channels, out_channels=out_channels, bias=bias, use_bn=use_bn, activation=None) def forward(self, x): x = self.conv1(x) if self.resize: x = self.pool(x) x = self.conv2(x) return x class ResNeStABottleneck(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, stride, bottleneck_factor=4): super(ResNeStABottleneck, self).__init__() self.resize = (stride > 1) mid_channels = out_channels // bottleneck_factor self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels) self.conv2 = saconv3x3_block( in_channels=mid_channels, out_channels=mid_channels) if self.resize: self.pool = nn.AvgPool2d( kernel_size=3, stride=stride, padding=1) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) if self.resize: x = self.pool(x) x = self.conv3(x) return x class ResNeStADownBlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(ResNeStADownBlock, self).__init__() self.pool = nn.AvgPool2d( kernel_size=stride, stride=stride, ceil_mode=True, count_include_pad=False) self.conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.pool(x) x = self.conv(x) return x class ResNeStAUnit(nn.Module): """ ResNeSt(A) unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool, default True Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, stride, bottleneck=True): super(ResNeStAUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) if bottleneck: self.body = ResNeStABottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride) else: self.body = ResNeStABlock( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_block = ResNeStADownBlock( in_channels=in_channels, out_channels=out_channels, stride=stride) self.activ = nn.ReLU(inplace=True) def forward(self, 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(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, dropout_rate=0.0, in_channels=3, in_size=(224, 224), num_classes=1000): super(ResNeStA, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", SEInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), ResNeStAUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AdaptiveAvgPool2d(output_size=1)) self.output = nn.Sequential() if dropout_rate > 0.0: self.output.add_module("dropout", nn.Dropout(p=dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resnesta(blocks, bottleneck=None, width_scale=1.0, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnesta(blocks=26, bottleneck=True, model_name="resnestabc26", **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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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 _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (resnestabc14, 224), (resnesta18, 224), (resnestabc26, 224), (resnesta50, 224), (resnesta101, 224), (resnesta152, 224), (resnesta200, 256), (resnesta269, 320), ] for model, size in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 != 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 = torch.randn(batch, 3, size, size) y = net(x) y.sum().backward() assert (tuple(y.size()) == (batch, 1000)) if __name__ == "__main__": _test()
17,572
30.892922
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/senet.py
""" SENet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block, SEBlock class SENetBottleneck(nn.Module): """ SENet bottleneck block for residual path in SENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width): super(SENetBottleneck, self).__init__() mid_channels = out_channels // 4 D = int(math.floor(mid_channels * (bottleneck_width / 64.0))) group_width = cardinality * D group_width2 = group_width // 2 self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=group_width2) self.conv2 = conv3x3_block( in_channels=group_width2, out_channels=group_width, stride=stride, groups=cardinality) self.conv3 = conv1x1_block( in_channels=group_width, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class SENetUnit(nn.Module): """ SENet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. cardinality: int Number of groups. bottleneck_width: int Width of bottleneck block. identity_conv3x3 : bool, default False Whether to use 3x3 convolution in the identity link. """ def __init__(self, in_channels, out_channels, stride, cardinality, bottleneck_width, identity_conv3x3): super(SENetUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = SENetBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width) 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, stride=stride, activation=None) else: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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(nn.Module): """ SENet 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): super(SEInitBlock, self).__init__() mid_channels = out_channels // 2 self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels) self.conv3 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) x = self.pool(x) return x class SENet(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, cardinality, bottleneck_width, in_channels=3, in_size=(224, 224), num_classes=1000): super(SENet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", SEInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() identity_conv3x3 = (i != 0) for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), SENetUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, cardinality=cardinality, bottleneck_width=bottleneck_width, identity_conv3x3=identity_conv3x3)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Sequential() self.output.add_module("dropout", nn.Dropout(p=0.2)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_senet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_senet(blocks=154, model_name="senet154", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ senet16, senet28, senet40, senet52, senet103, senet154, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
13,095
28.696145
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/diapreresnet_cifar.py
""" DIA-PreResNet for CIFAR/SVHN, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv3x3, DualPathSequential from .preresnet import PreResActivation from .diaresnet import DIAAttention from .diapreresnet import DIAPreResUnit class CIFARDIAPreResNet(nn.Module): """ 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARDIAPreResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", 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) attention = DIAAttention( in_x_features=channels_per_stage[0], in_h_features=channels_per_stage[0]) for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), DIAPreResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=False, attention=attention)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_diapreresnet_cifar(num_classes, blocks, bottleneck, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create DIA-PreResNet model for CIFAR with specific parameters. Parameters: ---------- num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def diapreresnet20_cifar10(num_classes=10, **kwargs): """ DIA-PreResNet-20 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="diapreresnet20_cifar10", **kwargs) def diapreresnet20_cifar100(num_classes=100, **kwargs): """ DIA-PreResNet-20 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="diapreresnet20_cifar100", **kwargs) def diapreresnet20_svhn(num_classes=10, **kwargs): """ DIA-PreResNet-20 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=20, bottleneck=False, model_name="diapreresnet20_svhn", **kwargs) def diapreresnet56_cifar10(num_classes=10, **kwargs): """ DIA-PreResNet-56 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="diapreresnet56_cifar10", **kwargs) def diapreresnet56_cifar100(num_classes=100, **kwargs): """ DIA-PreResNet-56 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="diapreresnet56_cifar100", **kwargs) def diapreresnet56_svhn(num_classes=10, **kwargs): """ DIA-PreResNet-56 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=56, bottleneck=False, model_name="diapreresnet56_svhn", **kwargs) def diapreresnet110_cifar10(num_classes=10, **kwargs): """ DIA-PreResNet-110 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="diapreresnet110_cifar10", **kwargs) def diapreresnet110_cifar100(num_classes=100, **kwargs): """ DIA-PreResNet-110 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="diapreresnet110_cifar100", **kwargs) def diapreresnet110_svhn(num_classes=10, **kwargs): """ DIA-PreResNet-110 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=110, bottleneck=False, model_name="diapreresnet110_svhn", **kwargs) def diapreresnet164bn_cifar10(num_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: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="diapreresnet164bn_cifar10", **kwargs) def diapreresnet164bn_cifar100(num_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: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="diapreresnet164bn_cifar100", **kwargs) def diapreresnet164bn_svhn(num_classes=10, **kwargs): """ DIA-PreResNet-164(BN) model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=164, bottleneck=True, model_name="diapreresnet164bn_svhn", **kwargs) def diapreresnet1001_cifar10(num_classes=10, **kwargs): """ DIA-PreResNet-1001 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="diapreresnet1001_cifar10", **kwargs) def diapreresnet1001_cifar100(num_classes=100, **kwargs): """ DIA-PreResNet-1001 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="diapreresnet1001_cifar100", **kwargs) def diapreresnet1001_svhn(num_classes=10, **kwargs): """ DIA-PreResNet-1001 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=1001, bottleneck=True, model_name="diapreresnet1001_svhn", **kwargs) def diapreresnet1202_cifar10(num_classes=10, **kwargs): """ DIA-PreResNet-1202 model for CIFAR-10 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="diapreresnet1202_cifar10", **kwargs) def diapreresnet1202_cifar100(num_classes=100, **kwargs): """ DIA-PreResNet-1202 model for CIFAR-100 from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="diapreresnet1202_cifar100", **kwargs) def diapreresnet1202_svhn(num_classes=10, **kwargs): """ DIA-PreResNet-1202 model for SVHN from 'DIANet: Dense-and-Implicit Attention Network,' https://arxiv.org/abs/1905.10671. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_diapreresnet_cifar(num_classes=num_classes, blocks=1202, bottleneck=False, model_name="diapreresnet1202_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
20,604
36.327899
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/simplepose_coco.py
""" SimplePose for COCO Keypoint, implemented in PyTorch. 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'] import os import torch import torch.nn as nn from .common import DeconvBlock, conv1x1, HeatmapMaxDetBlock from .resnet import resnet18, resnet50b, resnet101b, resnet152b from .resneta import resneta50b, resneta101b, resneta152b class SimplePose(nn.Module): """ 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. return_heatmap : bool, default False Whether to return only heatmap. 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, return_heatmap=False, in_channels=3, in_size=(256, 192), keypoints=17): super(SimplePose, self).__init__() assert (in_channels == 3) self.in_size = in_size self.keypoints = keypoints self.return_heatmap = return_heatmap self.backbone = backbone self.decoder = nn.Sequential() in_channels = backbone_out_channels for i, out_channels in enumerate(channels): self.decoder.add_module("unit{}".format(i + 1), DeconvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=4, stride=2, padding=1)) in_channels = out_channels self.decoder.add_module("final_block", conv1x1( in_channels=in_channels, out_channels=keypoints, bias=True)) self.heatmap_max_det = HeatmapMaxDetBlock() self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, 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, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [256, 256, 256] net = SimplePose( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def simplepose_resnet18_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=512, keypoints=keypoints, model_name="simplepose_resnet18_coco", **kwargs) def simplepose_resnet50b_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_resnet50b_coco", **kwargs) def simplepose_resnet101b_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet101b(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_resnet101b_coco", **kwargs) def simplepose_resnet152b_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet152b(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_resnet152b_coco", **kwargs) def simplepose_resneta50b_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resneta50b(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_resneta50b_coco", **kwargs) def simplepose_resneta101b_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resneta101b(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_resneta101b_coco", **kwargs) def simplepose_resneta152b_coco(pretrained_backbone=False, keypoints=17, **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. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resneta152b(pretrained=pretrained_backbone).features del backbone[-1] return get_simplepose(backbone=backbone, backbone_out_channels=2048, keypoints=keypoints, model_name="simplepose_resneta152b_coco", **kwargs) def _calc_width(net): import numpy as np 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 _test(): in_size = (256, 192) keypoints = 17 return_heatmap = False 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(batch, 3, in_size[0], in_size[1]) 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()
12,777
36.145349
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/vovnet.py
