repo
stringlengths
1
99
file
stringlengths
13
215
code
stringlengths
12
59.2M
file_length
int64
12
59.2M
avg_line_length
float64
3.82
1.48M
max_line_length
int64
12
2.51M
extension_type
stringclasses
1 value
fl-analysis
fl-analysis-master/src/subspace/keras_ext/engine_topology.py
# Copyright (c) 2018 Uber Technologies, Inc. # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. from tensorflow.keras.layers import Input from keras.engine.topology import Network as Container from .util import full_static_shape class LazyContainer(Container): '''Like Container. But lazy.''' def __init__(self, container_function, use_method_disposable=True): self._container_function = container_function self._lazy_has_run = False self.use_method_disposable = use_method_disposable # Delay rest of construction until first call def __call__(self, x, mask=None): if not self._lazy_has_run: # Make short-lived Input Layers for each x this was called with # TODO: handle tuple or list x x_shape = full_static_shape(x) # Uses var._keras_shape or var.get_shape() if self.use_method_disposable: inp_layer = Input(batch_shape=x_shape, dtype=x.dtype, name='tmp_input_from__%s' % x.name.replace('/','_').replace(':','_')) else: print('Warning: using non-disposable approach. May not work yet.') inp_layer = Input(tensor=x, batch_shape=x_shape, dtype=x.dtype, name='real_input_from__%s' % x.name.replace('/','_').replace(':','_')) # Call function of inputs to get output tensors outputs = self._container_function(inp_layer) # Initialize entire Container object here (finally) super(LazyContainer, self).__init__(inp_layer, outputs) self._lazy_has_run = True if not self.use_method_disposable: return outputs # Non-disposable mode: actually call the Container only the *second* and later times # Disposable mode: call the Container now ret = super(LazyContainer, self).__call__(x, mask=mask) return ret
3,056
45.318182
119
py
fl-analysis
fl-analysis-master/src/subspace/keras_ext/util.py
# Copyright (c) 2018 Uber Technologies, Inc. # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. import os import numpy as np import tensorflow as tf import tensorflow.keras.backend as K from tensorflow.keras.layers import Flatten, Input, Lambda from src.subspace.general.tfutil import tf_assert_gpu, hist_summaries_traintest ######################## # General Keras helpers ######################## def make_image_input_preproc(im_dims, dtype='float32', flatten_in=False, shift_in=None, name=None): '''Make an input for images and (optionally preprocess). Returns both the Input layer (which should be used as Model input) and the preproc version (which should be passed to the first layer of the model). If no preprocessing is done, the Input layer and preproc will be the same. ''' assert isinstance(im_dims, tuple) and len(im_dims) == 3, 'should be tuple of 3 dims (0,1,c)' assert dtype in ('float32', 'uint8'), 'unknown dtype' input_images = Input(shape=im_dims, dtype=dtype, name=name) preproc_images = input_images if dtype == 'uint8': preproc_images = Lambda(lambda x: K.cast(x, 'float32'))(preproc_images) if shift_in is not None: print('subtracting from each input:', shift_in) preproc_images = Lambda(lambda x: x - shift_in)(preproc_images) if flatten_in: preproc_images = Flatten()(preproc_images) return input_images, preproc_images def make_classlabel_input(n_label_vals): return Input(batch_shape=(None,), dtype='int64') def setup_session_and_seeds(seed, assert_gpu=True, mem_fraction=None): '''Start TF and register session with Keras''' # Use InteractiveSession instead of Session so the default session will be set if mem_fraction is not None: gpu_options = tf.compat.v1.GPUOptions(per_process_gpu_memory_fraction=mem_fraction) sess = tf.compat.v1.InteractiveSession(config=tf.compat.v1.ConfigProto(gpu_options=gpu_options)) else: sess = tf.compat.v1.InteractiveSession() K.set_session(sess) np.random.seed(seed) tf.compat.v1.set_random_seed(seed) print('Set numpy and tensorflow random seeds to: %s' % repr(seed)) print('My PID is %d' % os.getpid()) if assert_gpu: tf_assert_gpu(sess) return sess def add_act_summaries(model, quiet=False): tensors = [] if not quiet: print('\nActivations:') for layer in model.layers: for node in layer._inbound_nodes: for tensor in node.output_tensors: tensors.append(tensor) tdict = {tt.name: tt for tt in set(tensors)} for tname in sorted(tdict.keys()): hist_summaries_traintest(tdict[tname], name=tname + '__act') if not quiet: print(' ', tname, tdict[tname]) def get_model_tensors(model, with_layers_nodes=False): tensor_set = set() tensor_list = [] layer_list = [] node_list = [] for layer in model.layers: for node in layer._inbound_nodes: for tensor in node.output_tensors: if tensor not in tensor_set: # Make a list with deteministic order, but check membership using a fast set tensor_set.add(tensor) tensor_list.append(tensor) layer_list.append(layer) node_list.append(node) if with_layers_nodes: return tensor_list, layer_list, node_list else: return tensor_list def warn_misaligned_shapes(model): printed = False tlns = get_model_tensors(model, with_layers_nodes=True) for tln in zip(tlns[0], tlns[1], tlns[2]): tensor, layer, node = tln tf_shape = tuple(tensor.get_shape().as_list()) try: keras_shape = tensor._keras_shape except AttributeError: continue if tf_shape != keras_shape: if not printed: print('\nWarning: found the following tensor shape mismatches, may indicate problems.') print(' %-40s %-22s %-22s' % ('LAYER NAME', '', '')) print(' %-40s %-22s %-22s' % ('TENSOR NAME', 'KERAS SHAPE', 'TF SHAPE')) printed = True print(' %-40s %-22s %-22s' % (layer.name, '', '')) print(' %-40s %-22s %-22s' % (tensor.name, keras_shape, tf_shape)) def full_static_shape(var): '''Returns the most fully-specified static shape possible for var (at graph build time, not run time). Uses information in var.get_shape() as well as var._keras_shape. Raises an Exception if the two shapes are incompatible with each other. ''' try: tf_shape = [val.__int__() for val in var.get_shape()] except ValueError: raise Exception('Unclear why this would ever be encountered. If it pops up, debug here.') if not hasattr(var, '_keras_shape'): return tf_shape k_shape = var._keras_shape assert len(tf_shape) == len(k_shape), 'Shape lengths different; this probably should not occur' shape = [] for tt, kk in zip(tf_shape, k_shape): if tt == kk: shape.append(tt) else: if tt is None: shape.append(kk) elif kk is None: shape.append(tt) else: raise Exception('tf shape and Keras shape are contradictory: %s vs %s' % (tf_shape, k_shape)) return shape
6,451
38.10303
109
py
fl-analysis
fl-analysis-master/src/subspace/keras_ext/regularizers.py
# Copyright (c) 2018 Uber Technologies, Inc. # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. '''Custom Keras regularizers.''' import tensorflow.keras import tensorflow.keras.backend as K class WeightRegularizer(keras.regularizers.WeightRegularizer): '''Subclass of Keras WeightRegularizer that doesn't use K.in_train_phase, so that total loss can easily be compared between train and val modes. ''' def __init__(self, l1=0., l2=0.): self.l1 = K.cast_to_floatx(l1) self.l2 = K.cast_to_floatx(l2) self.uses_learning_phase = False self.p = None def get_loss(self): loss = 0.0 if self.l1: loss += K.sum(K.abs(self.p)) * self.l1 if self.l2: loss += K.sum(K.square(self.p)) * self.l2 return loss class WeightRegularizerMean(keras.regularizers.WeightRegularizer): '''Subclass of Keras WeightRegularizer that doesn't use K.in_train_phase, so that total loss can easily be compared between train and val modes. Uses mean instead of sum above. ''' def __init__(self, l1=0., l2=0.): self.l1 = K.cast_to_floatx(l1) self.l2 = K.cast_to_floatx(l2) self.uses_learning_phase = False self.p = None def get_loss(self): loss = 0.0 if self.l1: loss += K.mean(K.abs(self.p)) * self.l1 if self.l2: loss += K.mean(K.square(self.p)) * self.l2 return loss class ActivityRegularizer(keras.regularizers.ActivityRegularizer): '''Subclass of Keras ActivityRegularizer that doesn't use K.in_train_phase, so that total loss can easily be compared between train and val modes. ''' def __init__(self, l1=0., l2=0.): self.l1 = K.cast_to_floatx(l1) self.l2 = K.cast_to_floatx(l2) self.uses_learning_phase = False self.layer = None def get_loss(self): if self.layer is None: raise Exception('Need to call `set_layer` on ' 'ActivityRegularizer instance ' 'before calling the instance.') loss = 0.0 for i in range(len(self.layer.inbound_nodes)): output = self.layer.get_output_at(i) if self.l1: loss += K.sum(self.l1 * K.abs(output)) if self.l2: loss += K.sum(self.l2 * K.square(output)) return loss def l1(l=0.01): return WeightRegularizer(l1=l) def l2(l=0.01): return WeightRegularizer(l2=l) def l1l2(l1=0.01, l2=0.01): return WeightRegularizer(l1=l1, l2=l2) def activity_l1(l=0.01): return ActivityRegularizer(l1=l) def activity_l2(l=0.01): return ActivityRegularizer(l2=l) def activity_l1l2(l1=0.01, l2=0.01): return ActivityRegularizer(l1=l1, l2=l2)
3,830
30.925
80
py
fl-analysis
fl-analysis-master/src/subspace/keras_ext/rproj_layers.py
# Copyright (c) 2018 Uber Technologies, Inc. # # Permission is hereby granted, free of charge, to any person obtaining a copy # of this software and associated documentation files (the "Software"), to deal # in the Software without restriction, including without limitation the rights # to use, copy, modify, merge, publish, distribute, sublicense, and/or sell # copies of the Software, and to permit persons to whom the Software is # furnished to do so, subject to the following conditions: # # The above copyright notice and this permission notice shall be included in all # copies or substantial portions of the Software. # # THE SOFTWARE IS PROVIDED "AS IS", WITHOUT WARRANTY OF ANY KIND, EXPRESS OR # IMPLIED, INCLUDING BUT NOT LIMITED TO THE WARRANTIES OF MERCHANTABILITY, # FITNESS FOR A PARTICULAR PURPOSE AND NONINFRINGEMENT. IN NO EVENT SHALL THE # AUTHORS OR COPYRIGHT HOLDERS BE LIABLE FOR ANY CLAIM, DAMAGES OR OTHER # LIABILITY, WHETHER IN AN ACTION OF CONTRACT, TORT OR OTHERWISE, ARISING FROM, # OUT OF OR IN CONNECTION WITH THE SOFTWARE OR THE USE OR OTHER DEALINGS IN THE # SOFTWARE. import numpy as np import tensorflow as tf from tensorflow.keras.layers import Layer, InputSpec import tensorflow.keras.backend as K from tensorflow.python.keras.utils import conv_utils from src.subspace.keras_ext.rproj_layers_util import _convert_string_dtype # from keras.backend.tensorflow_backend import _convert_string_dtype from tensorflow.keras import regularizers, constraints, initializers, activations ########### # # Low Rank Basis Layers # # These layers are modified versions of standard Keras layers that # accept an OffsetCreator*Proj to create offsets from a weight basis # in a Dense/Sparse/Fastfood agnostic manner. # ########### class LowRankBasisLayer(Layer): '''Smarter version of Layer...''' def __init__(self, offset_creator_class, weight_basis, *args, **kwargs): super(LowRankBasisLayer, self).__init__(*args, **kwargs) # offset_creator is an object that creates theta offsets self.offset_creator = offset_creator_class() self.weight_basis = weight_basis # These may or may not be used by subclasses #self._basis_matrices = [] #self._basis_matrix_normalizers = [] # TODO check for use of basis_matrices @property def basis_matrices(self): print('USED HERE basis_matrices') return self._basis_matrices # TODO check for use of basis_matrix_normalizers @property def basis_matrix_normalizers(self): print('USED HERE basis_matrix_normalizers') return self._basis_matrix_normalizers def add_weight(self, name, shape, dtype=None, initializer=None, regularizer=None, trainable=True, constraint=None): '''Version of add_weight that creates a weight theta by instantiating theta_0 and then adding to it an offset from the member offset_creator. ''' initializer = initializers.get(initializer) if dtype is None: dtype = K.floatx() # Create Theta_0 value_0 = initializer(shape) theta_0 = tf.Variable(value_0, trainable=False, dtype=_convert_string_dtype(dtype), name='%s_theta0' % name) if isinstance(value_0, np.ndarray): theta_0._keras_shape = value_0.shape elif hasattr(value_0, 'get_shape'): theta_0._keras_shape = tuple(map(int, value_0.get_shape())) theta_0._uses_learning_phase = False # Call offset creator exec = self.offset_creator.create_theta_offset(self.weight_basis, theta_0.get_shape(), dtype=dtype, name=name) non_trainable_weights = exec.ww # if regularizer is not None: # self.add_loss(regularizer(theta)) # if constraint is not None: # self.constraints[theta] = constraint #self._base_thetas.append(theta_0) #self._basis_matrices.append(ww) #self._non_trainable_weights.extend([theta_0, ww]) self._non_trainable_weights.extend([theta_0] + [non_trainable_weights]) return theta_0, exec def add_non_trainable_weight(self, name, shape, dtype=None, initializer=None, regularizer=None, constraint=None): '''Adds a weight variable to the layer. # Arguments name: String, the name for the weight variable. shape: The shape tuple of the weight. dtype: The dtype of the weight. initializer: An Initializer instance (callable). regularizer: An optional Regularizer instance. trainable: A boolean, whether the weight should be trained via backprop or not (assuming that the layer itself is also trainable). constraint: An optional Constraint instance. # Returns The created weight variable. ''' initializer = initializers.get(initializer) if dtype is None: dtype = K.floatx() weight = tf.Variable(initializer(shape), dtype=dtype, name=name, trainable=False) # weight = K.variable(initializer(shape), dtype=dtype, name=name) if regularizer is not None: self.add_loss(regularizer(weight)) if constraint is not None: self.constraints[weight] = constraint self._non_trainable_weights.append(weight) return weight class RProjDense(LowRankBasisLayer): '''RProj version of Dense.''' def __init__(self, offset_creator_class, weight_basis, units, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs): if 'input_shape' not in kwargs and 'input_dim' in kwargs: kwargs['input_shape'] = (kwargs.pop('input_dim'),) super(RProjDense, self).__init__(offset_creator_class, weight_basis, **kwargs) self.units = units self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.activity_regularizer = regularizers.get(activity_regularizer) self.kernel_constraint = constraints.get(kernel_constraint) self.bias_constraint = constraints.get(bias_constraint) self.input_spec = InputSpec(min_ndim=2) self.supports_masking = True def build(self, input_shape): assert len(input_shape) >= 2 input_dim = input_shape[-1] self.kernel = self.add_weight(shape=(input_dim, self.units), initializer=self.kernel_initializer, name='kernel', regularizer=self.kernel_regularizer, constraint=self.kernel_constraint) if self.use_bias: self.bias = self.add_weight(shape=(self.units,), initializer=self.bias_initializer, name='bias', regularizer=self.bias_regularizer, constraint=self.bias_constraint) else: self.bias = None self.input_spec = InputSpec(min_ndim=2, axes={-1: input_dim}) self.built = True def call(self, inputs, **kwargs): kt0, eproj = self.kernel k = tf.add(kt0, eproj()) bt0, eprojb = self.bias b = tf.add(bt0, eprojb()) # Normal dense functionality output = K.dot(inputs, k) if self.use_bias: output = K.bias_add(output, b) if self.activation is not None: output = self.activation(output) return output def compute_output_shape(self, input_shape): assert input_shape and len(input_shape) >= 2 assert input_shape[-1] output_shape = list(input_shape) output_shape[-1] = self.units return tuple(output_shape) class _RProjConv(LowRankBasisLayer): '''Abstract nD convolution layer (private, used as implementation base). Only the intrinsic parameters (RProj) are Trainable.''' def __init__(self, offset_creator_class, weight_basis, rank, filters, kernel_size, strides=1, padding='valid', data_format=None, dilation_rate=1, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs): super(_RProjConv, self).__init__(offset_creator_class, weight_basis, **kwargs) self.rank = rank self.filters = filters self.kernel_size = conv_utils.normalize_tuple(kernel_size, rank, 'kernel_size') self.strides = conv_utils.normalize_tuple(strides, rank, 'strides') self.padding = conv_utils.normalize_padding(padding) self.data_format = conv_utils.normalize_data_format(data_format) self.dilation_rate = conv_utils.normalize_tuple(dilation_rate, rank, 'dilation_rate') self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.activity_regularizer = regularizers.get(activity_regularizer) self.kernel_constraint = constraints.get(kernel_constraint) self.bias_constraint = constraints.get(bias_constraint) self.input_spec = InputSpec(ndim=self.rank + 2) def build(self, input_shape): if self.data_format == 'channels_first': channel_axis = 1 else: channel_axis = -1 if input_shape[channel_axis] is None: raise ValueError('The channel dimension of the inputs ' 'should be defined. Found `None`.') input_dim = input_shape[channel_axis] kernel_shape = self.kernel_size + (input_dim, self.filters) self.kernel = self.add_weight(shape=kernel_shape, initializer=self.kernel_initializer, name='kernel', regularizer=self.kernel_regularizer, constraint=self.kernel_constraint) if self.use_bias: self.bias = self.add_weight(shape=(self.filters,), initializer=self.bias_initializer, name='bias', regularizer=self.bias_regularizer, constraint=self.bias_constraint) else: self.bias = None # Set input spec. self.input_spec = InputSpec(ndim=self.rank + 2, axes={channel_axis: input_dim}) self.built = True def call(self, inputs): if self.rank == 1: outputs = K.conv1d( inputs, self.kernel, strides=self.strides[0], padding=self.padding, data_format=self.data_format, dilation_rate=self.dilation_rate[0]) if self.rank == 2: outputs = K.conv2d( inputs, self.kernel, strides=self.strides, padding=self.padding, data_format=self.data_format, dilation_rate=self.dilation_rate) if self.rank == 3: outputs = K.conv3d( inputs, self.kernel, strides=self.strides, padding=self.padding, data_format=self.data_format, dilation_rate=self.dilation_rate) if self.use_bias: outputs = K.bias_add( outputs, self.bias, data_format=self.data_format) if self.activation is not None: return self.activation(outputs) return outputs def compute_output_shape(self, input_shape): if self.data_format == 'channels_last': space = input_shape[1:-1] new_space = [] for i in range(len(space)): new_dim = conv_utils.conv_output_length( space[i], self.kernel_size[i], padding=self.padding, stride=self.strides[i], dilation=self.dilation_rate[i]) new_space.append(new_dim) return (input_shape[0],) + tuple(new_space) + (self.filters,) if self.data_format == 'channels_first': space = input_shape[2:] new_space = [] for i in range(len(space)): new_dim = conv_utils.conv_output_length( space[i], self.kernel_size[i], padding=self.padding, stride=self.strides[i], dilation=self.dilation_rate[i]) new_space.append(new_dim) return (input_shape[0], self.filters) + tuple(new_space) class RProjConv2D(_RProjConv): '''Low Rank Basis Conv2D Filters if number of filters, output dimension is filters TODO: Documentation / unit tests ''' def __init__(self, offset_creator_class, weight_basis, filters, kernel_size, strides=(1, 1), padding='valid', data_format=None, dilation_rate=(1, 1), activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs): super(RProjConv2D, self).__init__( offset_creator_class=offset_creator_class, weight_basis=weight_basis, rank=2, filters=filters, kernel_size=kernel_size, strides=strides, padding=padding, data_format=data_format, dilation_rate=dilation_rate, activation=activation, use_bias=use_bias, kernel_initializer=kernel_initializer, bias_initializer=bias_initializer, kernel_regularizer=kernel_regularizer, bias_regularizer=bias_regularizer, activity_regularizer=activity_regularizer, kernel_constraint=kernel_constraint, bias_constraint=bias_constraint, **kwargs) self.input_spec = InputSpec(ndim=4) def build(self, input_shape): assert self.data_format != 'channels_first','only b01c supported' channel_axis = -1 if input_shape[channel_axis] is None: raise ValueError('The channel dimension of the inputs ' 'should be defined. Found `None`.') input_dim = input_shape[-1] self.units = self.filters kernel_shape = self.kernel_size + (input_dim, self.filters) self.kernel = self.add_weight(shape=kernel_shape, initializer=self.kernel_initializer, name='kernel', regularizer=self.kernel_regularizer, constraint=self.kernel_constraint) if self.use_bias: self.bias = self.add_weight(shape=(self.filters,), initializer=self.bias_initializer, name='bias', regularizer=self.bias_regularizer, constraint=self.bias_constraint) else: self.bias = None # Set input spec. self.input_spec = InputSpec(ndim=self.rank + 2, axes={channel_axis: input_dim}) self.built = True def call(self, inputs): assert self.rank == 2, 'only conv2d supported for now...' kt0, eproj = self.kernel k = tf.add(kt0, eproj()) bt0, eprojb = self.bias b = tf.add(bt0, eprojb()) if self.rank == 2: outputs = K.conv2d( inputs, k, strides=self.strides, padding=self.padding, data_format=self.data_format, dilation_rate=self.dilation_rate) if self.use_bias: outputs = K.bias_add( outputs, b, data_format=self.data_format) #if self.activation is not None: # assert False,'activation functions not supported' # return self.activation(outputs) return outputs def compute_output_shape(self, input_shape): if self.data_format == 'channels_last': space = input_shape[1:-1] new_space = [] for i in range(len(space)): new_dim = conv_utils.conv_output_length( space[i], self.kernel_size[i], padding=self.padding, stride=self.strides[i], dilation=self.dilation_rate[i]) new_space.append(new_dim) #self.filters*2 to accomodate LU representation return (input_shape[0],) + tuple(new_space) + (self.filters,) class RProjBatchNormalization(LowRankBasisLayer): '''RProj version of BatchNormalization.''' def __init__(self, offset_creator_class, weight_basis, axis=-1, momentum=0.99, epsilon=1e-3, center=True, scale=True, beta_initializer='zeros', gamma_initializer='ones', moving_mean_initializer='zeros', moving_variance_initializer='ones', beta_regularizer=None, gamma_regularizer=None, beta_constraint=None, gamma_constraint=None, **kwargs): super(RProjBatchNormalization, self).__init__(offset_creator_class, weight_basis, **kwargs) self.supports_masking = True self.axis = axis self.momentum = momentum self.epsilon = epsilon self.center = center self.scale = scale self.beta_initializer = initializers.get(beta_initializer) self.gamma_initializer = initializers.get(gamma_initializer) self.moving_mean_initializer = initializers.get(moving_mean_initializer) self.moving_variance_initializer = initializers.get(moving_variance_initializer) self.beta_regularizer = regularizers.get(beta_regularizer) self.gamma_regularizer = regularizers.get(gamma_regularizer) self.beta_constraint = constraints.get(beta_constraint) self.gamma_constraint = constraints.get(gamma_constraint) def build(self, input_shape): dim = input_shape[self.axis] if dim is None: raise ValueError('Axis ' + str(self.axis) + ' of ' 'input tensor should have a defined dimension ' 'but the layer received an input with shape ' + str(input_shape) + '.') self.input_spec = InputSpec(ndim=len(input_shape), axes={self.axis: dim}) shape = (dim,) if self.scale: self.gamma = self.add_weight(shape=shape, name='gamma', initializer=self.gamma_initializer, regularizer=self.gamma_regularizer, constraint=self.gamma_constraint) else: self.gamma = None if self.center: self.beta = self.add_weight(shape=shape, name='beta', initializer=self.beta_initializer, regularizer=self.beta_regularizer, constraint=self.beta_constraint) else: self.beta = None self.moving_mean = self.add_non_trainable_weight( shape=shape, name='moving_mean', initializer=self.moving_mean_initializer) self.moving_variance = self.add_non_trainable_weight( shape=shape, name='moving_variance', initializer=self.moving_variance_initializer) self.built = True def call(self, inputs, training=None): # training = self._get_training_value(training) input_shape = K.int_shape(inputs) # Prepare broadcasting shape. ndim = len(input_shape) reduction_axes = list(range(len(input_shape))) del reduction_axes[self.axis] broadcast_shape = [1] * len(input_shape) broadcast_shape[self.axis] = input_shape[self.axis] # Determines whether broadcasting is needed. needs_broadcasting = (sorted(reduction_axes) != list(range(ndim))[:-1]) # exec in call gamma_init, gamma_exec = self.gamma gamma = tf.add(gamma_init, gamma_exec()) beta_init, beta_exec = self.beta beta = tf.add(beta_init, beta_exec()) def normalize_inference(): if needs_broadcasting: # In this case we must explicitly broadcast all parameters. broadcast_moving_mean = K.reshape(self.moving_mean, broadcast_shape) broadcast_moving_variance = K.reshape(self.moving_variance, broadcast_shape) if self.center: broadcast_beta = K.reshape(beta, broadcast_shape) else: broadcast_beta = None if self.scale: broadcast_gamma = K.reshape(gamma, broadcast_shape) else: broadcast_gamma = None return K.batch_normalization( inputs, broadcast_moving_mean, broadcast_moving_variance, broadcast_beta, broadcast_gamma, epsilon=self.epsilon) else: return K.batch_normalization( inputs, self.moving_mean, self.moving_variance, beta, gamma, epsilon=self.epsilon) # If the learning phase is *static* and set to inference: # if tf.cond(training, tf.constant(True)): # if training in {0, False}: # return normalize_inference() # If the learning is either dynamic, or set to training: # print(inputs) # print(gamma, beta) normed_training, mean, variance = K.normalize_batch_in_training( inputs, gamma, beta, reduction_axes, epsilon=self.epsilon) self.add_update([K.moving_average_update(self.moving_mean, mean, self.momentum), K.moving_average_update(self.moving_variance, variance, self.momentum)], inputs) # Pick the normalized form corresponding to the training phase. return K.in_train_phase(normed_training, normalize_inference, training=training) # def _get_training_value(self, training=None): # print(training) # if training is None: # training = K.learning_phase() # # if isinstance(training, int): # training = bool(training) # return training # return training == tf.Tensor() # if base_layer_utils.is_in_keras_graph(): # training = math_ops.logical_and(training, self._get_trainable_var()) # else: # training = math_ops.logical_and(training, self.trainable) # return training class RProjLocallyConnected2D(LowRankBasisLayer): """Locally-connected layer for 2D inputs. The `LocallyConnected2D` layer works similarly to the `Conv2D` layer, except that weights are unshared, that is, a different set of filters is applied at each different patch of the input. """ # @interfaces.legacy_conv2d_support def __init__(self, offset_creator_class, weight_basis, filters, kernel_size, strides=(1, 1), padding='valid', data_format=None, activation=None, use_bias=True, kernel_initializer='glorot_uniform', bias_initializer='zeros', kernel_regularizer=None, bias_regularizer=None, activity_regularizer=None, kernel_constraint=None, bias_constraint=None, **kwargs): super(RProjLocallyConnected2D, self).__init__(offset_creator_class, weight_basis, **kwargs) self.filters = filters self.kernel_size = conv_utils.normalize_tuple(kernel_size, 2, 'kernel_size') self.strides = conv_utils.normalize_tuple(strides, 2, 'strides') self.padding = conv_utils.normalize_padding(padding) if self.padding != 'valid': raise ValueError('Invalid border mode for LocallyConnected2D ' '(only "valid" is supported): ' + padding) self.data_format = conv_utils.normalize_data_format(data_format) self.activation = activations.get(activation) self.use_bias = use_bias self.kernel_initializer = initializers.get(kernel_initializer) self.bias_initializer = initializers.get(bias_initializer) self.kernel_regularizer = regularizers.get(kernel_regularizer) self.bias_regularizer = regularizers.get(bias_regularizer) self.activity_regularizer = regularizers.get(activity_regularizer) self.kernel_constraint = constraints.get(kernel_constraint) self.bias_constraint = constraints.get(bias_constraint) self.input_spec = InputSpec(ndim=4) def build(self, input_shape): if self.data_format == 'channels_last': input_row, input_col = input_shape[1:-1] input_filter = input_shape[3] else: input_row, input_col = input_shape[2:] input_filter = input_shape[1] if input_row is None or input_col is None: raise ValueError('The spatial dimensions of the inputs to ' ' a LocallyConnected2D layer ' 'should be fully-defined, but layer received ' 'the inputs shape ' + str(input_shape)) output_row = conv_utils.conv_output_length(input_row, self.kernel_size[0], self.padding, self.strides[0]) output_col = conv_utils.conv_output_length(input_col, self.kernel_size[1], self.padding, self.strides[1]) self.output_row = output_row self.output_col = output_col self.kernel_shape = (output_row * output_col, self.kernel_size[0] * self.kernel_size[1] * input_filter, self.filters) self.kernel = self.add_weight(shape=self.kernel_shape, initializer=self.kernel_initializer, name='kernel', regularizer=self.kernel_regularizer, constraint=self.kernel_constraint) if self.use_bias: self.bias = self.add_weight(shape=(output_row, output_col, self.filters), initializer=self.bias_initializer, name='bias', regularizer=self.bias_regularizer, constraint=self.bias_constraint) else: self.bias = None if self.data_format == 'channels_first': self.input_spec = InputSpec(ndim=4, axes={1: input_filter}) else: self.input_spec = InputSpec(ndim=4, axes={-1: input_filter}) self.built = True def compute_output_shape(self, input_shape): if self.data_format == 'channels_first': rows = input_shape[2] cols = input_shape[3] elif self.data_format == 'channels_last': rows = input_shape[1] cols = input_shape[2] rows = conv_utils.conv_output_length(rows, self.kernel_size[0], self.padding, self.strides[0]) cols = conv_utils.conv_output_length(cols, self.kernel_size[1], self.padding, self.strides[1]) if self.data_format == 'channels_first': return (input_shape[0], self.filters, rows, cols) elif self.data_format == 'channels_last': return (input_shape[0], rows, cols, self.filters) def call(self, inputs): _, _, filters = self.kernel_shape output = K.local_conv2d(inputs, self.kernel, self.kernel_size, self.strides, (self.output_row, self.output_col), self.data_format) if self.use_bias: if self.data_format == 'channels_first' or self.data_format == 'channels_last': output = K.bias_add(output, self.bias, data_format=self.data_format) output = self.activation(output) return output def get_config(self): config = { 'filters': self.filters, 'kernel_size': self.kernel_size, 'strides': self.strides, 'padding': self.padding, 'data_format': self.data_format, 'activation': activations.serialize(self.activation), 'use_bias': self.use_bias, 'kernel_initializer': initializers.serialize(self.kernel_initializer), 'bias_initializer': initializers.serialize(self.bias_initializer), 'kernel_regularizer': regularizers.serialize(self.kernel_regularizer), 'bias_regularizer': regularizers.serialize(self.bias_regularizer), 'activity_regularizer': regularizers.serialize(self.activity_regularizer), 'kernel_constraint': constraints.serialize(self.kernel_constraint), 'bias_constraint': constraints.serialize(self.bias_constraint) } base_config = super(LocallyConnected2D, self).get_config() return dict(list(base_config.items()) + list(config.items()))
33,100
41.491656
116
py
fl-analysis
fl-analysis-master/src/data/emnist.py
import os import h5py import tensorflow as tf import numpy as np def load_data(only_digits=True, cache_dir=None): """Loads the Federated EMNIST dataset. Downloads and caches the dataset locally. If previously downloaded, tries to load the dataset from cache. This dataset is derived from the Leaf repository (https://github.com/TalwalkarLab/leaf) pre-processing of the Extended MNIST dataset, grouping examples by writer. Details about Leaf were published in "LEAF: A Benchmark for Federated Settings" https://arxiv.org/abs/1812.01097. *Note*: This dataset does not include some additional preprocessing that MNIST includes, such as size-normalization and centering. In the Federated EMNIST data, the value of 1.0 corresponds to the background, and 0.0 corresponds to the color of the digits themselves; this is the *inverse* of some MNIST representations, e.g. in [tensorflow_datasets] (https://github.com/tensorflow/datasets/blob/master/docs/datasets.md#mnist), where 0 corresponds to the background color, and 255 represents the color of the digit. Data set sizes: *only_digits=True*: 3,383 users, 10 label classes - train: 341,873 examples - test: 40,832 examples *only_digits=False*: 3,400 users, 62 label classes - train: 671,585 examples - test: 77,483 examples Rather than holding out specific users, each user's examples are split across _train_ and _test_ so that all users have at least one example in _train_ and one example in _test_. Writers that had less than 2 examples are excluded from the data set. The `tf.data.Datasets` returned by `tff.simulation.ClientData.create_tf_dataset_for_client` will yield `collections.OrderedDict` objects at each iteration, with the following keys and values: - `'pixels'`: a `tf.Tensor` with `dtype=tf.float32` and shape [28, 28], containing the pixels of the handwritten digit, with values in the range [0.0, 1.0]. - `'label'`: a `tf.Tensor` with `dtype=tf.int32` and shape [1], the class label of the corresponding pixels. Labels [0-9] correspond to the digits classes, labels [10-35] correspond to the uppercase classes (e.g., label 11 is 'B'), and labels [36-61] correspond to the lowercase classes (e.g., label 37 is 'b'). Args: only_digits: (Optional) whether to only include examples that are from the digits [0-9] classes. If `False`, includes lower and upper case characters, for a total of 62 class labels. cache_dir: (Optional) directory to cache the downloaded file. If `None`, caches in Keras' default cache directory. Returns: Tuple of (train, test) where the tuple elements are `tff.simulation.ClientData` objects. """ if only_digits: fileprefix = 'fed_emnist_digitsonly' sha256 = '55333deb8546765427c385710ca5e7301e16f4ed8b60c1dc5ae224b42bd5b14b' else: fileprefix = 'fed_emnist' sha256 = 'fe1ed5a502cea3a952eb105920bff8cffb32836b5173cb18a57a32c3606f3ea0' filename = fileprefix + '.tar.bz2' path = tf.keras.utils.get_file( filename, origin='https://storage.googleapis.com/tff-datasets-public/' + filename, file_hash=sha256, hash_algorithm='sha256', extract=True, archive_format='tar', cache_dir=cache_dir) dir_path = os.path.dirname(path) train_client_data = process_h5py(os.path.join(dir_path, fileprefix + '_train.h5')) test_client_data = process_h5py(os.path.join(dir_path, fileprefix + '_test.h5')) return train_client_data, test_client_data def process_h5py(filename): file = h5py.File(filename, 'r') drawers = file['examples'] out = [] for i, key in enumerate(drawers.keys()): out.append({ 'pixels': drawers[key]['pixels'].value, 'label': drawers[key]['label'].value}) return np.asarray(out)
3,842
43.172414
99
py
fl-analysis
fl-analysis-master/src/data/tf_data.py
import itertools import math import numpy as np import tensorflow as tf from tensorflow.python.keras.preprocessing.image import ImageDataGenerator from src.data import image_augmentation from src.data import emnist class Dataset: def __init__(self, x_train, y_train, batch_size=50, x_test=None, y_test=None): self.batch_size = batch_size # LIMIT = 5000 # for debugging remove this # x_train, y_train = x_train[:LIMIT], y_train[:LIMIT] self.x_train, self.y_train = self.shuffle(x_train, y_train) self.x_test, self.y_test = x_test, y_test self.x_aux, self.y_aux, self.mal_aux_labels = None, None, None self.x_aux_test, self.mal_aux_labels_test = None, None self.fg = tf.data.Dataset.from_tensor_slices((self.x_train, self.y_train)) def shuffle(self, x, y): perms = np.random.permutation(x.shape[0]) return x[perms, :], y[perms] def get_data(self): """Creates one batch of data. # This is a TERRIBLE way to load data... every epoch we get it in the same order !!! Yields:˚ tuple of two: input data batch and corresponding labels """ # count = int(self.x_train.shape[0] / self.batch_size) # if count == 0: # yield self.x_train, self.y_train # # return [(self.x_train, self.y_train)] # for bid in range(count): # Note: Unsafe if batch_size is small!!! # batch_x = self.x_train[bid * self.batch_size:(bid + 1) * self.batch_size] # batch_y = self.y_train[bid * self.batch_size:(bid + 1) * self.batch_size] # # yield batch_x, batch_y # bid = 0 shuffle_size = min(self.x_train.shape[0], 10000) return self.fg \ .shuffle(shuffle_size) \ .batch(self.batch_size, drop_remainder=True) \ .prefetch(tf.data.experimental.AUTOTUNE) def get_aux(self, mal_num_batch): """Creates one batch of data. Yields: tuple of two: input data batch and corresponding labels """ if int(self.x_aux.shape[0] / self.batch_size) < 1: yield self.x_aux, self.mal_aux_labels for bid in range(int(self.x_aux.shape[0] / self.batch_size)): batch_x = self.x_aux[bid * self.batch_size:(bid + 1) * self.batch_size] batch_y = self.mal_aux_labels[bid * self.batch_size:(bid + 1) * self.batch_size] yield batch_x, batch_y bid = 0 def get_data_with_aux(self, insert_aux_times, num_batches, pixel_pattern=None, noise_level=None): """Creates one batch of data with the AUX data inserted `insert_aux_times` per batch with malicious labels. :param insert_aux_times number of times aux should be inserted per batch. 1 for Bagdasaryan :param num_batches number of batches to generate. 200 for Bagdasaryan :param noise_level sigma of normal distribution noise to add to training samples Yields: tuple of two: input data batch and corresponding labels """ multiplier = max(float(insert_aux_times) / float(self.mal_aux_labels.shape[0]), 1) # potential multiplier if aux is smaller than insert number_of_mal_items = int(multiplier * num_batches) r1 = insert_aux_times r2 = self.batch_size - insert_aux_times normal_mult = max(float(num_batches) * float(self.batch_size) / self.x_train.shape[0], 1) normal_fg = self.fg \ .repeat(int(math.ceil(normal_mult * self.x_train.shape[0]))) \ .shuffle(self.x_train.shape[0]) \ .batch(r2, drop_remainder=True) \ if insert_aux_times == 0: return normal_fg mal_fb = tf.data.Dataset.from_tensor_slices((self.x_aux, self.mal_aux_labels)) \ .repeat(number_of_mal_items) \ .shuffle(number_of_mal_items) if noise_level is not None: # Add noise mal_fb = mal_fb.map(image_augmentation.add_noise(noise_level)) mal_fb = mal_fb.batch(r1, drop_remainder=True) zipped = tf.data.Dataset.zip((mal_fb, normal_fg)).map(lambda x, y: (tf.concat((x[0], y[0]), axis=0), tf.concat((x[1], y[1]), axis=0)) ) result = zipped.unbatch() return result.batch(self.batch_size, drop_remainder=True) \ .take(num_batches) def get_aux_test_generator(self, aux_size): if aux_size == 0: return tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .batch(self.batch_size, drop_remainder=False) \ .prefetch(tf.data.experimental.AUTOTUNE) return tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .repeat(aux_size) \ .batch(self.batch_size, drop_remainder=False) \ .prefetch(tf.data.experimental.AUTOTUNE) @staticmethod def keep_samples(x_train, y_train, number_of_samples): if number_of_samples == -1: return x_train, y_train perms = np.random.permutation(number_of_samples) return x_train[perms, :], y_train[perms] @staticmethod def keep_samples_iterative(x_train, y_train, number_of_samples): if number_of_samples == -1: return x_train, y_train perms = [np.random.permutation(min(number_of_samples, val.shape[0])) for val in x_train] return [val[perm, :] for val, perm in zip(x_train, perms)], \ [val[perm] for val, perm in zip(y_train, perms)] @staticmethod def apply_trigger(x_aux): triggersize = 4 trigger = np.ones((x_aux.shape[0], triggersize, triggersize, 1)) out = x_aux out[:, 0:triggersize, 0:triggersize, :] = trigger return out @staticmethod def get_mnist_dataset(number_of_samples): """MNIST dataset loader""" mnist = tf.keras.datasets.mnist (x_train, y_train), (x_test, y_test) = mnist.load_data() x_train, x_test = x_train / 255.0, x_test / 255.0 x_train, x_test = x_train.astype(np.float32), x_test.astype(np.float32) x_train, x_test = x_train[..., np.newaxis], x_test[..., np.newaxis] x_train, y_train = Dataset.keep_samples(x_train, y_train, number_of_samples) return (x_train, y_train), (x_test, y_test) @staticmethod def get_fmnist_dataset(number_of_samples): """Fashion MNIST dataset loader""" fmnist = tf.keras.datasets.fashion_mnist (x_train, y_train), (x_test, y_test) = fmnist.load_data() x_train, x_test = x_train / 255.0, x_test / 255.0 x_train, x_test = x_train.astype(np.float32), x_test.astype(np.float32) x_train, x_test = x_train[..., np.newaxis], x_test[..., np.newaxis] x_train, y_train = Dataset.keep_samples(x_train, y_train, number_of_samples) return (x_train, y_train), (x_test, y_test) @staticmethod def get_cifar10_dataset(number_of_samples): """Cifar10 dataset loader""" cifar = tf.keras.datasets.cifar10 (x_train, y_train), (x_test, y_test) = cifar.load_data() x_train, x_test = x_train / 255.0, x_test / 255.0 x_train, x_test = x_train.astype(np.float32), x_test.astype(np.float32) y_train, y_test = np.squeeze(y_train, axis=1), np.squeeze(y_test, axis=1) x_train, y_train = Dataset.keep_samples(x_train, y_train, number_of_samples) x_test, y_test = Dataset.keep_samples(x_test, y_test, -1) # Note: hardcoded # Subtract x_train_mean = np.mean(x_train, axis=0) x_train -= x_train_mean x_test -= x_train_mean return (x_train, y_train), (x_test, y_test) @staticmethod def get_emnist_dataset(number_of_samples, number_of_clients, normalize_mnist_data): """nonIID MNIST dataset loader""" train_dataset, test_dataset = emnist.load_data() x_train, y_train = np.array([1.0 - np.array(val['pixels']) for val in train_dataset]), \ np.array([np.array(val['label']).astype(np.uint8) for val in train_dataset]) x_test, y_test = np.array([1.0 - np.array(val['pixels']) for val in test_dataset]), \ np.array([np.array(val['label']).astype(np.uint8) for val in test_dataset]) if normalize_mnist_data: emnist_mean, emnist_std = 0.036910772, 0.16115953 x_train = np.array([(x - emnist_mean) / emnist_std for x in x_train]) x_test = np.array([(x - emnist_mean) / emnist_std for x in x_test]) # Randomly assign clients to buckets but keep them as client if number_of_clients < x_train.shape[0]: assignments = np.random.randint(0, number_of_clients, x_train.shape[0], dtype=np.uint16) new_x_train = [] new_y_train = [] new_x_test = [] new_y_test = [] for i in range(number_of_clients): new_x_train.append( np.concatenate(x_train[assignments == i], axis=0) ) new_y_train.append( np.concatenate(y_train[assignments == i], axis=0) ) new_x_test.append( np.concatenate(x_test[assignments == i], axis=0) ) new_y_test.append( np.concatenate(y_test[assignments == i], axis=0) ) # # new_x_train = np.concatenate(new_x_train, axis=0) # new_y_train = np.concatenate(new_y_train, axis=0) # new_x_test = np.concatenate(new_x_test, axis=0) # new_y_test = np.concatenate(new_y_test, axis=0) if number_of_samples == -1: number_of_samples_per_client = -1 else: number_of_samples_per_client = int(number_of_samples / float(number_of_clients)) x_train, y_train = Dataset.keep_samples_iterative(new_x_train, new_y_train, number_of_samples_per_client) x_test, y_test = Dataset.keep_samples_iterative(new_x_test, new_y_test, min(number_of_samples_per_client, 500)) elif number_of_clients > x_train.shape[0]: print(f"Number of clients {number_of_clients} is large than amount of EMNIST users {x_train.shape[0]}") else: print("Exactly using EMNIST as clients!") x_train, x_test = [val.astype(np.float32)[..., np.newaxis] for val in x_train], \ [val.astype(np.float32)[..., np.newaxis] for val in x_test] return (x_train, y_train), (x_test, y_test) class ImageGeneratorDataset(Dataset): def __init__(self, x_train, y_train, batch_size=50, x_test=None, y_test=None): super().__init__(x_train, y_train, batch_size, x_test, y_test) def get_aux_test_generator(self, aux_size): if aux_size == 0: return tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .batch(self.batch_size, drop_remainder=False) \ .prefetch(tf.data.experimental.AUTOTUNE) test_dataset = tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .repeat(aux_size) \ .batch(self.batch_size, drop_remainder=False) \ return test_dataset \ .map(image_augmentation.test_aux_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ .prefetch(tf.data.experimental.AUTOTUNE) def get_data(self): return self.fg\ .shuffle(self.x_train.shape[0]) \ .batch(self.batch_size, drop_remainder=True) \ .map(image_augmentation.augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ .prefetch(tf.data.experimental.AUTOTUNE) def get_aux(self, mal_num_batch): multiplier = max(float(self.batch_size) / float(self.mal_aux_labels.shape[0]), 1) # potential multiplier if aux is smaller than insert number_of_mal_items = int(multiplier * mal_num_batch) return tf.data.Dataset.from_tensor_slices((self.x_aux, self.mal_aux_labels)) \ .repeat(number_of_mal_items) \ .batch(self.batch_size, drop_remainder=False) \ .map(image_augmentation.train_aux_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ def get_data_with_aux(self, insert_aux_times, num_batches, noise_level=None): """Creates one batch of data with the AUX data inserted `insert_aux_times` per batch with malicious labels. :param insert_aux_times number of times aux should be inserted per batch. 1 for Bagdasaryan :param num_batches number of batches to generate. 200 for Bagdasaryan Yields: tuple of two: input data batch and corresponding labels """ # assert self.y_aux != [] and self.x_aux != [] and self.mal_aux_labels != [] multiplier = float(insert_aux_times) / float(self.mal_aux_labels.shape[0]) # potential multiplier if aux is smaller than insert number_of_mal_items = int(math.ceil(multiplier * num_batches)) r1 = insert_aux_times r2 = self.batch_size - insert_aux_times normal_mult = max(float(num_batches) * float(self.batch_size) / self.x_train.shape[0], 1) normal_fg = self.fg\ .repeat(int(math.ceil(normal_mult))) \ .shuffle(self.x_train.shape[0]) \ .batch(r2, drop_remainder=True) \ .map(image_augmentation.augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ if insert_aux_times == 0: print(f"Insert 0 {normal_mult}") return normal_fg mal_fb = tf.data.Dataset.from_tensor_slices((self.x_aux, self.mal_aux_labels)) \ .repeat(number_of_mal_items) \ .shuffle(number_of_mal_items * self.mal_aux_labels.shape[0]) mal_fb = mal_fb.batch(r1, drop_remainder=True) \ .map(image_augmentation.train_aux_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ if noise_level is not None: # Add noise mal_fb = mal_fb.map(image_augmentation.add_noise_batch(noise_level)) zipped = tf.data.Dataset.zip((mal_fb, normal_fg)).map(lambda x, y: (tf.concat((x[0], y[0]), axis=0), tf.concat((x[1], y[1]), axis=0)) ) result = zipped.unbatch() return result.batch(self.batch_size, drop_remainder=True)\ .take(num_batches) class PixelPatternDataset(ImageGeneratorDataset): def __init__(self, x_train, y_train, target_label, batch_size=50, x_test=None, y_test=None): super().__init__(x_train, y_train, batch_size, x_test, y_test) # Assign train set part (self.x_aux, self.y_aux) = \ (self.x_train, self.y_train) self.mal_aux_labels = np.repeat(target_label, self.y_aux.shape).astype(np.uint8) self.pixel_pattern = 'basic' def get_aux_test_generator(self, aux_size): if aux_size == 0: return tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .batch(self.batch_size, drop_remainder=False) \ .map(image_augmentation.add_pixel_pattern(self.pixel_pattern)) \ .prefetch(tf.data.experimental.AUTOTUNE) test_dataset = tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .repeat(aux_size) \ .batch(self.batch_size, drop_remainder=False) \ return test_dataset \ .map(image_augmentation.test_aux_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ .map(image_augmentation.add_pixel_pattern(self.pixel_pattern)) \ .prefetch(tf.data.experimental.AUTOTUNE) def get_data_with_aux(self, insert_aux_times, num_batches, noise_level=None): """Creates one batch of data with the AUX data inserted `insert_aux_times` per batch with malicious labels. :param insert_aux_times number of times aux should be inserted per batch. 1 for Bagdasaryan :param num_batches number of batches to generate. 200 for Bagdasaryan Yields: tuple of two: input data batch and corresponding labels """ # assert self.y_aux != [] and self.x_aux != [] and self.mal_aux_labels != [] multiplier = float(insert_aux_times) / float( self.mal_aux_labels.shape[0]) # potential multiplier if aux is smaller than insert number_of_mal_items = int(math.ceil(multiplier * num_batches)) r1 = insert_aux_times r2 = self.batch_size - insert_aux_times normal_mult = max(float(num_batches) * float(self.batch_size) / self.x_train.shape[0], 1) normal_fg = self.fg \ .repeat(int(math.ceil(normal_mult))) \ .shuffle(self.x_train.shape[0]) \ .batch(r2, drop_remainder=True) \ .map(image_augmentation.augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) if insert_aux_times == 0: return normal_fg mal_fb = tf.data.Dataset.from_tensor_slices((self.x_aux, self.mal_aux_labels)) \ .repeat(number_of_mal_items) \ .shuffle(number_of_mal_items * self.mal_aux_labels.shape[0]) mal_fb = mal_fb.batch(r1, drop_remainder=True) \ .map(image_augmentation.train_aux_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) if noise_level is not None: # Add noise mal_fb = mal_fb.map(image_augmentation.add_noise_batch(noise_level)) mal_fb = mal_fb.map(image_augmentation.add_pixel_pattern(self.pixel_pattern)) zipped = tf.data.Dataset.zip((mal_fb, normal_fg)).map(lambda x, y: (tf.concat((x[0], y[0]), axis=0), tf.concat((x[1], y[1]), axis=0)) ) result = zipped.unbatch() return result.batch(self.batch_size, drop_remainder=True)\ .take(num_batches) class GeneratorDataset(Dataset): def __init__(self, generator, batch_size): super().__init__([], [], 0, None, None) self.generator = generator self.batch_size = batch_size def get_data(self): return self.generator\ .batch(self.batch_size)\ .prefetch(tf.data.experimental.AUTOTUNE)
18,878
42.802784
135
py
fl-analysis
fl-analysis-master/src/data/image_augmentation.py
import tensorflow as tf import numpy as np from tensorflow.keras.preprocessing.image import apply_affine_transform def augment(image,label): image = tf.image.random_flip_left_right(image) image = tf.numpy_function(shift, [image], tf.float32) image = normalize(image) # debug(image, label) return image, label def test_augment(image,label): return normalize(image), label def train_aux_augment(image, label): image = tf.image.random_flip_left_right(image) image = tf.numpy_function(shift, [image], tf.float32) # image = tf.add(image, tf.random.normal(tf.shape(image), 0, 0.05)) return image, label def test_aux_augment(image, label): """Augmentation if aux test set is small""" return augment(image, label) # same as training def normalize(image): mean = [0.4914, 0.4822, 0.4465] std = [0.2023, 0.1994, 0.2010] # tf.print("Before:", tf.shape(image), tf.math.reduce_std(image)) # image = tf.image.per_image_standardization(image) # image = image - tf.reshape(mean, [1, 1, 1, 3]) # image = image / tf.reshape(std, [1, 1, 1, 3]) # tf.print("After:", tf.shape(image), tf.math.reduce_std(image)) return image def shift(images): return np.array([shift_single(i) for i in images]) def shift_single(image): """ Expects numpy, single image """ shape = image.shape tx = np.random.uniform(-0.1, 0.1) * shape[0] ty = np.random.uniform(-0.1, 0.1) * shape[1] image = apply_affine_transform(image, 0, tx, # tx ty, 0, 1, 1, row_axis=0, col_axis=1, channel_axis=2, fill_mode='nearest') return image def add_noise_batch(sigma): def cb(images, labels): images = images + tf.random.normal(tf.shape(images), mean=0, stddev=sigma) return images, labels return cb def add_pixel_pattern(pixel_pattern): triggersize = 4 def np_callback(images): trigger = np.ones((images.shape[0], triggersize, triggersize, images.shape[-1])) images[:, 0:triggersize, 0:triggersize, :] = trigger return images def cb(images, labels): # shape = tf.shape(images) # tf.print(shape) # print(shape) # trigger = tf.ones((shape[0], triggersize, triggersize, shape[-1])) # trigger = tf.ones((None, triggersize, triggersize, 3)) # tf.ones_like # d0 = shape[0] # tf.print(d0) # x = tf.constant(tf.float32, shape=[d0, triggersize, triggersize, 3]) # trigger = tf.ones_like(x) # images[:, 0:triggersize, 0:triggersize, :] = trigger # this callback is slower i think images = tf.numpy_function(np_callback, [images], tf.float32) return images, labels return cb def pixel_pattern_if_needed(needed): def no_op(images, labels): return images, labels if needed: return add_pixel_pattern(None) else: return no_op def debug(image, label): import matplotlib.pyplot as plt for i in range(image.shape[0]): plt.figure() plt.imshow(image[i] + 0.5) plt.title(f"Label: {label[i]}") plt.show()
3,414
28.695652
88
py
fl-analysis
fl-analysis-master/src/data/tf_data_global.py
from collections import defaultdict import numpy as np from tensorflow.python.keras.preprocessing.image import ImageDataGenerator import tensorflow as tf from src.data import image_augmentation import logging class GlobalDataset: """ A GlobalDataset represents a dataset as a whole. It has two purposes. - Client datasets are derived from it - Our global dataset is used for evaluation of the global model. `x_test`, `y_test` and the aux sets """ def __init__(self, x_test, y_test): self.x_test = x_test self.y_test = y_test self.x_train = [] self.y_train = [] self.x_aux_train, self.y_aux_train, self.mal_aux_labels_train = \ [], [], [] self.x_aux_test, self.y_aux_test, self.mal_aux_labels_test = \ [], [], [] self.test_generator = tf.data.Dataset.from_tensor_slices((x_test, y_test)) self.aux_test_generator = None def get_dataset_for_client(self, client_id): raise Exception("Not implemented") def get_normal_and_aux_dataset_for_client(self, client_id, aux_sample_size, attack_objective): raise Exception("Not implemented") def get_test_batch(self, batch_size, max_num_batches=-1): """Creates one batch of test data. Yields: tuple of two: input data batch and corresponding labels """ # count = min(int(self.x_test.shape[0] / batch_size), max_num_batches) # for bid in range(count): # batch_x = self.x_test[bid * batch_size:(bid + 1) * batch_size] # batch_y = self.y_test[bid * batch_size:(bid + 1) * batch_size] # # yield batch_x, batch_y # bid = 0 # Check here if non cifar? return self.test_generator.batch(batch_size) \ .take(max_num_batches) \ .map(image_augmentation.test_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ .prefetch(tf.data.experimental.AUTOTUNE) # TODO 0: Do we want to use num_backdoor_tasks or aux_sample_size ? def build_global_aux(self, mal_clients, num_backdoor_tasks, attack_objective, aux_sample_size, augment_size): """ Select backdoor tasks """ if np.count_nonzero(mal_clients) == 0: return # no aux assert np.count_nonzero(mal_clients) >= num_backdoor_tasks # assert we have less 'tasks' than clients data_x, data_y = self.x_train, self.y_train total_x_aux, total_y_aux, total_mal_aux_labels = [], [], [] if aux_sample_size == -1: aux_sample_size = 10000000 # fix as we reformat this total_aux_count = 0 num_tasks = 0 for i in range(len(data_x)): if num_tasks >= num_backdoor_tasks: break if total_aux_count >= aux_sample_size: print(f"Hit limit of {total_aux_count}/{aux_sample_size} samples!") break if mal_clients[i]: x_train_me, y_train_me = data_x[i], data_y[i] # Pick attack samples inds = np.where(y_train_me == attack_objective[0])[0] # Find all logging.debug(f"{i} Found {len(inds)} of class {attack_objective} to poison!") test_inds = np.ones(x_train_me.shape[0], dtype=bool) test_inds[inds] = False x_aux, y_aux = x_train_me[inds], y_train_me[inds] x_train_me, y_train_me = x_train_me[test_inds], y_train_me[test_inds] # randomly permute labels mal_labels = np.repeat(attack_objective[1], len(y_aux)) current_aux_count = y_aux.size if total_aux_count + current_aux_count > aux_sample_size: # constrain current_aux_count = aux_sample_size - total_aux_count # how many we have left x_aux = x_aux[:current_aux_count, :] y_aux = y_aux[:current_aux_count] mal_labels = mal_labels[:current_aux_count] total_x_aux.append(x_aux) total_y_aux.append(y_aux) total_mal_aux_labels.append(mal_labels) data_x[i], data_y[i] = x_train_me, y_train_me assert not np.any( data_y[i] == attack_objective[0]) # assert data_y doesnt contain any attack label total_aux_count += current_aux_count num_tasks += 1 # assert len(total_x_aux) == num_backdoor_tasks # not applicable with aux_sample_size self.x_aux_train = np.concatenate(total_x_aux) self.y_aux_train = np.concatenate(total_y_aux) self.mal_aux_labels_train = np.concatenate(total_mal_aux_labels).astype(np.uint8) # Assign train as test set for now ... ! Depends on how we want to implement the behavior self.x_aux_test = self.x_aux_train self.y_aux_test = self.y_aux_train self.mal_aux_labels_test = self.mal_aux_labels_train # self.build_aux_generator(augment_size) print(f"Got {len(self.x_aux_train)}/{aux_sample_size} samples for {num_backdoor_tasks} tasks!") # def build_aux_generator(self, augment_size): # # self.aux_test_generator = tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.y_aux_test)) # if augment_size == 0: # self.aux_test_generator = ImageDataGenerator() # else: # self.aux_test_generator = ImageDataGenerator( # # rotation_range=15, # horizontal_flip=True, # width_shift_range=0.1, # height_shift_range=0.1 # ) # self.aux_test_generator.fit(self.x_aux_test) def get_aux_generator(self, batch_size, aux_size, augment_cifar, attack_type, max_test_batches): if aux_size == 0: test_dataset = tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .batch(batch_size, drop_remainder=False) \ .map(image_augmentation.pixel_pattern_if_needed(attack_type == 'pixel_pattern'), num_parallel_calls=tf.data.experimental.AUTOTUNE) if max_test_batches is not None: test_dataset = test_dataset.shuffle(max_test_batches)\ .take(max_test_batches) return test_dataset \ .prefetch(tf.data.experimental.AUTOTUNE) test_dataset = tf.data.Dataset.from_tensor_slices((self.x_aux_test, self.mal_aux_labels_test)) \ .repeat(aux_size) \ .batch(batch_size, drop_remainder=False) \ .map(image_augmentation.pixel_pattern_if_needed(attack_type == 'pixel_pattern'), num_parallel_calls=tf.data.experimental.AUTOTUNE) if augment_cifar: return test_dataset\ .map(image_augmentation.test_aux_augment, num_parallel_calls=tf.data.experimental.AUTOTUNE) \ .prefetch(tf.data.experimental.AUTOTUNE) else: return test_dataset.prefetch(tf.data.experimental.AUTOTUNE) def get_full_dataset(self, size): x, y = np.concatenate(self.x_train), np.concatenate(self.y_train) perms = np.random.choice(x.shape[0], size, replace=False) x, y = x[perms, :], y[perms] return x, y class IIDGlobalDataset(GlobalDataset): def __init__(self, x_train, y_train, num_clients, x_test, y_test): super().__init__(x_test, y_test) self.num_clients = num_clients x_train, y_train = self.shuffle(x_train, y_train) # Add to list for client_id in range(num_clients): data_samples = int(x_train.shape[0] / self.num_clients) inds = (client_id * data_samples, (client_id + 1) * data_samples) x, y = x_train[inds[0]:inds[1]], y_train[inds[0]:inds[1]] self.x_train.append(x) self.y_train.append(y) def shuffle(self, x, y): perms = np.random.permutation(x.shape[0]) return x[perms, :], y[perms] def get_dataset_for_client(self, client_id): # dataset = tf.data.Dataset.from_tensor_slices((self.x_train[client_id], self.y_train[client_id])) # return dataset return self.x_train[client_id], self.y_train[client_id] class NonIIDGlobalDataset(GlobalDataset): def __init__(self, x_train, y_train, x_test, y_test, num_clients): """Expects x_train to be a list, x_test one array""" super().__init__(x_test, y_test) self.x_train, self.y_train = x_train, y_train def shuffle(self): raise Exception("Shuffling is not supported on a non-IID dataset!") def get_dataset_for_client(self, client_id): return self.x_train[client_id], self.y_train[client_id] class DirichletDistributionDivider(): """Divides dataset according to dirichlet distribution""" def __init__(self, x_train, y_train, train_aux, test_aux, exclude_aux, num_clients): """`train_aux` and `test_aux` should be indices for the `train` arrays.""" self.x_train = x_train self.y_train = y_train self.train_aux = train_aux self.test_aux = test_aux self.exclude_aux = exclude_aux self.num_clients = num_clients def build(self): alpha = 0.9 cifar_classes = {} for ind, x in enumerate(self.y_train): label = x if self.exclude_aux and (ind in self.train_aux or ind in self.test_aux): continue if label in cifar_classes: cifar_classes[label].append(ind) else: cifar_classes[label] = [ind] class_size = len(cifar_classes[0]) per_participant_list = defaultdict(list) no_classes = len(cifar_classes.keys()) for n in range(no_classes): np.random.shuffle(cifar_classes[n]) sampled_probabilities = class_size * np.random.dirichlet( np.array(self.num_clients * [alpha])) for user in range(self.num_clients): no_imgs = int(round(sampled_probabilities[user])) sampled_list = cifar_classes[n][:min(len(cifar_classes[n]), no_imgs)] per_participant_list[user].extend(sampled_list) cifar_classes[n] = cifar_classes[n][min(len(cifar_classes[n]), no_imgs):] per_participant_train_x = [self.x_train[ind] for _, ind in per_participant_list.items()] per_participant_train_y = [self.y_train[ind] for _, ind in per_participant_list.items()] for n in range(self.num_clients): perms = np.random.permutation(per_participant_train_x[n].shape[0]) per_participant_train_x[n] = per_participant_train_x[n][perms, :] per_participant_train_y[n] = per_participant_train_y[n][perms] return (per_participant_train_x, per_participant_train_y)
10,891
41.054054
146
py
fl-analysis
fl-analysis-master/src/model/resnet.py