""" VoVNet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn from .common import conv1x1_block, conv3x3_block, SequentialConcurrent class VoVUnit(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, branch_channels, num_branches, resize, use_residual): super(VoVUnit, self).__init__() self.resize = resize self.use_residual = use_residual if self.resize: self.pool = nn.MaxPool2d( kernel_size=3, stride=2, ceil_mode=True) self.branches = SequentialConcurrent() branch_in_channels = in_channels for i in range(num_branches): self.branches.add_module("branch{}".format(i + 1), conv3x3_block( in_channels=branch_in_channels, out_channels=branch_channels)) branch_in_channels = branch_channels self.concat_conv = conv1x1_block( in_channels=(in_channels + num_branches * branch_channels), out_channels=out_channels) def forward(self, 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(nn.Module): """ VoVNet 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): super(VoVInitBlock, self).__init__() mid_channels = out_channels // 2 self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=mid_channels, stride=2) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels) self.conv3 = conv3x3_block( in_channels=mid_channels, out_channels=out_channels, stride=2) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class VoVNet(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, branch_channels, num_branches, in_channels=3, in_size=(224, 224), num_classes=1000): super(VoVNet, self).__init__() self.in_size = in_size self.num_classes = num_classes init_block_channels = 128 self.features = nn.Sequential() self.features.add_module("init_block", VoVInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): use_residual = (j != 0) resize = (j == 0) and (i != 0) stage.add_module("unit{}".format(j + 1), 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)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight, mode="fan_out", nonlinearity="relu") if module.bias is not None: nn.init.constant_(module.bias, 0) elif isinstance(module, nn.BatchNorm2d): nn.init.constant_(module.weight, 1) nn.init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_vovnet(blocks, slim=False, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_vovnet(blocks=57, model_name="vovnet57", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ vovnet27s, vovnet39, vovnet57, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
10,220
29.601796
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/espnetv2.py
""" ESPNetv2 for ImageNet-1K, implemented in PyTorch. 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 import torch import torch.nn as nn import torch.nn.init as init from .common import conv3x3, conv1x1_block, conv3x3_block, DualPathSequential class PreActivation(nn.Module): """ PreResNet like pure pre-activation block without convolution layer. Parameters: ---------- in_channels : int Number of input channels. """ def __init__(self, in_channels): super(PreActivation, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.PReLU(num_parameters=in_channels) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class ShortcutBlock(nn.Module): """ ESPNetv2 shortcut block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(ShortcutBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, activation=(lambda: nn.PReLU(in_channels))) self.conv2 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class HierarchicalConcurrent(nn.Sequential): """ A container for hierarchical concatenation of modules on the base of the sequential container. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. """ def __init__(self, axis=1): super(HierarchicalConcurrent, self).__init__() self.axis = axis def forward(self, x): out = [] y_prev = None for module in self._modules.values(): y = module(x) if y_prev is not None: y += y_prev out.append(y) y_prev = y out = torch.cat(tuple(out), dim=self.axis) return out class ESPBlock(nn.Module): """ ESPNetv2 block (so-called EESP block). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the branch convolution layers. dilations : list of int Dilation values for branches. """ def __init__(self, in_channels, out_channels, stride, dilations): super(ESPBlock, self).__init__() num_branches = len(dilations) assert (out_channels % num_branches == 0) self.downsample = (stride != 1) mid_channels = out_channels // num_branches self.reduce_conv = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, groups=num_branches, activation=(lambda: nn.PReLU(mid_channels))) self.branches = HierarchicalConcurrent() for i in range(num_branches): self.branches.add_module("branch{}".format(i + 1), conv3x3( in_channels=mid_channels, out_channels=mid_channels, stride=stride, 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, activation=None) self.preactiv = PreActivation(in_channels=out_channels) if not self.downsample: self.activ = nn.PReLU(out_channels) def forward(self, 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(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, x0_channels, dilations): super(DownsampleBlock, self).__init__() inc_channels = out_channels - in_channels self.pool = nn.AvgPool2d( kernel_size=3, stride=2, padding=1) self.eesp = ESPBlock( in_channels=in_channels, out_channels=inc_channels, stride=2, dilations=dilations) self.shortcut_block = ShortcutBlock( in_channels=x0_channels, out_channels=out_channels) self.activ = nn.PReLU(out_channels) def forward(self, x, x0): y1 = self.pool(x) y2, _ = self.eesp(x, None) x = torch.cat((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(nn.Module): """ ESPNetv2 initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(ESPInitBlock, self).__init__() self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=2, activation=(lambda: nn.PReLU(out_channels))) self.pool = nn.AvgPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x, x0): x = self.conv(x) x0 = self.pool(x0) return x, x0 class ESPFinalBlock(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, final_groups): super(ESPFinalBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, groups=in_channels, activation=(lambda: nn.PReLU(in_channels))) self.conv2 = conv1x1_block( in_channels=in_channels, out_channels=out_channels, groups=final_groups, activation=(lambda: nn.PReLU(out_channels))) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class ESPNetv2(nn.Module): """ 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. 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. num_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, in_channels=3, in_size=(224, 224), num_classes=1000): super(ESPNetv2, self).__init__() self.in_size = in_size self.num_classes = num_classes x0_channels = in_channels self.features = DualPathSequential( return_two=False, first_ordinals=0, last_ordinals=2) self.features.add_module("init_block", ESPInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential() 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]) else: unit = ESPBlock( in_channels=in_channels, out_channels=out_channels, stride=1, dilations=dilations[i][j]) stage.add_module("unit{}".format(j + 1), unit) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", ESPFinalBlock( in_channels=in_channels, out_channels=final_block_channels, final_groups=final_block_groups)) in_channels = final_block_channels self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Sequential() self.output.add_module("dropout", nn.Dropout(p=dropout_rate)) self.output.add_module("fc", nn.Linear( in_features=in_channels, out_features=num_classes)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x, x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_espnetv2(width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_espnetv2(width_scale=2.0, model_name="espnetv2_w2", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ espnetv2_wd2, espnetv2_w1, espnetv2_w5d4, espnetv2_w3d2, espnetv2_w2, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) # assert (model != espnetv2_wd2 or weight_count == 1241332) # assert (model != espnetv2_w1 or weight_count == 1670072) # assert (model != espnetv2_w5d4 or weight_count == 1965440) # assert (model != espnetv2_w3d2 or weight_count == 2314856) # assert (model != espnetv2_w2 or weight_count == 3498136) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
17,203
30.336976
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/shufflenet.py
""" ShuffleNet for ImageNet-1K, implemented in PyTorch. 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 import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv3x3, depthwise_conv3x3, ChannelShuffle class ShuffleUnit(nn.Module): """ 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): super(ShuffleUnit, self).__init__() self.downsample = downsample mid_channels = out_channels // 4 if downsample: out_channels -= in_channels self.compress_conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels, groups=(1 if ignore_group else groups)) self.compress_bn1 = nn.BatchNorm2d(num_features=mid_channels) self.c_shuffle = ChannelShuffle( channels=mid_channels, groups=groups) self.dw_conv2 = depthwise_conv3x3( channels=mid_channels, stride=(2 if self.downsample else 1)) self.dw_bn2 = nn.BatchNorm2d(num_features=mid_channels) self.expand_conv3 = conv1x1( in_channels=mid_channels, out_channels=out_channels, groups=groups) self.expand_bn3 = nn.BatchNorm2d(num_features=out_channels) if downsample: self.avgpool = nn.AvgPool2d(kernel_size=3, stride=2, padding=1) self.activ = nn.ReLU(inplace=True) def forward(self, 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 = torch.cat((x, identity), dim=1) else: x = x + identity x = self.activ(x) return x class ShuffleInitBlock(nn.Module): """ 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): super(ShuffleInitBlock, self).__init__() self.conv = conv3x3( in_channels=in_channels, out_channels=out_channels, stride=2) self.bn = nn.BatchNorm2d(num_features=out_channels) self.activ = nn.ReLU(inplace=True) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.bn(x) x = self.activ(x) x = self.pool(x) return x class ShuffleNet(nn.Module): """ 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, groups, in_channels=3, in_size=(224, 224), num_classes=1000): super(ShuffleNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", 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.Sequential() for j, out_channels in enumerate(channels_per_stage): downsample = (j == 0) ignore_group = (i == 0) and (j == 0) stage.add_module("unit{}".format(j + 1), ShuffleUnit( in_channels=in_channels, out_channels=out_channels, groups=groups, downsample=downsample, ignore_group=ignore_group)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_shufflenet(groups, width_scale, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_shufflenet(groups=3, width_scale=0.25, model_name="shufflenet_g3_wd4", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
15,779
31.875
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/bamresnet.py
""" BAM-ResNet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv1x1_block, conv3x3_block from .resnet import ResInitBlock, ResUnit class DenseBlock(nn.Module): """ Standard dense block with Batch normalization and ReLU activation. Parameters: ---------- in_features : int Number of input features. out_features : int Number of output features. """ def __init__(self, in_features, out_features): super(DenseBlock, self).__init__() self.fc = nn.Linear( in_features=in_features, out_features=out_features) self.bn = nn.BatchNorm1d(num_features=out_features) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.fc(x) x = self.bn(x) x = self.activ(x) return x class ChannelGate(nn.Module): """ BAM channel gate block. Parameters: ---------- channels : int Number of input/output channels. reduction_ratio : int, default 16 Channel reduction ratio. num_layers : int, default 1 Number of dense blocks. """ def __init__(self, channels, reduction_ratio=16, num_layers=1): super(ChannelGate, self).__init__() mid_channels = channels // reduction_ratio self.pool = nn.AdaptiveAvgPool2d(output_size=(1, 1)) self.init_fc = DenseBlock( in_features=channels, out_features=mid_channels) self.main_fcs = nn.Sequential() for i in range(num_layers - 1): self.main_fcs.add_module("fc{}".format(i + 1), DenseBlock( in_features=mid_channels, out_features=mid_channels)) self.final_fc = nn.Linear( in_features=mid_channels, out_features=channels) def forward(self, x): input = x x = self.pool(x) x = x.view(x.size(0), -1) x = self.init_fc(x) x = self.main_fcs(x) x = self.final_fc(x) x = x.unsqueeze(2).unsqueeze(3).expand_as(input) return x class SpatialGate(nn.Module): """ BAM spatial gate block. Parameters: ---------- channels : int Number of input/output channels. 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, reduction_ratio=16, num_dil_convs=2, dilation=4): super(SpatialGate, self).__init__() mid_channels = channels // reduction_ratio self.init_conv = conv1x1_block( in_channels=channels, out_channels=mid_channels, stride=1, bias=True) self.dil_convs = nn.Sequential() for i in range(num_dil_convs): self.dil_convs.add_module("conv{}".format(i + 1), conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=1, padding=dilation, dilation=dilation, bias=True)) self.final_conv = conv1x1( in_channels=mid_channels, out_channels=1, stride=1, bias=True) def forward(self, x): input = x x = self.init_conv(x) x = self.dil_convs(x) x = self.final_conv(x) x = x.expand_as(input) return x class BamBlock(nn.Module): """ BAM attention block for BAM-ResNet. Parameters: ---------- channels : int Number of input/output channels. """ def __init__(self, channels): super(BamBlock, self).__init__() self.ch_att = ChannelGate(channels=channels) self.sp_att = SpatialGate(channels=channels) self.sigmoid = nn.Sigmoid() def forward(self, x): att = 1 + self.sigmoid(self.ch_att(x) * self.sp_att(x)) x = x * att return x class BamResUnit(nn.Module): """ BAM-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. """ def __init__(self, in_channels, out_channels, stride, bottleneck): super(BamResUnit, self).__init__() self.use_bam = (stride != 1) if self.use_bam: self.bam = BamBlock(channels=in_channels) self.res_unit = ResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, conv1_stride=False) def forward(self, x): if self.use_bam: x = self.bam(x) x = self.res_unit(x) return x class BamResNet(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(224, 224), num_classes=1000): super(BamResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), BamResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resnet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resnet(blocks=152, model_name="bam_resnet152", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ bam_resnet18, bam_resnet34, bam_resnet50, bam_resnet101, bam_resnet152, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
13,297
28.420354
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/resattnet.py