from __future__ import print_function import tensorflow as tf from tensorflow.keras.layers import Dense, Conv2D, BatchNormalization, Activation from tensorflow.keras.layers import AveragePooling2D, Input, Flatten from tensorflow.keras.optimizers import Adam from tensorflow.keras.callbacks import ModelCheckpoint, LearningRateScheduler from tensorflow.keras.callbacks import ReduceLROnPlateau from tensorflow.keras.preprocessing.image import ImageDataGenerator from tensorflow.keras.regularizers import l2 from tensorflow.keras.models import Model import numpy as np import os # # Training parameters # batch_size = 32 # orig paper trained all networks with batch_size=128 # epochs = 200 # data_augmentation = True # num_classes = 10 # # # Subtracting pixel mean improves accuracy # subtract_pixel_mean = True # # # Model parameter # # ---------------------------------------------------------------------------- # # | | 200-epoch | Orig Paper| 200-epoch | Orig Paper| sec/epoch # # Model | n | ResNet v1 | ResNet v1 | ResNet v2 | ResNet v2 | GTX1080Ti # # |v1(v2)| %Accuracy | %Accuracy | %Accuracy | %Accuracy | v1 (v2) # # ---------------------------------------------------------------------------- # # ResNet20 | 3 (2)| 92.16 | 91.25 | ----- | ----- | 35 (---) # # ResNet32 | 5(NA)| 92.46 | 92.49 | NA | NA | 50 ( NA) # # ResNet44 | 7(NA)| 92.50 | 92.83 | NA | NA | 70 ( NA) # # ResNet56 | 9 (6)| 92.71 | 93.03 | 93.01 | NA | 90 (100) # # ResNet110 |18(12)| 92.65 | 93.39+-.16| 93.15 | 93.63 | 165(180) # # ResNet164 |27(18)| ----- | 94.07 | ----- | 94.54 | ---(---) # # ResNet1001| (111)| ----- | 92.39 | ----- | 95.08+-.14| ---(---) # # --------------------------------------------------------------------------- # n = 3 # # # Model version # # Orig paper: version = 1 (ResNet v1), Improved ResNet: version = 2 (ResNet v2) # version = 1 # # # Computed depth from supplied model parameter n # if version == 1: # depth = n * 6 + 2 # elif version == 2: # depth = n * 9 + 2 # # # Model name, depth and version # model_type = 'ResNet%dv%d' % (depth, version) # # # Load the CIFAR10 data. # (x_train, y_train), (x_test, y_test) = cifar10.load_data() # # # Input image dimensions. # input_shape = x_train.shape[1:] # # # Normalize data. # x_train = x_train.astype('float32') / 255 # x_test = x_test.astype('float32') / 255 # # # If subtract pixel mean is enabled # if subtract_pixel_mean: # x_train_mean = np.mean(x_train, axis=0) # x_train -= x_train_mean # x_test -= x_train_mean # # print('x_train shape:', x_train.shape) # print(x_train.shape[0], 'train samples') # print(x_test.shape[0], 'test samples') # print('y_train shape:', y_train.shape) # # # Convert class vectors to binary class matrices. # y_train = keras.utils.to_categorical(y_train, num_classes) # y_test = keras.utils.to_categorical(y_test, num_classes) def lr_schedule(epoch): """Learning Rate Schedule Learning rate is scheduled to be reduced after 80, 120, 160, 180 epochs. Called automatically every epoch as part of callbacks during training. # Arguments epoch (int): The number of epochs # Returns lr (float32): learning rate """ lr = 1e-3 if epoch > 180: lr *= 0.5e-3 elif epoch > 160: lr *= 1e-3 elif epoch > 120: lr *= 1e-2 elif epoch > 80: lr *= 1e-1 print('Learning rate: ', lr) return lr def resnet_layer(inputs, num_filters=16, kernel_size=3, strides=1, activation='relu', batch_normalization=True, conv_first=True, name=None): """2D Convolution-Batch Normalization-Activation stack builder # Arguments inputs (tensor): input tensor from input image or previous layer num_filters (int): Conv2D number of filters kernel_size (int): Conv2D square kernel dimensions strides (int): Conv2D square stride dimensions activation (string): activation name batch_normalization (bool): whether to include batch normalization conv_first (bool): conv-bn-activation (True) or bn-activation-conv (False) # Returns x (tensor): tensor as input to the next layer """ conv = Conv2D(num_filters, kernel_size=kernel_size, strides=strides, padding='same', kernel_initializer='he_normal', kernel_regularizer=l2(1e-4), name=name) x = inputs if conv_first: x = conv(x) if batch_normalization: x = BatchNormalization()(x) if activation is not None: x = Activation(activation)(x) else: if batch_normalization: x = BatchNormalization()(x) if activation is not None: x = Activation(activation)(x) x = conv(x) return x def resnet_v1(input_shape, depth, num_classes=10): """ResNet Version 1 Model builder [a] Stacks of 2 x (3 x 3) Conv2D-BN-ReLU Last ReLU is after the shortcut connection. At the beginning of each stage, the feature map size is halved (downsampled) by a convolutional layer with strides=2, while the number of filters is doubled. Within each stage, the layers have the same number filters and the same number of filters. Features maps sizes: stage 0: 32x32, 16 stage 1: 16x16, 32 stage 2: 8x8, 64 The Number of parameters is approx the same as Table 6 of [a]: ResNet20 0.27M ResNet32 0.46M ResNet44 0.66M ResNet56 0.85M ResNet110 1.7M # Arguments input_shape (tensor): shape of input image tensor depth (int): number of core convolutional layers num_classes (int): number of classes (CIFAR10 has 10) # Returns model (Model): Keras model instance """ if (depth - 2) % 6 != 0: raise ValueError('depth should be 6n+2 (eg 20, 32, 44 in [a])') # Start model definition. num_filters = 16 num_res_blocks = int((depth - 2) / 6) inputs = Input(shape=input_shape) x = resnet_layer(inputs=inputs, num_filters = num_filters) # Instantiate the stack of residual units for stack in range(3): for res_block in range(num_res_blocks): strides = 1 if stack > 0 and res_block == 0: # first layer but not first stack strides = 2 # downsample y = resnet_layer(inputs=x, num_filters=num_filters, strides=strides, name=f"Conv2D_stack{stack}_res{res_block}_l0") y = resnet_layer(inputs=y, num_filters=num_filters, activation=None, name=f"Conv2D_stack{stack}_res{res_block}_l1") if stack > 0 and res_block == 0: # first layer but not first stack # linear projection residual shortcut connection to match # changed dims x = resnet_layer(inputs=x, num_filters=num_filters, kernel_size=1, strides=strides, activation=None, batch_normalization=False, name=f"Conv2D_stack{stack}_res{res_block}_l2") x = tf.keras.layers.add([x, y]) x = Activation('relu')(x) num_filters *= 2 # Add classifier on top. # v1 does not use BN after last shortcut connection-ReLU x = AveragePooling2D(pool_size=8)(x) y = Flatten()(x) outputs = Dense(num_classes, activation='softmax', kernel_initializer='he_normal')(y) # Instantiate model. model = Model(inputs=inputs, outputs=outputs) return model def resnet_v2(input_shape, depth, num_classes=10): """ResNet Version 2 Model builder [b] Stacks of (1 x 1)-(3 x 3)-(1 x 1) BN-ReLU-Conv2D or also known as bottleneck layer First shortcut connection per layer is 1 x 1 Conv2D. Second and onwards shortcut connection is identity. At the beginning of each stage, the feature map size is halved (downsampled) by a convolutional layer with strides=2, while the number of filter maps is doubled. Within each stage, the layers have the same number filters and the same filter map sizes. Features maps sizes: conv1 : 32x32, 16 stage 0: 32x32, 64 stage 1: 16x16, 128 stage 2: 8x8, 256 # Arguments input_shape (tensor): shape of input image tensor depth (int): number of core convolutional layers num_classes (int): number of classes (CIFAR10 has 10) # Returns model (Model): Keras model instance """ if (depth - 2) % 9 != 0: raise ValueError('depth should be 9n+2 (eg 56 or 110 in [b])') # Start model definition. num_filters_in = 32 num_res_blocks = int((depth - 2) / 9) inputs = Input(shape=input_shape) # v2 performs Conv2D with BN-ReLU on input before splitting into 2 paths x = resnet_layer(inputs=inputs, num_filters=num_filters_in, conv_first=True) # Instantiate the stack of residual units for stage in range(3): for res_block in range(num_res_blocks): activation = 'relu' batch_normalization = True strides = 1 if stage == 0: num_filters_out = num_filters_in * 4 if res_block == 0: # first layer and first stage activation = None batch_normalization = False else: num_filters_out = num_filters_in * 2 if res_block == 0: # first layer but not first stage strides = 2 # downsample # bottleneck residual unit y = resnet_layer(inputs=x, num_filters=num_filters_in, kernel_size=1, strides=strides, activation=activation, batch_normalization=batch_normalization, conv_first=False) y = resnet_layer(inputs=y, num_filters=num_filters_in, conv_first=False) y = resnet_layer(inputs=y, num_filters=num_filters_out, kernel_size=1, conv_first=False) if res_block == 0: # linear projection residual shortcut connection to match # changed dims x = resnet_layer(inputs=x, num_filters=num_filters_out, kernel_size=1, strides=strides, activation=None, batch_normalization=False) x = tf.keras.layers.add([x, y]) num_filters_in = num_filters_out # Add classifier on top. # v2 has BN-ReLU before Pooling x = BatchNormalization()(x) x = Activation('relu')(x) x = AveragePooling2D(pool_size=8)(x) y = Flatten()(x) outputs = Dense(num_classes, activation='softmax', kernel_initializer='glorot_uniform')(y) # Instantiate model. model = Model(inputs=inputs, outputs=outputs) return model # # if version == 2: # model = resnet_v2(input_shape=input_shape, depth=depth) # else: # model = resnet_v1(input_shape=input_shape, depth=depth) # # model.compile(loss='categorical_crossentropy', # optimizer=Adam(learning_rate=lr_schedule(0)), # metrics=['accuracy']) # model.summary() # print(model_type) # # # Prepare model model saving directory. # save_dir = os.path.join(os.getcwd(), 'saved_models') # model_name = 'cifar10_%s_model.{epoch:03d}.h5' % model_type # if not os.path.isdir(save_dir): # os.makedirs(save_dir) # filepath = os.path.join(save_dir, model_name) # # # Prepare callbacks for model saving and for learning rate adjustment. # checkpoint = ModelCheckpoint(filepath=filepath, # monitor='val_acc', # verbose=1, # save_best_only=True) # # lr_scheduler = LearningRateScheduler(lr_schedule) # # lr_reducer = ReduceLROnPlateau(factor=np.sqrt(0.1), # cooldown=0, # patience=5, # min_lr=0.5e-6) # # callbacks = [checkpoint, lr_reducer, lr_scheduler] # # # Run training, with or without data augmentation. # if not data_augmentation: # print('Not using data augmentation.') # model.fit(x_train, y_train, # batch_size=batch_size, # epochs=epochs, # validation_data=(x_test, y_test), # shuffle=True, # callbacks=callbacks) # else: # print('Using real-time data augmentation.') # # This will do preprocessing and realtime data augmentation: # datagen = ImageDataGenerator( # # set input mean to 0 over the dataset # featurewise_center=False, # # set each sample mean to 0 # samplewise_center=False, # # divide inputs by std of dataset # featurewise_std_normalization=False, # # divide each input by its std # samplewise_std_normalization=False, # # apply ZCA whitening # zca_whitening=False, # # epsilon for ZCA whitening # zca_epsilon=1e-06, # # randomly rotate images in the range (deg 0 to 180) # rotation_range=0, # # randomly shift images horizontally # width_shift_range=0.1, # # randomly shift images vertically # height_shift_range=0.1, # # set range for random shear # shear_range=0., # # set range for random zoom # zoom_range=0., # # set range for random channel shifts # channel_shift_range=0., # # set mode for filling points outside the input boundaries # fill_mode='nearest', # # value used for fill_mode = "constant" # cval=0., # # randomly flip images # horizontal_flip=True, # # randomly flip images # vertical_flip=False, # # set rescaling factor (applied before any other transformation) # rescale=None, # # set function that will be applied on each input # preprocessing_function=None, # # image data format, either "channels_first" or "channels_last" # data_format=None, # # fraction of images reserved for validation (strictly between 0 and 1) # validation_split=0.0) # # # Compute quantities required for featurewise normalization # # (std, mean, and principal components if ZCA whitening is applied). # datagen.fit(x_train) # # # Fit the model on the batches generated by datagen.flow(). # model.fit_generator(datagen.flow(x_train, y_train, batch_size=batch_size), # validation_data=(x_test, y_test), # epochs=epochs, verbose=1, workers=4, # callbacks=callbacks)
15,635
36.317422
81
py
fl-analysis
fl-analysis-master/src/model/modelc.py
import tensorflow.keras as keras from tensorflow.keras.regularizers import l2 from tensorflow.keras import layers def build_modelc(l2_reg): do = 0.2 model = keras.Sequential() # model.add(layers.Dropout(0.2, noise_shape=(32, 32, 3))) model.add(layers.Conv2D(filters=96, kernel_size=3, strides=1, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg), input_shape=(32, 32, 3))) model.add(layers.Conv2D(filters=96, kernel_size=3, strides=1, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) model.add(layers.Conv2D(filters=96, kernel_size=3, strides=2, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) # model.add(layers.Dropout(0.5)) model.add(layers.Conv2D(filters=192, kernel_size=3, strides=1, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) model.add(layers.Conv2D(filters=192, kernel_size=3, strides=1, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) model.add(layers.Conv2D(filters=192, kernel_size=3, strides=2, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) # model.add(layers.Dropout(0.5)) model.add(layers.Conv2D(filters=192, kernel_size=3, strides=1, kernel_initializer='he_normal', padding='same', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) model.add(layers.Conv2D(filters=192, kernel_size=1, strides=1, kernel_initializer='he_normal', padding='valid', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) model.add(layers.Conv2D(filters=10, kernel_size=1, strides=1, kernel_initializer='he_normal', padding='valid', activation='relu', kernel_regularizer=l2(l2_reg), bias_regularizer=l2(l2_reg))) model.add(layers.GlobalAveragePooling2D()) model.add(layers.Dense(units=10, activation='softmax')) return model
2,230
73.366667
218
py
fl-analysis
fl-analysis-master/src/model/lenet.py
import tensorflow.keras as keras from tensorflow.keras import layers from tensorflow.keras.regularizers import l2 def build_lenet5(input_shape=(32, 32, 3), l2_reg=None): do = 0.0 regularizer = l2(l2_reg) if l2_reg is not None else None model = keras.Sequential() model.add(layers.Conv2D(filters=6, kernel_size=5, strides=1, kernel_initializer='he_normal', padding='valid', activation='relu', kernel_regularizer=regularizer, bias_regularizer=regularizer, input_shape=input_shape)) model.add(layers.MaxPooling2D()) model.add(layers.Conv2D(filters=16, kernel_size=5, strides=1, kernel_initializer='he_normal', padding='valid', activation='relu', kernel_regularizer=regularizer, bias_regularizer=regularizer)) model.add(layers.MaxPooling2D()) model.add(layers.Flatten()) model.add(layers.Dropout(do)) model.add(layers.Dense(units=120, kernel_initializer='he_normal', kernel_regularizer=regularizer, bias_regularizer=regularizer, activation='relu')) model.add(layers.Dropout(do)) model.add(layers.Dense(units=84, kernel_initializer='he_normal', kernel_regularizer=regularizer, bias_regularizer=regularizer, activation='relu')) model.add(layers.Dropout(do)) model.add(layers.Dense(units=10, activation='softmax')) return model
1,288
46.740741
220
py
fl-analysis
fl-analysis-master/src/model/test_model.py
import tensorflow.keras as keras from tensorflow.keras import layers from tensorflow.keras.regularizers import l2 def build_test_model(input_shape=(32, 32, 3), l2_reg=None): do = 0.0 regularizer = l2(l2_reg) if l2_reg is not None else None model = keras.Sequential() model.add(layers.Conv2D(filters=6, kernel_size=5, strides=1, kernel_initializer='he_normal', padding='valid', activation='relu', kernel_regularizer=regularizer, bias_regularizer=regularizer, input_shape=input_shape)) model.add(layers.BatchNormalization()) model.add(layers.MaxPooling2D()) model.add(layers.Flatten()) model.add(layers.Dense(units=120, kernel_initializer='he_normal', kernel_regularizer=regularizer, bias_regularizer=regularizer, activation='relu')) model.add(layers.Dense(units=10, activation='softmax')) return model
848
37.590909
220
py
fl-analysis
fl-analysis-master/src/model/mobilenet.py
# Implementation by https://github.com/ruchi15/CNN-MobileNetV2-Cifar10 import tensorflow as tf import os import warnings import numpy as np from tensorflow.keras.layers import Input, Activation, Conv2D, Dense, Dropout, BatchNormalization, ReLU, \ DepthwiseConv2D, GlobalAveragePooling2D, GlobalMaxPooling2D, Add from tensorflow.keras.models import Model from keras import regularizers # define the filter size def _make_divisible(v, divisor, min_value=None): if min_value is None: min_value = divisor new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) if new_v < 0.9 * v: new_v += divisor return new_v # define the calculation of each 'inverted Res_Block' def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id): prefix = 'block_{}_'.format(block_id) in_channels = inputs.shape[-1] pointwise_conv_filters = int(filters * alpha) pointwise_filters = _make_divisible(pointwise_conv_filters, 8) x = inputs # Expand if block_id: x = Conv2D(expansion * in_channels, kernel_size=1, strides=1, padding='same', use_bias=False, activation=None, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name=prefix + 'expand')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name=prefix + 'expand_BN')(x) x = ReLU(6., name=prefix + 'expand_relu')(x) else: prefix = 'expanded_conv_' # Depthwise x = DepthwiseConv2D(kernel_size=3, strides=stride, activation=None, use_bias=False, padding='same', kernel_initializer="he_normal", depthwise_regularizer=regularizers.l2(4e-5), name=prefix + 'depthwise')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name=prefix + 'depthwise_BN')(x) x = ReLU(6., name=prefix + 'depthwise_relu')(x) # Project x = Conv2D(pointwise_filters, kernel_size=1, strides=1, padding='same', use_bias=False, activation=None, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name=prefix + 'project')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name=prefix + 'project_BN')(x) if in_channels == pointwise_filters and stride == 1: return Add(name=prefix + 'add')([inputs, x]) return x # Create Build def create_model(rows, cols, channels): # encoder - input alpha = 0.5 include_top = True model_input = tf.keras.Input(shape=(rows, cols, channels), name='input_image') x = model_input first_block_filters = _make_divisible(32 * alpha, 8) # model architechture x = Conv2D(first_block_filters, kernel_size=3, strides=1, padding='same', use_bias=False, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name='Conv1')(model_input) x = _inverted_res_block(x, filters=16, alpha=alpha, stride=1, expansion=1, block_id=0) x = _inverted_res_block(x, filters=24, alpha=alpha, stride=1, expansion=6, block_id=1) x = _inverted_res_block(x, filters=24, alpha=alpha, stride=1, expansion=6, block_id=2) x = _inverted_res_block(x, filters=32, alpha=alpha, stride=2, expansion=6, block_id=3) x = _inverted_res_block(x, filters=32, alpha=alpha, stride=1, expansion=6, block_id=4) x = _inverted_res_block(x, filters=32, alpha=alpha, stride=1, expansion=6, block_id=5) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=2, expansion=6, block_id=6) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=7) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=8) x = _inverted_res_block(x, filters=64, alpha=alpha, stride=1, expansion=6, block_id=9) x = Dropout(rate=0.25)(x) x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=10) x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=11) x = _inverted_res_block(x, filters=96, alpha=alpha, stride=1, expansion=6, block_id=12) x = Dropout(rate=0.25)(x) x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=13) x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=14) x = _inverted_res_block(x, filters=160, alpha=alpha, stride=1, expansion=6, block_id=15) x = Dropout(rate=0.25)(x) x = _inverted_res_block(x, filters=320, alpha=alpha, stride=1, expansion=6, block_id=16) x = Dropout(rate=0.25)(x) # define filter size (last block) if alpha > 1.0: last_block_filters = _make_divisible(1280 * alpha, 8) else: last_block_filters = 1280 x = Conv2D(last_block_filters, kernel_size=1, use_bias=False, kernel_initializer="he_normal", kernel_regularizer=regularizers.l2(4e-5), name='Conv_1')(x) x = BatchNormalization(epsilon=1e-3, momentum=0.999, name='Conv_1_bn')(x) x = ReLU(6., name='out_relu')(x) if include_top: x = GlobalAveragePooling2D(name='global_average_pool')(x) x = Dense(10, activation='softmax', use_bias=True, name='Logits')(x) else: pass # if pooling == 'avg': # x = GlobalAveragePooling2D()(x) # elif pooling == 'max': # x = GlobalMaxPooling2D()(x) # create model of MobileNetV2 (for CIFAR-10) model = Model(inputs=model_input, outputs=x, name='mobilenetv2_cifar10') # model.compile(optimizer=tf.keras.optimizers.Adam(lr_initial), loss='sparse_categorical_crossentropy', metrics=['accuracy']) # model.compile(optimizer=tf.keras.optimizers.Adam(lr_initial), loss='sparse_categorical_crossentropy', # metrics=['accuracy']) return model def mobilenetv2_cifar10(): # model = create_model(32, 32, 3) # # model.summary() # # return model # inputs = tf.keras.Input(shape=(32, 32, 3)) # resize_layer = tf.keras.layers.Lambda( # lambda image: tf.image.resize( # image, # (224, 224), # method=tf.image.ResizeMethod.BICUBIC, # preserve_aspect_ratio=True # ) # , input_shape=(32, 32, 3))(inputs) return tf.keras.applications.mobilenet_v2.MobileNetV2( input_shape=(32, 32, 3), alpha=0.5, include_top=True, weights=None, input_tensor=None, pooling=None, classes=10 )
6,395
41.357616
129
py
fl-analysis
fl-analysis-master/src/model/stacked_lstm.py
import tensorflow as tf # class StackedLSTM(tf.keras.Model): # def __init__(self, vocab_size, embedding_dim, n_hidden): # super().__init__(self) # self.embedding = tf.keras.layers.Embedding(vocab_size, embedding_dim) # # rnn_cells = [tf.keras.layers.LSTMCell(n_hidden) for _ in range(2)] # stacked_lstm = tf.keras.layers.StackedRNNCells(rnn_cells) # self.lstm_layer = tf.keras.layers.RNN(stacked_lstm) # # self.dense = tf.keras.layers.Dense(vocab_size) # # def call(self, inputs, states=None, return_state=False, training=False): # x = inputs # x = self.embedding(x, training=training) # if states is None: # states = self.lstm_layer.get_initial_state(x) # x, states = self.lstm_layer(x, initial_state=states, training=training) # x = self.dense(x, training=training) # # if return_state: # return x, states # else: # return x # # # # def build_stacked_lstm(): # model = StackedLSTM(80, 8, 256) # model.call(tf.keras.layers.Input(shape=(80), name="test_prefix")) # # model.build(input_shape=(None, 80)) # model.summary() # return model from tensorflow.keras.layers import Embedding, Dense, LSTM from tensorflow.keras import Sequential def build_stacked_lstm(): vocab_size, embedding_dim, n_hidden = 80, 8, 256 model = Sequential() model.add(Embedding(vocab_size, embedding_dim)) # rnn_cells = [tf.keras.layers.LSTMCell(n_hidden) for _ in range(2)] # stacked_lstm = tf.keras.layers.StackedRNNCells(rnn_cells) # lstm_layer = tf.keras.layers.RNN(stacked_lstm) model.add(LSTM(n_hidden, return_sequences=True)) model.add(LSTM(n_hidden, return_sequences=False)) model.add(Dense(vocab_size, activation='softmax')) return model
1,762
31.054545
77
py
houghnet
houghnet-master/src/main.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import sys import os sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__)))) import src._init_paths import os import torch import torch.utils.data from src.lib.opts import opts from src.lib.models.model import create_model, load_model, save_model from src.lib.models.data_parallel import DataParallel from src.lib.logger import Logger from src.lib.datasets.dataset_factory import get_dataset from src.lib.trains.train_factory import train_factory def main(opt): torch.manual_seed(opt.seed) torch.backends.cudnn.benchmark = not opt.not_cuda_benchmark and not opt.test # torch.backends.cudnn.enabled = False Dataset = get_dataset(opt.dataset, opt.task) opt = opts().update_dataset_info_and_set_heads(opt, Dataset) print(opt) logger = Logger(opt) os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpus_str opt.device = torch.device('cuda' if opt.gpus[0] >= 0 else 'cpu') print('Creating model...') model = create_model(opt.arch, opt.heads, opt.head_conv, opt.region_num, opt.vote_field_size) optimizer = torch.optim.Adam(model.parameters(), opt.lr) start_epoch = 0 if opt.load_model != '': model, optimizer, start_epoch = load_model( model, opt.load_model, optimizer, opt.resume, opt.lr, opt.lr_step) Trainer = train_factory[opt.task] trainer = Trainer(opt, model, optimizer) trainer.set_device(opt.gpus, opt.chunk_sizes, opt.device) print('Setting up data...') val_loader = torch.utils.data.DataLoader( Dataset(opt, 'val'), batch_size=1, shuffle=False, num_workers=1, pin_memory=True ) if opt.test: _, preds = trainer.val(0, val_loader) val_loader.dataset.run_eval(preds, opt.save_dir) return train_loader = torch.utils.data.DataLoader( Dataset(opt, 'train'), #train batch_size=opt.batch_size, shuffle=True, num_workers=opt.num_workers, pin_memory=True, drop_last=True ) print('Starting training...') best = 1e10 for epoch in range(start_epoch + 1, opt.num_epochs + 1): mark = epoch if opt.save_all else 'last' log_dict_train, _ = trainer.train(epoch, train_loader) logger.write('epoch: {} |'.format(epoch)) for k, v in log_dict_train.items(): logger.scalar_summary('train_{}'.format(k), v, epoch) logger.write('{} {:8f} | '.format(k, v)) if opt.val_intervals > 0 and epoch % opt.val_intervals == 0: save_model(os.path.join(opt.save_dir, 'model_{}.pth'.format(mark)), epoch, model, optimizer) with torch.no_grad(): log_dict_val, preds = trainer.val(epoch, val_loader) for k, v in log_dict_val.items(): logger.scalar_summary('val_{}'.format(k), v, epoch) logger.write('{} {:8f} | '.format(k, v)) if log_dict_val[opt.metric] < best: best = log_dict_val[opt.metric] save_model(os.path.join(opt.save_dir, 'model_best.pth'), epoch, model) else: save_model(os.path.join(opt.save_dir, 'model_last.pth'), epoch, model, optimizer) logger.write('\n') if epoch in opt.lr_step: save_model(os.path.join(opt.save_dir, 'model_{}.pth'.format(epoch)), epoch, model, optimizer) lr = opt.lr * (0.1 ** (opt.lr_step.index(epoch) + 1)) print('Drop LR to', lr) for param_group in optimizer.param_groups: param_group['lr'] = lr logger.close() if __name__ == '__main__': opt = opts().parse() main(opt)
3,606
32.398148
78
py
houghnet
houghnet-master/src/test.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import sys import os sys.path.append(os.path.dirname(os.path.dirname(os.path.dirname(__file__)))) import src._init_paths import os import json import cv2 import numpy as np import time from progress.bar import Bar import torch # from src.lib.external.nms import soft_nms from src.lib.opts import opts from src.lib.logger import Logger from src.lib.utils.utils import AverageMeter from src.lib.datasets.dataset_factory import dataset_factory from src.lib.detectors.detector_factory import detector_factory class PrefetchDataset(torch.utils.data.Dataset): def __init__(self, opt, dataset, pre_process_func): self.images = dataset.images self.load_image_func = dataset.coco.loadImgs self.img_dir = dataset.img_dir self.pre_process_func = pre_process_func self.opt = opt def __getitem__(self, index): img_id = self.images[index] img_info = self.load_image_func(ids=[img_id])[0] img_path = os.path.join(self.img_dir, img_info['file_name']) image = cv2.imread(img_path) images, meta = {}, {} for scale in opt.test_scales: if opt.task == 'ddd': images[scale], meta[scale] = self.pre_process_func( image, scale, img_info['calib']) else: images[scale], meta[scale] = self.pre_process_func(image, scale) return img_id, {'images': images, 'image': image, 'meta': meta} def __len__(self): return len(self.images) def prefetch_test(opt): os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpus_str Dataset = dataset_factory[opt.dataset] opt = opts().update_dataset_info_and_set_heads(opt, Dataset) print(opt) Logger(opt) Detector = detector_factory[opt.task] split = 'val' if not opt.trainval else 'test' dataset = Dataset(opt, split) detector = Detector(opt) data_loader = torch.utils.data.DataLoader( PrefetchDataset(opt, dataset, detector.pre_process), batch_size=1, shuffle=False, num_workers=1, pin_memory=True) results = {} num_iters = len(dataset) bar = Bar('{}'.format(opt.exp_id), max=num_iters) time_stats = ['tot', 'load', 'pre', 'net', 'dec', 'post', 'merge'] avg_time_stats = {t: AverageMeter() for t in time_stats} for ind, (img_id, pre_processed_images) in enumerate(data_loader): ret = detector.run(pre_processed_images) results[img_id.numpy().astype(np.int32)[0]] = ret['results'] Bar.suffix = '[{0}/{1}]|Tot: {total:} |ETA: {eta:} '.format( ind, num_iters, total=bar.elapsed_td, eta=bar.eta_td) for t in avg_time_stats: avg_time_stats[t].update(ret[t]) Bar.suffix = Bar.suffix + '|{} {tm.val:.3f}s ({tm.avg:.3f}s) '.format( t, tm = avg_time_stats[t]) bar.next() bar.finish() for t in avg_time_stats: print('|{} {tm.val:.3f}s ({tm.avg:.3f}s) '.format(t, tm=avg_time_stats[t])) dataset.run_eval(results, opt.save_dir) def test(opt): os.environ['CUDA_VISIBLE_DEVICES'] = opt.gpus_str Dataset = dataset_factory[opt.dataset] opt = opts().update_dataset_info_and_set_heads(opt, Dataset) print(opt) Logger(opt) Detector = detector_factory[opt.task] split = 'val' if not opt.trainval else 'test' dataset = Dataset(opt, split) detector = Detector(opt) results = {} num_iters = len(dataset) bar = Bar('{}'.format(opt.exp_id), max=num_iters) time_stats = ['tot', 'load', 'pre', 'net', 'dec', 'post', 'merge'] avg_time_stats = {t: AverageMeter() for t in time_stats} for ind in range(num_iters): img_id = dataset.images[ind] img_info = dataset.coco.loadImgs(ids=[img_id])[0] img_path = os.path.join(dataset.img_dir, img_info['file_name']) if opt.task == 'ddd': ret = detector.run(img_path, img_info['calib']) else: ret = detector.run(img_path) results[img_id] = ret['results'] Bar.suffix = '[{0}/{1}]|Tot: {total:} |ETA: {eta:} '.format( ind, num_iters, total=bar.elapsed_td, eta=bar.eta_td) for t in avg_time_stats: avg_time_stats[t].update(ret[t]) Bar.suffix = Bar.suffix + '|{} {:.3f} '.format(t, avg_time_stats[t].avg) bar.next() bar.finish() dataset.run_eval(results, opt.save_dir) if __name__ == '__main__': opt = opts().parse() if opt.not_prefetch_test: test(opt) else: prefetch_test(opt)
4,351
32.476923
79
py
houghnet
houghnet-master/src/tools/convert_hourglass_weight.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function MODEL_PATH = '../../models/ExtremeNet_500000.pkl' OUT_PATH = '../../models/ExtremeNet_500000.pth' import torch state_dict = torch.load(MODEL_PATH) key_map = {'t_heats': 'hm_t', 'l_heats': 'hm_l', 'b_heats': 'hm_b', \ 'r_heats': 'hm_r', 'ct_heats': 'hm_c', \ 't_regrs': 'reg_t', 'l_regrs': 'reg_l', \ 'b_regrs': 'reg_b', 'r_regrs': 'reg_r'} out = {} for k in state_dict.keys(): changed = False for m in key_map.keys(): if m in k: if 'ct_heats' in k and m == 't_heats': continue new_k = k.replace(m, key_map[m]) out[new_k] = state_dict[k] changed = True print('replace {} to {}'.format(k, new_k)) if not changed: out[k] = state_dict[k] data = {'epoch': 0, 'state_dict': out} torch.save(data, OUT_PATH)
905
28.225806
69
py
houghnet
houghnet-master/src/tools/voc_eval_lib/model/config.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import os.path as osp import numpy as np # `pip install easydict` if you don't have it from easydict import EasyDict as edict __C = edict() # Consumers can get config by: # from fast_rcnn_config import cfg cfg = __C # # Training options # __C.TRAIN = edict() # Initial learning rate __C.TRAIN.LEARNING_RATE = 0.001 # Momentum __C.TRAIN.MOMENTUM = 0.9 # Weight decay, for regularization __C.TRAIN.WEIGHT_DECAY = 0.0001 # Factor for reducing the learning rate __C.TRAIN.GAMMA = 0.1 # Step size for reducing the learning rate, currently only support one step __C.TRAIN.STEPSIZE = [30000] # Iteration intervals for showing the loss during training, on command line interface __C.TRAIN.DISPLAY = 10 # Whether to double the learning rate for bias __C.TRAIN.DOUBLE_BIAS = True # Whether to initialize the weights with truncated normal distribution __C.TRAIN.TRUNCATED = False # Whether to have weight decay on bias as well __C.TRAIN.BIAS_DECAY = False # Whether to add ground truth boxes to the pool when sampling regions __C.TRAIN.USE_GT = False # Whether to use aspect-ratio grouping of training images, introduced merely for saving # GPU memory __C.TRAIN.ASPECT_GROUPING = False # The number of snapshots kept, older ones are deleted to save space __C.TRAIN.SNAPSHOT_KEPT = 3 # The time interval for saving tensorflow summaries __C.TRAIN.SUMMARY_INTERVAL = 180 # Scale to use during training (can list multiple scales) # The scale is the pixel size of an image's shortest side __C.TRAIN.SCALES = (600,) # Max pixel size of the longest side of a scaled input image __C.TRAIN.MAX_SIZE = 1000 # Images to use per minibatch __C.TRAIN.IMS_PER_BATCH = 1 # Minibatch size (number of regions of interest [ROIs]) __C.TRAIN.BATCH_SIZE = 128 # Fraction of minibatch that is labeled foreground (i.e. class > 0) __C.TRAIN.FG_FRACTION = 0.25 # Overlap threshold for a ROI to be considered foreground (if >= FG_THRESH) __C.TRAIN.FG_THRESH = 0.5 # Overlap threshold for a ROI to be considered background (class = 0 if # overlap in [LO, HI)) __C.TRAIN.BG_THRESH_HI = 0.5 __C.TRAIN.BG_THRESH_LO = 0.1 # Use horizontally-flipped images during training? __C.TRAIN.USE_FLIPPED = True # Train bounding-box regressors __C.TRAIN.BBOX_REG = True # Overlap required between a ROI and ground-truth box in order for that ROI to # be used as a bounding-box regression training example __C.TRAIN.BBOX_THRESH = 0.5 # Iterations between snapshots __C.TRAIN.SNAPSHOT_ITERS = 5000 # solver.prototxt specifies the snapshot path prefix, this adds an optional # infix to yield the path: <prefix>[_<infix>]_iters_XYZ.caffemodel __C.TRAIN.SNAPSHOT_PREFIX = 'res101_faster_rcnn' # Normalize the targets (subtract empirical mean, divide by empirical stddev) __C.TRAIN.BBOX_NORMALIZE_TARGETS = True # Deprecated (inside weights) __C.TRAIN.BBOX_INSIDE_WEIGHTS = (1.0, 1.0, 1.0, 1.0) # Normalize the targets using "precomputed" (or made up) means and stdevs # (BBOX_NORMALIZE_TARGETS must also be True) __C.TRAIN.BBOX_NORMALIZE_TARGETS_PRECOMPUTED = True __C.TRAIN.BBOX_NORMALIZE_MEANS = (0.0, 0.0, 0.0, 0.0) __C.TRAIN.BBOX_NORMALIZE_STDS = (0.1, 0.1, 0.2, 0.2) # Train using these proposals __C.TRAIN.PROPOSAL_METHOD = 'gt' # Make minibatches from images that have similar aspect ratios (i.e. both # tall and thin or both short and wide) in order to avoid wasting computation # on zero-padding. # Use RPN to detect objects __C.TRAIN.HAS_RPN = True # IOU >= thresh: positive example __C.TRAIN.RPN_POSITIVE_OVERLAP = 0.7 # IOU < thresh: negative example __C.TRAIN.RPN_NEGATIVE_OVERLAP = 0.3 # If an anchor satisfied by positive and negative conditions set to negative __C.TRAIN.RPN_CLOBBER_POSITIVES = False # Max number of foreground examples __C.TRAIN.RPN_FG_FRACTION = 0.5 # Total number of examples __C.TRAIN.RPN_BATCHSIZE = 256 # NMS threshold used on RPN proposals __C.TRAIN.RPN_NMS_THRESH = 0.7 # Number of top scoring boxes to keep before apply NMS to RPN proposals __C.TRAIN.RPN_PRE_NMS_TOP_N = 12000 # Number of top scoring boxes to keep after applying NMS to RPN proposals __C.TRAIN.RPN_POST_NMS_TOP_N = 2000 # Deprecated (outside weights) __C.TRAIN.RPN_BBOX_INSIDE_WEIGHTS = (1.0, 1.0, 1.0, 1.0) # Give the positive RPN examples weight of p * 1 / {num positives} # and give negatives a weight of (1 - p) # Set to -1.0 to use uniform example weighting __C.TRAIN.RPN_POSITIVE_WEIGHT = -1.0 # Whether to use all ground truth bounding boxes for training, # For COCO, setting USE_ALL_GT to False will exclude boxes that are flagged as ''iscrowd'' __C.TRAIN.USE_ALL_GT = True # # Testing options # __C.TEST = edict() # Scale to use during testing (can NOT list multiple scales) # The scale is the pixel size of an image's shortest side __C.TEST.SCALES = (600,) # Max pixel size of the longest side of a scaled input image __C.TEST.MAX_SIZE = 1000 # Overlap threshold used for non-maximum suppression (suppress boxes with # IoU >= this threshold) __C.TEST.NMS = 0.3 # Experimental: treat the (K+1) units in the cls_score layer as linear # predictors (trained, eg, with one-vs-rest SVMs). __C.TEST.SVM = False # Test using bounding-box regressors __C.TEST.BBOX_REG = True # Propose boxes __C.TEST.HAS_RPN = False # Test using these proposals __C.TEST.PROPOSAL_METHOD = 'gt' ## NMS threshold used on RPN proposals __C.TEST.RPN_NMS_THRESH = 0.7 # Number of top scoring boxes to keep before apply NMS to RPN proposals __C.TEST.RPN_PRE_NMS_TOP_N = 6000 # Number of top scoring boxes to keep after applying NMS to RPN proposals __C.TEST.RPN_POST_NMS_TOP_N = 300 # Proposal height and width both need to be greater than RPN_MIN_SIZE (at orig image scale) # __C.TEST.RPN_MIN_SIZE = 16 # Testing mode, default to be 'nms', 'top' is slower but better # See report for details __C.TEST.MODE = 'nms' # Only useful when TEST.MODE is 'top', specifies the number of top proposals to select __C.TEST.RPN_TOP_N = 5000 # # ResNet options # __C.RESNET = edict() # Option to set if max-pooling is appended after crop_and_resize. # if true, the region will be resized to a square of 2xPOOLING_SIZE, # then 2x2 max-pooling is applied; otherwise the region will be directly # resized to a square of POOLING_SIZE __C.RESNET.MAX_POOL = False # Number of fixed blocks during training, by default the first of all 4 blocks is fixed # Range: 0 (none) to 3 (all) __C.RESNET.FIXED_BLOCKS = 1 # # MobileNet options # __C.MOBILENET = edict() # Whether to regularize the depth-wise filters during training __C.MOBILENET.REGU_DEPTH = False # Number of fixed layers during training, by default the bottom 5 of 14 layers is fixed # Range: 0 (none) to 12 (all) __C.MOBILENET.FIXED_LAYERS = 5 # Weight decay for the mobilenet weights __C.MOBILENET.WEIGHT_DECAY = 0.00004 # Depth multiplier __C.MOBILENET.DEPTH_MULTIPLIER = 1. # # MISC # # Pixel mean values (BGR order) as a (1, 1, 3) array # We use the same pixel mean for all networks even though it's not exactly what # they were trained with __C.PIXEL_MEANS = np.array([[[102.9801, 115.9465, 122.7717]]]) # For reproducibility __C.RNG_SEED = 3 # Root directory of project __C.ROOT_DIR = osp.abspath(osp.join(osp.dirname(__file__), '..', '..', '..', '..')) # Data directory __C.DATA_DIR = osp.abspath(osp.join(__C.ROOT_DIR, 'data')) # Name (or path to) the matlab executable __C.MATLAB = 'matlab' # Place outputs under an experiments directory __C.EXP_DIR = 'default' # Use GPU implementation of non-maximum suppression __C.USE_GPU_NMS = True # Use an end-to-end tensorflow model. # Note: models in E2E tensorflow mode have only been tested in feed-forward mode, # but these models are exportable to other tensorflow instances as GraphDef files. __C.USE_E2E_TF = True # Default pooling mode, only 'crop' is available __C.POOLING_MODE = 'crop' # Size of the pooled region after RoI pooling __C.POOLING_SIZE = 7 # Anchor scales for RPN __C.ANCHOR_SCALES = [8,16,32] # Anchor ratios for RPN __C.ANCHOR_RATIOS = [0.5,1,2] # Number of filters for the RPN layer __C.RPN_CHANNELS = 512 def get_output_dir(imdb, weights_filename): """Return the directory where experimental artifacts are placed. If the directory does not exist, it is created. A canonical path is built using the name from an imdb and a network (if not None). """ outdir = osp.abspath(osp.join(__C.ROOT_DIR, 'output', __C.EXP_DIR, imdb.name)) if weights_filename is None: weights_filename = 'default' outdir = osp.join(outdir, weights_filename) if not os.path.exists(outdir): os.makedirs(outdir) return outdir def get_output_tb_dir(imdb, weights_filename): """Return the directory where tensorflow summaries are placed. If the directory does not exist, it is created. A canonical path is built using the name from an imdb and a network (if not None). """ outdir = osp.abspath(osp.join(__C.ROOT_DIR, 'tensorboard', __C.EXP_DIR, imdb.name)) if weights_filename is None: weights_filename = 'default' outdir = osp.join(outdir, weights_filename) if not os.path.exists(outdir): os.makedirs(outdir) return outdir def _merge_a_into_b(a, b): """Merge config dictionary a into config dictionary b, clobbering the options in b whenever they are also specified in a. """ if type(a) is not edict: return for k, v in a.items(): # a must specify keys that are in b if k not in b: raise KeyError('{} is not a valid config key'.format(k)) # the types must match, too old_type = type(b[k]) if old_type is not type(v): if isinstance(b[k], np.ndarray): v = np.array(v, dtype=b[k].dtype) else: raise ValueError(('Type mismatch ({} vs. {}) ' 'for config key: {}').format(type(b[k]), type(v), k)) # recursively merge dicts if type(v) is edict: try: _merge_a_into_b(a[k], b[k]) except: print(('Error under config key: {}'.format(k))) raise else: b[k] = v def cfg_from_file(filename): """Load a config file and merge it into the default options.""" import yaml with open(filename, 'r') as f: yaml_cfg = edict(yaml.load(f)) _merge_a_into_b(yaml_cfg, __C) def cfg_from_list(cfg_list): """Set config keys via list (e.g., from command line).""" from ast import literal_eval assert len(cfg_list) % 2 == 0 for k, v in zip(cfg_list[0::2], cfg_list[1::2]): key_list = k.split('.') d = __C for subkey in key_list[:-1]: assert subkey in d d = d[subkey] subkey = key_list[-1] assert subkey in d try: value = literal_eval(v) except: # handle the case when v is a string literal value = v assert type(value) == type(d[subkey]), \ 'type {} does not match original type {}'.format( type(value), type(d[subkey])) d[subkey] = value
11,010
27.378866
91
py
houghnet
houghnet-master/src/lib/logger.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function # Code referenced from https://gist.github.com/gyglim/1f8dfb1b5c82627ae3efcfbbadb9f514 import os import time import sys import torch USE_TENSORBOARD = True try: import tensorboardX print('Using tensorboardX') except: USE_TENSORBOARD = False class Logger(object): def __init__(self, opt): """Create a summary writer logging to log_dir.""" if not os.path.exists(opt.save_dir): os.makedirs(opt.save_dir) if not os.path.exists(opt.debug_dir): os.makedirs(opt.debug_dir) time_str = time.strftime('%Y-%m-%d-%H-%M') args = dict((name, getattr(opt, name)) for name in dir(opt) if not name.startswith('_')) file_name = os.path.join(opt.save_dir, 'opt.txt') with open(file_name, 'wt') as opt_file: opt_file.write('==> torch version: {}\n'.format(torch.__version__)) opt_file.write('==> cudnn version: {}\n'.format( torch.backends.cudnn.version())) opt_file.write('==> Cmd:\n') opt_file.write(str(sys.argv)) opt_file.write('\n==> Opt:\n') for k, v in sorted(args.items()): opt_file.write(' %s: %s\n' % (str(k), str(v))) log_dir = opt.save_dir + '/logs_{}'.format(time_str) if USE_TENSORBOARD: self.writer = tensorboardX.SummaryWriter(log_dir=log_dir) else: if not os.path.exists(os.path.dirname(log_dir)): os.mkdir(os.path.dirname(log_dir)) if not os.path.exists(log_dir): os.mkdir(log_dir) self.log = open(log_dir + '/log.txt', 'w') try: os.system('cp {}/opt.txt {}/'.format(opt.save_dir, log_dir)) except: pass self.start_line = True def write(self, txt): if self.start_line: time_str = time.strftime('%Y-%m-%d-%H-%M') self.log.write('{}: {}'.format(time_str, txt)) else: self.log.write(txt) self.start_line = False if '\n' in txt: self.start_line = True self.log.flush() def close(self): self.log.close() def scalar_summary(self, tag, value, step): """Log a scalar variable.""" if USE_TENSORBOARD: self.writer.add_scalar(tag, value, step)
2,228
29.534247
86
py
houghnet
houghnet-master/src/lib/detectors/exdet.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import src._init_paths import os import cv2 import numpy as np from progress.bar import Bar import time import torch from src.lib.external.nms import soft_nms from src.lib.models.decode import exct_decode, agnex_ct_decode from src.lib.models.utils import flip_tensor from src.lib.utils.image import get_affine_transform, transform_preds from src.lib.utils.post_process import ctdet_post_process from src.lib.utils.debugger import Debugger from .base_detector import BaseDetector class ExdetDetector(BaseDetector): def __init__(self, opt): super(ExdetDetector, self).__init__(opt) self.decode = agnex_ct_decode if opt.agnostic_ex else exct_decode def process(self, images, return_time=False): with torch.no_grad(): torch.cuda.synchronize() output = self.model(images)[-1] t_heat = output['hm_t'].sigmoid_() l_heat = output['hm_l'].sigmoid_() b_heat = output['hm_b'].sigmoid_() r_heat = output['hm_r'].sigmoid_() c_heat = output['hm_c'].sigmoid_() torch.cuda.synchronize() forward_time = time.time() if self.opt.reg_offset: dets = self.decode(t_heat, l_heat, b_heat, r_heat, c_heat, output['reg_t'], output['reg_l'], output['reg_b'], output['reg_r'], K=self.opt.K, scores_thresh=self.opt.scores_thresh, center_thresh=self.opt.center_thresh, aggr_weight=self.opt.aggr_weight) else: dets = self.decode(t_heat, l_heat, b_heat, r_heat, c_heat, K=self.opt.K, scores_thresh=self.opt.scores_thresh, center_thresh=self.opt.center_thresh, aggr_weight=self.opt.aggr_weight) if return_time: return output, dets, forward_time else: return output, dets def debug(self, debugger, images, dets, output, scale=1): detection = dets.detach().cpu().numpy().copy() detection[:, :, :4] *= self.opt.down_ratio for i in range(1): inp_height, inp_width = images.shape[2], images.shape[3] pred_hm = np.zeros((inp_height, inp_width, 3), dtype=np.uint8) img = images[i].detach().cpu().numpy().transpose(1, 2, 0) img = ((img * self.std + self.mean) * 255).astype(np.uint8) parts = ['t', 'l', 'b', 'r', 'c'] for p in parts: tag = 'hm_{}'.format(p) pred = debugger.gen_colormap( output[tag][i].detach().cpu().numpy(), (inp_height, inp_width)) if p != 'c': pred_hm = np.maximum(pred_hm, pred) else: debugger.add_blend_img( img, pred, 'pred_{}_{:.1f}'.format(p, scale)) debugger.add_blend_img(img, pred_hm, 'pred_{:.1f}'.format(scale)) debugger.add_img(img, img_id='out_{:.1f}'.format(scale)) for k in range(len(detection[i])): # print('detection', detection[i, k, 4], detection[i, k]) if detection[i, k, 4] > 0.01: # print('detection', detection[i, k, 4], detection[i, k]) debugger.add_coco_bbox(detection[i, k, :4], detection[i, k, -1], detection[i, k, 4], img_id='out_{:.1f}'.format(scale)) def post_process(self, dets, meta, scale=1): out_width, out_height = meta['out_width'], meta['out_height'] dets = dets.detach().cpu().numpy().reshape(2, -1, 14) dets[1, :, [0, 2]] = out_width - dets[1, :, [2, 0]] dets = dets.reshape(1, -1, 14) dets[0, :, 0:2] = transform_preds( dets[0, :, 0:2], meta['c'], meta['s'], (out_width, out_height)) dets[0, :, 2:4] = transform_preds( dets[0, :, 2:4], meta['c'], meta['s'], (out_width, out_height)) dets[:, :, 0:4] /= scale return dets[0] def merge_outputs(self, detections): detections = np.concatenate( [detection for detection in detections], axis=0).astype(np.float32) classes = detections[..., -1] keep_inds = (detections[:, 4] > 0) detections = detections[keep_inds] classes = classes[keep_inds] results = {} for j in range(self.num_classes): keep_inds = (classes == j) results[j + 1] = detections[keep_inds][:, 0:7].astype(np.float32) soft_nms(results[j + 1], Nt=0.5, method=2) results[j + 1] = results[j + 1][:, 0:5] scores = np.hstack([ results[j][:, -1] for j in range(1, self.num_classes + 1) ]) if len(scores) > self.max_per_image: kth = len(scores) - self.max_per_image thresh = np.partition(scores, kth)[kth] for j in range(1, self.num_classes + 1): keep_inds = (results[j][:, -1] >= thresh) results[j] = results[j][keep_inds] return results def show_results(self, debugger, image, results): debugger.add_img(image, img_id='exdet') for j in range(1, self.num_classes + 1): for bbox in results[j]: if bbox[4] > self.opt.vis_thresh: debugger.add_coco_bbox(bbox[:4], j - 1, bbox[4], img_id='exdet') debugger.show_all_imgs(pause=self.pause)
5,149
37.721805
80
py
houghnet
houghnet-master/src/lib/detectors/ctdet.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import cv2 import numpy as np from progress.bar import Bar import time import torch from src.lib.external.nms import soft_nms from src.lib.models.decode import ctdet_decode from src.lib.models.utils import flip_tensor from src.lib.utils.image import get_affine_transform from src.lib.utils.post_process import ctdet_post_process from src.lib.utils.debugger import Debugger from .base_detector import BaseDetector class CtdetDetector(BaseDetector): def __init__(self, opt): super(CtdetDetector, self).__init__(opt) def process(self, images, return_time=False): with torch.no_grad(): output = self.model(images)[-1] hm = output['hm'].sigmoid_() wh = output['wh'] reg = output['reg'] if self.opt.reg_offset else None if self.opt.flip_test: hm = (hm[0:1] + flip_tensor(hm[1:2])) / 2 wh = (wh[0:1] + flip_tensor(wh[1:2])) / 2 reg = reg[0:1] if reg is not None else None torch.cuda.synchronize() forward_time = time.time() dets = ctdet_decode(hm, wh, reg=reg, K=self.opt.K) if return_time: return output, dets, forward_time else: return output, dets def post_process(self, dets, meta, scale=1): dets = dets.detach().cpu().numpy() dets = dets.reshape(1, -1, dets.shape[2]) dets = ctdet_post_process( dets.copy(), [meta['c']], [meta['s']], meta['out_height'], meta['out_width'], self.opt.num_classes) for j in range(1, self.num_classes + 1): dets[0][j] = np.array(dets[0][j], dtype=np.float32).reshape(-1, 5) dets[0][j][:, :4] /= scale return dets[0] def merge_outputs(self, detections): results = {} for j in range(1, self.num_classes + 1): results[j] = np.concatenate( [detection[j] for detection in detections], axis=0).astype(np.float32) if len(self.scales) > 1 or self.opt.nms: soft_nms(results[j], Nt=0.5, method=2) scores = np.hstack( [results[j][:, 4] for j in range(1, self.num_classes + 1)]) if len(scores) > self.max_per_image: kth = len(scores) - self.max_per_image thresh = np.partition(scores, kth)[kth] for j in range(1, self.num_classes + 1): keep_inds = (results[j][:, 4] >= thresh) results[j] = results[j][keep_inds] return results def debug(self, debugger, images, dets, output, scale=1): detection = dets.detach().cpu().numpy().copy() detection[:, :, :4] *= self.opt.down_ratio for i in range(1): img = images[i].detach().cpu().numpy().transpose(1, 2, 0) img = ((img * self.std + self.mean) * 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][i].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm_{:.1f}'.format(scale)) debugger.add_img(img, img_id='out_pred_{:.1f}'.format(scale)) for k in range(len(dets[i])): if detection[i, k, 4] > self.opt.center_thresh: debugger.add_coco_bbox(detection[i, k, :4], detection[i, k, -1], detection[i, k, 4], img_id='out_pred_{:.1f}'.format(scale)) def show_results(self, debugger, image, results): debugger.add_img(image, img_id='ctdet') for j in range(1, self.num_classes + 1): for bbox in results[j]: if bbox[4] > self.opt.vis_thresh: debugger.add_coco_bbox(bbox[:4], j - 1, bbox[4], img_id='ctdet') debugger.show_all_imgs(pause=self.pause)
3,566
37.771739
78
py
houghnet
houghnet-master/src/lib/detectors/ddd.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import cv2 import numpy as np from progress.bar import Bar import time import torch from src.lib.external.nms import soft_nms from src.lib.models.decode import ddd_decode from src.lib.models.utils import flip_tensor from src.lib.utils.image import get_affine_transform from src.lib.utils.post_process import ddd_post_process from src.lib.utils.debugger import Debugger from src.lib.utils.ddd_utils import compute_box_3d, project_to_image, alpha2rot_y from src.lib.utils.ddd_utils import draw_box_3d, unproject_2d_to_3d from .base_detector import BaseDetector class DddDetector(BaseDetector): def __init__(self, opt): super(DddDetector, self).__init__(opt) self.calib = np.array([[707.0493, 0, 604.0814, 45.75831], [0, 707.0493, 180.5066, -0.3454157], [0, 0, 1., 0.004981016]], dtype=np.float32) def pre_process(self, image, scale, calib=None): height, width = image.shape[0:2] inp_height, inp_width = self.opt.input_h, self.opt.input_w c = np.array([width / 2, height / 2], dtype=np.float32) if self.opt.keep_res: s = np.array([inp_width, inp_height], dtype=np.int32) else: s = np.array([width, height], dtype=np.int32) trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height]) resized_image = image #cv2.resize(image, (width, height)) inp_image = cv2.warpAffine( resized_image, trans_input, (inp_width, inp_height), flags=cv2.INTER_LINEAR) inp_image = (inp_image.astype(np.float32) / 255.) inp_image = (inp_image - self.mean) / self.std images = inp_image.transpose(2, 0, 1)[np.newaxis, ...] calib = np.array(calib, dtype=np.float32) if calib is not None \ else self.calib images = torch.from_numpy(images) meta = {'c': c, 's': s, 'out_height': inp_height // self.opt.down_ratio, 'out_width': inp_width // self.opt.down_ratio, 'calib': calib} return images, meta def process(self, images, return_time=False): with torch.no_grad(): torch.cuda.synchronize() output = self.model(images)[-1] output['hm'] = output['hm'].sigmoid_() output['dep'] = 1. / (output['dep'].sigmoid() + 1e-6) - 1. wh = output['wh'] if self.opt.reg_bbox else None reg = output['reg'] if self.opt.reg_offset else None torch.cuda.synchronize() forward_time = time.time() dets = ddd_decode(output['hm'], output['rot'], output['dep'], output['dim'], wh=wh, reg=reg, K=self.opt.K) if return_time: return output, dets, forward_time else: return output, dets def post_process(self, dets, meta, scale=1): dets = dets.detach().cpu().numpy() detections = ddd_post_process( dets.copy(), [meta['c']], [meta['s']], [meta['calib']], self.opt) self.this_calib = meta['calib'] return detections[0] def merge_outputs(self, detections): results = detections[0] for j in range(1, self.num_classes + 1): if len(results[j] > 0): keep_inds = (results[j][:, -1] > self.opt.peak_thresh) results[j] = results[j][keep_inds] return results def debug(self, debugger, images, dets, output, scale=1): dets = dets.detach().cpu().numpy() img = images[0].detach().cpu().numpy().transpose(1, 2, 0) img = ((img * self.std + self.mean) * 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm') debugger.add_ct_detection( img, dets[0], show_box=self.opt.reg_bbox, center_thresh=self.opt.vis_thresh, img_id='det_pred') def show_results(self, debugger, image, results): debugger.add_3d_detection( image, results, self.this_calib, center_thresh=self.opt.vis_thresh, img_id='add_pred') debugger.add_bird_view( results, center_thresh=self.opt.vis_thresh, img_id='bird_pred') debugger.show_all_imgs(pause=self.pause)
4,110
37.783019
81
py
houghnet
houghnet-master/src/lib/detectors/multi_pose.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import cv2 import numpy as np from progress.bar import Bar import time import torch from src.lib.external.nms import soft_nms_39 from src.lib.models.decode import multi_pose_decode from src.lib.models.utils import flip_tensor, flip_lr_off, flip_lr from src.lib.utils.image import get_affine_transform from src.lib.utils.post_process import multi_pose_post_process from src.lib.utils.debugger import Debugger from .base_detector import BaseDetector class MultiPoseDetector(BaseDetector): def __init__(self, opt): super(MultiPoseDetector, self).__init__(opt) self.flip_idx = opt.flip_idx def process(self, images, return_time=False): with torch.no_grad(): torch.cuda.synchronize() output = self.model(images)[-1] output['hm'] = output['hm'].sigmoid_() if self.opt.hm_hp and not self.opt.mse_loss: output['hm_hp'] = output['hm_hp'].sigmoid_() reg = output['reg'] if self.opt.reg_offset else None hm_hp = output['hm_hp'] if self.opt.hm_hp else None hp_offset = output['hp_offset'] if self.opt.reg_hp_offset else None torch.cuda.synchronize() forward_time = time.time() if self.opt.flip_test: output['hm'] = (output['hm'][0:1] + flip_tensor(output['hm'][1:2])) / 2 output['wh'] = (output['wh'][0:1] + flip_tensor(output['wh'][1:2])) / 2 output['hps'] = (output['hps'][0:1] + flip_lr_off(output['hps'][1:2], self.flip_idx)) / 2 hm_hp = (hm_hp[0:1] + flip_lr(hm_hp[1:2], self.flip_idx)) / 2 \ if hm_hp is not None else None reg = reg[0:1] if reg is not None else None hp_offset = hp_offset[0:1] if hp_offset is not None else None dets = multi_pose_decode( output['hm'], output['wh'], output['hps'], reg=reg, hm_hp=hm_hp, hp_offset=hp_offset, K=self.opt.K) if return_time: return output, dets, forward_time else: return output, dets def post_process(self, dets, meta, scale=1): dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets = multi_pose_post_process( dets.copy(), [meta['c']], [meta['s']], meta['out_height'], meta['out_width']) for j in range(1, self.num_classes + 1): dets[0][j] = np.array(dets[0][j], dtype=np.float32).reshape(-1, 39) # import pdb; pdb.set_trace() dets[0][j][:, :4] /= scale dets[0][j][:, 5:] /= scale return dets[0] def merge_outputs(self, detections): results = {} results[1] = np.concatenate( [detection[1] for detection in detections], axis=0).astype(np.float32) if self.opt.nms or len(self.opt.test_scales) > 1: soft_nms_39(results[1], Nt=0.5, method=2) results[1] = results[1].tolist() return results def debug(self, debugger, images, dets, output, scale=1): dets = dets.detach().cpu().numpy().copy() dets[:, :, :4] *= self.opt.down_ratio dets[:, :, 5:39] *= self.opt.down_ratio img = images[0].detach().cpu().numpy().transpose(1, 2, 0) img = np.clip((( img * self.std + self.mean) * 255.), 0, 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm') if self.opt.hm_hp: pred = debugger.gen_colormap_hp( output['hm_hp'][0].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hmhp') def show_results(self, debugger, image, results): debugger.add_img(image, img_id='multi_pose') for bbox in results[1]: if bbox[4] > self.opt.vis_thresh: debugger.add_coco_bbox(bbox[:4], 0, bbox[4], img_id='multi_pose') debugger.add_coco_hp(bbox[5:39], img_id='multi_pose') debugger.show_all_imgs(pause=self.pause)
3,850
37.89899
79
py
houghnet
houghnet-master/src/lib/detectors/base_detector.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import cv2 import numpy as np from progress.bar import Bar import time import torch from src.lib.models.model import create_model, load_model from src.lib.utils.image import get_affine_transform from src.lib.utils.debugger import Debugger class BaseDetector(object): def __init__(self, opt): if opt.gpus[0] >= 0: opt.device = torch.device('cuda') else: opt.device = torch.device('cpu') print('Creating model...') self.model = create_model(opt.arch, opt.heads, opt.head_conv, opt.region_num, opt.vote_field_size, opt.model_v1) self.model = load_model(self.model, opt.load_model) self.model = self.model.to(opt.device) self.model.eval() self.mean = np.array(opt.mean, dtype=np.float32).reshape(1, 1, 3) self.std = np.array(opt.std, dtype=np.float32).reshape(1, 1, 3) self.max_per_image = 100 self.num_classes = opt.num_classes self.scales = opt.test_scales self.opt = opt self.pause = True def pre_process(self, image, scale, meta=None): height, width = image.shape[0:2] new_height = int(height * scale) new_width = int(width * scale) if self.opt.fix_res: inp_height, inp_width = self.opt.input_h, self.opt.input_w c = np.array([new_width / 2., new_height / 2.], dtype=np.float32) s = max(height, width) * 1.0 else: inp_height = (new_height | self.opt.pad) + 1 inp_width = (new_width | self.opt.pad) + 1 c = np.array([new_width // 2, new_height // 2], dtype=np.float32) s = np.array([inp_width, inp_height], dtype=np.float32) trans_input = get_affine_transform(c, s, 0, [inp_width, inp_height]) resized_image = cv2.resize(image, (new_width, new_height)) inp_image = cv2.warpAffine( resized_image, trans_input, (inp_width, inp_height), flags=cv2.INTER_LINEAR) inp_image = ((inp_image / 255. - self.mean) / self.std).astype(np.float32) images = inp_image.transpose(2, 0, 1).reshape(1, 3, inp_height, inp_width) if self.opt.flip_test: images = np.concatenate((images, images[:, :, :, ::-1]), axis=0) images = torch.from_numpy(images) meta = {'c': c, 's': s, 'out_height': inp_height // self.opt.down_ratio, 'out_width': inp_width // self.opt.down_ratio} return images, meta def process(self, images, return_time=False): raise NotImplementedError def post_process(self, dets, meta, scale=1): raise NotImplementedError def merge_outputs(self, detections): raise NotImplementedError def debug(self, debugger, images, dets, output, scale=1): raise NotImplementedError def show_results(self, debugger, image, results): raise NotImplementedError def run(self, image_or_path_or_tensor, meta=None): load_time, pre_time, net_time, dec_time, post_time = 0, 0, 0, 0, 0 merge_time, tot_time = 0, 0 debugger = Debugger(dataset=self.opt.dataset, ipynb=(self.opt.debug==3), theme=self.opt.debugger_theme) start_time = time.time() pre_processed = False if isinstance(image_or_path_or_tensor, np.ndarray): image = image_or_path_or_tensor elif type(image_or_path_or_tensor) == type (''): image = cv2.imread(image_or_path_or_tensor) else: image = image_or_path_or_tensor['image'][0].numpy() pre_processed_images = image_or_path_or_tensor pre_processed = True loaded_time = time.time() load_time += (loaded_time - start_time) img_h, img_w = image.shape[:2] detections = [] for scale in self.scales: scale_start_time = time.time() if not pre_processed: images, meta = self.pre_process(image, scale, meta) else: # import pdb; pdb.set_trace() images = pre_processed_images['images'][scale][0] meta = pre_processed_images['meta'][scale] meta = {k: v.numpy()[0] for k, v in meta.items()} meta['img_size'] = (img_h, img_w) images = images.to(self.opt.device) torch.cuda.synchronize() pre_process_time = time.time() pre_time += pre_process_time - scale_start_time output, dets, forward_time = self.process(images, return_time=True) torch.cuda.synchronize() net_time += forward_time - pre_process_time decode_time = time.time() dec_time += decode_time - forward_time if self.opt.debug >= 2: self.debug(debugger, images, dets, output, scale) dets = self.post_process(dets, meta, scale) torch.cuda.synchronize() post_process_time = time.time() post_time += post_process_time - decode_time detections.append(dets) results = self.merge_outputs(detections) torch.cuda.synchronize() end_time = time.time() merge_time += end_time - post_process_time tot_time += end_time - start_time if self.opt.debug >= 1: self.show_results(debugger, image, results) return {'results': results, 'tot': tot_time, 'load': load_time, 'pre': pre_time, 'net': net_time, 'dec': dec_time, 'post': post_time, 'merge': merge_time}
5,206
34.910345
116
py
houghnet
houghnet-master/src/lib/detectors/ctseg.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import cv2 import numpy as np from progress.bar import Bar import time import torch from pycocotools import mask as mask_utils try: from external.nms import soft_nms except: print('NMS not imported! If you need it,' ' do \n cd $CenterNet_ROOT/src/lib/external \n make') from models.decode import ctseg_decode from models.utils import flip_tensor from utils.image import get_affine_transform from utils.post_process import ctseg_post_process from utils.debugger import Debugger from src.lib.models.losses import SegLoss from .base_detector import BaseDetector class CtsegDetector(BaseDetector): def __init__(self, opt): super(CtsegDetector, self).__init__(opt) self.seg_model = SegLoss(opt.seg_feat_channel) def process(self, images, return_time=False): with torch.no_grad(): output = self.model(images)[-1] hm = output['hm'].sigmoid_() wh = output['wh'] shape_feat = output['shape'] saliency = output['saliency'] reg = output['reg'] if self.opt.reg_offset else None if self.opt.flip_test: hm = (hm[0:1] + flip_tensor(hm[1:2])) / 2 wh = (wh[0:1] + flip_tensor(wh[1:2])) / 2 reg = reg[0:1] if reg is not None else None saliency = (saliency[0:1] + flip_tensor(saliency[1:2])) / 2 shape_feat = shape_feat[0:1] if shape_feat is not None else None # assert not self.opt.flip_test,"not support flip_test" torch.cuda.synchronize() forward_time = time.time() dets, masks = ctseg_decode(hm, wh, shape_feat, saliency, self.seg_model, reg=reg, cat_spec_wh=self.opt.cat_spec_wh, K=self.opt.K) if return_time: return output, (dets,masks), forward_time else: return output, (dets,masks) def post_process(self, det_seg, meta, scale=1): assert scale == 1, "not support scale != 1" dets,seg = det_seg dets = dets.detach().cpu().numpy() seg = seg.detach().cpu().numpy() dets = dets.reshape(1, -1, dets.shape[2]) dets = ctseg_post_process( dets.copy(),seg.copy(), [meta['c']], [meta['s']], meta['out_height'], meta['out_width'],*meta['img_size'], self.opt.num_classes) return dets[0] def merge_outputs(self, detections): return detections[0] def show_results(self, debugger, image, results): debugger.add_img(image, img_id='ctseg') for j in range(1, self.num_classes + 1): for i in range(len(results[j]['boxs'])): bbox=results[j]['boxs'][i] mask = mask_utils.decode(results[j]['pred_mask'][i]) if bbox[4] > self.opt.vis_thresh: debugger.add_coco_bbox(bbox[:4], j - 1, bbox[4], img_id='ctseg') debugger.add_coco_seg(mask,img_id='ctseg') debugger.show_all_imgs(pause=self.pause)
3,141
38.275
127
py
houghnet