""" ResAttNet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from .common import conv1x1, conv7x7_block, pre_conv1x1_block, pre_conv3x3_block, Hourglass class PreResBottleneck(nn.Module): """ PreResNet bottleneck block for residual path in PreResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(PreResBottleneck, self).__init__() mid_channels = out_channels // 4 self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, return_preact=True) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride) self.conv3 = pre_conv1x1_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x, x_pre_activ = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x, x_pre_activ class ResBlock(nn.Module): """ Residual block with pre-activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride=1): super(ResBlock, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = PreResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, stride=stride) def forward(self, 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(nn.Module): """ Interpolation block. Parameters: ---------- scale_factor : float Multiplier for spatial size. """ def __init__(self, scale_factor): super(InterpolationBlock, self).__init__() self.scale_factor = scale_factor def forward(self, x): return F.interpolate( input=x, scale_factor=self.scale_factor, mode="bilinear", align_corners=True) class DoubleSkipBlock(nn.Module): """ Double skip connection block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(DoubleSkipBlock, self).__init__() self.skip1 = ResBlock( in_channels=in_channels, out_channels=out_channels) def forward(self, x): x = x + self.skip1(x) return x class ResBlockSequence(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, length): super(ResBlockSequence, self).__init__() self.blocks = nn.Sequential() for i in range(length): self.blocks.add_module("block{}".format(i + 1), ResBlock( in_channels=in_channels, out_channels=out_channels)) def forward(self, x): x = self.blocks(x) return x class DownAttBlock(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, length): super(DownAttBlock, self).__init__() self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) self.res_blocks = ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=length) def forward(self, x): x = self.pool(x) x = self.res_blocks(x) return x class UpAttBlock(nn.Module): """ 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. scale_factor : float Multiplier for spatial size. """ def __init__(self, in_channels, out_channels, length, scale_factor): super(UpAttBlock, self).__init__() self.res_blocks = ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=length) self.upsample = InterpolationBlock(scale_factor) def forward(self, x): x = self.res_blocks(x) x = self.upsample(x) return x class MiddleAttBlock(nn.Module): """ Middle sub-block for attention block. Parameters: ---------- channels : int Number of input/output channels. """ def __init__(self, channels): super(MiddleAttBlock, self).__init__() self.conv1 = pre_conv1x1_block( in_channels=channels, out_channels=channels) self.conv2 = pre_conv1x1_block( in_channels=channels, out_channels=channels) self.sigmoid = nn.Sigmoid() def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.sigmoid(x) return x class AttBlock(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, hourglass_depth, att_scales): super(AttBlock, self).__init__() assert (len(att_scales) == 3) scale_factor = 2 scale_p, scale_t, scale_r = att_scales self.init_blocks = ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=scale_p) down_seq = nn.Sequential() up_seq = nn.Sequential() skip_seq = nn.Sequential() for i in range(hourglass_depth): down_seq.add_module("down{}".format(i + 1), DownAttBlock( in_channels=in_channels, out_channels=out_channels, length=scale_r)) up_seq.add_module("up{}".format(i + 1), UpAttBlock( in_channels=in_channels, out_channels=out_channels, length=scale_r, scale_factor=scale_factor)) if i == 0: skip_seq.add_module("skip1", ResBlockSequence( in_channels=in_channels, out_channels=out_channels, length=scale_t)) else: skip_seq.add_module("skip{}".format(i + 1), DoubleSkipBlock( in_channels=in_channels, out_channels=out_channels)) 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) self.final_block = ResBlock( in_channels=in_channels, out_channels=out_channels) def forward(self, 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(nn.Module): """ ResAttNet 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): super(ResAttInitBlock, self).__init__() self.conv = conv7x7_block( in_channels=in_channels, out_channels=out_channels, stride=2) self.pool = nn.MaxPool2d( kernel_size=3, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.pool(x) return x class PreActivation(nn.Module): """ 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. """ def __init__(self, in_channels): super(PreActivation, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class ResAttNet(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, attentions, att_scales, in_channels=3, in_size=(224, 224), num_classes=1000): super(ResAttNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResAttInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): hourglass_depth = len(channels) - 1 - i stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 1 if (i == 0) or (j != 0) else 2 if attentions[i][j]: stage.add_module("unit{}".format(j + 1), AttBlock( in_channels=in_channels, out_channels=out_channels, hourglass_depth=hourglass_depth, att_scales=att_scales)) else: stage.add_module("unit{}".format(j + 1), ResBlock( in_channels=in_channels, out_channels=out_channels, stride=stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_resattnet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_resattnet(blocks=452, model_name="resattnet452", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ resattnet56, resattnet92, resattnet128, resattnet164, resattnet200, resattnet236, resattnet452, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
20,035
28.464706
117
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/centernet.py
""" CenterNet for ImageNet-1K, implemented in PyTorch. 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 import torch import torch.nn as nn from .common import conv1x1, conv3x3_block, DeconvBlock, Concurrent from .resnet import resnet18, resnet50b, resnet101b class CenterNetDecoderUnit(nn.Module): """ CenterNet decoder unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(CenterNetDecoderUnit, self).__init__() self.conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bias=True) self.deconv = DeconvBlock( in_channels=out_channels, out_channels=out_channels, kernel_size=4, stride=2, padding=1) def forward(self, x): x = self.conv(x) x = self.deconv(x) return x class CenterNetHeadBlock(nn.Module): """ 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): super(CenterNetHeadBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=in_channels, bias=True, use_bn=False) self.conv2 = conv1x1( in_channels=in_channels, out_channels=out_channels, bias=True) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class CenterNetHeatmapBlock(nn.Module): """ 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): super(CenterNetHeatmapBlock, self).__init__() self.do_nms = do_nms self.head = CenterNetHeadBlock( in_channels=in_channels, out_channels=out_channels) self.sigmoid = nn.Sigmoid() if self.do_nms: self.pool = nn.MaxPool2d( kernel_size=3, stride=1, padding=1) def forward(self, x): x = self.head(x) x = self.sigmoid(x) if self.do_nms: y = self.pool(x) x = x * (y == x) else: eps = 1e-4 x = x.clamp(min=eps, max=(1.0 - eps)) return x class CenterNetHeatmapMaxDet(nn.Module): """ CenterNet decoder for heads (heatmap, wh, reg). Parameters: ---------- topk : int, default 40 Keep only `topk` detections. scale : int, default is 4 Downsampling scale factor. """ def __init__(self, topk=40, scale=4): super(CenterNetHeatmapMaxDet, self).__init__() self.topk = topk self.scale = scale def forward(self, x): heatmap = x[:, :-4] wh = x[:, -4:-2] reg = x[:, -2:] batch, _, out_h, out_w = heatmap.shape scores, indices = heatmap.view((batch, -1)).topk(k=self.topk) topk_classes = (indices / (out_h * out_w)).type(torch.float32) topk_indices = indices.fmod(out_h * out_w) topk_ys = (topk_indices / out_w).type(torch.float32) topk_xs = topk_indices.fmod(out_w).type(torch.float32) center = reg.permute(0, 2, 3, 1).view((batch, -1, 2)) wh = wh.permute(0, 2, 3, 1).view((batch, -1, 2)) xs = torch.gather(center[:, :, 0], dim=-1, index=topk_indices) ys = torch.gather(center[:, :, 1], dim=-1, index=topk_indices) topk_xs = topk_xs + xs topk_ys = topk_ys + ys w = torch.gather(wh[:, :, 0], dim=-1, index=topk_indices) h = torch.gather(wh[:, :, 1], dim=-1, index=topk_indices) half_w = 0.5 * w half_h = 0.5 * h bboxes = torch.stack((topk_xs - half_w, topk_ys - half_h, topk_xs + half_w, topk_ys + half_h), dim=-1) bboxes = bboxes * self.scale topk_classes = topk_classes.unsqueeze(dim=-1) scores = scores.unsqueeze(dim=-1) result = torch.cat((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(nn.Module): """ 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. 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. num_classes : int, default 80 Number of classification classes. """ def __init__(self, backbone, backbone_out_channels, channels, return_heatmap=False, topk=40, in_channels=3, in_size=(512, 512), num_classes=80): super(CenterNet, self).__init__() self.in_size = in_size self.in_channels = in_channels self.return_heatmap = return_heatmap self.backbone = backbone self.decoder = nn.Sequential() in_channels = backbone_out_channels for i, out_channels in enumerate(channels): self.decoder.add_module("unit{}".format(i + 1), CenterNetDecoderUnit( in_channels=in_channels, out_channels=out_channels)) in_channels = out_channels heads = Concurrent() heads.add_module("heapmap_block", CenterNetHeatmapBlock( in_channels=in_channels, out_channels=num_classes, do_nms=(not self.return_heatmap))) heads.add_module("wh_block", CenterNetHeadBlock( in_channels=in_channels, out_channels=2)) heads.add_module("reg_block", CenterNetHeadBlock( in_channels=in_channels, out_channels=2)) self.decoder.add_module("heads", heads) if not self.return_heatmap: self.heatmap_max_det = CenterNetHeatmapMaxDet( topk=topk, scale=4) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, 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, num_classes, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. Returns: ------- nn.Module A network. """ channels = [256, 128, 64] net = CenterNet( backbone=backbone, backbone_out_channels=backbone_out_channels, channels=channels, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def centernet_resnet18_voc(pretrained_backbone=False, num_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. num_classes : int, default 20 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features del backbone[-1] return get_centernet(backbone=backbone, backbone_out_channels=512, num_classes=num_classes, model_name="centernet_resnet18_voc", **kwargs) def centernet_resnet18_coco(pretrained_backbone=False, num_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. num_classes : int, default 80 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet18(pretrained=pretrained_backbone).features del backbone[-1] return get_centernet(backbone=backbone, backbone_out_channels=512, num_classes=num_classes, model_name="centernet_resnet18_coco", **kwargs) def centernet_resnet50b_voc(pretrained_backbone=False, num_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. num_classes : int, default 20 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features del backbone[-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, num_classes=num_classes, model_name="centernet_resnet50b_voc", **kwargs) def centernet_resnet50b_coco(pretrained_backbone=False, num_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. num_classes : int, default 80 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features del backbone[-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, num_classes=num_classes, model_name="centernet_resnet50b_coco", **kwargs) def centernet_resnet101b_voc(pretrained_backbone=False, num_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. num_classes : int, default 20 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet101b(pretrained=pretrained_backbone).features del backbone[-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, num_classes=num_classes, model_name="centernet_resnet101b_voc", **kwargs) def centernet_resnet101b_coco(pretrained_backbone=False, num_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. num_classes : int, default 80 Number of classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet101b(pretrained=pretrained_backbone).features del backbone[-1] return get_centernet(backbone=backbone, backbone_out_channels=2048, num_classes=num_classes, model_name="centernet_resnet101b_coco", **kwargs) def _calc_width(net): import numpy as np 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 _test(): 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(batch, 3, in_size[0], in_size[1]) 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()
16,535
32.204819
118
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/xdensenet_cifar.py
""" X-DenseNet for CIFAR/SVHN, implemented in PyTorch. 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 import torch import torch.nn as nn import torch.nn.init as init from .common import conv3x3 from .preresnet import PreResActivation from .densenet import TransitionBlock from .xdensenet import pre_xconv3x3_block, XDenseUnit class XDenseSimpleUnit(nn.Module): """ X-DenseNet simple unit for CIFAR. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. 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, dropout_rate, expand_ratio): super(XDenseSimpleUnit, self).__init__() self.use_dropout = (dropout_rate != 0.0) inc_channels = out_channels - in_channels self.conv = pre_xconv3x3_block( in_channels=in_channels, out_channels=inc_channels, expand_ratio=expand_ratio) if self.use_dropout: self.dropout = nn.Dropout(p=dropout_rate) def forward(self, x): identity = x x = self.conv(x) if self.use_dropout: x = self.dropout(x) x = torch.cat((identity, x), dim=1) return x class CIFARXDenseNet(nn.Module): """ 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, dropout_rate=0.0, expand_ratio=2, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARXDenseNet, self).__init__() self.in_size = in_size self.num_classes = num_classes unit_class = XDenseUnit if bottleneck else XDenseSimpleUnit self.features = nn.Sequential() self.features.add_module("init_block", 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.Sequential() if i != 0: stage.add_module("trans{}".format(i + 1), TransitionBlock( in_channels=in_channels, out_channels=(in_channels // 2))) in_channels = in_channels // 2 for j, out_channels in enumerate(channels_per_stage): stage.add_module("unit{}".format(j + 1), unit_class( in_channels=in_channels, out_channels=out_channels, dropout_rate=dropout_rate, expand_ratio=expand_ratio)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("post_activ", PreResActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_xdensenet_cifar(num_classes, blocks, growth_rate, bottleneck, expand_ratio=2, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create X-DenseNet model for CIFAR with specific parameters. Parameters: ---------- num_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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ assert (num_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, num_classes=num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def xdensenet40_2_k24_bc_cifar10(num_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: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(num_classes=num_classes, blocks=40, growth_rate=24, bottleneck=True, model_name="xdensenet40_2_k24_bc_cifar10", **kwargs) def xdensenet40_2_k24_bc_cifar100(num_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: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(num_classes=num_classes, blocks=40, growth_rate=24, bottleneck=True, model_name="xdensenet40_2_k24_bc_cifar100", **kwargs) def xdensenet40_2_k24_bc_svhn(num_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: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(num_classes=num_classes, blocks=40, growth_rate=24, bottleneck=True, model_name="xdensenet40_2_k24_bc_svhn", **kwargs) def xdensenet40_2_k36_bc_cifar10(num_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: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(num_classes=num_classes, blocks=40, growth_rate=36, bottleneck=True, model_name="xdensenet40_2_k36_bc_cifar10", **kwargs) def xdensenet40_2_k36_bc_cifar100(num_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: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(num_classes=num_classes, blocks=40, growth_rate=36, bottleneck=True, model_name="xdensenet40_2_k36_bc_cifar100", **kwargs) def xdensenet40_2_k36_bc_svhn(num_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: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_xdensenet_cifar(num_classes=num_classes, blocks=40, growth_rate=36, bottleneck=True, model_name="xdensenet40_2_k36_bc_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch 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, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