houghnet-master/src/lib/models/decode.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import torch.nn as nn from .utils import _gather_feat, _tranpose_and_gather_feat from detectron2.structures import Boxes # Each row is (x1, y1, x2, y2). from detectron2.layers import paste_masks_in_image from detectron2.utils.memory import retry_if_cuda_oom def _nms(heat, kernel=3): pad = (kernel - 1) // 2 hmax = nn.functional.max_pool2d( heat, (kernel, kernel), stride=1, padding=pad) keep = (hmax == heat).float() return heat * keep def _left_aggregate(heat): ''' heat: batchsize x channels x h x w ''' shape = heat.shape heat = heat.reshape(-1, heat.shape[3]) heat = heat.transpose(1, 0).contiguous() ret = heat.clone() for i in range(1, heat.shape[0]): inds = (heat[i] >= heat[i - 1]) ret[i] += ret[i - 1] * inds.float() return (ret - heat).transpose(1, 0).reshape(shape) def _right_aggregate(heat): ''' heat: batchsize x channels x h x w ''' shape = heat.shape heat = heat.reshape(-1, heat.shape[3]) heat = heat.transpose(1, 0).contiguous() ret = heat.clone() for i in range(heat.shape[0] - 2, -1, -1): inds = (heat[i] >= heat[i +1]) ret[i] += ret[i + 1] * inds.float() return (ret - heat).transpose(1, 0).reshape(shape) def _top_aggregate(heat): ''' heat: batchsize x channels x h x w ''' heat = heat.transpose(3, 2) shape = heat.shape heat = heat.reshape(-1, heat.shape[3]) heat = heat.transpose(1, 0).contiguous() ret = heat.clone() for i in range(1, heat.shape[0]): inds = (heat[i] >= heat[i - 1]) ret[i] += ret[i - 1] * inds.float() return (ret - heat).transpose(1, 0).reshape(shape).transpose(3, 2) def _bottom_aggregate(heat): ''' heat: batchsize x channels x h x w ''' heat = heat.transpose(3, 2) shape = heat.shape heat = heat.reshape(-1, heat.shape[3]) heat = heat.transpose(1, 0).contiguous() ret = heat.clone() for i in range(heat.shape[0] - 2, -1, -1): inds = (heat[i] >= heat[i + 1]) ret[i] += ret[i + 1] * inds.float() return (ret - heat).transpose(1, 0).reshape(shape).transpose(3, 2) def _h_aggregate(heat, aggr_weight=0.1): return aggr_weight * _left_aggregate(heat) + \ aggr_weight * _right_aggregate(heat) + heat def _v_aggregate(heat, aggr_weight=0.1): return aggr_weight * _top_aggregate(heat) + \ aggr_weight * _bottom_aggregate(heat) + heat ''' # Slow for large number of categories def _topk(scores, K=40): batch, cat, height, width = scores.size() topk_scores, topk_inds = torch.topk(scores.view(batch, -1), K) topk_clses = (topk_inds / (height * width)).int() topk_inds = topk_inds % (height * width) topk_ys = (topk_inds / width).int().float() topk_xs = (topk_inds % width).int().float() return topk_scores, topk_inds, topk_clses, topk_ys, topk_xs ''' def _topk_channel(scores, K=40): batch, cat, height, width = scores.size() topk_scores, topk_inds = torch.topk(scores.view(batch, cat, -1), K) topk_inds = topk_inds % (height * width) topk_ys = (topk_inds / width).int().float() topk_xs = (topk_inds % width).int().float() return topk_scores, topk_inds, topk_ys, topk_xs def _topk(scores, K=40): batch, cat, height, width = scores.size() topk_scores, topk_inds = torch.topk(scores.view(batch, cat, -1), K) topk_inds = topk_inds % (height * width) topk_ys = (topk_inds / width).int().float() topk_xs = (topk_inds % width).int().float() topk_score, topk_ind = torch.topk(topk_scores.view(batch, -1), K) topk_clses = (topk_ind / K).int() topk_inds = _gather_feat( topk_inds.view(batch, -1, 1), topk_ind).view(batch, K) topk_ys = _gather_feat(topk_ys.view(batch, -1, 1), topk_ind).view(batch, K) topk_xs = _gather_feat(topk_xs.view(batch, -1, 1), topk_ind).view(batch, K) return topk_score, topk_inds, topk_clses, topk_ys, topk_xs def agnex_ct_decode( t_heat, l_heat, b_heat, r_heat, ct_heat, t_regr=None, l_regr=None, b_regr=None, r_regr=None, K=40, scores_thresh=0.1, center_thresh=0.1, aggr_weight=0.0, num_dets=1000 ): batch, cat, height, width = t_heat.size() ''' t_heat = torch.sigmoid(t_heat) l_heat = torch.sigmoid(l_heat) b_heat = torch.sigmoid(b_heat) r_heat = torch.sigmoid(r_heat) ct_heat = torch.sigmoid(ct_heat) ''' if aggr_weight > 0: t_heat = _h_aggregate(t_heat, aggr_weight=aggr_weight) l_heat = _v_aggregate(l_heat, aggr_weight=aggr_weight) b_heat = _h_aggregate(b_heat, aggr_weight=aggr_weight) r_heat = _v_aggregate(r_heat, aggr_weight=aggr_weight) # perform nms on heatmaps t_heat = _nms(t_heat) l_heat = _nms(l_heat) b_heat = _nms(b_heat) r_heat = _nms(r_heat) t_heat[t_heat > 1] = 1 l_heat[l_heat > 1] = 1 b_heat[b_heat > 1] = 1 r_heat[r_heat > 1] = 1 t_scores, t_inds, _, t_ys, t_xs = _topk(t_heat, K=K) l_scores, l_inds, _, l_ys, l_xs = _topk(l_heat, K=K) b_scores, b_inds, _, b_ys, b_xs = _topk(b_heat, K=K) r_scores, r_inds, _, r_ys, r_xs = _topk(r_heat, K=K) ct_heat_agn, ct_clses = torch.max(ct_heat, dim=1, keepdim=True) # import pdb; pdb.set_trace() t_ys = t_ys.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) t_xs = t_xs.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) l_ys = l_ys.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) l_xs = l_xs.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) b_ys = b_ys.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) b_xs = b_xs.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) r_ys = r_ys.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) r_xs = r_xs.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) box_ct_xs = ((l_xs + r_xs + 0.5) / 2).long() box_ct_ys = ((t_ys + b_ys + 0.5) / 2).long() ct_inds = box_ct_ys * width + box_ct_xs ct_inds = ct_inds.view(batch, -1) ct_heat_agn = ct_heat_agn.view(batch, -1, 1) ct_clses = ct_clses.view(batch, -1, 1) ct_scores = _gather_feat(ct_heat_agn, ct_inds) clses = _gather_feat(ct_clses, ct_inds) t_scores = t_scores.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) l_scores = l_scores.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) b_scores = b_scores.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) r_scores = r_scores.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) ct_scores = ct_scores.view(batch, K, K, K, K) scores = (t_scores + l_scores + b_scores + r_scores + 2 * ct_scores) / 6 # reject boxes based on classes top_inds = (t_ys > l_ys) + (t_ys > b_ys) + (t_ys > r_ys) top_inds = (top_inds > 0) left_inds = (l_xs > t_xs) + (l_xs > b_xs) + (l_xs > r_xs) left_inds = (left_inds > 0) bottom_inds = (b_ys < t_ys) + (b_ys < l_ys) + (b_ys < r_ys) bottom_inds = (bottom_inds > 0) right_inds = (r_xs < t_xs) + (r_xs < l_xs) + (r_xs < b_xs) right_inds = (right_inds > 0) sc_inds = (t_scores < scores_thresh) + (l_scores < scores_thresh) + \ (b_scores < scores_thresh) + (r_scores < scores_thresh) + \ (ct_scores < center_thresh) sc_inds = (sc_inds > 0) scores = scores - sc_inds.float() scores = scores - top_inds.float() scores = scores - left_inds.float() scores = scores - bottom_inds.float() scores = scores - right_inds.float() scores = scores.view(batch, -1) scores, inds = torch.topk(scores, num_dets) scores = scores.unsqueeze(2) if t_regr is not None and l_regr is not None \ and b_regr is not None and r_regr is not None: t_regr = _tranpose_and_gather_feat(t_regr, t_inds) t_regr = t_regr.view(batch, K, 1, 1, 1, 2) l_regr = _tranpose_and_gather_feat(l_regr, l_inds) l_regr = l_regr.view(batch, 1, K, 1, 1, 2) b_regr = _tranpose_and_gather_feat(b_regr, b_inds) b_regr = b_regr.view(batch, 1, 1, K, 1, 2) r_regr = _tranpose_and_gather_feat(r_regr, r_inds) r_regr = r_regr.view(batch, 1, 1, 1, K, 2) t_xs = t_xs + t_regr[..., 0] t_ys = t_ys + t_regr[..., 1] l_xs = l_xs + l_regr[..., 0] l_ys = l_ys + l_regr[..., 1] b_xs = b_xs + b_regr[..., 0] b_ys = b_ys + b_regr[..., 1] r_xs = r_xs + r_regr[..., 0] r_ys = r_ys + r_regr[..., 1] else: t_xs = t_xs + 0.5 t_ys = t_ys + 0.5 l_xs = l_xs + 0.5 l_ys = l_ys + 0.5 b_xs = b_xs + 0.5 b_ys = b_ys + 0.5 r_xs = r_xs + 0.5 r_ys = r_ys + 0.5 bboxes = torch.stack((l_xs, t_ys, r_xs, b_ys), dim=5) bboxes = bboxes.view(batch, -1, 4) bboxes = _gather_feat(bboxes, inds) clses = clses.contiguous().view(batch, -1, 1) clses = _gather_feat(clses, inds).float() t_xs = t_xs.contiguous().view(batch, -1, 1) t_xs = _gather_feat(t_xs, inds).float() t_ys = t_ys.contiguous().view(batch, -1, 1) t_ys = _gather_feat(t_ys, inds).float() l_xs = l_xs.contiguous().view(batch, -1, 1) l_xs = _gather_feat(l_xs, inds).float() l_ys = l_ys.contiguous().view(batch, -1, 1) l_ys = _gather_feat(l_ys, inds).float() b_xs = b_xs.contiguous().view(batch, -1, 1) b_xs = _gather_feat(b_xs, inds).float() b_ys = b_ys.contiguous().view(batch, -1, 1) b_ys = _gather_feat(b_ys, inds).float() r_xs = r_xs.contiguous().view(batch, -1, 1) r_xs = _gather_feat(r_xs, inds).float() r_ys = r_ys.contiguous().view(batch, -1, 1) r_ys = _gather_feat(r_ys, inds).float() detections = torch.cat([bboxes, scores, t_xs, t_ys, l_xs, l_ys, b_xs, b_ys, r_xs, r_ys, clses], dim=2) return detections def exct_decode( t_heat, l_heat, b_heat, r_heat, ct_heat, t_regr=None, l_regr=None, b_regr=None, r_regr=None, K=40, scores_thresh=0.1, center_thresh=0.1, aggr_weight=0.0, num_dets=1000 ): batch, cat, height, width = t_heat.size() ''' t_heat = torch.sigmoid(t_heat) l_heat = torch.sigmoid(l_heat) b_heat = torch.sigmoid(b_heat) r_heat = torch.sigmoid(r_heat) ct_heat = torch.sigmoid(ct_heat) ''' if aggr_weight > 0: t_heat = _h_aggregate(t_heat, aggr_weight=aggr_weight) l_heat = _v_aggregate(l_heat, aggr_weight=aggr_weight) b_heat = _h_aggregate(b_heat, aggr_weight=aggr_weight) r_heat = _v_aggregate(r_heat, aggr_weight=aggr_weight) # perform nms on heatmaps t_heat = _nms(t_heat) l_heat = _nms(l_heat) b_heat = _nms(b_heat) r_heat = _nms(r_heat) t_heat[t_heat > 1] = 1 l_heat[l_heat > 1] = 1 b_heat[b_heat > 1] = 1 r_heat[r_heat > 1] = 1 t_scores, t_inds, t_clses, t_ys, t_xs = _topk(t_heat, K=K) l_scores, l_inds, l_clses, l_ys, l_xs = _topk(l_heat, K=K) b_scores, b_inds, b_clses, b_ys, b_xs = _topk(b_heat, K=K) r_scores, r_inds, r_clses, r_ys, r_xs = _topk(r_heat, K=K) t_ys = t_ys.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) t_xs = t_xs.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) l_ys = l_ys.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) l_xs = l_xs.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) b_ys = b_ys.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) b_xs = b_xs.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) r_ys = r_ys.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) r_xs = r_xs.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) t_clses = t_clses.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) l_clses = l_clses.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) b_clses = b_clses.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) r_clses = r_clses.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) box_ct_xs = ((l_xs + r_xs + 0.5) / 2).long() box_ct_ys = ((t_ys + b_ys + 0.5) / 2).long() ct_inds = t_clses.long() * (height * width) + box_ct_ys * width + box_ct_xs ct_inds = ct_inds.view(batch, -1) ct_heat = ct_heat.view(batch, -1, 1) ct_scores = _gather_feat(ct_heat, ct_inds) t_scores = t_scores.view(batch, K, 1, 1, 1).expand(batch, K, K, K, K) l_scores = l_scores.view(batch, 1, K, 1, 1).expand(batch, K, K, K, K) b_scores = b_scores.view(batch, 1, 1, K, 1).expand(batch, K, K, K, K) r_scores = r_scores.view(batch, 1, 1, 1, K).expand(batch, K, K, K, K) ct_scores = ct_scores.view(batch, K, K, K, K) scores = (t_scores + l_scores + b_scores + r_scores + 2 * ct_scores) / 6 # reject boxes based on classes cls_inds = (t_clses != l_clses) + (t_clses != b_clses) + \ (t_clses != r_clses) cls_inds = (cls_inds > 0) top_inds = (t_ys > l_ys) + (t_ys > b_ys) + (t_ys > r_ys) top_inds = (top_inds > 0) left_inds = (l_xs > t_xs) + (l_xs > b_xs) + (l_xs > r_xs) left_inds = (left_inds > 0) bottom_inds = (b_ys < t_ys) + (b_ys < l_ys) + (b_ys < r_ys) bottom_inds = (bottom_inds > 0) right_inds = (r_xs < t_xs) + (r_xs < l_xs) + (r_xs < b_xs) right_inds = (right_inds > 0) sc_inds = (t_scores < scores_thresh) + (l_scores < scores_thresh) + \ (b_scores < scores_thresh) + (r_scores < scores_thresh) + \ (ct_scores < center_thresh) sc_inds = (sc_inds > 0) scores = scores - sc_inds.float() scores = scores - cls_inds.float() scores = scores - top_inds.float() scores = scores - left_inds.float() scores = scores - bottom_inds.float() scores = scores - right_inds.float() scores = scores.view(batch, -1) scores, inds = torch.topk(scores, num_dets) scores = scores.unsqueeze(2) if t_regr is not None and l_regr is not None \ and b_regr is not None and r_regr is not None: t_regr = _tranpose_and_gather_feat(t_regr, t_inds) t_regr = t_regr.view(batch, K, 1, 1, 1, 2) l_regr = _tranpose_and_gather_feat(l_regr, l_inds) l_regr = l_regr.view(batch, 1, K, 1, 1, 2) b_regr = _tranpose_and_gather_feat(b_regr, b_inds) b_regr = b_regr.view(batch, 1, 1, K, 1, 2) r_regr = _tranpose_and_gather_feat(r_regr, r_inds) r_regr = r_regr.view(batch, 1, 1, 1, K, 2) t_xs = t_xs + t_regr[..., 0] t_ys = t_ys + t_regr[..., 1] l_xs = l_xs + l_regr[..., 0] l_ys = l_ys + l_regr[..., 1] b_xs = b_xs + b_regr[..., 0] b_ys = b_ys + b_regr[..., 1] r_xs = r_xs + r_regr[..., 0] r_ys = r_ys + r_regr[..., 1] else: t_xs = t_xs + 0.5 t_ys = t_ys + 0.5 l_xs = l_xs + 0.5 l_ys = l_ys + 0.5 b_xs = b_xs + 0.5 b_ys = b_ys + 0.5 r_xs = r_xs + 0.5 r_ys = r_ys + 0.5 bboxes = torch.stack((l_xs, t_ys, r_xs, b_ys), dim=5) bboxes = bboxes.view(batch, -1, 4) bboxes = _gather_feat(bboxes, inds) clses = t_clses.contiguous().view(batch, -1, 1) clses = _gather_feat(clses, inds).float() t_xs = t_xs.contiguous().view(batch, -1, 1) t_xs = _gather_feat(t_xs, inds).float() t_ys = t_ys.contiguous().view(batch, -1, 1) t_ys = _gather_feat(t_ys, inds).float() l_xs = l_xs.contiguous().view(batch, -1, 1) l_xs = _gather_feat(l_xs, inds).float() l_ys = l_ys.contiguous().view(batch, -1, 1) l_ys = _gather_feat(l_ys, inds).float() b_xs = b_xs.contiguous().view(batch, -1, 1) b_xs = _gather_feat(b_xs, inds).float() b_ys = b_ys.contiguous().view(batch, -1, 1) b_ys = _gather_feat(b_ys, inds).float() r_xs = r_xs.contiguous().view(batch, -1, 1) r_xs = _gather_feat(r_xs, inds).float() r_ys = r_ys.contiguous().view(batch, -1, 1) r_ys = _gather_feat(r_ys, inds).float() detections = torch.cat([bboxes, scores, t_xs, t_ys, l_xs, l_ys, b_xs, b_ys, r_xs, r_ys, clses], dim=2) return detections def ddd_decode(heat, rot, depth, dim, wh=None, reg=None, K=40): batch, cat, height, width = heat.size() # heat = torch.sigmoid(heat) # perform nms on heatmaps heat = _nms(heat) scores, inds, clses, ys, xs = _topk(heat, K=K) if reg is not None: reg = _tranpose_and_gather_feat(reg, inds) reg = reg.view(batch, K, 2) xs = xs.view(batch, K, 1) + reg[:, :, 0:1] ys = ys.view(batch, K, 1) + reg[:, :, 1:2] else: xs = xs.view(batch, K, 1) + 0.5 ys = ys.view(batch, K, 1) + 0.5 rot = _tranpose_and_gather_feat(rot, inds) rot = rot.view(batch, K, 8) depth = _tranpose_and_gather_feat(depth, inds) depth = depth.view(batch, K, 1) dim = _tranpose_and_gather_feat(dim, inds) dim = dim.view(batch, K, 3) clses = clses.view(batch, K, 1).float() scores = scores.view(batch, K, 1) xs = xs.view(batch, K, 1) ys = ys.view(batch, K, 1) if wh is not None: wh = _tranpose_and_gather_feat(wh, inds) wh = wh.view(batch, K, 2) detections = torch.cat( [xs, ys, scores, rot, depth, dim, wh, clses], dim=2) else: detections = torch.cat( [xs, ys, scores, rot, depth, dim, clses], dim=2) return detections def ctdet_decode(heat, wh, reg=None, cat_spec_wh=False, K=100): batch, cat, height, width = heat.size() # heat = torch.sigmoid(heat) # perform nms on heatmaps heat = _nms(heat) scores, inds, clses, ys, xs = _topk(heat, K=K) if reg is not None: reg = _tranpose_and_gather_feat(reg, inds) reg = reg.view(batch, K, 2) xs = xs.view(batch, K, 1) + reg[:, :, 0:1] ys = ys.view(batch, K, 1) + reg[:, :, 1:2] else: xs = xs.view(batch, K, 1) + 0.5 ys = ys.view(batch, K, 1) + 0.5 wh = _tranpose_and_gather_feat(wh, inds) if cat_spec_wh: wh = wh.view(batch, K, cat, 2) clses_ind = clses.view(batch, K, 1, 1).expand(batch, K, 1, 2).long() wh = wh.gather(2, clses_ind).view(batch, K, 2) else: wh = wh.view(batch, K, 2) clses = clses.view(batch, K, 1).float() scores = scores.view(batch, K, 1) bboxes = torch.cat([xs - wh[..., 0:1] / 2, ys - wh[..., 1:2] / 2, xs + wh[..., 0:1] / 2, ys + wh[..., 1:2] / 2], dim=2) detections = torch.cat([bboxes, scores, clses], dim=2) return detections def multi_pose_decode( heat, wh, kps, reg=None, hm_hp=None, hp_offset=None, K=100): batch, cat, height, width = heat.size() num_joints = kps.shape[1] // 2 # heat = torch.sigmoid(heat) # perform nms on heatmaps heat = _nms(heat) scores, inds, clses, ys, xs = _topk(heat, K=K) kps = _tranpose_and_gather_feat(kps, inds) kps = kps.view(batch, K, num_joints * 2) kps[..., ::2] += xs.view(batch, K, 1).expand(batch, K, num_joints) kps[..., 1::2] += ys.view(batch, K, 1).expand(batch, K, num_joints) if reg is not None: reg = _tranpose_and_gather_feat(reg, inds) reg = reg.view(batch, K, 2) xs = xs.view(batch, K, 1) + reg[:, :, 0:1] ys = ys.view(batch, K, 1) + reg[:, :, 1:2] else: xs = xs.view(batch, K, 1) + 0.5 ys = ys.view(batch, K, 1) + 0.5 wh = _tranpose_and_gather_feat(wh, inds) wh = wh.view(batch, K, 2) clses = clses.view(batch, K, 1).float() scores = scores.view(batch, K, 1) bboxes = torch.cat([xs - wh[..., 0:1] / 2, ys - wh[..., 1:2] / 2, xs + wh[..., 0:1] / 2, ys + wh[..., 1:2] / 2], dim=2) if hm_hp is not None: hm_hp = _nms(hm_hp) thresh = 0.1 kps = kps.view(batch, K, num_joints, 2).permute( 0, 2, 1, 3).contiguous() # b x J x K x 2 reg_kps = kps.unsqueeze(3).expand(batch, num_joints, K, K, 2) hm_score, hm_inds, hm_ys, hm_xs = _topk_channel(hm_hp, K=K) # b x J x K if hp_offset is not None: hp_offset = _tranpose_and_gather_feat( hp_offset, hm_inds.view(batch, -1)) hp_offset = hp_offset.view(batch, num_joints, K, 2) hm_xs = hm_xs + hp_offset[:, :, :, 0] hm_ys = hm_ys + hp_offset[:, :, :, 1] else: hm_xs = hm_xs + 0.5 hm_ys = hm_ys + 0.5 mask = (hm_score > thresh).float() hm_score = (1 - mask) * -1 + mask * hm_score hm_ys = (1 - mask) * (-10000) + mask * hm_ys hm_xs = (1 - mask) * (-10000) + mask * hm_xs hm_kps = torch.stack([hm_xs, hm_ys], dim=-1).unsqueeze( 2).expand(batch, num_joints, K, K, 2) dist = (((reg_kps - hm_kps) ** 2).sum(dim=4) ** 0.5) min_dist, min_ind = dist.min(dim=3) # b x J x K hm_score = hm_score.gather(2, min_ind).unsqueeze(-1) # b x J x K x 1 min_dist = min_dist.unsqueeze(-1) min_ind = min_ind.view(batch, num_joints, K, 1, 1).expand( batch, num_joints, K, 1, 2) hm_kps = hm_kps.gather(3, min_ind) hm_kps = hm_kps.view(batch, num_joints, K, 2) l = bboxes[:, :, 0].view(batch, 1, K, 1).expand(batch, num_joints, K, 1) t = bboxes[:, :, 1].view(batch, 1, K, 1).expand(batch, num_joints, K, 1) r = bboxes[:, :, 2].view(batch, 1, K, 1).expand(batch, num_joints, K, 1) b = bboxes[:, :, 3].view(batch, 1, K, 1).expand(batch, num_joints, K, 1) mask = (hm_kps[..., 0:1] < l) + (hm_kps[..., 0:1] > r) + \ (hm_kps[..., 1:2] < t) + (hm_kps[..., 1:2] > b) + \ (hm_score < thresh) + (min_dist > (torch.max(b - t, r - l) * 0.3)) mask = (mask > 0).float().expand(batch, num_joints, K, 2) kps = (1 - mask) * hm_kps + mask * kps kps = kps.permute(0, 2, 1, 3).contiguous().view( batch, K, num_joints * 2) detections = torch.cat([bboxes, scores, kps, clses], dim=2) return detections def ctseg_decode(heat, wh, shape_feat, saliency, seg_model, reg=None, cat_spec_wh=False, K=100): batch, cat, height, width = heat.size() # heat = torch.sigmoid(heat) # perform nms on heatmaps heat = _nms(heat) scores, inds, clses, ys, xs = _topk(heat, K=K) selected = scores > 0.05 if selected.sum() == 0: selected[0, 0] = 1 scores = scores[selected].unsqueeze(dim=0) inds = inds[selected].unsqueeze(dim=0) clses = clses[selected].unsqueeze(dim=0) ys = ys[selected].unsqueeze(dim=0) xs = xs[selected].unsqueeze(dim=0) K_ = scores.shape[1] if reg is not None: reg = _tranpose_and_gather_feat(reg, inds) reg = reg.view(batch, K_, 2) xs = xs.view(batch, K_, 1) + reg[:, :, 0:1] ys = ys.view(batch, K_, 1) + reg[:, :, 1:2] else: xs = xs.view(batch, K_, 1) + 0.5 ys = ys.view(batch, K_, 1) + 0.5 wh = _tranpose_and_gather_feat(wh, inds) if cat_spec_wh: wh = wh.view(batch, K_, cat, 2) clses_ind = clses.view(batch, K_, 1, 1).expand(batch, K_, 1, 2).long() wh = wh.gather(2, clses_ind).view(batch, K_, 2) else: wh = wh.view(batch, K_, 2) clses = clses.view(batch, K_, 1).float() scores = scores.view(batch, K_, 1) bboxes = torch.cat([xs - wh[..., 0:1] / 2, ys - wh[..., 1:2] / 2, xs + wh[..., 0:1] / 2, ys + wh[..., 1:2] / 2], dim=2) detections = torch.cat([bboxes, scores, clses], dim=2) h, w = shape_feat.size(-2), shape_feat.size(-1) local_shapes = _tranpose_and_gather_feat(shape_feat, inds) attns = torch.reshape(local_shapes, (1, -1, seg_model.attn_size, seg_model.attn_size)) saliency_list = [] boxes_list = [] saliency_list.append(saliency) for i in range(1): boxes_list.append(Boxes(bboxes[i, :, :] * 4.)) rois = seg_model.pooler(saliency_list, boxes_list) pred_mask_logits = seg_model.merge_bases(rois, attns) pred_mask_logits = pred_mask_logits.view( -1, 1, seg_model.pooler_resolution, seg_model.pooler_resolution) boxes_list[0].scale(0.25, 0.25) pred_masks = retry_if_cuda_oom(paste_masks_in_image)( pred_mask_logits[:, 0, :, :], # N, 1, M, M boxes_list[0], (h, w), threshold=0.5, ) pred_masks = torch.unsqueeze(pred_masks, dim=0) return detections, pred_masks
24,358
36.824534
96
py
houghnet
houghnet-master/src/lib/models/losses.py
# ------------------------------------------------------------------------------ # Portions of this code are from # CornerNet (https://github.com/princeton-vl/CornerNet) # Copyright (c) 2018, University of Michigan # Licensed under the BSD 3-Clause License # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import torch.nn as nn from .utils import _tranpose_and_gather_feat import torch.nn.functional as F from detectron2.structures import Boxes from detectron2.modeling.poolers import ROIPooler from detectron2.layers import ROIAlign, ROIAlignRotated, cat def _slow_neg_loss(pred, gt): '''focal loss from CornerNet''' pos_inds = gt.eq(1) neg_inds = gt.lt(1) neg_weights = torch.pow(1 - gt[neg_inds], 4) loss = 0 pos_pred = pred[pos_inds] neg_pred = pred[neg_inds] pos_loss = torch.log(pos_pred) * torch.pow(1 - pos_pred, 2) neg_loss = torch.log(1 - neg_pred) * torch.pow(neg_pred, 2) * neg_weights num_pos = pos_inds.float().sum() pos_loss = pos_loss.sum() neg_loss = neg_loss.sum() if pos_pred.nelement() == 0: loss = loss - neg_loss else: loss = loss - (pos_loss + neg_loss) / num_pos return loss def _neg_loss(pred, gt): ''' Modified focal loss. Exactly the same as CornerNet. Runs faster and costs a little bit more memory Arguments: pred (batch x c x h x w) gt_regr (batch x c x h x w) ''' pos_inds = gt.eq(1).float() neg_inds = gt.lt(1).float() neg_weights = torch.pow(1 - gt, 4) loss = 0 pos_loss = torch.log(pred) * torch.pow(1 - pred, 2) * pos_inds neg_loss = torch.log(1 - pred) * torch.pow(pred, 2) * neg_weights * neg_inds num_pos = pos_inds.float().sum() pos_loss = pos_loss.sum() neg_loss = neg_loss.sum() if num_pos == 0: loss = loss - neg_loss else: loss = loss - (pos_loss + neg_loss) / num_pos return loss def _not_faster_neg_loss(pred, gt): pos_inds = gt.eq(1).float() neg_inds = gt.lt(1).float() num_pos = pos_inds.float().sum() neg_weights = torch.pow(1 - gt, 4) loss = 0 trans_pred = pred * neg_inds + (1 - pred) * pos_inds weight = neg_weights * neg_inds + pos_inds all_loss = torch.log(1 - trans_pred) * torch.pow(trans_pred, 2) * weight all_loss = all_loss.sum() if num_pos > 0: all_loss /= num_pos loss -= all_loss return loss def _slow_reg_loss(regr, gt_regr, mask): num = mask.float().sum() mask = mask.unsqueeze(2).expand_as(gt_regr) regr = regr[mask] gt_regr = gt_regr[mask] regr_loss = nn.functional.smooth_l1_loss(regr, gt_regr, size_average=False) regr_loss = regr_loss / (num + 1e-4) return regr_loss def _reg_loss(regr, gt_regr, mask): ''' L1 regression loss Arguments: regr (batch x max_objects x dim) gt_regr (batch x max_objects x dim) mask (batch x max_objects) ''' num = mask.float().sum() mask = mask.unsqueeze(2).expand_as(gt_regr).float() regr = regr * mask gt_regr = gt_regr * mask regr_loss = nn.functional.smooth_l1_loss(regr, gt_regr, size_average=False) regr_loss = regr_loss / (num + 1e-4) return regr_loss class FocalLoss(nn.Module): '''nn.Module warpper for focal loss''' def __init__(self): super(FocalLoss, self).__init__() self.neg_loss = _neg_loss def forward(self, out, target): return self.neg_loss(out, target) class RegLoss(nn.Module): '''Regression loss for an output tensor Arguments: output (batch x dim x h x w) mask (batch x max_objects) ind (batch x max_objects) target (batch x max_objects x dim) ''' def __init__(self): super(RegLoss, self).__init__() def forward(self, output, mask, ind, target): pred = _tranpose_and_gather_feat(output, ind) loss = _reg_loss(pred, target, mask) return loss class RegL1Loss(nn.Module): def __init__(self): super(RegL1Loss, self).__init__() def forward(self, output, mask, ind, target): pred = _tranpose_and_gather_feat(output, ind) mask = mask.unsqueeze(2).expand_as(pred).float() # loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean') loss = F.l1_loss(pred * mask, target * mask, size_average=False) loss = loss / (mask.sum() + 1e-4) return loss class NormRegL1Loss(nn.Module): def __init__(self): super(NormRegL1Loss, self).__init__() def forward(self, output, mask, ind, target): pred = _tranpose_and_gather_feat(output, ind) mask = mask.unsqueeze(2).expand_as(pred).float() # loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean') pred = pred / (target + 1e-4) target = target * 0 + 1 loss = F.l1_loss(pred * mask, target * mask, size_average=False) loss = loss / (mask.sum() + 1e-4) return loss class RegWeightedL1Loss(nn.Module): def __init__(self): super(RegWeightedL1Loss, self).__init__() def forward(self, output, mask, ind, target): pred = _tranpose_and_gather_feat(output, ind) mask = mask.float() # loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean') loss = F.l1_loss(pred * mask, target * mask, size_average=False) loss = loss / (mask.sum() + 1e-4) return loss class L1Loss(nn.Module): def __init__(self): super(L1Loss, self).__init__() def forward(self, output, mask, ind, target): pred = _tranpose_and_gather_feat(output, ind) mask = mask.unsqueeze(2).expand_as(pred).float() loss = F.l1_loss(pred * mask, target * mask, reduction='elementwise_mean') return loss class BinRotLoss(nn.Module): def __init__(self): super(BinRotLoss, self).__init__() def forward(self, output, mask, ind, rotbin, rotres): pred = _tranpose_and_gather_feat(output, ind) loss = compute_rot_loss(pred, rotbin, rotres, mask) return loss def compute_res_loss(output, target): return F.smooth_l1_loss(output, target, reduction='elementwise_mean') # TODO: weight def compute_bin_loss(output, target, mask): mask = mask.expand_as(output) output = output * mask.float() return F.cross_entropy(output, target, reduction='elementwise_mean') def compute_rot_loss(output, target_bin, target_res, mask): # output: (B, 128, 8) [bin1_cls[0], bin1_cls[1], bin1_sin, bin1_cos, # bin2_cls[0], bin2_cls[1], bin2_sin, bin2_cos] # target_bin: (B, 128, 2) [bin1_cls, bin2_cls] # target_res: (B, 128, 2) [bin1_res, bin2_res] # mask: (B, 128, 1) # import pdb; pdb.set_trace() output = output.view(-1, 8) target_bin = target_bin.view(-1, 2) target_res = target_res.view(-1, 2) mask = mask.view(-1, 1) loss_bin1 = compute_bin_loss(output[:, 0:2], target_bin[:, 0], mask) loss_bin2 = compute_bin_loss(output[:, 4:6], target_bin[:, 1], mask) loss_res = torch.zeros_like(loss_bin1) if target_bin[:, 0].nonzero().shape[0] > 0: idx1 = target_bin[:, 0].nonzero()[:, 0] valid_output1 = torch.index_select(output, 0, idx1.long()) valid_target_res1 = torch.index_select(target_res, 0, idx1.long()) loss_sin1 = compute_res_loss( valid_output1[:, 2], torch.sin(valid_target_res1[:, 0])) loss_cos1 = compute_res_loss( valid_output1[:, 3], torch.cos(valid_target_res1[:, 0])) loss_res += loss_sin1 + loss_cos1 if target_bin[:, 1].nonzero().shape[0] > 0: idx2 = target_bin[:, 1].nonzero()[:, 0] valid_output2 = torch.index_select(output, 0, idx2.long()) valid_target_res2 = torch.index_select(target_res, 0, idx2.long()) loss_sin2 = compute_res_loss( valid_output2[:, 6], torch.sin(valid_target_res2[:, 1])) loss_cos2 = compute_res_loss( valid_output2[:, 7], torch.cos(valid_target_res2[:, 1])) loss_res += loss_sin2 + loss_cos2 return loss_bin1 + loss_bin2 + loss_res class SegLoss(nn.Module): def __init__(self,feat_channel): super(SegLoss, self).__init__() self.attn_size = 14 self.pooler_resolution = 56 self.num_gpus = 4 self.pooler = ROIPooler( output_size=self.pooler_resolution, scales=[0.25], sampling_ratio=1, pooler_type='ROIAlignV2', canonical_level=2) def forward(self, saliency, shape, gtboxes, reg_mask, ind, instance_mask, center_target=None, cat_mask=None): batch_size = saliency.size(0) local_shapes = _tranpose_and_gather_feat(shape, ind) attns = torch.reshape(local_shapes, (batch_size,-1, self.attn_size, self.attn_size )) saliency_list = [] boxes_list = [] reg_mask_list = [] saliency_list.append(saliency) for i in range(batch_size): boxes_list.append(Boxes(gtboxes[i,:,:]*4.)) reg_mask_list.append(reg_mask[i]) center_target = _tranpose_and_gather_feat(center_target, ind)[cat_mask.to(dtype=bool)] # num_obj = reg_mask.sum() reg_mask = cat(reg_mask_list, dim=0) rois = self.pooler(saliency_list, boxes_list) pred_mask_logits = self.merge_bases(rois, attns) gt_masks = [] for i, instances_per_image in enumerate(boxes_list): if len(instances_per_image.tensor) == 0: continue instances_per_image.scale(0.25, 0.25) gt_mask_per_image = self.crop_and_resize(instance_mask[i, :, :, :], instances_per_image.tensor, self.pooler_resolution ).to(device=pred_mask_logits.device) gt_masks.append(gt_mask_per_image) gt_masks = cat(gt_masks, dim=0) N = gt_masks.size(0) gt_masks = gt_masks.view(N, -1) loss_denorm = max(center_target.sum()/ self.num_gpus, 1e-6) # num_rois = pred_mask_logits.size(1) # true_mask = torch.repeat_interleave(reg_mask.unsqueeze(dim=1), # repeats=num_rois, dim=1) reg_mask = reg_mask.to(dtype=bool) mask_losses = F.binary_cross_entropy( pred_mask_logits[reg_mask], gt_masks.to(dtype=torch.float32)[reg_mask], reduction="none") mask_loss = ((mask_losses.mean(dim=-1)*center_target).sum() / loss_denorm) # mask_loss = mask_loss / num_obj return mask_loss def merge_bases(self, rois, coeffs, location_to_inds=None): # merge predictions # N = coeffs.size(0) if location_to_inds is not None: rois = rois[location_to_inds] N, B, H, W = rois.size() coeffs = coeffs.view(N, -1, self.attn_size, self.attn_size) coeffs = F.interpolate(coeffs, (H, W), mode='bilinear').sigmoid() #.softmax(dim=1) masks_preds = (rois.sigmoid() * coeffs ).sum(dim=1) return masks_preds.view(N, -1) def crop_and_resize(self, instance_mask, boxes, mask_size): """ Crop each bitmask by the given box, and resize results to (mask_size, mask_size). This can be used to prepare training targets for Mask R-CNN. It has less reconstruction error compared to rasterization with polygons. However we observe no difference in accuracy, but BitMasks requires more memory to store all the masks. Args: boxes (Tensor): Nx4 tensor storing the boxes for each mask mask_size (int): the size of the rasterized mask. Returns: Tensor: A bool tensor of shape (N, mask_size, mask_size), where N is the number of predicted boxes for this image. """ assert len(boxes) == len(instance_mask), "{} != {}".format(len(boxes), len(instance_mask)) device = instance_mask.device batch_inds = torch.arange(len(boxes), device=device).to(dtype=boxes.dtype)[:, None] rois = torch.cat([batch_inds, boxes], dim=1) # Nx5 bit_masks = instance_mask.to(dtype=torch.float32) rois = rois.to(device=device) output = ( ROIAlign((mask_size, mask_size), 1.0, 0, aligned=True) .forward(bit_masks[:, None, :, :], rois) .squeeze(1) ) output = output >= 0.5 return output
12,420
35.212828
103
py
houghnet
houghnet-master/src/lib/models/data_parallel.py