12,852
33.831978
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/revnet.py
""" RevNet for ImageNet-1K, implemented in PyTorch. Original paper: 'The Reversible Residual Network: Backpropagation Without Storing Activations,' https://arxiv.org/abs/1707.04585. """ __all__ = ['RevNet', 'revnet38', 'revnet110', 'revnet164'] import os from contextlib import contextmanager import torch import torch.nn as nn import torch.nn.init as init from torch.autograd import Variable from .common import conv1x1, conv3x3, conv1x1_block, conv3x3_block, pre_conv1x1_block, pre_conv3x3_block use_context_mans = int( torch.__version__[0]) * 100 + int(torch.__version__[2]) - (1 if 'a' in torch.__version__ else 0) > 3 @contextmanager def set_grad_enabled(grad_mode): if not use_context_mans: yield else: with torch.set_grad_enabled(grad_mode) as c: yield [c] class ReversibleBlockFunction(torch.autograd.Function): """ RevNet reversible block function. """ @staticmethod def forward(ctx, x, fm, gm, *params): with torch.no_grad(): x1, x2 = torch.chunk(x, chunks=2, dim=1) x1 = x1.contiguous() x2 = x2.contiguous() y1 = x1 + fm(x2) y2 = x2 + gm(y1) y = torch.cat((y1, y2), dim=1) x1.set_() x2.set_() y1.set_() y2.set_() del x1, x2, y1, y2 ctx.save_for_backward(x, y) ctx.fm = fm ctx.gm = gm return y @staticmethod def backward(ctx, grad_y): fm = ctx.fm gm = ctx.gm x, y = ctx.saved_variables y1, y2 = torch.chunk(y, chunks=2, dim=1) y1 = y1.contiguous() y2 = y2.contiguous() with torch.no_grad(): y1_z = Variable(y1.data, requires_grad=True) x2 = y2 - gm(y1_z) x1 = y1 - fm(x2) with set_grad_enabled(True): x1_ = Variable(x1.data, requires_grad=True) x2_ = Variable(x2.data, requires_grad=True) y1_ = x1_ + fm.forward(x2_) y2_ = x2_ + gm(y1_) y = torch.cat((y1_, y2_), dim=1) dd = torch.autograd.grad(y, (x1_, x2_) + tuple(gm.parameters()) + tuple(fm.parameters()), grad_y) gm_params_len = len([p for p in gm.parameters()]) gm_params_grads = dd[2:2 + gm_params_len] fm_params_grads = dd[2 + gm_params_len:] grad_x = torch.cat((dd[0], dd[1]), dim=1) y1_.detach_() y2_.detach_() del y1_, y2_ x.data.set_(torch.cat((x1, x2), dim=1).data.contiguous()) return (grad_x, None, None) + fm_params_grads + gm_params_grads class ReversibleBlock(nn.Module): """ RevNet reversible block. Parameters: ---------- fm : nn.Module Fm-function. gm : nn.Module Gm-function. """ def __init__(self, fm, gm): super(ReversibleBlock, self).__init__() self.gm = gm self.fm = fm self.rev_funct = ReversibleBlockFunction.apply def forward(self, x): assert (x.shape[1] % 2 == 0) params = [w for w in self.fm.parameters()] + [w for w in self.gm.parameters()] y = self.rev_funct(x, self.fm, self.gm, *params) x.data.set_() return y def inverse(self, y): assert (y.shape[1] % 2 == 0) y1, y2 = torch.chunk(y, chunks=2, dim=1) y1 = y1.contiguous() y2 = y2.contiguous() x2 = y2 - self.gm(y1) x1 = y1 - self.fm(x2) x = torch.cat((x1, x2), dim=1) return x class RevResBlock(nn.Module): """ Simple RevNet block for residual path in RevNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. preactivate : bool Whether use pre-activation for the first convolution block. """ def __init__(self, in_channels, out_channels, stride, preactivate): super(RevResBlock, self).__init__() if preactivate: self.conv1 = pre_conv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride) else: self.conv1 = conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride) self.conv2 = pre_conv3x3_block( in_channels=out_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class RevResBottleneck(nn.Module): """ RevNet bottleneck block for residual path in RevNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. preactivate : bool Whether use pre-activation for the first convolution block. bottleneck_factor : int, default 4 Bottleneck factor. """ def __init__(self, in_channels, out_channels, stride, preactivate, bottleneck_factor=4): super(RevResBottleneck, self).__init__() mid_channels = out_channels // bottleneck_factor if preactivate: self.conv1 = pre_conv1x1_block( in_channels=in_channels, out_channels=mid_channels) else: self.conv1 = conv1x1( in_channels=in_channels, out_channels=mid_channels) self.conv2 = pre_conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride) self.conv3 = pre_conv1x1_block( in_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class RevUnit(nn.Module): """ RevNet unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bottleneck : bool Whether to use a bottleneck or simple block in units. preactivate : bool Whether use pre-activation for the first convolution block. """ def __init__(self, in_channels, out_channels, stride, bottleneck, preactivate): super(RevUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) body_class = RevResBottleneck if bottleneck else RevResBlock if (not self.resize_identity) and (stride == 1): assert (in_channels % 2 == 0) assert (out_channels % 2 == 0) in_channels2 = in_channels // 2 out_channels2 = out_channels // 2 gm = body_class( in_channels=in_channels2, out_channels=out_channels2, stride=1, preactivate=preactivate) fm = body_class( in_channels=in_channels2, out_channels=out_channels2, stride=1, preactivate=preactivate) self.body = ReversibleBlock(gm, fm) else: self.body = body_class( in_channels=in_channels, out_channels=out_channels, stride=stride, preactivate=preactivate) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) def forward(self, x): if self.resize_identity: identity = self.identity_conv(x) x = self.body(x) x = x + identity else: x = self.body(x) return x class RevPostActivation(nn.Module): """ RevNet specific post-activation block. Parameters: ---------- in_channels : int Number of input channels. """ def __init__(self, in_channels): super(RevPostActivation, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class RevNet(nn.Module): """ RevNet model from 'The Reversible Residual Network: Backpropagation Without Storing Activations,' https://arxiv.org/abs/1707.04585. 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, bottleneck, in_channels=3, in_size=(224, 224), num_classes=1000): super(RevNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 preactivate = (j != 0) or (i != 0) stage.add_module("unit{}".format(j + 1), RevUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, bottleneck=bottleneck, preactivate=preactivate)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_postactiv", RevPostActivation(in_channels=in_channels)) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=56, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_revnet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ if blocks == 38: layers = [3, 3, 3] channels_per_layers = [32, 64, 112] bottleneck = False elif blocks == 110: layers = [9, 9, 9] channels_per_layers = [32, 64, 128] bottleneck = False elif blocks == 164: layers = [9, 9, 9] channels_per_layers = [128, 256, 512] bottleneck = True else: raise ValueError("Unsupported RevNet with number of blocks: {}".format(blocks)) init_block_channels = 32 channels = [[ci] * li for (ci, li) in zip(channels_per_layers, layers)] net = RevNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def revnet38(**kwargs): """ RevNet-38 model from 'The Reversible Residual Network: Backpropagation Without Storing Activations,' https://arxiv.org/abs/1707.04585. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_revnet(blocks=38, model_name="revnet38", **kwargs) def revnet110(**kwargs): """ RevNet-110 model from 'The Reversible Residual Network: Backpropagation Without Storing Activations,' https://arxiv.org/abs/1707.04585. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_revnet(blocks=110, model_name="revnet110", **kwargs) def revnet164(**kwargs): """ RevNet-164 model from 'The Reversible Residual Network: Backpropagation Without Storing Activations,' https://arxiv.org/abs/1707.04585. Parameters: ---------- pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_revnet(blocks=164, model_name="revnet164", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ revnet38, revnet110, revnet164, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != revnet38 or weight_count == 685864) assert (model != revnet110 or weight_count == 1982600) assert (model != revnet164 or weight_count == 2491656) x = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
15,590
28.142056
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/ntsnet_cub.py
""" NTS-Net for CUB-200-2011, implemented in PyTorch. 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 torch import torch.nn as nn import torch.nn.functional as F import torch.nn.init as init from .common import conv1x1, conv3x3, Flatten 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(nn.Module): """ Navigator branch block for Navigator unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(NavigatorBranch, self).__init__() mid_channels = 128 self.down_conv = conv3x3( in_channels=in_channels, out_channels=mid_channels, stride=stride, bias=True) self.activ = nn.ReLU(inplace=False) self.tidy_conv = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) self.flatten = Flatten() def forward(self, x): y = self.down_conv(x) y = self.activ(y) z = self.tidy_conv(y) z = self.flatten(z) return z, y class NavigatorUnit(nn.Module): """ Navigator init. """ def __init__(self): super(NavigatorUnit, self).__init__() self.branch1 = NavigatorBranch( in_channels=2048, out_channels=6, stride=1) self.branch2 = NavigatorBranch( in_channels=128, out_channels=6, stride=2) self.branch3 = NavigatorBranch( in_channels=128, out_channels=9, stride=2) def forward(self, x): t1, x = self.branch1(x) t2, x = self.branch2(x) t3, _ = self.branch3(x) return torch.cat((t1, t2, t3), dim=1) class NTSNet(nn.Module): """ 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, backbone, aux=False, top_n=4, in_channels=3, in_size=(448, 448), num_classes=200): super(NTSNet, self).__init__() assert (in_channels > 0) self.in_size = in_size self.num_classes = num_classes pad_side = 224 pad_width = (pad_side, pad_side, pad_side, pad_side) self.top_n = top_n self.aux = aux self.num_cat = 4 _, 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) self.backbone = backbone self.backbone_tail = nn.Sequential() self.backbone_tail.add_module("final_pool", nn.AdaptiveAvgPool2d(1)) self.backbone_tail.add_module("flatten", Flatten()) self.backbone_tail.add_module("dropout", nn.Dropout(p=0.5)) self.backbone_classifier = nn.Linear( in_features=(512 * 4), out_features=num_classes) self.pad = nn.ZeroPad2d(padding=pad_width) self.navigator_unit = NavigatorUnit() self.concat_net = nn.Linear( in_features=(2048 * (self.num_cat + 1)), out_features=num_classes) if self.aux: self.partcls_net = nn.Linear( in_features=(512 * 4), out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, 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.detach().cpu().numpy()] 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_index = torch.from_numpy(top_n_index).long().to(x.device) top_n_prob = torch.gather(rpn_score, dim=1, index=top_n_index) batch = x.size(0) part_imgs = torch.zeros(batch, self.top_n, 3, 224, 224, dtype=x.dtype, device=x.device) x_pad = self.pad(x) 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.interpolate( input=x_pad[i:i + 1, :, y0:y1, x0:x1], size=(224, 224), mode="bilinear", align_corners=True) part_imgs = part_imgs.view(batch * self.top_n, 3, 224, 224) part_features = self.backbone_tail(self.backbone(part_imgs.detach())) part_feature = part_features.view(batch, self.top_n, -1) part_feature = part_feature[:, :self.num_cat, :].contiguous() part_feature = part_feature.view(batch, -1) raw_features = self.backbone_tail(raw_pre_features.detach()) concat_out = torch.cat((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).view(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, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ backbone = resnet50b(pretrained=pretrained_backbone).features del backbone[-1] return get_ntsnet(backbone=backbone, aux=aux, model_name="ntsnet_cub", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False aux = True models = [ ntsnet_cub, ] for model in models: net = model(pretrained=pretrained, aux=aux) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(5, 3, 448, 448) ys = net(x) y = ys[0] if aux else ys y.sum().backward() assert (tuple(y.size()) == (5, 200)) if __name__ == "__main__": _test()
14,019
32.54067
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/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(num_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: ---------- num_classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_proxylessnas(num_classes=num_classes, version="cpu", model_name="proxylessnas_cpu_cub", **kwargs) def proxylessnas_gpu_cub(num_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: ---------- num_classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_proxylessnas(num_classes=num_classes, version="gpu", model_name="proxylessnas_gpu_cub", **kwargs) def proxylessnas_mobile_cub(num_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: ---------- num_classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_proxylessnas(num_classes=num_classes, version="mobile", model_name="proxylessnas_mobile_cub", **kwargs) def proxylessnas_mobile14_cub(num_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: ---------- num_classes : int, default 200 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_proxylessnas(num_classes=num_classes, version="mobile14", model_name="proxylessnas_mobile14_cub", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ proxylessnas_cpu_cub, proxylessnas_gpu_cub, proxylessnas_mobile_cub, proxylessnas_mobile14_cub, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(14, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (14, 200)) if __name__ == "__main__": _test()