import torch from torch.nn.modules import Module from torch.nn.parallel.scatter_gather import gather from torch.nn.parallel.replicate import replicate from torch.nn.parallel.parallel_apply import parallel_apply from .scatter_gather import scatter_kwargs class _DataParallel(Module): r"""Implements data parallelism at the module level. This container parallelizes the application of the given module by splitting the input across the specified devices by chunking in the batch dimension. In the forward pass, the module is replicated on each device, and each replica handles a portion of the input. During the backwards pass, gradients from each replica are summed into the original module. The batch size should be larger than the number of GPUs used. It should also be an integer multiple of the number of GPUs so that each chunk is the same size (so that each GPU processes the same number of samples). See also: :ref:`cuda-nn-dataparallel-instead` Arbitrary positional and keyword inputs are allowed to be passed into DataParallel EXCEPT Tensors. All variables will be scattered on dim specified (default 0). Primitive types will be broadcasted, but all other types will be a shallow copy and can be corrupted if written to in the model's forward pass. Args: module: module to be parallelized device_ids: CUDA devices (default: all devices) output_device: device location of output (default: device_ids[0]) Example:: >>> net = torch.nn.DataParallel(model, device_ids=[0, 1, 2]) >>> output = net(input_var) """ # TODO: update notes/cuda.rst when this class handles 8+ GPUs well def __init__(self, module, device_ids=None, output_device=None, dim=0, chunk_sizes=None): super(_DataParallel, self).__init__() if not torch.cuda.is_available(): self.module = module self.device_ids = [] return if device_ids is None: device_ids = list(range(torch.cuda.device_count())) if output_device is None: output_device = device_ids[0] self.dim = dim self.module = module self.device_ids = device_ids self.chunk_sizes = chunk_sizes self.output_device = output_device if len(self.device_ids) == 1: self.module.cuda(device_ids[0]) def forward(self, *inputs, **kwargs): if not self.device_ids: return self.module(*inputs, **kwargs) inputs, kwargs = self.scatter(inputs, kwargs, self.device_ids, self.chunk_sizes) if len(self.device_ids) == 1: return self.module(*inputs[0], **kwargs[0]) replicas = self.replicate(self.module, self.device_ids[:len(inputs)]) outputs = self.parallel_apply(replicas, inputs, kwargs) return self.gather(outputs, self.output_device) def replicate(self, module, device_ids): return replicate(module, device_ids) def scatter(self, inputs, kwargs, device_ids, chunk_sizes): return scatter_kwargs(inputs, kwargs, device_ids, dim=self.dim, chunk_sizes=self.chunk_sizes) def parallel_apply(self, replicas, inputs, kwargs): return parallel_apply(replicas, inputs, kwargs, self.device_ids[:len(replicas)]) def gather(self, outputs, output_device): return gather(outputs, output_device, dim=self.dim) def data_parallel(module, inputs, device_ids=None, output_device=None, dim=0, module_kwargs=None): r"""Evaluates module(input) in parallel across the GPUs given in device_ids. This is the functional version of the DataParallel module. Args: module: the module to evaluate in parallel inputs: inputs to the module device_ids: GPU ids on which to replicate module output_device: GPU location of the output Use -1 to indicate the CPU. (default: device_ids[0]) Returns: a Variable containing the result of module(input) located on output_device """ if not isinstance(inputs, tuple): inputs = (inputs,) if device_ids is None: device_ids = list(range(torch.cuda.device_count())) if output_device is None: output_device = device_ids[0] inputs, module_kwargs = scatter_kwargs(inputs, module_kwargs, device_ids, dim) if len(device_ids) == 1: return module(*inputs[0], **module_kwargs[0]) used_device_ids = device_ids[:len(inputs)] replicas = replicate(module, used_device_ids) outputs = parallel_apply(replicas, inputs, module_kwargs, used_device_ids) return gather(outputs, output_device, dim) def DataParallel(module, device_ids=None, output_device=None, dim=0, chunk_sizes=None): if chunk_sizes is None: return torch.nn.DataParallel(module, device_ids, output_device, dim) standard_size = True for i in range(1, len(chunk_sizes)): if chunk_sizes[i] != chunk_sizes[0]: standard_size = False if standard_size: return torch.nn.DataParallel(module, device_ids, output_device, dim) return _DataParallel(module, device_ids, output_device, dim, chunk_sizes)
5,176
39.445313
101
py
houghnet
houghnet-master/src/lib/models/utils.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import torch.nn as nn def _sigmoid(x): y = torch.clamp(x.sigmoid_(), min=1e-4, max=1-1e-4) return y def _gather_feat(feat, ind, mask=None): dim = feat.size(2) ind = ind.unsqueeze(2).expand(ind.size(0), ind.size(1), dim) feat = feat.gather(1, ind) if mask is not None: mask = mask.unsqueeze(2).expand_as(feat) feat = feat[mask] feat = feat.view(-1, dim) return feat def _tranpose_and_gather_feat(feat, ind): feat = feat.permute(0, 2, 3, 1).contiguous() feat = feat.view(feat.size(0), -1, feat.size(3)) feat = _gather_feat(feat, ind) return feat def flip_tensor(x): return torch.flip(x, [3]) # tmp = x.detach().cpu().numpy()[..., ::-1].copy() # return torch.from_numpy(tmp).to(x.device) def flip_lr(x, flip_idx): tmp = x.detach().cpu().numpy()[..., ::-1].copy() shape = tmp.shape for e in flip_idx: tmp[:, e[0], ...], tmp[:, e[1], ...] = \ tmp[:, e[1], ...].copy(), tmp[:, e[0], ...].copy() return torch.from_numpy(tmp.reshape(shape)).to(x.device) def flip_lr_off(x, flip_idx): tmp = x.detach().cpu().numpy()[..., ::-1].copy() shape = tmp.shape tmp = tmp.reshape(tmp.shape[0], 17, 2, tmp.shape[2], tmp.shape[3]) tmp[:, :, 0, :, :] *= -1 for e in flip_idx: tmp[:, e[0], ...], tmp[:, e[1], ...] = \ tmp[:, e[1], ...].copy(), tmp[:, e[0], ...].copy() return torch.from_numpy(tmp.reshape(shape)).to(x.device)
1,570
30.42
65
py
houghnet
houghnet-master/src/lib/models/model.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import torch.nn as nn from .networks.msra_resnet import get_pose_net from .networks.resnet_dcn import get_pose_net as get_pose_net_dcn from .networks.large_hourglass import get_large_hourglass_net from .networks.houghnet_resnet import get_houghnet_net from .networks.houghnet_dcn import get_houghnet_dcn_net from .networks.houghnet_large_hourglass import get_houghnet_large_hourglass_net from .networks.pose_dla_dcn_hough import get_pose_net as get_dlahough_dcn from .networks.pose_dla_dcn import get_pose_net as get_dla_dcn _model_factory = { 'res': [get_pose_net, get_houghnet_net], 'resdcn': [get_pose_net_dcn, get_houghnet_dcn_net], 'hourglass': [get_large_hourglass_net, get_houghnet_large_hourglass_net], 'dla': [get_dla_dcn, get_dlahough_dcn] } def create_model(arch, heads, head_conv, region_num=0, vote_field_size=0, model_v1=False): num_layers = int(arch[arch.find('_') + 1:]) if '_' in arch else 0 arch = arch[:arch.find('_')] if '_' in arch else arch get_model = _model_factory[arch][1] model = get_model(num_layers=num_layers, heads=heads, head_conv=head_conv, region_num=region_num, vote_field_size=vote_field_size, model_v1=model_v1) return model def load_model(model, model_path, optimizer=None, resume=False, lr=None, lr_step=None): start_epoch = 0 checkpoint = torch.load(model_path, map_location=lambda storage, loc: storage) print('loaded {}, epoch {}'.format(model_path, checkpoint['epoch'])) state_dict_ = checkpoint['state_dict'] state_dict = {} # convert data_parallal to model for k in state_dict_: if k.startswith('module') and not k.startswith('module_list'): state_dict[k[7:]] = state_dict_[k] else: state_dict[k] = state_dict_[k] model_state_dict = model.state_dict() # check loaded parameters and created model parameters for k in state_dict: if k in model_state_dict: if state_dict[k].shape != model_state_dict[k].shape: print('Skip loading parameter {}, required shape{}, ' \ 'loaded shape{}.'.format( k, model_state_dict[k].shape, state_dict[k].shape)) state_dict[k] = model_state_dict[k] else: print('Drop parameter {}.'.format(k)) for k in model_state_dict: if not (k in state_dict): print('No param {}.'.format(k)) state_dict[k] = model_state_dict[k] model.load_state_dict(state_dict, strict=False) # resume optimizer parameters if optimizer is not None and resume: if 'optimizer' in checkpoint: optimizer.load_state_dict(checkpoint['optimizer']) start_epoch = checkpoint['epoch'] start_lr = lr for step in lr_step: if start_epoch >= step: start_lr *= 0.1 for param_group in optimizer.param_groups: param_group['lr'] = start_lr print('Resumed optimizer with start lr', start_lr) else: print('No optimizer parameters in checkpoint.') if optimizer is not None: return model, optimizer, start_epoch else: return model def save_model(path, epoch, model, optimizer=None): if isinstance(model, torch.nn.DataParallel): state_dict = model.module.state_dict() else: state_dict = model.state_dict() data = {'epoch': epoch, 'state_dict': state_dict} if not (optimizer is None): data['optimizer'] = optimizer.state_dict() torch.save(data, path)
3,752
37.295918
96
py
houghnet
houghnet-master/src/lib/models/scatter_gather.py
import torch from torch.autograd import Variable from torch.nn.parallel._functions import Scatter, Gather def scatter(inputs, target_gpus, dim=0, chunk_sizes=None): r""" Slices variables into approximately equal chunks and distributes them across given GPUs. Duplicates references to objects that are not variables. Does not support Tensors. """ def scatter_map(obj): if isinstance(obj, Variable): return Scatter.apply(target_gpus, chunk_sizes, dim, obj) assert not torch.is_tensor(obj), "Tensors not supported in scatter." if isinstance(obj, tuple): return list(zip(*map(scatter_map, obj))) if isinstance(obj, list): return list(map(list, zip(*map(scatter_map, obj)))) if isinstance(obj, dict): return list(map(type(obj), zip(*map(scatter_map, obj.items())))) return [obj for targets in target_gpus] return scatter_map(inputs) def scatter_kwargs(inputs, kwargs, target_gpus, dim=0, chunk_sizes=None): r"""Scatter with support for kwargs dictionary""" inputs = scatter(inputs, target_gpus, dim, chunk_sizes) if inputs else [] kwargs = scatter(kwargs, target_gpus, dim, chunk_sizes) if kwargs else [] if len(inputs) < len(kwargs): inputs.extend([() for _ in range(len(kwargs) - len(inputs))]) elif len(kwargs) < len(inputs): kwargs.extend([{} for _ in range(len(inputs) - len(kwargs))]) inputs = tuple(inputs) kwargs = tuple(kwargs) return inputs, kwargs
1,535
38.384615
77
py
houghnet
houghnet-master/src/lib/models/networks/resnet_dcn.py
# ------------------------------------------------------------------------------ # Copyright (c) Microsoft # Licensed under the MIT License. # Written by Bin Xiao (Bin.Xiao@microsoft.com) # Modified by Dequan Wang and Xingyi Zhou # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import math import logging import torch import torch.nn as nn from .DCNv2.dcn_v2 import DCN import torch.utils.model_zoo as model_zoo BN_MOMENTUM = 0.1 logger = logging.getLogger(__name__) model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', } def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out def fill_up_weights(up): w = up.weight.data f = math.ceil(w.size(2) / 2) c = (2 * f - 1 - f % 2) / (2. * f) for i in range(w.size(2)): for j in range(w.size(3)): w[0, 0, i, j] = \ (1 - math.fabs(i / f - c)) * (1 - math.fabs(j / f - c)) for c in range(1, w.size(0)): w[c, 0, :, :] = w[0, 0, :, :] def fill_fc_weights(layers): for m in layers.modules(): if isinstance(m, nn.Conv2d): nn.init.normal_(m.weight, std=0.001) # torch.nn.init.kaiming_normal_(m.weight.data, nonlinearity='relu') # torch.nn.init.xavier_normal_(m.weight.data) if m.bias is not None: nn.init.constant_(m.bias, 0) class PoseResNet(nn.Module): def __init__(self, block, layers, heads, head_conv): self.inplanes = 64 self.heads = heads self.deconv_with_bias = False super(PoseResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) # used for deconv layers self.deconv_layers = self._make_deconv_layer( 3, [256, 128, 64], [4, 4, 4], ) for head in self.heads: classes = self.heads[head] if head_conv > 0: if 'hm' in head: fc = nn.Sequential( nn.Conv2d(64, head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, classes, kernel_size=4, stride=1, padding=33, dilation=22, bias=True) ) fc[-1].bias.data.fill_(-2.19) else: fc = nn.Sequential( nn.Conv2d(64, head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, classes, kernel_size=1, stride=1, padding=0, bias=True)) fill_fc_weights(fc) else: fc = nn.Conv2d(64, classes, kernel_size=1, stride=1, padding=0, bias=True) if 'hm' in head: fc.bias.data.fill_(-2.19) else: fill_fc_weights(fc) self.__setattr__(head, fc) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion, momentum=BN_MOMENTUM), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def _get_deconv_cfg(self, deconv_kernel, index): if deconv_kernel == 4: padding = 1 output_padding = 0 elif deconv_kernel == 3: padding = 1 output_padding = 1 elif deconv_kernel == 2: padding = 0 output_padding = 0 return deconv_kernel, padding, output_padding def _make_deconv_layer(self, num_layers, num_filters, num_kernels): assert num_layers == len(num_filters), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' assert num_layers == len(num_kernels), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' layers = [] for i in range(num_layers): kernel, padding, output_padding = \ self._get_deconv_cfg(num_kernels[i], i) planes = num_filters[i] fc = DCN(self.inplanes, planes, kernel_size=(3,3), stride=1, padding=1, dilation=1, deformable_groups=1) # fc = nn.Conv2d(self.inplanes, planes, # kernel_size=3, stride=1, # padding=1, dilation=1, bias=False) # fill_fc_weights(fc) up = nn.ConvTranspose2d( in_channels=planes, out_channels=planes, kernel_size=kernel, stride=2, padding=padding, output_padding=output_padding, bias=self.deconv_with_bias) fill_up_weights(up) layers.append(fc) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) layers.append(up) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) self.inplanes = planes return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.deconv_layers(x) ret = {} for head in self.heads: ret[head] = self.__getattr__(head)(x) return [ret] def init_weights(self, num_layers): if 1: url = model_urls['resnet{}'.format(num_layers)] pretrained_state_dict = model_zoo.load_url(url) print('=> loading pretrained model {}'.format(url)) self.load_state_dict(pretrained_state_dict, strict=False) print('=> init deconv weights from normal distribution') for name, m in self.deconv_layers.named_modules(): if isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) resnet_spec = {18: (BasicBlock, [2, 2, 2, 2]), 34: (BasicBlock, [3, 4, 6, 3]), 50: (Bottleneck, [3, 4, 6, 3]), 101: (Bottleneck, [3, 4, 23, 3]), 152: (Bottleneck, [3, 8, 36, 3])} def get_pose_net(num_layers, heads, head_conv=256): block_class, layers = resnet_spec[num_layers] model = PoseResNet(block_class, layers, heads, head_conv=head_conv) model.init_weights(num_layers) return model
10,495
34.221477
80
py
houghnet
houghnet-master/src/lib/models/networks/houghnet_large_hourglass.py
# ------------------------------------------------------------------------------ # This code is base on # CornerNet (https://github.com/princeton-vl/CornerNet) # Copyright (c) 2018, University of Michigan # Licensed under the BSD 3-Clause License # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import re import torch import torch.nn as nn from src.lib.models.networks.hough_module import Hough class convolution(nn.Module): def __init__(self, k, inp_dim, out_dim, stride=1, with_bn=True): super(convolution, self).__init__() pad = (k - 1) // 2 self.conv = nn.Conv2d(inp_dim, out_dim, (k, k), padding=(pad, pad), stride=(stride, stride), bias=not with_bn) self.bn = nn.BatchNorm2d(out_dim) if with_bn else nn.Sequential() self.relu = nn.ReLU(inplace=True) def forward(self, x): conv = self.conv(x) bn = self.bn(conv) relu = self.relu(bn) return relu class fully_connected(nn.Module): def __init__(self, inp_dim, out_dim, with_bn=True): super(fully_connected, self).__init__() self.with_bn = with_bn self.linear = nn.Linear(inp_dim, out_dim) if self.with_bn: self.bn = nn.BatchNorm1d(out_dim) self.relu = nn.ReLU(inplace=True) def forward(self, x): linear = self.linear(x) bn = self.bn(linear) if self.with_bn else linear relu = self.relu(bn) return relu class residual(nn.Module): def __init__(self, k, inp_dim, out_dim, stride=1, with_bn=True): super(residual, self).__init__() self.conv1 = nn.Conv2d(inp_dim, out_dim, (3, 3), padding=(1, 1), stride=(stride, stride), bias=False) self.bn1 = nn.BatchNorm2d(out_dim) self.relu1 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(out_dim, out_dim, (3, 3), padding=(1, 1), bias=False) self.bn2 = nn.BatchNorm2d(out_dim) self.skip = nn.Sequential( nn.Conv2d(inp_dim, out_dim, (1, 1), stride=(stride, stride), bias=False), nn.BatchNorm2d(out_dim) ) if stride != 1 or inp_dim != out_dim else nn.Sequential() self.relu = nn.ReLU(inplace=True) def forward(self, x): conv1 = self.conv1(x) bn1 = self.bn1(conv1) relu1 = self.relu1(bn1) conv2 = self.conv2(relu1) bn2 = self.bn2(conv2) skip = self.skip(x) return self.relu(bn2 + skip) def make_layer(k, inp_dim, out_dim, modules, layer=convolution, **kwargs): layers = [layer(k, inp_dim, out_dim, **kwargs)] for _ in range(1, modules): layers.append(layer(k, out_dim, out_dim, **kwargs)) return nn.Sequential(*layers) def make_layer_revr(k, inp_dim, out_dim, modules, layer=convolution, **kwargs): layers = [] for _ in range(modules - 1): layers.append(layer(k, inp_dim, inp_dim, **kwargs)) layers.append(layer(k, inp_dim, out_dim, **kwargs)) return nn.Sequential(*layers) class MergeUp(nn.Module): def forward(self, up1, up2): return up1 + up2 def make_merge_layer(dim): return MergeUp() # def make_pool_layer(dim): # return nn.MaxPool2d(kernel_size=2, stride=2) def make_pool_layer(dim): return nn.Sequential() def make_unpool_layer(dim): return nn.Upsample(scale_factor=2) def make_kp_layer(cnv_dim, curr_dim, out_dim): return nn.Sequential( convolution(3, cnv_dim, curr_dim, with_bn=False), nn.Conv2d(curr_dim, out_dim, (1, 1)) ) def make_inter_layer(dim): return residual(3, dim, dim) def make_cnv_layer(inp_dim, out_dim): return convolution(3, inp_dim, out_dim) class kp_module(nn.Module): def __init__( self, n, dims, modules, layer=residual, make_up_layer=make_layer, make_low_layer=make_layer, make_hg_layer=make_layer, make_hg_layer_revr=make_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, **kwargs ): super(kp_module, self).__init__() self.n = n curr_mod = modules[0] next_mod = modules[1] curr_dim = dims[0] next_dim = dims[1] self.up1 = make_up_layer( 3, curr_dim, curr_dim, curr_mod, layer=layer, **kwargs ) self.max1 = make_pool_layer(curr_dim) self.low1 = make_hg_layer( 3, curr_dim, next_dim, curr_mod, layer=layer, **kwargs ) self.low2 = kp_module( n - 1, dims[1:], modules[1:], layer=layer, make_up_layer=make_up_layer, make_low_layer=make_low_layer, make_hg_layer=make_hg_layer, make_hg_layer_revr=make_hg_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, **kwargs ) if self.n > 1 else \ make_low_layer( 3, next_dim, next_dim, next_mod, layer=layer, **kwargs ) self.low3 = make_hg_layer_revr( 3, next_dim, curr_dim, curr_mod, layer=layer, **kwargs ) self.up2 = make_unpool_layer(curr_dim) self.merge = make_merge_layer(curr_dim) def forward(self, x): up1 = self.up1(x) max1 = self.max1(x) low1 = self.low1(max1) low2 = self.low2(low1) low3 = self.low3(low2) up2 = self.up2(low3) return self.merge(up1, up2) class exkp(nn.Module): def __init__( self, region_num, vote_field_size, model_v1, n, nstack, dims, modules, heads, pre=None, cnv_dim=256, make_tl_layer=None, make_br_layer=None, make_cnv_layer=make_cnv_layer, make_heat_layer=make_kp_layer, make_tag_layer=make_kp_layer, make_regr_layer=make_kp_layer, make_up_layer=make_layer, make_low_layer=make_layer, make_hg_layer=make_layer, make_hg_layer_revr=make_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, make_inter_layer=make_inter_layer, kp_layer=residual ): super(exkp, self).__init__() self.nstack = nstack self.heads = heads self.region_num = region_num self.vote_field_size = vote_field_size self.deconv_filter_padding = int(self.vote_field_size / 2) curr_dim = dims[0] self.pre = nn.Sequential( convolution(7, 3, 128, stride=2), residual(3, 128, 256, stride=2) ) if pre is None else pre self.kps = nn.ModuleList([ kp_module( n, dims, modules, layer=kp_layer, make_up_layer=make_up_layer, make_low_layer=make_low_layer, make_hg_layer=make_hg_layer, make_hg_layer_revr=make_hg_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer ) for _ in range(nstack) ]) self.cnvs = nn.ModuleList([ make_cnv_layer(curr_dim, cnv_dim) for _ in range(nstack) ]) self.inters = nn.ModuleList([ make_inter_layer(curr_dim) for _ in range(nstack - 1) ]) self.inters_ = nn.ModuleList([ nn.Sequential( nn.Conv2d(curr_dim, curr_dim, (1, 1), bias=False), nn.BatchNorm2d(curr_dim) ) for _ in range(nstack - 1) ]) self.cnvs_ = nn.ModuleList([ nn.Sequential( nn.Conv2d(cnv_dim, curr_dim, (1, 1), bias=False), nn.BatchNorm2d(curr_dim) ) for _ in range(nstack - 1) ]) self.voting_heads = list(heads['voting_heads']) del heads['voting_heads'] voting = False ## keypoint heatmaps for head in heads.keys(): if 'hm' in head: module = nn.ModuleList([ make_heat_layer( cnv_dim, curr_dim, heads[head]) for _ in range(nstack) ]) self.__setattr__(head, module) for heat in self.__getattr__(head): heat[-1].bias.data.fill_(-2.19) for voting_head in self.voting_heads: if re.fullmatch(head, voting_head): voting = True if voting: for heat in self.__getattr__(head): heat[-1].bias.data.fill_(0) heat[-1].weight.data.fill_(0) out_classes = int(heads[head] / self.region_num) hough_voting = Hough(region_num=self.region_num, vote_field_size=self.vote_field_size, num_classes=out_classes, model_v1=model_v1) self.__setattr__('voting_' + head, hough_voting) voting = False else: module = nn.ModuleList([ make_regr_layer( cnv_dim, curr_dim, heads[head]) for _ in range(nstack) ]) self.__setattr__(head, module) self.relu = nn.ReLU(inplace=True) # deconv def forward(self, image): # print('image shape', image.shape) inter = self.pre(image) outs = [] for ind in range(self.nstack): kp_, cnv_ = self.kps[ind], self.cnvs[ind] kp = kp_(inter) cnv = cnv_(kp) out = {} for head in self.heads: layer = self.__getattr__(head)[ind] if head in self.voting_heads: voting_map_hm = layer(cnv) out[head] = self.__getattr__('voting_' + head)(voting_map_hm) else: y = layer(cnv) out[head] = y outs.append(out) if ind < self.nstack - 1: inter = self.inters_[ind](inter) + self.cnvs_[ind](cnv) inter = self.relu(inter) inter = self.inters[ind](inter) return outs def make_hg_layer(kernel, dim0, dim1, mod, layer=convolution, **kwargs): layers = [layer(kernel, dim0, dim1, stride=2)] layers += [layer(kernel, dim1, dim1) for _ in range(mod - 1)] return nn.Sequential(*layers) class HourglassNet(exkp): def __init__(self, heads, region_num, vote_field_size, model_v1, num_stacks=2): n = 5 dims = [256, 256, 384, 384, 384, 512] modules = [2, 2, 2, 2, 2, 4] super(HourglassNet, self).__init__( region_num, vote_field_size, model_v1, n, num_stacks, dims, modules, heads, make_tl_layer=None, make_br_layer=None, make_pool_layer=make_pool_layer, make_hg_layer=make_hg_layer, kp_layer=residual, cnv_dim=256 ) def get_houghnet_large_hourglass_net(num_layers, heads, head_conv, region_num, vote_field_size, model_v1): model = HourglassNet(heads, region_num, vote_field_size, model_v1, 2) return model
11,454
34.030581
118
py
houghnet
houghnet-master/src/lib/models/networks/houghnet_dcn.py
# ------------------------------------------------------------------------------ # Copyright (c) Microsoft # Licensed under the MIT License. # Written by Bin Xiao (Bin.Xiao@microsoft.com) # Modified by Dequan Wang and Xingyi Zhou # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import re import math import logging import torch.nn as nn from .DCNv2.dcn_v2 import DCN import torch.utils.model_zoo as model_zoo from src.lib.models.networks.hough_module import Hough BN_MOMENTUM = 0.1 logger = logging.getLogger(__name__) model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', } def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out def fill_up_weights(up): w = up.weight.data f = math.ceil(w.size(2) / 2) c = (2 * f - 1 - f % 2) / (2. * f) for i in range(w.size(2)): for j in range(w.size(3)): w[0, 0, i, j] = \ (1 - math.fabs(i / f - c)) * (1 - math.fabs(j / f - c)) for c in range(1, w.size(0)): w[c, 0, :, :] = w[0, 0, :, :] def fill_fc_weights(layers): for m in layers.modules(): if isinstance(m, nn.Conv2d): nn.init.normal_(m.weight, std=0.001) # torch.nn.init.kaiming_normal_(m.weight.data, nonlinearity='relu') # torch.nn.init.xavier_normal_(m.weight.data) if m.bias is not None: nn.init.constant_(m.bias, 0) class HoughNetDcnNet(nn.Module): def __init__(self, block, layers, heads, region_num, vote_field_size, model_v1, head_conv): self.inplanes = 64 self.deconv_with_bias = False self.region_num = region_num self.vote_field_size = vote_field_size # self.deconv_filter_padding = int(self.vote_field_size / 2) super(HoughNetDcnNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) # used for deconv layers self.deconv_layers = self._make_deconv_layer( 3, [256, 128, 64], [4, 4, 4], ) self.voting_heads = list(heads['voting_heads']) del heads['voting_heads'] voting = False self.heads = heads for head in self.heads: classes = self.heads[head] if head_conv > 0: fc = nn.Sequential( nn.Conv2d(64, head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, classes, kernel_size=1, stride=1, padding=0, bias=True)) if 'hm' in head: fc[-1].bias.data.fill_(-2.19) else: fill_fc_weights(fc) for voting_head in self.voting_heads: if re.fullmatch(head, voting_head): voting = True if voting: fc[-1].bias.data.fill_(0) fc[-1].weight.data.fill_(0) out_classes = int(classes / self.region_num) hough_voting = Hough(region_num=self.region_num, vote_field_size=self.vote_field_size, num_classes=out_classes, model_v1=model_v1) # self.hough_voting_heads.update({head:hough_voting}) self.__setattr__('voting_' + head, hough_voting) voting = False else: fc = nn.Conv2d(64, classes, kernel_size=1, stride=1, padding=0, bias=True) if 'hm' in head: fc.bias.data.fill_(-2.19) else: fill_fc_weights(fc) self.__setattr__(head, fc) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion, momentum=BN_MOMENTUM), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def _get_deconv_cfg(self, deconv_kernel, index): if deconv_kernel == 4: padding = 1 output_padding = 0 elif deconv_kernel == 3: padding = 1 output_padding = 1 elif deconv_kernel == 2: padding = 0 output_padding = 0 return deconv_kernel, padding, output_padding def _make_deconv_layer(self, num_layers, num_filters, num_kernels): assert num_layers == len(num_filters), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' assert num_layers == len(num_kernels), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' layers = [] for i in range(num_layers): kernel, padding, output_padding = \ self._get_deconv_cfg(num_kernels[i], i) planes = num_filters[i] fc = DCN(self.inplanes, planes, kernel_size=(3,3), stride=1, padding=1, dilation=1, deformable_groups=1) # fc = nn.Conv2d(self.inplanes, planes, # kernel_size=3, stride=1, # padding=1, dilation=1, bias=False) # fill_fc_weights(fc) up = nn.ConvTranspose2d( in_channels=planes, out_channels=planes, kernel_size=kernel, stride=2, padding=padding, output_padding=output_padding, bias=self.deconv_with_bias) fill_up_weights(up) layers.append(fc) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) layers.append(up) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) self.inplanes = planes return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.deconv_layers(x) ret = {} for head in self.heads: if head in self.voting_heads: voting_map_hm = self.__getattr__(head)(x) ret[head] = self.__getattr__('voting_' + head)(voting_map_hm) else: ret[head] = self.__getattr__(head)(x) return [ret] def init_weights(self, num_layers): if 1: url = model_urls['resnet{}'.format(num_layers)] pretrained_state_dict = model_zoo.load_url(url) print('=> loading pretrained model {}'.format(url)) self.load_state_dict(pretrained_state_dict, strict=False) print('=> init deconv weights from normal distribution') for name, m in self.deconv_layers.named_modules(): if isinstance(m, nn.BatchNorm2d): nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) resnet_spec = {18: (BasicBlock, [2, 2, 2, 2]), 34: (BasicBlock, [3, 4, 6, 3]), 50: (Bottleneck, [3, 4, 6, 3]), 101: (Bottleneck, [3, 4, 23, 3]), 152: (Bottleneck, [3, 8, 36, 3])} def get_houghnet_dcn_net(num_layers, heads, region_num, vote_field_size, model_v1, head_conv=256): block_class, layers = resnet_spec[num_layers] model = HoughNetDcnNet(block_class, layers, heads, region_num, vote_field_size, model_v1, head_conv=head_conv) model.init_weights(num_layers) return model
11,486
34.673913
112
py
houghnet
houghnet-master/src/lib/models/networks/pose_dla_dcn.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import math import logging import numpy as np from os.path import join import torch from torch import nn import torch.nn.functional as F import torch.utils.model_zoo as model_zoo from .DCNv2.dcn_v2 import DCN BN_MOMENTUM = 0.1 logger = logging.getLogger(__name__) def get_model_url(data='imagenet', name='dla34', hash='ba72cf86'): return join('http://dl.yf.io/dla/models', data, '{}-{}.pth'.format(name, hash)) def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): def __init__(self, inplanes, planes, stride=1, dilation=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=dilation, bias=False, dilation=dilation) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 2 def __init__(self, inplanes, planes, stride=1, dilation=1): super(Bottleneck, self).__init__() expansion = Bottleneck.expansion bottle_planes = planes // expansion self.conv1 = nn.Conv2d(inplanes, bottle_planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(bottle_planes, bottle_planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn2 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(bottle_planes, planes, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out class BottleneckX(nn.Module): expansion = 2 cardinality = 32 def __init__(self, inplanes, planes, stride=1, dilation=1): super(BottleneckX, self).__init__() cardinality = BottleneckX.cardinality # dim = int(math.floor(planes * (BottleneckV5.expansion / 64.0))) # bottle_planes = dim * cardinality bottle_planes = planes * cardinality // 32 self.conv1 = nn.Conv2d(inplanes, bottle_planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(bottle_planes, bottle_planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation, groups=cardinality) self.bn2 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(bottle_planes, planes, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out class Root(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, residual): super(Root, self).__init__() self.conv = nn.Conv2d( in_channels, out_channels, 1, stride=1, bias=False, padding=(kernel_size - 1) // 2) self.bn = nn.BatchNorm2d(out_channels, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.residual = residual def forward(self, *x): children = x x = self.conv(torch.cat(x, 1)) x = self.bn(x) if self.residual: x += children[0] x = self.relu(x) return x class Tree(nn.Module): def __init__(self, levels, block, in_channels, out_channels, stride=1, level_root=False, root_dim=0, root_kernel_size=1, dilation=1, root_residual=False): super(Tree, self).__init__() if root_dim == 0: root_dim = 2 * out_channels if level_root: root_dim += in_channels if levels == 1: self.tree1 = block(in_channels, out_channels, stride, dilation=dilation) self.tree2 = block(out_channels, out_channels, 1, dilation=dilation) else: self.tree1 = Tree(levels - 1, block, in_channels, out_channels, stride, root_dim=0, root_kernel_size=root_kernel_size, dilation=dilation, root_residual=root_residual) self.tree2 = Tree(levels - 1, block, out_channels, out_channels, root_dim=root_dim + out_channels, root_kernel_size=root_kernel_size, dilation=dilation, root_residual=root_residual) if levels == 1: self.root = Root(root_dim, out_channels, root_kernel_size, root_residual) self.level_root = level_root self.root_dim = root_dim self.downsample = None self.project = None self.levels = levels if stride > 1: self.downsample = nn.MaxPool2d(stride, stride=stride) if in_channels != out_channels: self.project = nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False), nn.BatchNorm2d(out_channels, momentum=BN_MOMENTUM) ) def forward(self, x, residual=None, children=None): children = [] if children is None else children bottom = self.downsample(x) if self.downsample else x residual = self.project(bottom) if self.project else bottom if self.level_root: children.append(bottom) x1 = self.tree1(x, residual) if self.levels == 1: x2 = self.tree2(x1) x = self.root(x2, x1, *children) else: children.append(x1) x = self.tree2(x1, children=children) return x class DLA(nn.Module): def __init__(self, levels, channels, num_classes=1000, block=BasicBlock, residual_root=False, linear_root=False): super(DLA, self).__init__() self.channels = channels self.num_classes = num_classes self.base_layer = nn.Sequential( nn.Conv2d(3, channels[0], kernel_size=7, stride=1, padding=3, bias=False), nn.BatchNorm2d(channels[0], momentum=BN_MOMENTUM), nn.ReLU(inplace=True)) self.level0 = self._make_conv_level( channels[0], channels[0], levels[0]) self.level1 = self._make_conv_level( channels[0], channels[1], levels[1], stride=2) self.level2 = Tree(levels[2], block, channels[1], channels[2], 2, level_root=False, root_residual=residual_root) self.level3 = Tree(levels[3], block, channels[2], channels[3], 2, level_root=True, root_residual=residual_root) self.level4 = Tree(levels[4], block, channels[3], channels[4], 2, level_root=True, root_residual=residual_root) self.level5 = Tree(levels[5], block, channels[4], channels[5], 2, level_root=True, root_residual=residual_root) # for m in self.modules(): # if isinstance(m, nn.Conv2d): # n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels # m.weight.data.normal_(0, math.sqrt(2. / n)) # elif isinstance(m, nn.BatchNorm2d): # m.weight.data.fill_(1) # m.bias.data.zero_() def _make_level(self, block, inplanes, planes, blocks, stride=1): downsample = None if stride != 1 or inplanes != planes: downsample = nn.Sequential( nn.MaxPool2d(stride, stride=stride), nn.Conv2d(inplanes, planes, kernel_size=1, stride=1, bias=False), nn.BatchNorm2d(planes, momentum=BN_MOMENTUM), ) layers = [] layers.append(block(inplanes, planes, stride, downsample=downsample)) for i in range(1, blocks): layers.append(block(inplanes, planes)) return nn.Sequential(*layers) def _make_conv_level(self, inplanes, planes, convs, stride=1, dilation=1): modules = [] for i in range(convs): modules.extend([ nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride if i == 0 else 1, padding=dilation, bias=False, dilation=dilation), nn.BatchNorm2d(planes, momentum=BN_MOMENTUM), nn.ReLU(inplace=True)]) inplanes = planes return nn.Sequential(*modules) def forward(self, x): y = [] x = self.base_layer(x) for i in range(6): x = getattr(self, 'level{}'.format(i))(x) y.append(x) return y def load_pretrained_model(self, data='imagenet', name='dla34', hash='ba72cf86'): # fc = self.fc if name.endswith('.pth'): model_weights = torch.load(data + name) else: model_url = get_model_url(data, name, hash) model_weights = model_zoo.load_url(model_url) num_classes = len(model_weights[list(model_weights.keys())[-1]]) self.fc = nn.Conv2d( self.channels[-1], num_classes, kernel_size=1, stride=1, padding=0, bias=True) self.load_state_dict(model_weights) # self.fc = fc def dla34(pretrained=True, **kwargs): # DLA-34 model = DLA([1, 1, 1, 2, 2, 1], [16, 32, 64, 128, 256, 512], block=BasicBlock, **kwargs) if pretrained: model.load_pretrained_model(data='imagenet', name='dla34', hash='ba72cf86') return model class Identity(nn.Module): def __init__(self): super(Identity, self).__init__() def forward(self, x): return x def fill_fc_weights(layers): for m in layers.modules(): if isinstance(m, nn.Conv2d): if m.bias is not None: nn.init.constant_(m.bias, 0) def fill_up_weights(up): w = up.weight.data f = math.ceil(w.size(2) / 2) c = (2 * f - 1 - f % 2) / (2. * f) for i in range(w.size(2)): for j in range(w.size(3)): w[0, 0, i, j] = \ (1 - math.fabs(i / f - c)) * (1 - math.fabs(j / f - c)) for c in range(1, w.size(0)): w[c, 0, :, :] = w[0, 0, :, :] class DeformConv(nn.Module): def __init__(self, chi, cho): super(DeformConv, self).__init__() self.actf = nn.Sequential( nn.BatchNorm2d(cho, momentum=BN_MOMENTUM), nn.ReLU(inplace=True) ) self.conv = DCN(chi, cho, kernel_size=(3,3), stride=1, padding=1, dilation=1, deformable_groups=1) def forward(self, x): x = self.conv(x) x = self.actf(x) return x class IDAUp(nn.Module): def __init__(self, o, channels, up_f): super(IDAUp, self).__init__() for i in range(1, len(channels)): c = channels[i] f = int(up_f[i]) proj = DeformConv(c, o) node = DeformConv(o, o) up = nn.ConvTranspose2d(o, o, f * 2, stride=f, padding=f // 2, output_padding=0, groups=o, bias=False) fill_up_weights(up) setattr(self, 'proj_' + str(i), proj) setattr(self, 'up_' + str(i), up) setattr(self, 'node_' + str(i), node) def forward(self, layers, startp, endp): for i in range(startp + 1, endp): upsample = getattr(self, 'up_' + str(i - startp)) project = getattr(self, 'proj_' + str(i - startp)) layers[i] = upsample(project(layers[i])) node = getattr(self, 'node_' + str(i - startp)) layers[i] = node(layers[i] + layers[i - 1]) class DLAUp(nn.Module): def __init__(self, startp, channels, scales, in_channels=None): super(DLAUp, self).__init__() self.startp = startp if in_channels is None: in_channels = channels self.channels = channels channels = list(channels) scales = np.array(scales, dtype=int) for i in range(len(channels) - 1): j = -i - 2 setattr(self, 'ida_{}'.format(i), IDAUp(channels[j], in_channels[j:], scales[j:] // scales[j])) scales[j + 1:] = scales[j] in_channels[j + 1:] = [channels[j] for _ in channels[j + 1:]] def forward(self, layers): out = [layers[-1]] # start with 32 for i in range(len(layers) - self.startp - 1): ida = getattr(self, 'ida_{}'.format(i)) ida(layers, len(layers) -i - 2, len(layers)) out.insert(0, layers[-1]) return out class Interpolate(nn.Module): def __init__(self, scale, mode): super(Interpolate, self).__init__() self.scale = scale self.mode = mode def forward(self, x): x = F.interpolate(x, scale_factor=self.scale, mode=self.mode, align_corners=False) return x class DLASeg(nn.Module): def __init__(self, base_name, heads, pretrained, down_ratio, final_kernel, last_level, head_conv, out_channel=0): super(DLASeg, self).__init__() assert down_ratio in [2, 4, 8, 16] self.first_level = int(np.log2(down_ratio)) self.last_level = last_level self.base = globals()[base_name](pretrained=pretrained) channels = self.base.channels scales = [2 ** i for i in range(len(channels[self.first_level:]))] self.dla_up = DLAUp(self.first_level, channels[self.first_level:], scales) if out_channel == 0: out_channel = channels[self.first_level] self.ida_up = IDAUp(out_channel, channels[self.first_level:self.last_level], [2 ** i for i in range(self.last_level - self.first_level)]) self.heads = heads for head in self.heads: classes = self.heads[head] if head_conv > 0: fc = nn.Sequential( nn.Conv2d(channels[self.first_level], head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, classes, kernel_size=final_kernel, stride=1, padding=final_kernel // 2, bias=True)) if 'hm' in head: fc[-1].bias.data.fill_(-2.19) else: fill_fc_weights(fc) else: fc = nn.Conv2d(channels[self.first_level], classes, kernel_size=final_kernel, stride=1, padding=final_kernel // 2, bias=True) if 'hm' in head: fc.bias.data.fill_(-2.19) else: fill_fc_weights(fc) self.__setattr__(head, fc) def forward(self, x): x = self.base(x) x = self.dla_up(x) y = [] for i in range(self.last_level - self.first_level): y.append(x[i].clone()) self.ida_up(y, 0, len(y)) z = {} for head in self.heads: z[head] = self.__getattr__(head)(y[-1]) return [z] def get_pose_net(num_layers, heads, head_conv=256, down_ratio=4): model = DLASeg('dla{}'.format(num_layers), heads, pretrained=True, down_ratio=down_ratio, final_kernel=1, last_level=5, head_conv=head_conv) return model
17,594
34.617409
106
py
houghnet
houghnet-master/src/lib/models/networks/msra_resnet.py
# ------------------------------------------------------------------------------ # Copyright (c) Microsoft # Licensed under the MIT License. # Written by Bin Xiao (Bin.Xiao@microsoft.com) # Modified by Xingyi Zhou # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import os import torch import torch.nn as nn import torch.utils.model_zoo as model_zoo BN_MOMENTUM = 0.1 model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', } def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class PoseResNet(nn.Module): def __init__(self, block, layers, heads, head_conv, **kwargs): self.inplanes = 64 self.deconv_with_bias = False self.heads = heads super(PoseResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) # used for deconv layers self.deconv_layers = self._make_deconv_layer( 3, [256, 256, 256], [4, 4, 4], ) # self.final_layer = [] for head in sorted(self.heads): num_output = self.heads[head] if head_conv > 0: fc = nn.Sequential( nn.Conv2d(256, head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, num_output, kernel_size=1, stride=1, padding=0)) else: fc = nn.Conv2d( in_channels=256, out_channels=num_output, kernel_size=1, stride=1, padding=0 ) self.__setattr__(head, fc) # self.final_layer = nn.ModuleList(self.final_layer) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion, momentum=BN_MOMENTUM), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def _get_deconv_cfg(self, deconv_kernel, index): if deconv_kernel == 4: padding = 1 output_padding = 0 elif deconv_kernel == 3: padding = 1 output_padding = 1 elif deconv_kernel == 2: padding = 0 output_padding = 0 return deconv_kernel, padding, output_padding def _make_deconv_layer(self, num_layers, num_filters, num_kernels): assert num_layers == len(num_filters), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' assert num_layers == len(num_kernels), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' layers = [] for i in range(num_layers): kernel, padding, output_padding = \ self._get_deconv_cfg(num_kernels[i], i) planes = num_filters[i] layers.append( nn.ConvTranspose2d( in_channels=self.inplanes, out_channels=planes, kernel_size=kernel, stride=2, padding=padding, output_padding=output_padding, bias=self.deconv_with_bias)) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) self.inplanes = planes return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.deconv_layers(x) ret = {} for head in self.heads: ret[head] = self.__getattr__(head)(x) return [ret] def init_weights(self, num_layers, pretrained=True): if pretrained: # print('=> init resnet deconv weights from normal distribution') for _, m in self.deconv_layers.named_modules(): if isinstance(m, nn.ConvTranspose2d): # print('=> init {}.weight as normal(0, 0.001)'.format(name)) # print('=> init {}.bias as 0'.format(name)) nn.init.normal_(m.weight, std=0.001) if self.deconv_with_bias: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d): # print('=> init {}.weight as 1'.format(name)) # print('=> init {}.bias as 0'.format(name)) nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) # print('=> init final conv weights from normal distribution') for head in self.heads: final_layer = self.__getattr__(head) for i, m in enumerate(final_layer.modules()): if isinstance(m, nn.Conv2d): # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') # print('=> init {}.weight as normal(0, 0.001)'.format(name)) # print('=> init {}.bias as 0'.format(name)) if m.weight.shape[0] == self.heads[head]: if 'hm' in head: nn.init.constant_(m.bias, -2.19) else: nn.init.normal_(m.weight, std=0.001) nn.init.constant_(m.bias, 0) #pretrained_state_dict = torch.load(pretrained) url = model_urls['resnet{}'.format(num_layers)] pretrained_state_dict = model_zoo.load_url(url) print('=> loading pretrained model {}'.format(url)) self.load_state_dict(pretrained_state_dict, strict=False) else: print('=> imagenet pretrained model dose not exist') print('=> please download it first') raise ValueError('imagenet pretrained model does not exist') resnet_spec = {18: (BasicBlock, [2, 2, 2, 2]), 34: (BasicBlock, [3, 4, 6, 3]), 50: (Bottleneck, [3, 4, 6, 3]), 101: (Bottleneck, [3, 4, 23, 3]), 152: (Bottleneck, [3, 8, 36, 3])} def get_pose_net(num_layers, heads, head_conv): block_class, layers = resnet_spec[num_layers] model = PoseResNet(block_class, layers, heads, head_conv=head_conv) model.init_weights(num_layers, pretrained=True) return model
10,167
35.185053
94
py
houghnet
houghnet-master/src/lib/models/networks/large_hourglass.py
# ------------------------------------------------------------------------------ # This code is base on # CornerNet (https://github.com/princeton-vl/CornerNet) # Copyright (c) 2018, University of Michigan # Licensed under the BSD 3-Clause License # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import numpy as np import torch import torch.nn as nn class convolution(nn.Module): def __init__(self, k, inp_dim, out_dim, stride=1, with_bn=True): super(convolution, self).__init__() pad = (k - 1) // 2 self.conv = nn.Conv2d(inp_dim, out_dim, (k, k), padding=(pad, pad), stride=(stride, stride), bias=not with_bn) self.bn = nn.BatchNorm2d(out_dim) if with_bn else nn.Sequential() self.relu = nn.ReLU(inplace=True) def forward(self, x): conv = self.conv(x) bn = self.bn(conv) relu = self.relu(bn) return relu class fully_connected(nn.Module): def __init__(self, inp_dim, out_dim, with_bn=True): super(fully_connected, self).__init__() self.with_bn = with_bn self.linear = nn.Linear(inp_dim, out_dim) if self.with_bn: self.bn = nn.BatchNorm1d(out_dim) self.relu = nn.ReLU(inplace=True) def forward(self, x): linear = self.linear(x) bn = self.bn(linear) if self.with_bn else linear relu = self.relu(bn) return relu class residual(nn.Module): def __init__(self, k, inp_dim, out_dim, stride=1, with_bn=True): super(residual, self).__init__() self.conv1 = nn.Conv2d(inp_dim, out_dim, (3, 3), padding=(1, 1), stride=(stride, stride), bias=False) self.bn1 = nn.BatchNorm2d(out_dim) self.relu1 = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(out_dim, out_dim, (3, 3), padding=(1, 1), bias=False) self.bn2 = nn.BatchNorm2d(out_dim) self.skip = nn.Sequential( nn.Conv2d(inp_dim, out_dim, (1, 1), stride=(stride, stride), bias=False), nn.BatchNorm2d(out_dim) ) if stride != 1 or inp_dim != out_dim else nn.Sequential() self.relu = nn.ReLU(inplace=True) def forward(self, x): conv1 = self.conv1(x) bn1 = self.bn1(conv1) relu1 = self.relu1(bn1) conv2 = self.conv2(relu1) bn2 = self.bn2(conv2) skip = self.skip(x) return self.relu(bn2 + skip) def make_layer(k, inp_dim, out_dim, modules, layer=convolution, **kwargs): layers = [layer(k, inp_dim, out_dim, **kwargs)] for _ in range(1, modules): layers.append(layer(k, out_dim, out_dim, **kwargs)) return nn.Sequential(*layers) def make_layer_revr(k, inp_dim, out_dim, modules, layer=convolution, **kwargs): layers = [] for _ in range(modules - 1): layers.append(layer(k, inp_dim, inp_dim, **kwargs)) layers.append(layer(k, inp_dim, out_dim, **kwargs)) return nn.Sequential(*layers) class MergeUp(nn.Module): def forward(self, up1, up2): return up1 + up2 def make_merge_layer(dim): return MergeUp() # def make_pool_layer(dim): # return nn.MaxPool2d(kernel_size=2, stride=2) def make_pool_layer(dim): return nn.Sequential() def make_unpool_layer(dim): return nn.Upsample(scale_factor=2) def make_kp_layer(cnv_dim, curr_dim, out_dim): return nn.Sequential( convolution(3, cnv_dim, curr_dim, with_bn=False), nn.Conv2d(curr_dim, out_dim, (1, 1)) ) def make_inter_layer(dim): return residual(3, dim, dim) def make_cnv_layer(inp_dim, out_dim): return convolution(3, inp_dim, out_dim) class kp_module(nn.Module): def __init__( self, n, dims, modules, layer=residual, make_up_layer=make_layer, make_low_layer=make_layer, make_hg_layer=make_layer, make_hg_layer_revr=make_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, **kwargs ): super(kp_module, self).__init__() self.n = n curr_mod = modules[0] next_mod = modules[1] curr_dim = dims[0] next_dim = dims[1] self.up1 = make_up_layer( 3, curr_dim, curr_dim, curr_mod, layer=layer, **kwargs ) self.max1 = make_pool_layer(curr_dim) self.low1 = make_hg_layer( 3, curr_dim, next_dim, curr_mod, layer=layer, **kwargs ) self.low2 = kp_module( n - 1, dims[1:], modules[1:], layer=layer, make_up_layer=make_up_layer, make_low_layer=make_low_layer, make_hg_layer=make_hg_layer, make_hg_layer_revr=make_hg_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, **kwargs ) if self.n > 1 else \ make_low_layer( 3, next_dim, next_dim, next_mod, layer=layer, **kwargs ) self.low3 = make_hg_layer_revr( 3, next_dim, curr_dim, curr_mod, layer=layer, **kwargs ) self.up2 = make_unpool_layer(curr_dim) self.merge = make_merge_layer(curr_dim) def forward(self, x): up1 = self.up1(x) max1 = self.max1(x) low1 = self.low1(max1) low2 = self.low2(low1) low3 = self.low3(low2) up2 = self.up2(low3) return self.merge(up1, up2) class exkp(nn.Module): def __init__( self, n, nstack, dims, modules, heads, pre=None, cnv_dim=256, make_tl_layer=None, make_br_layer=None, make_cnv_layer=make_cnv_layer, make_heat_layer=make_kp_layer, make_tag_layer=make_kp_layer, make_regr_layer=make_kp_layer, make_up_layer=make_layer, make_low_layer=make_layer, make_hg_layer=make_layer, make_hg_layer_revr=make_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer, make_inter_layer=make_inter_layer, kp_layer=residual ): super(exkp, self).__init__() self.nstack = nstack self.heads = heads curr_dim = dims[0] self.pre = nn.Sequential( convolution(7, 3, 128, stride=2), residual(3, 128, 256, stride=2) ) if pre is None else pre self.kps = nn.ModuleList([ kp_module( n, dims, modules, layer=kp_layer, make_up_layer=make_up_layer, make_low_layer=make_low_layer, make_hg_layer=make_hg_layer, make_hg_layer_revr=make_hg_layer_revr, make_pool_layer=make_pool_layer, make_unpool_layer=make_unpool_layer, make_merge_layer=make_merge_layer ) for _ in range(nstack) ]) self.cnvs = nn.ModuleList([ make_cnv_layer(curr_dim, cnv_dim) for _ in range(nstack) ]) self.inters = nn.ModuleList([ make_inter_layer(curr_dim) for _ in range(nstack - 1) ]) self.inters_ = nn.ModuleList([ nn.Sequential( nn.Conv2d(curr_dim, curr_dim, (1, 1), bias=False), nn.BatchNorm2d(curr_dim) ) for _ in range(nstack - 1) ]) self.cnvs_ = nn.ModuleList([ nn.Sequential( nn.Conv2d(cnv_dim, curr_dim, (1, 1), bias=False), nn.BatchNorm2d(curr_dim) ) for _ in range(nstack - 1) ]) ## keypoint heatmaps for head in heads.keys(): if 'hm' in head: module = nn.ModuleList([ make_heat_layer( cnv_dim, curr_dim, heads[head]) for _ in range(nstack) ]) self.__setattr__(head, module) for heat in self.__getattr__(head): heat[-1].bias.data.fill_(-2.19) else: module = nn.ModuleList([ make_regr_layer( cnv_dim, curr_dim, heads[head]) for _ in range(nstack) ]) self.__setattr__(head, module) self.relu = nn.ReLU(inplace=True) def forward(self, image): # print('image shape', image.shape) inter = self.pre(image) outs = [] for ind in range(self.nstack): kp_, cnv_ = self.kps[ind], self.cnvs[ind] kp = kp_(inter) cnv = cnv_(kp) out = {} for head in self.heads: layer = self.__getattr__(head)[ind] y = layer(cnv) out[head] = y outs.append(out) if ind < self.nstack - 1: inter = self.inters_[ind](inter) + self.cnvs_[ind](cnv) inter = self.relu(inter) inter = self.inters[ind](inter) return outs def make_hg_layer(kernel, dim0, dim1, mod, layer=convolution, **kwargs): layers = [layer(kernel, dim0, dim1, stride=2)] layers += [layer(kernel, dim1, dim1) for _ in range(mod - 1)] return nn.Sequential(*layers) class HourglassNet(exkp): def __init__(self, heads, num_stacks=2): n = 5 dims = [256, 256, 384, 384, 384, 512] modules = [2, 2, 2, 2, 2, 4] super(HourglassNet, self).__init__( n, num_stacks, dims, modules, heads, make_tl_layer=None, make_br_layer=None, make_pool_layer=make_pool_layer, make_hg_layer=make_hg_layer, kp_layer=residual, cnv_dim=256 ) def get_large_hourglass_net(num_layers, heads, head_conv): model = HourglassNet(heads, 2) return model
9,942
32.033223
118
py
houghnet
houghnet-master/src/lib/models/networks/houghnet_resnet.py
# ------------------------------------------------------------------------------ # Copyright (c) Microsoft # Licensed under the MIT License. # Written by Bin Xiao (Bin.Xiao@microsoft.com) # Modified by Xingyi Zhou # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import re import torch import torch.nn as nn import torch.utils.model_zoo as model_zoo from src.lib.models.networks.hough_module import Hough BN_MOMENTUM = 0.1 model_urls = { 'resnet18': 'https://download.pytorch.org/models/resnet18-5c106cde.pth', 'resnet34': 'https://download.pytorch.org/models/resnet34-333f7ec4.pth', 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'resnet101': 'https://download.pytorch.org/models/resnet101-5d3b4d8f.pth', 'resnet152': 'https://download.pytorch.org/models/resnet152-b121ed2d.pth', } def conv3x3(in_planes, out_planes, stride=1): """3x3 convolution with padding""" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes * self.expansion, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride def forward(self, x): residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) if self.downsample is not None: residual = self.downsample(x) out += residual out = self.relu(out) return out def fill_fc_weights(layers): for m in layers.modules(): if isinstance(m, nn.Conv2d): nn.init.normal_(m.weight, std=0.001) # torch.nn.init.kaiming_normal_(m.weight.data, nonlinearity='relu') # torch.nn.init.xavier_normal_(m.weight.data) if m.bias is not None: nn.init.constant_(m.bias, 0) class HoughNetResNet(nn.Module): def __init__(self, block, layers, heads, region_num, vote_field_size, model_v1, head_conv, **kwargs): self.inplanes = 64 self.deconv_with_bias = False self.heads = heads self.region_num = region_num self.vote_field_size = vote_field_size super(HoughNetResNet, self).__init__() self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = nn.BatchNorm2d(64, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1) self.layer1 = self._make_layer(block, 64, layers[0]) self.layer2 = self._make_layer(block, 128, layers[1], stride=2) self.layer3 = self._make_layer(block, 256, layers[2], stride=2) self.layer4 = self._make_layer(block, 512, layers[3], stride=2) # used for deconv layers self.deconv_layers = self._make_deconv_layer2( 3, [256, 128, 64], [4, 4, 4], ) # self.final_layer = [] self.voting_heads = list(heads['voting_heads']) del heads['voting_heads'] voting = False self.heads = heads for head in sorted(self.heads): num_output = self.heads[head] if head_conv > 0: fc = nn.Sequential( nn.Conv2d(64, head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, num_output, kernel_size=1, stride=1, padding=0)) for voting_head in self.voting_heads: if re.fullmatch(head, voting_head): voting = True if voting: out_classes = int(num_output / self.region_num) hough_voting = Hough(region_num=self.region_num, vote_field_size=self.vote_field_size, num_classes=out_classes, model_v1=model_v1) self.__setattr__('voting_' + head, hough_voting) voting = False else: fc = nn.Conv2d( in_channels=64, out_channels=num_output, kernel_size=1, stride=1, padding=0 ) self.__setattr__(head, fc) def _make_layer(self, block, planes, blocks, stride=1): downsample = None if stride != 1 or self.inplanes != planes * block.expansion: downsample = nn.Sequential( nn.Conv2d(self.inplanes, planes * block.expansion, kernel_size=1, stride=stride, bias=False), nn.BatchNorm2d(planes * block.expansion, momentum=BN_MOMENTUM), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes)) return nn.Sequential(*layers) def _get_deconv_cfg(self, deconv_kernel, index): if deconv_kernel == 4: padding = 1 output_padding = 0 elif deconv_kernel == 3: padding = 1 output_padding = 1 elif deconv_kernel == 2: padding = 0 output_padding = 0 return deconv_kernel, padding, output_padding def _make_deconv_layer(self, num_layers, num_filters, num_kernels): assert num_layers == len(num_filters), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' assert num_layers == len(num_kernels), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' layers = [] for i in range(num_layers): kernel, padding, output_padding = \ self._get_deconv_cfg(num_kernels[i], i) planes = num_filters[i] layers.append( nn.ConvTranspose2d( in_channels=self.inplanes, out_channels=planes, kernel_size=kernel, stride=2, padding=padding, output_padding=output_padding, bias=self.deconv_with_bias)) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) self.inplanes = planes return nn.Sequential(*layers) def _make_deconv_layer2(self, num_layers, num_filters, num_kernels): assert num_layers == len(num_filters), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' assert num_layers == len(num_kernels), \ 'ERROR: num_deconv_layers is different len(num_deconv_filters)' layers = [] for i in range(num_layers): kernel, padding, output_padding = \ self._get_deconv_cfg(num_kernels[i], i) planes = num_filters[i] fc = nn.Conv2d(self.inplanes, planes, kernel_size=3, stride=1, padding=1, dilation=1, bias=False) fill_fc_weights(fc) up = nn.ConvTranspose2d( in_channels=planes, out_channels=planes, kernel_size=kernel, stride=2, padding=padding, output_padding=output_padding, bias=self.deconv_with_bias) layers.append(fc) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) layers.append(up) layers.append(nn.BatchNorm2d(planes, momentum=BN_MOMENTUM)) layers.append(nn.ReLU(inplace=True)) self.inplanes = planes return nn.Sequential(*layers) def forward(self, x): x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) x = self.layer4(x) x = self.deconv_layers(x) ret = {} for head in self.heads: if head in self.voting_heads: voting_map_hm = self.__getattr__(head)(x) ret[head] = self.__getattr__('voting_' + head)(voting_map_hm) else: ret[head] = self.__getattr__(head)(x) return [ret] def init_weights(self, num_layers, pretrained=True): if pretrained: # print('=> init resnet deconv weights from normal distribution') for _, m in self.deconv_layers.named_modules(): if isinstance(m, nn.ConvTranspose2d): # print('=> init {}.weight as normal(0, 0.001)'.format(name)) # print('=> init {}.bias as 0'.format(name)) nn.init.normal_(m.weight, std=0.001) if self.deconv_with_bias: nn.init.constant_(m.bias, 0) elif isinstance(m, nn.BatchNorm2d): # print('=> init {}.weight as 1'.format(name)) # print('=> init {}.bias as 0'.format(name)) nn.init.constant_(m.weight, 1) nn.init.constant_(m.bias, 0) # print('=> init final conv weights from normal distribution') for head in self.heads: final_layer = self.__getattr__(head) for i, m in enumerate(final_layer.modules()): if isinstance(m, nn.Conv2d): # nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu') # print('=> init {}.weight as normal(0, 0.001)'.format(name)) # print('=> init {}.bias as 0'.format(name)) if m.weight.shape[0] == self.heads[head]: if 'hm' in head: nn.init.constant_(m.weight, 0) nn.init.constant_(m.bias, 0) else: nn.init.normal_(m.weight, std=0.001) nn.init.constant_(m.bias, 0) #pretrained_state_dict = torch.load(pretrained) url = model_urls['resnet{}'.format(num_layers)] pretrained_state_dict = model_zoo.load_url(url) print('=> loading pretrained model {}'.format(url)) self.load_state_dict(pretrained_state_dict, strict=False) else: print('=> imagenet pretrained model dose not exist') print('=> please download it first') raise ValueError('imagenet pretrained model does not exist') resnet_spec = {18: (BasicBlock, [2, 2, 2, 2]), 34: (BasicBlock, [3, 4, 6, 3]), 50: (Bottleneck, [3, 4, 6, 3]), 101: (Bottleneck, [3, 4, 23, 3]), 152: (Bottleneck, [3, 8, 36, 3])} def get_houghnet_net(num_layers, heads, head_conv, region_num, model_v1, vote_field_size): block_class, layers = resnet_spec[num_layers] model = HoughNetResNet(block_class, layers, heads, region_num, vote_field_size, model_v1, head_conv=head_conv) model.init_weights(num_layers, pretrained=True) return model
13,125
36.289773
113
py
houghnet