4,155
32.788618
120
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/ibnresnet.py
""" IBN-ResNet for ImageNet-1K, implemented in PyTorch. 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 import torch.nn as nn import torch.nn.init as init from .common import conv1x1_block, conv3x3_block, IBN from .resnet import ResInitBlock class IBNConvBlock(nn.Module): """ 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_ibn : bool, default False Whether use Instance-Batch Normalization. activate : bool, default True Whether activate the convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False, use_ibn=False, activate=True): super(IBNConvBlock, self).__init__() self.activate = activate self.use_ibn = use_ibn self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) if self.use_ibn: self.ibn = IBN(channels=out_channels) else: self.bn = nn.BatchNorm2d(num_features=out_channels) if self.activate: self.activ = nn.ReLU(inplace=True) def forward(self, 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, stride=1, groups=1, bias=False, use_ibn=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. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. use_ibn : bool, default False Whether use Instance-Batch Normalization. activate : bool, default True Whether activate the convolution block. """ return IBNConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=0, groups=groups, bias=bias, use_ibn=use_ibn, activate=activate) class IBNResBottleneck(nn.Module): """ 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. stride : 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. """ def __init__(self, in_channels, out_channels, stride, conv1_ibn): super(IBNResBottleneck, self).__init__() mid_channels = out_channels // 4 self.conv1 = ibn_conv1x1_block( in_channels=in_channels, out_channels=mid_channels, use_ibn=conv1_ibn) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class IBNResUnit(nn.Module): """ IBN-ResNet unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. """ def __init__(self, in_channels, out_channels, stride, conv1_ibn): super(IBNResUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = IBNResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, conv1_ibn=conv1_ibn) if self.resize_identity: self.identity_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, stride=stride, activation=None) self.activ = nn.ReLU(inplace=True) def forward(self, 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(nn.Module): """ 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. 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. num_classes : int, default 1000 Number of classification classes. """ def __init__(self, channels, init_block_channels, in_channels=3, in_size=(224, 224), num_classes=1000): super(IBNResNet, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = nn.Sequential() self.features.add_module("init_block", ResInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 conv1_ibn = (out_channels < 2048) stage.add_module("unit{}".format(j + 1), IBNResUnit( in_channels=in_channels, out_channels=out_channels, stride=stride, conv1_ibn=conv1_ibn)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_pool", nn.AvgPool2d( kernel_size=7, stride=1)) self.output = nn.Linear( in_features=in_channels, out_features=num_classes) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = x.view(x.size(0), -1) x = self.output(x) return x def get_ibnresnet(blocks, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_ibnresnet(blocks=152, model_name="ibn_resnet152", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ ibn_resnet50, ibn_resnet101, ibn_resnet152, ] for model in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 224, 224) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, 1000)) if __name__ == "__main__": _test()
12,570
29.002387
115
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/common.py
""" Common routines for models in PyTorch. """ __all__ = ['round_channels', 'Identity', 'BreakBlock', 'Swish', 'HSigmoid', 'HSwish', 'get_activation_layer', 'SelectableDense', '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', 'AsymConvBlock', 'asym_conv3x3_block', 'DeconvBlock', 'deconv3x3_block', 'NormActivation', 'InterpolationBlock', 'ChannelShuffle', 'ChannelShuffle2', 'SEBlock', 'SABlock', 'SAConvBlock', 'saconv3x3_block', 'DucBlock', 'IBN', 'DualPathSequential', 'Concurrent', 'SequentialConcurrent', 'ParametricSequential', 'ParametricConcurrent', 'Hourglass', 'SesquialteralHourglass', 'MultiOutputSequential', 'ParallelConcurent', 'DualPathParallelConcurent', 'Flatten', 'HeatmapMaxDetBlock'] import math from inspect import isfunction import torch import torch.nn as nn import torch.nn.functional as F from torch.nn.parameter import Parameter 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 Identity(nn.Module): """ Identity block. """ def __init__(self): super(Identity, self).__init__() def forward(self, x): return x def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) class BreakBlock(nn.Module): """ Break coonnection block for hourglass. """ def __init__(self): super(BreakBlock, self).__init__() def forward(self, x): return None def __repr__(self): return '{name}()'.format(name=self.__class__.__name__) class Swish(nn.Module): """ Swish activation function from 'Searching for Activation Functions,' https://arxiv.org/abs/1710.05941. """ def forward(self, x): return x * torch.sigmoid(x) class HSigmoid(nn.Module): """ Approximated sigmoid function, so-called hard-version of sigmoid from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. """ def forward(self, x): return F.relu6(x + 3.0, inplace=True) / 6.0 class HSwish(nn.Module): """ H-Swish activation function from 'Searching for MobileNetV3,' https://arxiv.org/abs/1905.02244. Parameters: ---------- inplace : bool Whether to use inplace version of the module. """ def __init__(self, inplace=False): super(HSwish, self).__init__() self.inplace = inplace def forward(self, x): return x * F.relu6(x + 3.0, inplace=self.inplace) / 6.0 def get_activation_layer(activation): """ Create activation layer from string/function. Parameters: ---------- activation : function, or str, or nn.Module Activation function or name of activation function. Returns: ------- nn.Module Activation layer. """ assert (activation is not None) if isfunction(activation): return activation() elif isinstance(activation, str): if activation == "relu": return nn.ReLU(inplace=True) elif activation == "relu6": return nn.ReLU6(inplace=True) elif activation == "swish": return Swish() elif activation == "hswish": return HSwish(inplace=True) elif activation == "sigmoid": return nn.Sigmoid() elif activation == "hsigmoid": return HSigmoid() elif activation == "identity": return Identity() else: raise NotImplementedError() else: assert (isinstance(activation, nn.Module)) return activation class SelectableDense(nn.Module): """ Selectable dense layer. Parameters: ---------- in_features : int Number of input features. out_features : int Number of output features. bias : bool, default False Whether the layer uses a bias vector. num_options : int, default 1 Number of selectable options. """ def __init__(self, in_features, out_features, bias=False, num_options=1): super(SelectableDense, self).__init__() self.in_features = in_features self.out_features = out_features self.use_bias = bias self.num_options = num_options self.weight = Parameter(torch.Tensor(num_options, out_features, in_features)) if bias: self.bias = Parameter(torch.Tensor(num_options, out_features)) else: self.register_parameter("bias", None) def forward(self, x, indices): weight = torch.index_select(self.weight, dim=0, index=indices) x = x.unsqueeze(-1) x = weight.bmm(x) x = x.squeeze(dim=-1) if self.use_bias: bias = torch.index_select(self.bias, dim=0, index=indices) x += bias return x def extra_repr(self): return "in_features={}, out_features={}, bias={}, num_options={}".format( self.in_features, self.out_features, self.use_bias, self.num_options) class DenseBlock(nn.Module): """ Standard dense block with Batch normalization and activation. Parameters: ---------- in_features : int Number of input features. out_features : int Number of output features. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, in_features, out_features, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(DenseBlock, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.fc = nn.Linear( in_features=in_features, out_features=out_features, bias=bias) if self.use_bn: self.bn = nn.BatchNorm1d( num_features=out_features, eps=bn_eps) if self.activate: self.activ = get_activation_layer(activation) def forward(self, x): x = self.fc(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x class ConvBlock1d(nn.Module): """ 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(ConvBlock1d, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.conv = nn.Conv1d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) if self.use_bn: self.bn = nn.BatchNorm1d( num_features=out_channels, eps=bn_eps) if self.activate: self.activ = get_activation_layer(activation) def forward(self, 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, stride=1, groups=1, bias=False): """ Convolution 1x1 layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, groups=groups, bias=bias) def conv3x3(in_channels, out_channels, stride=1, padding=1, dilation=1, groups=1, bias=False): """ Convolution 3x3 layer. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : 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. bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) def depthwise_conv3x3(channels, stride=1, padding=1, dilation=1, 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. bias : bool, default False Whether the layer uses a bias vector. """ return nn.Conv2d( in_channels=channels, out_channels=channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, groups=channels, bias=bias) class ConvBlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(ConvBlock, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.use_pad = (isinstance(padding, (list, tuple)) and (len(padding) == 4)) if self.use_pad: self.pad = nn.ZeroPad2d(padding=padding) padding = 0 self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias) if self.use_bn: self.bn = nn.BatchNorm2d( num_features=out_channels, eps=bn_eps) if self.activate: self.activ = get_activation_layer(activation) def forward(self, x): if self.use_pad: x = self.pad(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, stride=1, padding=0, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 1x1 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 int, default 0 Padding value for convolution layer. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=stride, padding=padding, groups=groups, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def conv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 3x3 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def conv5x5_block(in_channels, out_channels, stride=1, padding=2, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 5x5 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=5, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def conv7x7_block(in_channels, out_channels, stride=1, padding=3, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 7x7 version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. padding : int, or tuple/list of 2 int, or tuple/list of 4 int, default 1 Strides of the convolution. padding : int or tuple/list of 2 int, default 3 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=7, stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def dwconv_block(in_channels, out_channels, kernel_size, stride=1, padding=1, dilation=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 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. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return ConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=out_channels, bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) def dwconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 3x3 depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, bn_eps=bn_eps, activation=activation) def dwconv5x5_block(in_channels, out_channels, stride=1, padding=2, dilation=1, bias=False, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): """ 5x5 depthwise version of the standard convolution block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 int, default 2 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return dwconv_block( in_channels=in_channels, out_channels=out_channels, kernel_size=5, stride=stride, padding=padding, dilation=dilation, bias=bias, bn_eps=bn_eps, activation=activation) class DwsConvBlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 int Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. 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_eps : float, default 1e-5 Small float added to variance in Batch norm. dw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the depthwise convolution block. pw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the pointwise convolution block. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, bias=False, dw_use_bn=True, pw_use_bn=True, bn_eps=1e-5, dw_activation=(lambda: nn.ReLU(inplace=True)), pw_activation=(lambda: nn.ReLU(inplace=True))): super(DwsConvBlock, self).__init__() self.dw_conv = dwconv_block( in_channels=in_channels, out_channels=in_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=dw_use_bn, bn_eps=bn_eps, activation=dw_activation) self.pw_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=bias, use_bn=pw_use_bn, bn_eps=bn_eps, activation=pw_activation) def forward(self, x): x = self.dw_conv(x) x = self.pw_conv(x) return x def dwsconv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, bn_eps=1e-5, dw_activation=(lambda: nn.ReLU(inplace=True)), pw_activation=(lambda: nn.ReLU(inplace=True)), **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. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 int, default 1 Padding value for convolution layer. dilation : int or tuple/list of 2 int, default 1 Dilation value for convolution layer. bias : bool, default False Whether the layer uses a bias vector. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. dw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the depthwise convolution block. pw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the pointwise convolution block. """ return DwsConvBlock( in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=stride, padding=padding, dilation=dilation, bias=bias, bn_eps=bn_eps, dw_activation=dw_activation, pw_activation=pw_activation, **kwargs) class PreConvBlock(nn.Module): """ 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. 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, stride, padding, dilation=1, bias=False, use_bn=True, return_preact=False, activate=True): super(PreConvBlock, self).__init__() self.return_preact = return_preact self.activate = activate self.use_bn = use_bn if self.use_bn: self.bn = nn.BatchNorm2d(num_features=in_channels) if self.activate: self.activ = nn.ReLU(inplace=True) self.conv = nn.Conv2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, bias=bias) def forward(self, 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, stride=1, bias=False, use_bn=True, 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. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. 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, stride=stride, padding=0, bias=bias, use_bn=use_bn, return_preact=return_preact, activate=activate) def pre_conv3x3_block(in_channels, out_channels, stride=1, padding=1, dilation=1, bias=False, use_bn=True, 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. stride : 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. 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, stride=stride, padding=padding, dilation=dilation, bias=bias, use_bn=use_bn, return_preact=return_preact, activate=activate) class AsymConvBlock(nn.Module): """ Asymmetric separable convolution block. Parameters: ---------- channels : int Number of input/output channels. kernel_size : int Convolution window size. padding : int Padding value for convolution layer. dilation : int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. lw_use_bn : bool, default True Whether to use BatchNorm layer (leftwise convolution block). rw_use_bn : bool, default True Whether to use BatchNorm layer (rightwise convolution block). bn_eps : float, default 1e-5 Small float added to variance in Batch norm. lw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the leftwise convolution block. rw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the rightwise convolution block. """ def __init__(self, channels, kernel_size, padding, dilation=1, groups=1, bias=False, lw_use_bn=True, rw_use_bn=True, bn_eps=1e-5, lw_activation=(lambda: nn.ReLU(inplace=True)), rw_activation=(lambda: nn.ReLU(inplace=True))): super(AsymConvBlock, self).__init__() self.lw_conv = ConvBlock( in_channels=channels, out_channels=channels, kernel_size=(kernel_size, 1), stride=1, padding=(padding, 0), dilation=(dilation, 1), groups=groups, bias=bias, use_bn=lw_use_bn, bn_eps=bn_eps, activation=lw_activation) self.rw_conv = ConvBlock( in_channels=channels, out_channels=channels, kernel_size=(1, kernel_size), stride=1, padding=(0, padding), dilation=(1, dilation), groups=groups, bias=bias, use_bn=rw_use_bn, bn_eps=bn_eps, activation=rw_activation) def forward(self, x): x = self.lw_conv(x) x = self.rw_conv(x) return x def asym_conv3x3_block(padding=1, **kwargs): """ 3x3 asymmetric separable convolution block. Parameters: ---------- channels : int Number of input/output channels. padding : int, default 1 Padding value for convolution layer. dilation : int, default 1 Dilation value for convolution layer. groups : int, default 1 Number of groups. bias : bool, default False Whether the layer uses a bias vector. lw_use_bn : bool, default True Whether to use BatchNorm layer (leftwise convolution block). rw_use_bn : bool, default True Whether to use BatchNorm layer (rightwise convolution block). bn_eps : float, default 1e-5 Small float added to variance in Batch norm. lw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the leftwise convolution block. rw_activation : function or str or None, default nn.ReLU(inplace=True) Activation function after the rightwise convolution block. """ return AsymConvBlock( kernel_size=3, padding=padding, **kwargs) class DeconvBlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the deconvolution. padding : int or tuple/list of 2 int Padding value for deconvolution layer. ext_padding : tuple/list of 4 int, default None Extra 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, in_channels, out_channels, kernel_size, stride, padding, ext_padding=None, out_padding=0, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(DeconvBlock, self).__init__() self.activate = (activation is not None) self.use_bn = use_bn self.use_pad = (ext_padding is not None) if self.use_pad: self.pad = nn.ZeroPad2d(padding=ext_padding) self.conv = nn.ConvTranspose2d( in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size, stride=stride, padding=padding, output_padding=out_padding, dilation=dilation, groups=groups, bias=bias) if self.use_bn: self.bn = nn.BatchNorm2d( num_features=out_channels, eps=bn_eps) if self.activate: self.activ = get_activation_layer(activation) def forward(self, x): if self.use_pad: x = self.pad(x) x = self.conv(x) if self.use_bn: x = self.bn(x) if self.activate: x = self.activ(x) return x def deconv3x3_block(padding=1, out_padding=1, **kwargs): """ 3x3 version of the deconvolution block with batch normalization and activation. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the deconvolution. padding : int or tuple/list of 2 int, default 1 Padding value for deconvolution layer. ext_padding : tuple/list of 4 int, default None Extra padding value for deconvolution layer. out_padding : int or tuple/list of 2 int, default 1 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ return DeconvBlock( kernel_size=3, padding=padding, out_padding=out_padding, **kwargs) class NormActivation(nn.Module): """ 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_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) Activation function or name of activation function. """ def __init__(self, in_channels, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True))): super(NormActivation, self).__init__() self.bn = nn.BatchNorm2d( num_features=in_channels, eps=bn_eps) self.activ = get_activation_layer(activation) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class InterpolationBlock(nn.Module): """ Interpolation upsampling 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. mode : str, default 'bilinear' Algorithm used for upsampling. align_corners : bool, default True Whether to align the corner pixels of the input and output tensors. up : bool, default True Whether to upsample or downsample. """ def __init__(self, scale_factor, out_size=None, mode="bilinear", align_corners=True, up=True): super(InterpolationBlock, self).__init__() self.scale_factor = scale_factor self.out_size = out_size self.mode = mode self.align_corners = align_corners self.up = up def forward(self, x, size=None): if (self.mode == "bilinear") or (size is not None): out_size = self.calc_out_size(x) if size is None else size return F.interpolate( input=x, size=out_size, mode=self.mode, align_corners=self.align_corners) else: return F.interpolate( input=x, scale_factor=self.scale_factor, mode=self.mode, align_corners=self.align_corners) 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}, mode={mode}, align_corners={align_corners}, up={up})' # noqa return s.format( name=self.__class__.__name__, scale_factor=self.scale_factor, out_size=self.out_size, mode=self.mode, align_corners=self.align_corners, up=self.up) def calc_flops(self, x): assert (x.shape[0] == 1) if self.mode == "bilinear": num_flops = 9 * x.numel() else: num_flops = 4 * x.numel() 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 : Tensor Input tensor. groups : int Number of groups. Returns: ------- Tensor Resulted tensor. """ batch, channels, height, width = x.size() # assert (channels % groups == 0) channels_per_group = channels // groups x = x.view(batch, groups, channels_per_group, height, width) x = torch.transpose(x, 1, 2).contiguous() x = x.view(batch, channels, height, width) return x class ChannelShuffle(nn.Module): """ 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): super(ChannelShuffle, self).__init__() # assert (channels % groups == 0) if channels % groups != 0: raise ValueError("channels must be divisible by groups") self.groups = groups def forward(self, 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, groups): """ 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 : Tensor Input tensor. groups : int Number of groups. Returns: ------- Tensor Resulted tensor. """ batch, channels, height, width = x.size() # assert (channels % groups == 0) channels_per_group = channels // groups x = x.view(batch, channels_per_group, groups, height, width) x = torch.transpose(x, 1, 2).contiguous() x = x.view(batch, channels, height, width) return x class ChannelShuffle2(nn.Module): """ 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): super(ChannelShuffle2, self).__init__() # assert (channels % groups == 0) if channels % groups != 0: raise ValueError("channels must be divisible by groups") self.groups = groups def forward(self, x): return channel_shuffle2(x, self.groups) class SEBlock(nn.Module): """ 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 nn.Module, default 'relu' Activation function after the first convolution. out_activation : function, or str, or nn.Module, 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.ReLU(inplace=True)), out_activation=(lambda: nn.Sigmoid())): super(SEBlock, self).__init__() self.use_conv = use_conv if mid_channels is None: mid_channels = channels // reduction if not round_mid else round_channels(float(channels) / reduction) self.pool = nn.AdaptiveAvgPool2d(output_size=1) if use_conv: self.conv1 = conv1x1( in_channels=channels, out_channels=mid_channels, bias=True) else: self.fc1 = nn.Linear( in_features=channels, out_features=mid_channels) self.activ = get_activation_layer(mid_activation) if use_conv: self.conv2 = conv1x1( in_channels=mid_channels, out_channels=channels, bias=True) else: self.fc2 = nn.Linear( in_features=mid_channels, out_features=channels) self.sigmoid = get_activation_layer(out_activation) def forward(self, x): w = self.pool(x) if not self.use_conv: w = w.view(x.size(0), -1) 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.unsqueeze(2).unsqueeze(3) x = x * w return x class SABlock(nn.Module): """ 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_eps : float, default 1e-5 Small float added to variance in Batch norm. """ def __init__(self, out_channels, groups, radix, reduction=4, min_channels=32, use_conv=True, bn_eps=1e-5): super(SABlock, self).__init__() self.groups = groups self.radix = radix self.use_conv = use_conv in_channels = out_channels * radix mid_channels = max(in_channels // reduction, min_channels) self.pool = nn.AdaptiveAvgPool2d(output_size=1) if use_conv: self.conv1 = conv1x1( in_channels=out_channels, out_channels=mid_channels, bias=True) else: self.fc1 = nn.Linear( in_features=out_channels, out_features=mid_channels) self.bn = nn.BatchNorm2d( num_features=mid_channels, eps=bn_eps) self.activ = nn.ReLU(inplace=True) if use_conv: self.conv2 = conv1x1( in_channels=mid_channels, out_channels=in_channels, bias=True) else: self.fc2 = nn.Linear( in_features=mid_channels, out_features=in_channels) self.softmax = nn.Softmax(dim=1) def forward(self, x): batch, channels, height, width = x.size() x = x.view(batch, self.radix, channels // self.radix, height, width) w = x.sum(dim=1) w = self.pool(w) if not self.use_conv: w = w.view(x.size(0), -1) 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.view(batch, self.groups, self.radix, -1) w = torch.transpose(w, 1, 2).contiguous() w = self.softmax(w) w = w.view(batch, self.radix, -1, 1, 1) x = x * w x = x.sum(dim=1) return x class SAConvBlock(nn.Module): """ 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. stride : int or tuple/list of 2 int Strides of the convolution. padding : int, or tuple/list of 2 int, or tuple/list of 4 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. bias : bool, default False Whether the layer uses a bias vector. use_bn : bool, default True Whether to use BatchNorm layer. bn_eps : float, default 1e-5 Small float added to variance in Batch norm. activation : function or str or None, default nn.ReLU(inplace=True) 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, stride, padding, dilation=1, groups=1, bias=False, use_bn=True, bn_eps=1e-5, activation=(lambda: nn.ReLU(inplace=True)), radix=2, reduction=4, min_channels=32, use_conv=True): super(SAConvBlock, self).__init__() self.conv = ConvBlock( in_channels=in_channels, out_channels=(out_channels * radix), kernel_size=kernel_size, stride=stride, padding=padding, dilation=dilation, groups=(groups * radix), bias=bias, use_bn=use_bn, bn_eps=bn_eps, activation=activation) self.att = SABlock( out_channels=out_channels, groups=groups, radix=radix, reduction=reduction, min_channels=min_channels, use_conv=use_conv, bn_eps=bn_eps) def forward(self, x): x = self.conv(x) x = self.att(x) return x def saconv3x3_block(in_channels, out_channels, stride=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. stride : 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, stride=stride, padding=padding, **kwargs) class DucBlock(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels, scale_factor): super(DucBlock, self).__init__() mid_channels = (scale_factor * scale_factor) * out_channels self.conv = conv3x3_block( in_channels=in_channels, out_channels=mid_channels) self.pix_shuffle = nn.PixelShuffle(upscale_factor=scale_factor) def forward(self, x): x = self.conv(x) x = self.pix_shuffle(x) return x class IBN(nn.Module): """ 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. 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, first_fraction=0.5, inst_first=True): super(IBN, self).__init__() 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.InstanceNorm2d( num_features=h1_channels, affine=True) self.batch_norm = nn.BatchNorm2d(num_features=h2_channels) else: self.batch_norm = nn.BatchNorm2d(num_features=h1_channels) self.inst_norm = nn.InstanceNorm2d( num_features=h2_channels, affine=True) def forward(self, x): x1, x2 = torch.split(x, split_size_or_sections=self.split_sections, dim=1) if self.inst_first: x1 = self.inst_norm(x1.contiguous()) x2 = self.batch_norm(x2.contiguous()) else: x1 = self.batch_norm(x1.contiguous()) x2 = self.inst_norm(x2.contiguous()) x = torch.cat((x1, x2), dim=1) return x class DualPathSequential(nn.Sequential): """ A sequential container for modules with dual inputs/outputs. Modules 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 modules with single input/output. last_ordinals : int, default 0 Number of the final modules with single input/output. dual_path_scheme : function Scheme of dual path response for a module. dual_path_scheme_ordinal : function Scheme of dual path response for an ordinal module. """ def __init__(self, return_two=True, first_ordinals=0, last_ordinals=0, dual_path_scheme=(lambda module, x1, x2: module(x1, x2)), dual_path_scheme_ordinal=(lambda module, x1, x2: (module(x1), x2))): super(DualPathSequential, self).__init__() 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 forward(self, x1, x2=None): length = len(self._modules.values()) for i, module in enumerate(self._modules.values()): if (i < self.first_ordinals) or (i >= length - self.last_ordinals): x1, x2 = self.dual_path_scheme_ordinal(module, x1, x2) else: x1, x2 = self.dual_path_scheme(module, x1, x2) if self.return_two: return x1, x2 else: return x1 class Concurrent(nn.Sequential): """ A container for concatenation of modules 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. """ def __init__(self, axis=1, stack=False, merge_type=None): super(Concurrent, self).__init__() assert (merge_type is None) or (merge_type in ["cat", "stack", "sum"]) self.axis = axis if merge_type is not None: self.merge_type = merge_type else: self.merge_type = "stack" if stack else "cat" def forward(self, x): out = [] for module in self._modules.values(): out.append(module(x)) if self.merge_type == "stack": out = torch.stack(tuple(out), dim=self.axis) elif self.merge_type == "cat": out = torch.cat(tuple(out), dim=self.axis) elif self.merge_type == "sum": out = torch.stack(tuple(out), dim=self.axis).sum(self.axis) else: raise NotImplementedError() return out class SequentialConcurrent(nn.Sequential): """ A sequential container with concatenated outputs. Modules 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): super(SequentialConcurrent, self).__init__() self.axis = axis self.stack = stack self.cat_input = cat_input def forward(self, x): out = [x] if self.cat_input else [] for module in self._modules.values(): x = module(x) out.append(x) if self.stack: out = torch.stack(tuple(out), dim=self.axis) else: out = torch.cat(tuple(out), dim=self.axis) return out class ParametricSequential(nn.Sequential): """ A sequential container for modules with parameters. Modules will be executed in the order they are added. """ def __init__(self, *args): super(ParametricSequential, self).__init__(*args) def forward(self, x, **kwargs): for module in self._modules.values(): x = module(x, **kwargs) return x class ParametricConcurrent(nn.Sequential): """ A container for concatenation of modules with parameters. Parameters: ---------- axis : int, default 1 The axis on which to concatenate the outputs. """ def __init__(self, axis=1): super(ParametricConcurrent, self).__init__() self.axis = axis def forward(self, x, **kwargs): out = [] for module in self._modules.values(): out.append(module(x, **kwargs)) out = torch.cat(tuple(out), dim=self.axis) return out class Hourglass(nn.Module): """ A hourglass module. Parameters: ---------- down_seq : nn.Sequential Down modules as sequential. up_seq : nn.Sequential Up modules as sequential. skip_seq : nn.Sequential 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): super(Hourglass, self).__init__() 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) self.down_seq = down_seq self.up_seq = up_seq self.skip_seq = skip_seq def _merge(self, x, y): if y is not None: if self.merge_type == "cat": x = torch.cat((x, y), dim=1) elif self.merge_type == "add": x = x + y return x def forward(self, x, **kwargs): y = None down_outs = [x] for down_module in self.down_seq._modules.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(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(nn.Module): """ 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"): super(SesquialteralHourglass, self).__init__() 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) 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, x, y): if y is not None: if self.merge_type == "cat": x = torch.cat((x, y), dim=1) elif self.merge_type == "add": x = x + y return x def forward(self, x, **kwargs): 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(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(x, y) return x class MultiOutputSequential(nn.Sequential): """ A sequential container with multiple outputs. Modules 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): super(MultiOutputSequential, self).__init__() self.multi_output = multi_output self.dual_output = dual_output self.return_last = return_last def forward(self, x): outs = [] for module in self._modules.values(): x = module(x) if hasattr(module, "do_output") and module.do_output: outs.append(x) elif hasattr(module, "do_output2") and module.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.Sequential): """ A sequential container with multiple inputs and single/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"): super(ParallelConcurent, self).__init__() assert (merge_type is None) or (merge_type in ["list", "cat", "stack", "sum"]) self.axis = axis self.merge_type = merge_type def forward(self, x): out = [] for module, xi in zip(self._modules.values(), x): out.append(module(xi)) if self.merge_type == "list": pass elif self.merge_type == "stack": out = torch.stack(tuple(out), dim=self.axis) elif self.merge_type == "cat": out = torch.cat(tuple(out), dim=self.axis) elif self.merge_type == "sum": out = torch.stack(tuple(out), dim=self.axis).sum(self.axis) else: raise NotImplementedError() return out class DualPathParallelConcurent(nn.Sequential): """ A sequential container with multiple dual-path inputs and single/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"): super(DualPathParallelConcurent, self).__init__() assert (merge_type is None) or (merge_type in ["list", "cat", "stack", "sum"]) self.axis = axis self.merge_type = merge_type def forward(self, x1, x2): x1_out = [] x2_out = [] for module, x1i, x2i in zip(self._modules.values(), x1, x2): y1i, y2i = module(x1i, x2i) x1_out.append(y1i) x2_out.append(y2i) if self.merge_type == "list": pass elif self.merge_type == "stack": x1_out = torch.stack(tuple(x1_out), dim=self.axis) x2_out = torch.stack(tuple(x2_out), dim=self.axis) elif self.merge_type == "cat": x1_out = torch.cat(tuple(x1_out), dim=self.axis) x2_out = torch.cat(tuple(x2_out), dim=self.axis) elif self.merge_type == "sum": x1_out = torch.stack(tuple(x1_out), dim=self.axis).sum(self.axis) x2_out = torch.stack(tuple(x2_out), dim=self.axis).sum(self.axis) else: raise NotImplementedError() return x1_out, x2_out class Flatten(nn.Module): """ Simple flatten module. """ def forward(self, x): return x.view(x.size(0), -1) class HeatmapMaxDetBlock(nn.Module): """ Heatmap maximum detector block (for human pose estimation task). """ def __init__(self): super(HeatmapMaxDetBlock, self).__init__() def forward(self, x): heatmap = x vector_dim = 2 batch = heatmap.shape[0] channels = heatmap.shape[1] in_size = x.shape[2:] heatmap_vector = heatmap.view(batch, channels, -1) scores, indices = heatmap_vector.max(dim=vector_dim, keepdims=True) scores_mask = (scores > 0.0).float() pts_x = (indices % in_size[1]) * scores_mask pts_y = (indices // in_size[1]) * scores_mask pts = torch.cat((pts_x, pts_y, scores), dim=vector_dim) for b in range(batch): for k in range(channels): hm = heatmap[b, k, :, :] px = int(pts[b, k, 0]) py = int(pts[b, k, 1]) 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 @staticmethod def calc_flops(x): assert (x.shape[0] == 1) num_flops = x.numel() + 26 * x.shape[1] num_macs = 0 return num_flops, num_macs
74,363
30.902188
130
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/lwopenpose_cmupan.py
""" Lightweight OpenPose 2D/3D for CMU Panoptic, implemented in PyTorch. 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 import torch from torch import nn from .common import conv1x1, conv1x1_block, conv3x3_block, dwsconv3x3_block class LwopResBottleneck(nn.Module): """ Bottleneck block for residual path in the residual unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. bias : bool, default True Whether the layer uses a bias vector. 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, stride, bias=True, bottleneck_factor=2, squeeze_out=False): super(LwopResBottleneck, self).__init__() mid_channels = out_channels // bottleneck_factor if squeeze_out else in_channels // bottleneck_factor self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bias=bias) self.conv2 = conv3x3_block( in_channels=mid_channels, out_channels=mid_channels, stride=stride, bias=bias) self.conv3 = conv1x1_block( in_channels=mid_channels, out_channels=out_channels, bias=bias, activation=None) def forward(self, x): x = self.conv1(x) x = self.conv2(x) x = self.conv3(x) return x class LwopResUnit(nn.Module): """ ResNet-like residual unit with residual connection. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int, default 1 Strides of the convolution. bias : bool, default True Whether the layer uses a bias vector. 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, stride=1, bias=True, bottleneck_factor=2, squeeze_out=False, activate=False): super(LwopResUnit, self).__init__() self.activate = activate self.resize_identity = (in_channels != out_channels) or (stride != 1) self.body = LwopResBottleneck( in_channels=in_channels, out_channels=out_channels, stride=stride, bias=bias, 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, stride=stride, bias=bias, activation=None) if self.activate: self.activ = nn.ReLU(inplace=True) def forward(self, 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(nn.Module): """ Lightweight OpenPose 2D/3D specific encoder final block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(LwopEncoderFinalBlock, self).__init__() self.pre_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=True, use_bn=False) self.body = nn.Sequential() for i in range(3): self.body.add_module("block{}".format(i + 1), dwsconv3x3_block( in_channels=out_channels, out_channels=out_channels, dw_use_bn=False, pw_use_bn=False, dw_activation=(lambda: nn.ELU(inplace=True)), pw_activation=(lambda: nn.ELU(inplace=True)))) self.post_conv = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bias=True, use_bn=False) def forward(self, x): x = self.pre_conv(x) x = x + self.body(x) x = self.post_conv(x) return x class LwopRefinementBlock(nn.Module): """ 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. """ def __init__(self, in_channels, out_channels): super(LwopRefinementBlock, self).__init__() self.pre_conv = conv1x1_block( in_channels=in_channels, out_channels=out_channels, bias=True, use_bn=False) self.body = nn.Sequential() self.body.add_module("block1", conv3x3_block( in_channels=out_channels, out_channels=out_channels, bias=True)) self.body.add_module("block2", conv3x3_block( in_channels=out_channels, out_channels=out_channels, padding=2, dilation=2, bias=True)) def forward(self, x): x = self.pre_conv(x) x = x + self.body(x) return x class LwopDecoderBend(nn.Module): """ 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. """ def __init__(self, in_channels, mid_channels, out_channels): super(LwopDecoderBend, self).__init__() self.conv1 = conv1x1_block( in_channels=in_channels, out_channels=mid_channels, bias=True, use_bn=False) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class LwopDecoderInitBlock(nn.Module): """ Lightweight OpenPose 2D/3D specific decoder init block. Parameters: ---------- in_channels : int Number of input channels. keypoints : int Number of keypoints. """ def __init__(self, in_channels, keypoints): super(LwopDecoderInitBlock, self).__init__() num_heatmap = keypoints num_paf = 2 * keypoints bend_mid_channels = 512 self.body = nn.Sequential() for i in range(3): self.body.add_module("block{}".format(i + 1), conv3x3_block( in_channels=in_channels, out_channels=in_channels, bias=True, use_bn=False)) self.heatmap_bend = LwopDecoderBend( in_channels=in_channels, mid_channels=bend_mid_channels, out_channels=num_heatmap) self.paf_bend = LwopDecoderBend( in_channels=in_channels, mid_channels=bend_mid_channels, out_channels=num_paf) def forward(self, x): y = self.body(x) heatmap = self.heatmap_bend(y) paf = self.paf_bend(y) y = torch.cat((x, heatmap, paf), dim=1) return y class LwopDecoderUnit(nn.Module): """ Lightweight OpenPose 2D/3D specific decoder init. Parameters: ---------- in_channels : int Number of input channels. keypoints : int Number of keypoints. """ def __init__(self, in_channels, keypoints): super(LwopDecoderUnit, self).__init__() num_heatmap = keypoints num_paf = 2 * keypoints self.features_channels = in_channels - num_heatmap - num_paf self.body = nn.Sequential() for i in range(5): self.body.add_module("block{}".format(i + 1), LwopRefinementBlock( in_channels=in_channels, out_channels=self.features_channels)) in_channels = self.features_channels self.heatmap_bend = LwopDecoderBend( in_channels=self.features_channels, mid_channels=self.features_channels, out_channels=num_heatmap) self.paf_bend = LwopDecoderBend( in_channels=self.features_channels, mid_channels=self.features_channels, out_channels=num_paf) def forward(self, x): features = x[:, :self.features_channels] y = self.body(x) heatmap = self.heatmap_bend(y) paf = self.paf_bend(y) y = torch.cat((features, heatmap, paf), dim=1) return y class LwopDecoderFeaturesBend(nn.Module): """ 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. """ def __init__(self, in_channels, mid_channels, out_channels): super(LwopDecoderFeaturesBend, self).__init__() self.body = nn.Sequential() for i in range(2): self.body.add_module("block{}".format(i + 1), LwopRefinementBlock( in_channels=in_channels, out_channels=mid_channels)) in_channels = mid_channels self.features_bend = LwopDecoderBend( in_channels=mid_channels, mid_channels=mid_channels, out_channels=out_channels) def forward(self, x): x = self.body(x) x = self.features_bend(x) return x class LwopDecoderFinalBlock(nn.Module): """ 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. """ def __init__(self, in_channels, keypoints, bottleneck_factor, calc_3d_features): super(LwopDecoderFinalBlock, self).__init__() 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: self.body = nn.Sequential() for i in range(5): self.body.add_module("block{}".format(i + 1), LwopResUnit( in_channels=in_channels, out_channels=features_in_channels, bottleneck_factor=bottleneck_factor)) in_channels = features_in_channels self.features_bend = LwopDecoderFeaturesBend( in_channels=features_in_channels, mid_channels=features_in_channels, out_channels=features_out_channels) def forward(self, x): heatmap_paf_2d = x[:, -self.num_heatmap_paf:] if not self.calc_3d_features: return heatmap_paf_2d x = self.body(x) x = self.features_bend(x) y = torch.cat((heatmap_paf_2d, x), dim=1) return y class LwOpenPose(nn.Module): """ 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. 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, in_channels=3, in_size=(368, 368), keypoints=19): super(LwOpenPose, self).__init__() 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 self.encoder = nn.Sequential() backbone = nn.Sequential() backbone.add_module("init_block", conv3x3_block( in_channels=in_channels, out_channels=encoder_init_block_channels, stride=2)) in_channels = encoder_init_block_channels for i, channels_per_stage in enumerate(encoder_channels): stage = nn.Sequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 padding = encoder_paddings[i][j] stage.add_module("unit{}".format(j + 1), dwsconv3x3_block( in_channels=in_channels, out_channels=out_channels, stride=stride, padding=padding, dilation=padding)) in_channels = out_channels backbone.add_module("stage{}".format(i + 1), stage) self.encoder.add_module("backbone", backbone) self.encoder.add_module("final_block", LwopEncoderFinalBlock( in_channels=in_channels, out_channels=encoder_final_block_channels)) in_channels = encoder_final_block_channels self.decoder = nn.Sequential() self.decoder.add_module("init_block", LwopDecoderInitBlock( in_channels=in_channels, keypoints=keypoints)) in_channels = encoder_final_block_channels + num_heatmap_paf for i in range(refinement_units): self.decoder.add_module("unit{}".format(i + 1), LwopDecoderUnit( in_channels=in_channels, keypoints=keypoints)) self.decoder.add_module("final_block", LwopDecoderFinalBlock( in_channels=in_channels, keypoints=keypoints, bottleneck_factor=2, calc_3d_features=calc_3d_features)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, 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, root=os.path.join("~", ".torch", "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. root : str, default '~/.torch/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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) 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. root : str, default '~/.torch/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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_lwopenpose(calc_3d_features=True, keypoints=keypoints, model_name="lwopenpose3d_mobilenet_cmupan_coco", **kwargs) def _calc_width(net): import numpy as np 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 _test(): 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) # net.train() net.eval() weight_count = _calc_width(net) 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 = 1 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() if model_dim == "2d": assert (tuple(y.size()) == (batch, 3 * keypoints, in_size[0] // 8, in_size[0] // 8)) else: assert (tuple(y.size()) == (batch, 6 * keypoints, in_size[0] // 8, in_size[0] // 8)) if __name__ == "__main__": _test()
21,152
31.643519
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/rir_cifar.py
""" RiR for CIFAR/SVHN, implemented in PyTorch. 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 import torch import torch.nn as nn import torch.nn.init as init from .common import conv1x1, conv3x3, conv1x1_block, conv3x3_block, DualPathSequential class PostActivation(nn.Module): """ Pure pre-activation block without convolution layer. Parameters: ---------- in_channels : int Number of input channels. """ def __init__(self, in_channels): super(PostActivation, self).__init__() self.bn = nn.BatchNorm2d(num_features=in_channels) self.activ = nn.ReLU(inplace=True) def forward(self, x): x = self.bn(x) x = self.activ(x) return x class RiRUnit(nn.Module): """ RiR unit. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. stride : int or tuple/list of 2 int Strides of the convolution. """ def __init__(self, in_channels, out_channels, stride): super(RiRUnit, self).__init__() self.resize_identity = (in_channels != out_channels) or (stride != 1) self.res_pass_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride) self.trans_pass_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride) self.res_cross_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride) self.trans_cross_conv = conv3x3( in_channels=in_channels, out_channels=out_channels, stride=stride) self.res_postactiv = PostActivation(in_channels=out_channels) self.trans_postactiv = PostActivation(in_channels=out_channels) if self.resize_identity: self.identity_conv = conv1x1( in_channels=in_channels, out_channels=out_channels, stride=stride) def forward(self, 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(nn.Module): """ RiR initial block. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. """ def __init__(self, in_channels, out_channels): super(RiRInitBlock, self).__init__() self.res_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels) self.trans_conv = conv3x3_block( in_channels=in_channels, out_channels=out_channels) def forward(self, x, _): x_res = self.res_conv(x) x_trans = self.trans_conv(x) return x_res, x_trans class RiRFinalBlock(nn.Module): """ RiR final block. """ def __init__(self): super(RiRFinalBlock, self).__init__() def forward(self, x_res, x_trans): x = torch.cat((x_res, x_trans), dim=1) return x, None class CIFARRiR(nn.Module): """ 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. 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. num_classes : int, default 10 Number of classification classes. """ def __init__(self, channels, init_block_channels, final_block_channels, in_channels=3, in_size=(32, 32), num_classes=10): super(CIFARRiR, self).__init__() self.in_size = in_size self.num_classes = num_classes self.features = DualPathSequential( return_two=False, first_ordinals=0, last_ordinals=0) self.features.add_module("init_block", RiRInitBlock( in_channels=in_channels, out_channels=init_block_channels)) in_channels = init_block_channels for i, channels_per_stage in enumerate(channels): stage = DualPathSequential() for j, out_channels in enumerate(channels_per_stage): stride = 2 if (j == 0) and (i != 0) else 1 stage.add_module("unit{}".format(j + 1), RiRUnit( in_channels=in_channels, out_channels=out_channels, stride=stride)) in_channels = out_channels self.features.add_module("stage{}".format(i + 1), stage) self.features.add_module("final_block", RiRFinalBlock()) in_channels = final_block_channels self.output = nn.Sequential() self.output.add_module("final_conv", conv1x1_block( in_channels=in_channels, out_channels=num_classes, activation=None)) self.output.add_module("final_pool", nn.AvgPool2d( kernel_size=8, stride=1)) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): init.kaiming_uniform_(module.weight) if module.bias is not None: init.constant_(module.bias, 0) def forward(self, x): x = self.features(x) x = self.output(x) x = x.view(x.size(0), -1) return x def get_rir_cifar(num_classes, model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create RiR 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. root : str, default '~/.torch/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, num_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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def rir_cifar10(num_classes=10, **kwargs): """ RiR model for CIFAR-10 from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_rir_cifar(num_classes=num_classes, model_name="rir_cifar10", **kwargs) def rir_cifar100(num_classes=100, **kwargs): """ RiR model for CIFAR-100 from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- num_classes : int, default 100 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_rir_cifar(num_classes=num_classes, model_name="rir_cifar100", **kwargs) def rir_svhn(num_classes=10, **kwargs): """ RiR model for SVHN from 'Resnet in Resnet: Generalizing Residual Architectures,' https://arxiv.org/abs/1603.08029. Parameters: ---------- num_classes : int, default 10 Number of classification classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_rir_cifar(num_classes=num_classes, model_name="rir_svhn", **kwargs) def _calc_width(net): import numpy as np 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 _test(): import torch pretrained = False models = [ (rir_cifar10, 10), (rir_cifar100, 100), (rir_svhn, 10), ] for model, num_classes in models: net = model(pretrained=pretrained) # net.train() net.eval() weight_count = _calc_width(net) 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 = torch.randn(1, 3, 32, 32) y = net(x) y.sum().backward() assert (tuple(y.size()) == (1, num_classes)) if __name__ == "__main__": _test()
10,658
29.454286
119
py
imgclsmob
imgclsmob-master/pytorch/pytorchcv/models/unet.py
""" U-Net for image segmentation, implemented in PyTorch. Original paper: 'U-Net: Convolutional Networks for Biomedical Image Segmentation,' https://arxiv.org/abs/1505.04597. """ __all__ = ['UNet', 'unet_cityscapes'] import os import torch import torch.nn as nn from .common import conv1x1, conv3x3_block, InterpolationBlock, Hourglass, Identity class UNetBlock(nn.Module): """ U-Net specific base block (double convolution). Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. """ def __init__(self, in_channels, out_channels, bias): super(UNetBlock, self).__init__() self.conv1 = conv3x3_block( in_channels=in_channels, out_channels=out_channels, bias=bias) self.conv2 = conv3x3_block( in_channels=out_channels, out_channels=out_channels, bias=bias) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class UNetDownStage(nn.Module): """ U-Net specific downscale (encoder) stage. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. """ def __init__(self, in_channels, out_channels, bias): super(UNetDownStage, self).__init__() self.pool = nn.MaxPool2d(kernel_size=2) self.conv = UNetBlock( in_channels=in_channels, out_channels=out_channels, bias=bias) def forward(self, x): x = self.pool(x) x = self.conv(x) return x class UNetUpStage(nn.Module): """ U-Net specific upscale (decoder) stage. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. """ def __init__(self, in_channels, out_channels, bias): super(UNetUpStage, self).__init__() self.conv = UNetBlock( in_channels=in_channels, out_channels=out_channels, bias=bias) self.up = InterpolationBlock( scale_factor=2, align_corners=True) def forward(self, x): x = self.conv(x) x = self.up(x) return x class UNetHead(nn.Module): """ U-Net specific head. Parameters: ---------- in_channels : int Number of input channels. out_channels : int Number of output channels. bias : bool Whether the layer uses a bias vector. """ def __init__(self, in_channels, out_channels, bias): super(UNetHead, self).__init__() mid_channels = in_channels // 2 self.conv1 = UNetBlock( in_channels=in_channels, out_channels=mid_channels, bias=bias) self.conv2 = conv1x1( in_channels=mid_channels, out_channels=out_channels, bias=True) def forward(self, x): x = self.conv1(x) x = self.conv2(x) return x class UNet(nn.Module): """ U-Net model from 'U-Net: Convolutional Networks for Biomedical Image Segmentation,' https://arxiv.org/abs/1505.04597. Parameters: ---------- channels : list of list of int Number of output channels for each stage in encoder and decoder. init_block_channels : int Number of output channels for the initial unit. 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. num_classes : int, default 19 Number of segmentation classes. """ def __init__(self, channels, init_block_channels, aux=False, fixed_size=False, in_channels=3, in_size=(1024, 2048), num_classes=19): super(UNet, self).__init__() 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.num_classes = num_classes self.fixed_size = fixed_size bias = True self.stem = UNetBlock( in_channels=in_channels, out_channels=init_block_channels, bias=bias) in_channels = init_block_channels down_seq = nn.Sequential() skip_seq = nn.Sequential() for i, out_channels in enumerate(channels[0]): down_seq.add_module("down{}".format(i + 1), UNetDownStage( in_channels=in_channels, out_channels=out_channels, bias=bias)) in_channels = out_channels skip_seq.add_module("skip{}".format(i + 1), Identity()) up_seq = nn.Sequential() for i, out_channels in enumerate(channels[1]): if i == 0: up_seq.add_module("down{}".format(i + 1), InterpolationBlock( scale_factor=2, align_corners=True)) else: up_seq.add_module("down{}".format(i + 1), UNetUpStage( in_channels=(2 * in_channels), out_channels=out_channels, bias=bias)) in_channels = out_channels up_seq = up_seq[::-1] self.hg = Hourglass( down_seq=down_seq, up_seq=up_seq, skip_seq=skip_seq, merge_type="cat") self.head = UNetHead( in_channels=(2 * in_channels), out_channels=num_classes, bias=True) self._init_params() def _init_params(self): for name, module in self.named_modules(): if isinstance(module, nn.Conv2d): nn.init.kaiming_uniform_(module.weight) if module.bias is not None: nn.init.constant_(module.bias, 0) def forward(self, x): x = self.stem(x) x = self.hg(x) x = self.head(x) return x def get_unet(model_name=None, pretrained=False, root=os.path.join("~", ".torch", "models"), **kwargs): """ Create U-Net 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. root : str, default '~/.torch/models' Location for keeping the model parameters. """ channels = [[128, 256, 512, 512], [512, 256, 128, 64]] init_block_channels = 64 net = UNet( 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 download_model download_model( net=net, model_name=model_name, local_model_store_dir_path=root) return net def unet_cityscapes(num_classes=19, **kwargs): """ U-Net model for Cityscapes from 'U-Net: Convolutional Networks for Biomedical Image Segmentation,' https://arxiv.org/abs/1505.04597. Parameters: ---------- num_classes : int, default 19 Number of segmentation classes. pretrained : bool, default False Whether to load the pretrained weights for model. root : str, default '~/.torch/models' Location for keeping the model parameters. """ return get_unet(num_classes=num_classes, model_name="unet_cityscapes", **kwargs) def _calc_width(net): import numpy as np 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 _test(): pretrained = False fixed_size = True in_size = (1024, 2048) classes = 19 models = [ unet_cityscapes, ] for model in models: net = model(pretrained=pretrained, in_size=in_size, fixed_size=fixed_size) # net.train() net.eval() weight_count = _calc_width(net) print("m={}, {}".format(model.__name__, weight_count)) assert (model != unet_cityscapes or weight_count == 13396499) batch = 4 x = torch.randn(batch, 3, in_size[0], in_size[1]) y = net(x) # y.sum().backward() assert (tuple(y.size()) == (batch, classes, in_size[0], in_size[1])) if __name__ == "__main__": _test()
9,378
27.335347
115
py