houghnet-master/src/lib/models/networks/pose_dla_dcn_hough.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import math import logging import numpy as np from os.path import join import torch from torch import nn import torch.nn.functional as F import torch.utils.model_zoo as model_zoo from src.lib.models.networks.hough_module import Hough import re from .DCNv2.dcn_v2 import DCN BN_MOMENTUM = 0.1 logger = logging.getLogger(__name__) def get_model_url(data='imagenet', name='dla34', hash='ba72cf86'): return join('http://dl.yf.io/dla/models', data, '{}-{}.pth'.format(name, hash)) def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): def __init__(self, inplanes, planes, stride=1, dilation=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn1 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=dilation, bias=False, dilation=dilation) self.bn2 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 2 def __init__(self, inplanes, planes, stride=1, dilation=1): super(Bottleneck, self).__init__() expansion = Bottleneck.expansion bottle_planes = planes // expansion self.conv1 = nn.Conv2d(inplanes, bottle_planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(bottle_planes, bottle_planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn2 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(bottle_planes, planes, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out class BottleneckX(nn.Module): expansion = 2 cardinality = 32 def __init__(self, inplanes, planes, stride=1, dilation=1): super(BottleneckX, self).__init__() cardinality = BottleneckX.cardinality # dim = int(math.floor(planes * (BottleneckV5.expansion / 64.0))) # bottle_planes = dim * cardinality bottle_planes = planes * cardinality // 32 self.conv1 = nn.Conv2d(inplanes, bottle_planes, kernel_size=1, bias=False) self.bn1 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv2 = nn.Conv2d(bottle_planes, bottle_planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation, groups=cardinality) self.bn2 = nn.BatchNorm2d(bottle_planes, momentum=BN_MOMENTUM) self.conv3 = nn.Conv2d(bottle_planes, planes, kernel_size=1, bias=False) self.bn3 = nn.BatchNorm2d(planes, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out class Root(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, residual): super(Root, self).__init__() self.conv = nn.Conv2d( in_channels, out_channels, 1, stride=1, bias=False, padding=(kernel_size - 1) // 2) self.bn = nn.BatchNorm2d(out_channels, momentum=BN_MOMENTUM) self.relu = nn.ReLU(inplace=True) self.residual = residual def forward(self, *x): children = x x = self.conv(torch.cat(x, 1)) x = self.bn(x) if self.residual: x += children[0] x = self.relu(x) return x class Tree(nn.Module): def __init__(self, levels, block, in_channels, out_channels, stride=1, level_root=False, root_dim=0, root_kernel_size=1, dilation=1, root_residual=False): super(Tree, self).__init__() if root_dim == 0: root_dim = 2 * out_channels if level_root: root_dim += in_channels if levels == 1: self.tree1 = block(in_channels, out_channels, stride, dilation=dilation) self.tree2 = block(out_channels, out_channels, 1, dilation=dilation) else: self.tree1 = Tree(levels - 1, block, in_channels, out_channels, stride, root_dim=0, root_kernel_size=root_kernel_size, dilation=dilation, root_residual=root_residual) self.tree2 = Tree(levels - 1, block, out_channels, out_channels, root_dim=root_dim + out_channels, root_kernel_size=root_kernel_size, dilation=dilation, root_residual=root_residual) if levels == 1: self.root = Root(root_dim, out_channels, root_kernel_size, root_residual) self.level_root = level_root self.root_dim = root_dim self.downsample = None self.project = None self.levels = levels if stride > 1: self.downsample = nn.MaxPool2d(stride, stride=stride) if in_channels != out_channels: self.project = nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False), nn.BatchNorm2d(out_channels, momentum=BN_MOMENTUM) ) def forward(self, x, residual=None, children=None): children = [] if children is None else children bottom = self.downsample(x) if self.downsample else x residual = self.project(bottom) if self.project else bottom if self.level_root: children.append(bottom) x1 = self.tree1(x, residual) if self.levels == 1: x2 = self.tree2(x1) x = self.root(x2, x1, *children) else: children.append(x1) x = self.tree2(x1, children=children) return x class DLA(nn.Module): def __init__(self, levels, channels, num_classes=1000, block=BasicBlock, residual_root=False, linear_root=False): super(DLA, self).__init__() self.channels = channels self.num_classes = num_classes self.base_layer = nn.Sequential( nn.Conv2d(3, channels[0], kernel_size=7, stride=1, padding=3, bias=False), nn.BatchNorm2d(channels[0], momentum=BN_MOMENTUM), nn.ReLU(inplace=True)) self.level0 = self._make_conv_level( channels[0], channels[0], levels[0]) self.level1 = self._make_conv_level( channels[0], channels[1], levels[1], stride=2) self.level2 = Tree(levels[2], block, channels[1], channels[2], 2, level_root=False, root_residual=residual_root) self.level3 = Tree(levels[3], block, channels[2], channels[3], 2, level_root=True, root_residual=residual_root) self.level4 = Tree(levels[4], block, channels[3], channels[4], 2, level_root=True, root_residual=residual_root) self.level5 = Tree(levels[5], block, channels[4], channels[5], 2, level_root=True, root_residual=residual_root) # for m in self.modules(): # if isinstance(m, nn.Conv2d): # n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels # m.weight.data.normal_(0, math.sqrt(2. / n)) # elif isinstance(m, nn.BatchNorm2d): # m.weight.data.fill_(1) # m.bias.data.zero_() def _make_level(self, block, inplanes, planes, blocks, stride=1): downsample = None if stride != 1 or inplanes != planes: downsample = nn.Sequential( nn.MaxPool2d(stride, stride=stride), nn.Conv2d(inplanes, planes, kernel_size=1, stride=1, bias=False), nn.BatchNorm2d(planes, momentum=BN_MOMENTUM), ) layers = [] layers.append(block(inplanes, planes, stride, downsample=downsample)) for i in range(1, blocks): layers.append(block(inplanes, planes)) return nn.Sequential(*layers) def _make_conv_level(self, inplanes, planes, convs, stride=1, dilation=1): modules = [] for i in range(convs): modules.extend([ nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride if i == 0 else 1, padding=dilation, bias=False, dilation=dilation), nn.BatchNorm2d(planes, momentum=BN_MOMENTUM), nn.ReLU(inplace=True)]) inplanes = planes return nn.Sequential(*modules) def forward(self, x): y = [] x = self.base_layer(x) for i in range(6): x = getattr(self, 'level{}'.format(i))(x) y.append(x) return y def load_pretrained_model(self, data='imagenet', name='dla34', hash='ba72cf86'): # fc = self.fc if name.endswith('.pth'): model_weights = torch.load(data + name) else: model_url = get_model_url(data, name, hash) model_weights = model_zoo.load_url(model_url) num_classes = len(model_weights[list(model_weights.keys())[-1]]) self.fc = nn.Conv2d( self.channels[-1], num_classes, kernel_size=1, stride=1, padding=0, bias=True) self.load_state_dict(model_weights) # self.fc = fc def dla34(pretrained=True, **kwargs): # DLA-34 model = DLA([1, 1, 1, 2, 2, 1], [16, 32, 64, 128, 256, 512], block=BasicBlock, **kwargs) if pretrained: model.load_pretrained_model(data='imagenet', name='dla34', hash='ba72cf86') return model class Identity(nn.Module): def __init__(self): super(Identity, self).__init__() def forward(self, x): return x def fill_fc_weights(layers): for m in layers.modules(): if isinstance(m, nn.Conv2d): if m.bias is not None: nn.init.constant_(m.bias, 0) def fill_up_weights(up): w = up.weight.data f = math.ceil(w.size(2) / 2) c = (2 * f - 1 - f % 2) / (2. * f) for i in range(w.size(2)): for j in range(w.size(3)): w[0, 0, i, j] = \ (1 - math.fabs(i / f - c)) * (1 - math.fabs(j / f - c)) for c in range(1, w.size(0)): w[c, 0, :, :] = w[0, 0, :, :] class DeformConv(nn.Module): def __init__(self, chi, cho): super(DeformConv, self).__init__() self.actf = nn.Sequential( nn.BatchNorm2d(cho, momentum=BN_MOMENTUM), nn.ReLU(inplace=True) ) self.conv = DCN(chi, cho, kernel_size=(3,3), stride=1, padding=1, dilation=1, deformable_groups=1) def forward(self, x): x = self.conv(x) x = self.actf(x) return x class IDAUp(nn.Module): def __init__(self, o, channels, up_f): super(IDAUp, self).__init__() for i in range(1, len(channels)): c = channels[i] f = int(up_f[i]) proj = DeformConv(c, o) node = DeformConv(o, o) up = nn.ConvTranspose2d(o, o, f * 2, stride=f, padding=f // 2, output_padding=0, groups=o, bias=False) fill_up_weights(up) setattr(self, 'proj_' + str(i), proj) setattr(self, 'up_' + str(i), up) setattr(self, 'node_' + str(i), node) def forward(self, layers, startp, endp): for i in range(startp + 1, endp): upsample = getattr(self, 'up_' + str(i - startp)) project = getattr(self, 'proj_' + str(i - startp)) layers[i] = upsample(project(layers[i])) node = getattr(self, 'node_' + str(i - startp)) layers[i] = node(layers[i] + layers[i - 1]) class DLAUp(nn.Module): def __init__(self, startp, channels, scales, in_channels=None): super(DLAUp, self).__init__() self.startp = startp if in_channels is None: in_channels = channels self.channels = channels channels = list(channels) scales = np.array(scales, dtype=int) for i in range(len(channels) - 1): j = -i - 2 setattr(self, 'ida_{}'.format(i), IDAUp(channels[j], in_channels[j:], scales[j:] // scales[j])) scales[j + 1:] = scales[j] in_channels[j + 1:] = [channels[j] for _ in channels[j + 1:]] def forward(self, layers): out = [layers[-1]] # start with 32 for i in range(len(layers) - self.startp - 1): ida = getattr(self, 'ida_{}'.format(i)) ida(layers, len(layers) -i - 2, len(layers)) out.insert(0, layers[-1]) return out class Interpolate(nn.Module): def __init__(self, scale, mode): super(Interpolate, self).__init__() self.scale = scale self.mode = mode def forward(self, x): x = F.interpolate(x, scale_factor=self.scale, mode=self.mode, align_corners=False) return x class DLASegHough(nn.Module): def __init__(self, base_name, heads, pretrained, down_ratio, final_kernel, last_level, region_num, vote_field_size, model_v1, head_conv, out_channel=0): super(DLASegHough, self).__init__() assert down_ratio in [2, 4, 8, 16] self.region_num = region_num self.vote_field_size = vote_field_size # self.num_classes = int(heads['hm_hp'] / region_num) self.first_level = int(np.log2(down_ratio)) self.last_level = last_level self.base = globals()[base_name](pretrained=pretrained) channels = self.base.channels scales = [2 ** i for i in range(len(channels[self.first_level:]))] self.dla_up = DLAUp(self.first_level, channels[self.first_level:], scales) if out_channel == 0: out_channel = channels[self.first_level] self.ida_up = IDAUp(out_channel, channels[self.first_level:self.last_level], [2 ** i for i in range(self.last_level - self.first_level)]) self.voting_heads = list(heads['voting_heads']) del heads['voting_heads'] voting = False self.heads = heads for head in self.heads: classes = self.heads[head] if head_conv > 0: fc = nn.Sequential( nn.Conv2d(channels[self.first_level], head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, classes, kernel_size=final_kernel, stride=1, padding=final_kernel // 2, bias=True)) if 'hm' in head: fc[-1].bias.data.fill_(-2.19) else: fill_fc_weights(fc) for voting_head in self.voting_heads: if re.fullmatch(head, voting_head): voting = True if voting: fc[-1].bias.data.fill_(0) fc[-1].weight.data.fill_(0) out_classes = int(classes / self.region_num) hough_voting = Hough(region_num=self.region_num, vote_field_size=self.vote_field_size, num_classes=out_classes, model_v1=model_v1) self.__setattr__('voting_' + head, hough_voting) voting = False else: fc = nn.Conv2d(channels[self.first_level], classes, kernel_size=final_kernel, stride=1, padding=final_kernel // 2, bias=True) if 'hm' in head: fc.bias.data.fill_(-2.19) else: fill_fc_weights(fc) self.__setattr__(head, fc) def forward(self, x): x = self.base(x) x = self.dla_up(x) y = [] for i in range(self.last_level - self.first_level): y.append(x[i].clone()) self.ida_up(y, 0, len(y)) z = {} for head in self.heads: if head in self.voting_heads: voting_map_hm = self.__getattr__(head)(y[-1]) z[head] = self.__getattr__('voting_' + head)(voting_map_hm) else: z[head] = self.__getattr__(head)(y[-1]) return [z] def get_pose_net(num_layers, heads, region_num, vote_field_size, head_conv=256, down_ratio=4, model_v1=False): model = DLASegHough('dla{}'.format(num_layers), heads, pretrained=True, down_ratio=down_ratio, final_kernel=1, last_level=5, region_num=region_num, vote_field_size=vote_field_size, model_v1=model_v1, head_conv=head_conv) return model
19,013
35.28626
111
py
houghnet
houghnet-master/src/lib/models/networks/dlav0.py
#!/usr/bin/env python # -*- coding: utf-8 -*- from __future__ import absolute_import from __future__ import division from __future__ import print_function import math from os.path import join import torch from torch import nn import torch.utils.model_zoo as model_zoo import numpy as np BatchNorm = nn.BatchNorm2d def get_model_url(data='imagenet', name='dla34', hash='ba72cf86'): return join('http://dl.yf.io/dla/models', data, '{}-{}.pth'.format(name, hash)) def conv3x3(in_planes, out_planes, stride=1): "3x3 convolution with padding" return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=1, bias=False) class BasicBlock(nn.Module): def __init__(self, inplanes, planes, stride=1, dilation=1): super(BasicBlock, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn1 = BatchNorm(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=dilation, bias=False, dilation=dilation) self.bn2 = BatchNorm(planes) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 2 def __init__(self, inplanes, planes, stride=1, dilation=1): super(Bottleneck, self).__init__() expansion = Bottleneck.expansion bottle_planes = planes // expansion self.conv1 = nn.Conv2d(inplanes, bottle_planes, kernel_size=1, bias=False) self.bn1 = BatchNorm(bottle_planes) self.conv2 = nn.Conv2d(bottle_planes, bottle_planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation) self.bn2 = BatchNorm(bottle_planes) self.conv3 = nn.Conv2d(bottle_planes, planes, kernel_size=1, bias=False) self.bn3 = BatchNorm(planes) self.relu = nn.ReLU(inplace=True) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out class BottleneckX(nn.Module): expansion = 2 cardinality = 32 def __init__(self, inplanes, planes, stride=1, dilation=1): super(BottleneckX, self).__init__() cardinality = BottleneckX.cardinality # dim = int(math.floor(planes * (BottleneckV5.expansion / 64.0))) # bottle_planes = dim * cardinality bottle_planes = planes * cardinality // 32 self.conv1 = nn.Conv2d(inplanes, bottle_planes, kernel_size=1, bias=False) self.bn1 = BatchNorm(bottle_planes) self.conv2 = nn.Conv2d(bottle_planes, bottle_planes, kernel_size=3, stride=stride, padding=dilation, bias=False, dilation=dilation, groups=cardinality) self.bn2 = BatchNorm(bottle_planes) self.conv3 = nn.Conv2d(bottle_planes, planes, kernel_size=1, bias=False) self.bn3 = BatchNorm(planes) self.relu = nn.ReLU(inplace=True) self.stride = stride def forward(self, x, residual=None): if residual is None: residual = x out = self.conv1(x) out = self.bn1(out) out = self.relu(out) out = self.conv2(out) out = self.bn2(out) out = self.relu(out) out = self.conv3(out) out = self.bn3(out) out += residual out = self.relu(out) return out class Root(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, residual): super(Root, self).__init__() self.conv = nn.Conv2d( in_channels, out_channels, 1, stride=1, bias=False, padding=(kernel_size - 1) // 2) self.bn = BatchNorm(out_channels) self.relu = nn.ReLU(inplace=True) self.residual = residual def forward(self, *x): children = x x = self.conv(torch.cat(x, 1)) x = self.bn(x) if self.residual: x += children[0] x = self.relu(x) return x class Tree(nn.Module): def __init__(self, levels, block, in_channels, out_channels, stride=1, level_root=False, root_dim=0, root_kernel_size=1, dilation=1, root_residual=False): super(Tree, self).__init__() if root_dim == 0: root_dim = 2 * out_channels if level_root: root_dim += in_channels if levels == 1: self.tree1 = block(in_channels, out_channels, stride, dilation=dilation) self.tree2 = block(out_channels, out_channels, 1, dilation=dilation) else: self.tree1 = Tree(levels - 1, block, in_channels, out_channels, stride, root_dim=0, root_kernel_size=root_kernel_size, dilation=dilation, root_residual=root_residual) self.tree2 = Tree(levels - 1, block, out_channels, out_channels, root_dim=root_dim + out_channels, root_kernel_size=root_kernel_size, dilation=dilation, root_residual=root_residual) if levels == 1: self.root = Root(root_dim, out_channels, root_kernel_size, root_residual) self.level_root = level_root self.root_dim = root_dim self.downsample = None self.project = None self.levels = levels if stride > 1: self.downsample = nn.MaxPool2d(stride, stride=stride) if in_channels != out_channels: self.project = nn.Sequential( nn.Conv2d(in_channels, out_channels, kernel_size=1, stride=1, bias=False), BatchNorm(out_channels) ) def forward(self, x, residual=None, children=None): children = [] if children is None else children bottom = self.downsample(x) if self.downsample else x residual = self.project(bottom) if self.project else bottom if self.level_root: children.append(bottom) x1 = self.tree1(x, residual) if self.levels == 1: x2 = self.tree2(x1) x = self.root(x2, x1, *children) else: children.append(x1) x = self.tree2(x1, children=children) return x class DLA(nn.Module): def __init__(self, levels, channels, num_classes=1000, block=BasicBlock, residual_root=False, return_levels=False, pool_size=7, linear_root=False): super(DLA, self).__init__() self.channels = channels self.return_levels = return_levels self.num_classes = num_classes self.base_layer = nn.Sequential( nn.Conv2d(3, channels[0], kernel_size=7, stride=1, padding=3, bias=False), BatchNorm(channels[0]), nn.ReLU(inplace=True)) self.level0 = self._make_conv_level( channels[0], channels[0], levels[0]) self.level1 = self._make_conv_level( channels[0], channels[1], levels[1], stride=2) self.level2 = Tree(levels[2], block, channels[1], channels[2], 2, level_root=False, root_residual=residual_root) self.level3 = Tree(levels[3], block, channels[2], channels[3], 2, level_root=True, root_residual=residual_root) self.level4 = Tree(levels[4], block, channels[3], channels[4], 2, level_root=True, root_residual=residual_root) self.level5 = Tree(levels[5], block, channels[4], channels[5], 2, level_root=True, root_residual=residual_root) self.avgpool = nn.AvgPool2d(pool_size) self.fc = nn.Conv2d(channels[-1], num_classes, kernel_size=1, stride=1, padding=0, bias=True) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, BatchNorm): m.weight.data.fill_(1) m.bias.data.zero_() def _make_level(self, block, inplanes, planes, blocks, stride=1): downsample = None if stride != 1 or inplanes != planes: downsample = nn.Sequential( nn.MaxPool2d(stride, stride=stride), nn.Conv2d(inplanes, planes, kernel_size=1, stride=1, bias=False), BatchNorm(planes), ) layers = [] layers.append(block(inplanes, planes, stride, downsample=downsample)) for i in range(1, blocks): layers.append(block(inplanes, planes)) return nn.Sequential(*layers) def _make_conv_level(self, inplanes, planes, convs, stride=1, dilation=1): modules = [] for i in range(convs): modules.extend([ nn.Conv2d(inplanes, planes, kernel_size=3, stride=stride if i == 0 else 1, padding=dilation, bias=False, dilation=dilation), BatchNorm(planes), nn.ReLU(inplace=True)]) inplanes = planes return nn.Sequential(*modules) def forward(self, x): y = [] x = self.base_layer(x) for i in range(6): x = getattr(self, 'level{}'.format(i))(x) y.append(x) if self.return_levels: return y else: x = self.avgpool(x) x = self.fc(x) x = x.view(x.size(0), -1) return x def load_pretrained_model(self, data='imagenet', name='dla34', hash='ba72cf86'): fc = self.fc if name.endswith('.pth'): model_weights = torch.load(data + name) else: model_url = get_model_url(data, name, hash) model_weights = model_zoo.load_url(model_url) num_classes = len(model_weights[list(model_weights.keys())[-1]]) self.fc = nn.Conv2d( self.channels[-1], num_classes, kernel_size=1, stride=1, padding=0, bias=True) self.load_state_dict(model_weights) self.fc = fc def dla34(pretrained, **kwargs): # DLA-34 model = DLA([1, 1, 1, 2, 2, 1], [16, 32, 64, 128, 256, 512], block=BasicBlock, **kwargs) if pretrained: model.load_pretrained_model(data='imagenet', name='dla34', hash='ba72cf86') return model def dla46_c(pretrained=None, **kwargs): # DLA-46-C Bottleneck.expansion = 2 model = DLA([1, 1, 1, 2, 2, 1], [16, 32, 64, 64, 128, 256], block=Bottleneck, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla46_c') return model def dla46x_c(pretrained=None, **kwargs): # DLA-X-46-C BottleneckX.expansion = 2 model = DLA([1, 1, 1, 2, 2, 1], [16, 32, 64, 64, 128, 256], block=BottleneckX, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla46x_c') return model def dla60x_c(pretrained, **kwargs): # DLA-X-60-C BottleneckX.expansion = 2 model = DLA([1, 1, 1, 2, 3, 1], [16, 32, 64, 64, 128, 256], block=BottleneckX, **kwargs) if pretrained: model.load_pretrained_model(data='imagenet', name='dla60x_c', hash='b870c45c') return model def dla60(pretrained=None, **kwargs): # DLA-60 Bottleneck.expansion = 2 model = DLA([1, 1, 1, 2, 3, 1], [16, 32, 128, 256, 512, 1024], block=Bottleneck, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla60') return model def dla60x(pretrained=None, **kwargs): # DLA-X-60 BottleneckX.expansion = 2 model = DLA([1, 1, 1, 2, 3, 1], [16, 32, 128, 256, 512, 1024], block=BottleneckX, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla60x') return model def dla102(pretrained=None, **kwargs): # DLA-102 Bottleneck.expansion = 2 model = DLA([1, 1, 1, 3, 4, 1], [16, 32, 128, 256, 512, 1024], block=Bottleneck, residual_root=True, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla102') return model def dla102x(pretrained=None, **kwargs): # DLA-X-102 BottleneckX.expansion = 2 model = DLA([1, 1, 1, 3, 4, 1], [16, 32, 128, 256, 512, 1024], block=BottleneckX, residual_root=True, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla102x') return model def dla102x2(pretrained=None, **kwargs): # DLA-X-102 64 BottleneckX.cardinality = 64 model = DLA([1, 1, 1, 3, 4, 1], [16, 32, 128, 256, 512, 1024], block=BottleneckX, residual_root=True, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla102x2') return model def dla169(pretrained=None, **kwargs): # DLA-169 Bottleneck.expansion = 2 model = DLA([1, 1, 2, 3, 5, 1], [16, 32, 128, 256, 512, 1024], block=Bottleneck, residual_root=True, **kwargs) if pretrained is not None: model.load_pretrained_model(pretrained, 'dla169') return model def set_bn(bn): global BatchNorm BatchNorm = bn dla.BatchNorm = bn class Identity(nn.Module): def __init__(self): super(Identity, self).__init__() def forward(self, x): return x def fill_up_weights(up): w = up.weight.data f = math.ceil(w.size(2) / 2) c = (2 * f - 1 - f % 2) / (2. * f) for i in range(w.size(2)): for j in range(w.size(3)): w[0, 0, i, j] = \ (1 - math.fabs(i / f - c)) * (1 - math.fabs(j / f - c)) for c in range(1, w.size(0)): w[c, 0, :, :] = w[0, 0, :, :] class IDAUp(nn.Module): def __init__(self, node_kernel, out_dim, channels, up_factors): super(IDAUp, self).__init__() self.channels = channels self.out_dim = out_dim for i, c in enumerate(channels): if c == out_dim: proj = Identity() else: proj = nn.Sequential( nn.Conv2d(c, out_dim, kernel_size=1, stride=1, bias=False), BatchNorm(out_dim), nn.ReLU(inplace=True)) f = int(up_factors[i]) if f == 1: up = Identity() else: up = nn.ConvTranspose2d( out_dim, out_dim, f * 2, stride=f, padding=f // 2, output_padding=0, groups=out_dim, bias=False) fill_up_weights(up) setattr(self, 'proj_' + str(i), proj) setattr(self, 'up_' + str(i), up) for i in range(1, len(channels)): node = nn.Sequential( nn.Conv2d(out_dim * 2, out_dim, kernel_size=node_kernel, stride=1, padding=node_kernel // 2, bias=False), BatchNorm(out_dim), nn.ReLU(inplace=True)) setattr(self, 'node_' + str(i), node) for m in self.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, BatchNorm): m.weight.data.fill_(1) m.bias.data.zero_() def forward(self, layers): assert len(self.channels) == len(layers), \ '{} vs {} layers'.format(len(self.channels), len(layers)) layers = list(layers) for i, l in enumerate(layers): upsample = getattr(self, 'up_' + str(i)) project = getattr(self, 'proj_' + str(i)) layers[i] = upsample(project(l)) x = layers[0] y = [] for i in range(1, len(layers)): node = getattr(self, 'node_' + str(i)) x = node(torch.cat([x, layers[i]], 1)) y.append(x) return x, y class DLAUp(nn.Module): def __init__(self, channels, scales=(1, 2, 4, 8, 16), in_channels=None): super(DLAUp, self).__init__() if in_channels is None: in_channels = channels self.channels = channels channels = list(channels) scales = np.array(scales, dtype=int) for i in range(len(channels) - 1): j = -i - 2 setattr(self, 'ida_{}'.format(i), IDAUp(3, channels[j], in_channels[j:], scales[j:] // scales[j])) scales[j + 1:] = scales[j] in_channels[j + 1:] = [channels[j] for _ in channels[j + 1:]] def forward(self, layers): layers = list(layers) assert len(layers) > 1 for i in range(len(layers) - 1): ida = getattr(self, 'ida_{}'.format(i)) x, y = ida(layers[-i - 2:]) layers[-i - 1:] = y return x def fill_fc_weights(layers): for m in layers.modules(): if isinstance(m, nn.Conv2d): nn.init.normal_(m.weight, std=0.001) # torch.nn.init.kaiming_normal_(m.weight.data, nonlinearity='relu') # torch.nn.init.xavier_normal_(m.weight.data) if m.bias is not None: nn.init.constant_(m.bias, 0) class DLASeg(nn.Module): def __init__(self, base_name, heads, pretrained=True, down_ratio=4, head_conv=256): super(DLASeg, self).__init__() assert down_ratio in [2, 4, 8, 16] self.heads = heads self.first_level = int(np.log2(down_ratio)) self.base = globals()[base_name]( pretrained=pretrained, return_levels=True) channels = self.base.channels scales = [2 ** i for i in range(len(channels[self.first_level:]))] self.dla_up = DLAUp(channels[self.first_level:], scales=scales) ''' self.fc = nn.Sequential( nn.Conv2d(channels[self.first_level], classes, kernel_size=1, stride=1, padding=0, bias=True) ) ''' for head in self.heads: classes = self.heads[head] if head_conv > 0: fc = nn.Sequential( nn.Conv2d(channels[self.first_level], head_conv, kernel_size=3, padding=1, bias=True), nn.ReLU(inplace=True), nn.Conv2d(head_conv, classes, kernel_size=1, stride=1, padding=0, bias=True)) if 'hm' in head: fc[-1].bias.data.fill_(-2.19) else: fill_fc_weights(fc) else: fc = nn.Conv2d(channels[self.first_level], classes, kernel_size=1, stride=1, padding=0, bias=True) if 'hm' in head: fc.bias.data.fill_(-2.19) else: fill_fc_weights(fc) self.__setattr__(head, fc) ''' up_factor = 2 ** self.first_level if up_factor > 1: up = nn.ConvTranspose2d(classes, classes, up_factor * 2, stride=up_factor, padding=up_factor // 2, output_padding=0, groups=classes, bias=False) fill_up_weights(up) up.weight.requires_grad = False else: up = Identity() self.up = up self.softmax = nn.LogSoftmax(dim=1) for m in self.fc.modules(): if isinstance(m, nn.Conv2d): n = m.kernel_size[0] * m.kernel_size[1] * m.out_channels m.weight.data.normal_(0, math.sqrt(2. / n)) elif isinstance(m, BatchNorm): m.weight.data.fill_(1) m.bias.data.zero_() ''' def forward(self, x): x = self.base(x) x = self.dla_up(x[self.first_level:]) # x = self.fc(x) # y = self.softmax(self.up(x)) ret = {} for head in self.heads: ret[head] = self.__getattr__(head)(x) return [ret] ''' def optim_parameters(self, memo=None): for param in self.base.parameters(): yield param for param in self.dla_up.parameters(): yield param for param in self.fc.parameters(): yield param ''' ''' def dla34up(classes, pretrained_base=None, **kwargs): model = DLASeg('dla34', classes, pretrained_base=pretrained_base, **kwargs) return model def dla60up(classes, pretrained_base=None, **kwargs): model = DLASeg('dla60', classes, pretrained_base=pretrained_base, **kwargs) return model def dla102up(classes, pretrained_base=None, **kwargs): model = DLASeg('dla102', classes, pretrained_base=pretrained_base, **kwargs) return model def dla169up(classes, pretrained_base=None, **kwargs): model = DLASeg('dla169', classes, pretrained_base=pretrained_base, **kwargs) return model ''' def get_pose_net(num_layers, heads, add_conv=256, down_ratio=4): model = DLASeg('dla{}'.format(num_layers), heads, pretrained=True, down_ratio=down_ratio, head_conv=head_conv) return model
22,681
34.003086
86
py
houghnet
houghnet-master/src/lib/models/networks/hough_module.py
import torch import torch.nn as nn import numpy as np PI = np.pi class Hough(nn.Module): def __init__(self, angle=90, R2_list=[4, 64, 256, 1024], num_classes=80, region_num=9, vote_field_size=17, voting_map_size_w=128, voting_map_size_h=128, model_v1=False): super(Hough, self).__init__() self.angle = angle self.R2_list = R2_list self.region_num = region_num self.num_classes = num_classes self.vote_field_size = vote_field_size self.deconv_filter_padding = int(self.vote_field_size / 2) self.voting_map_size_w = voting_map_size_w self.voting_map_size_h = voting_map_size_h self.model_v1 = model_v1 self.deconv_filters = self._prepare_deconv_filters() def _prepare_deconv_filters(self): half_w = int(self.voting_map_size_w / 2) half_h = int(self.voting_map_size_h / 2) vote_center = torch.tensor([half_h, half_w]).cuda() logmap_onehot = self.calculate_logmap((self.voting_map_size_h, self.voting_map_size_w), vote_center) weights = logmap_onehot / \ torch.clamp(torch.sum(torch.sum(logmap_onehot, dim=0), dim=0).float(), min=1.0) start_x = half_h - int(self.vote_field_size/2) stop_x = half_h + int(self.vote_field_size/2) + 1 start_y = half_w - int(self.vote_field_size/2) stop_y = half_w + int(self.vote_field_size/2) + 1 '''This if-block only applies for my two pretrained models. Please ignore this for your own trainings.''' if self.model_v1 and self.region_num==17 and self.vote_field_size==65: start_x -=1 stop_x -=1 start_y -=1 stop_y -=1 deconv_filters = weights[start_x:stop_x, start_y:stop_y,:].permute(2,0,1).view(self.region_num, 1, self.vote_field_size, self.vote_field_size) W = nn.Parameter(deconv_filters.repeat(self.num_classes, 1, 1, 1)) W.requires_grad = False layers = [] deconv_kernel = nn.ConvTranspose2d( in_channels=self.region_num*self.num_classes, out_channels=1*self.num_classes, kernel_size=self.vote_field_size, padding=self.deconv_filter_padding, groups=self.num_classes, bias=False) with torch.no_grad(): deconv_kernel.weight = W layers.append(deconv_kernel) return nn.Sequential(*layers) def generate_grid(self, h, w): x = torch.arange(0, w).float().cuda() y = torch.arange(0, h).float().cuda() grid = torch.stack([x.repeat(h), y.repeat(w, 1).t().contiguous().view(-1)], 1) return grid.repeat(1, 1).view(-1, 2) def calculate_logmap(self, im_size, center, angle=90, R2_list=[4, 64, 256, 1024]): points = self.generate_grid(im_size[0], im_size[1]) # [x,y] total_angles = 360 / angle # check inside which circle y_dif = points[:, 1].cuda() - center[0].float() x_dif = points[:, 0].cuda() - center[1].float() xdif_2 = x_dif * x_dif ydif_2 = y_dif * y_dif sum_of_squares = xdif_2 + ydif_2 # find angle arc_angle = (torch.atan2(y_dif, x_dif) * 180 / PI).long() arc_angle[arc_angle < 0] += 360 angle_id = (arc_angle / angle).long() + 1 c_region = torch.ones(xdif_2.shape, dtype=torch.long).cuda() * len(R2_list) for i in range(len(R2_list) - 1, -1, -1): region = R2_list[i] c_region[(sum_of_squares) <= region] = i results = angle_id + (c_region - 1) * total_angles results[results < 0] = 0 results.view(im_size[0], im_size[1]) logmap = results.view(im_size[0], im_size[1]) logmap_onehot = torch.nn.functional.one_hot(logmap.long(), num_classes=17).float() logmap_onehot = logmap_onehot[:, :, :self.region_num] return logmap_onehot def forward(self, voting_map, targets=None): if self.model_v1: batch_size, channels, width, height = voting_map.shape voting_map = voting_map.view(batch_size, self.region_num, self.num_classes, width, height) voting_map = voting_map.permute(0, 2, 1, 3, 4) voting_map = voting_map.reshape(batch_size, -1, width, height) heatmap = self.deconv_filters(voting_map) return heatmap
4,491
34.370079
113
py
houghnet
houghnet-master/src/lib/models/networks/DCNv2/test.py
#!/usr/bin/env python from __future__ import absolute_import from __future__ import print_function from __future__ import division import time import torch import torch.nn as nn from torch.autograd import gradcheck from dcn_v2 import dcn_v2_conv, DCNv2, DCN from dcn_v2 import dcn_v2_pooling, DCNv2Pooling, DCNPooling deformable_groups = 1 N, inC, inH, inW = 2, 2, 4, 4 outC = 2 kH, kW = 3, 3 def conv_identify(weight, bias): weight.data.zero_() bias.data.zero_() o, i, h, w = weight.shape y = h//2 x = w//2 for p in range(i): for q in range(o): if p == q: weight.data[q, p, y, x] = 1.0 def check_zero_offset(): conv_offset = nn.Conv2d(inC, deformable_groups * 2 * kH * kW, kernel_size=(kH, kW), stride=(1, 1), padding=(1, 1), bias=True).cuda() conv_mask = nn.Conv2d(inC, deformable_groups * 1 * kH * kW, kernel_size=(kH, kW), stride=(1, 1), padding=(1, 1), bias=True).cuda() dcn_v2 = DCNv2(inC, outC, (kH, kW), stride=1, padding=1, dilation=1, deformable_groups=deformable_groups).cuda() conv_offset.weight.data.zero_() conv_offset.bias.data.zero_() conv_mask.weight.data.zero_() conv_mask.bias.data.zero_() conv_identify(dcn_v2.weight, dcn_v2.bias) input = torch.randn(N, inC, inH, inW).cuda() offset = conv_offset(input) mask = conv_mask(input) mask = torch.sigmoid(mask) output = dcn_v2(input, offset, mask) output *= 2 d = (input - output).abs().max() if d < 1e-10: print('Zero offset passed') else: print('Zero offset failed') print(input) print(output) def check_gradient_dconv(): input = torch.rand(N, inC, inH, inW).cuda() * 0.01 input.requires_grad = True offset = torch.randn(N, deformable_groups * 2 * kW * kH, inH, inW).cuda() * 2 # offset.data.zero_() # offset.data -= 0.5 offset.requires_grad = True mask = torch.rand(N, deformable_groups * 1 * kW * kH, inH, inW).cuda() # mask.data.zero_() mask.requires_grad = True mask = torch.sigmoid(mask) weight = torch.randn(outC, inC, kH, kW).cuda() weight.requires_grad = True bias = torch.rand(outC).cuda() bias.requires_grad = True stride = 1 padding = 1 dilation = 1 print('check_gradient_dconv: ', gradcheck(dcn_v2_conv, (input, offset, mask, weight, bias, stride, padding, dilation, deformable_groups), eps=1e-3, atol=1e-4, rtol=1e-2)) def check_pooling_zero_offset(): input = torch.randn(2, 16, 64, 64).cuda().zero_() input[0, :, 16:26, 16:26] = 1. input[1, :, 10:20, 20:30] = 2. rois = torch.tensor([ [0, 65, 65, 103, 103], [1, 81, 41, 119, 79], ]).cuda().float() pooling = DCNv2Pooling(spatial_scale=1.0 / 4, pooled_size=7, output_dim=16, no_trans=True, group_size=1, trans_std=0.0).cuda() out = pooling(input, rois, input.new()) s = ', '.join(['%f' % out[i, :, :, :].mean().item() for i in range(rois.shape[0])]) print(s) dpooling = DCNv2Pooling(spatial_scale=1.0 / 4, pooled_size=7, output_dim=16, no_trans=False, group_size=1, trans_std=0.0).cuda() offset = torch.randn(20, 2, 7, 7).cuda().zero_() dout = dpooling(input, rois, offset) s = ', '.join(['%f' % dout[i, :, :, :].mean().item() for i in range(rois.shape[0])]) print(s) def check_gradient_dpooling(): input = torch.randn(2, 3, 5, 5).cuda() * 0.01 N = 4 batch_inds = torch.randint(2, (N, 1)).cuda().float() x = torch.rand((N, 1)).cuda().float() * 15 y = torch.rand((N, 1)).cuda().float() * 15 w = torch.rand((N, 1)).cuda().float() * 10 h = torch.rand((N, 1)).cuda().float() * 10 rois = torch.cat((batch_inds, x, y, x + w, y + h), dim=1) offset = torch.randn(N, 2, 3, 3).cuda() input.requires_grad = True offset.requires_grad = True spatial_scale = 1.0 / 4 pooled_size = 3 output_dim = 3 no_trans = 0 group_size = 1 trans_std = 0.0 sample_per_part = 4 part_size = pooled_size print('check_gradient_dpooling:', gradcheck(dcn_v2_pooling, (input, rois, offset, spatial_scale, pooled_size, output_dim, no_trans, group_size, part_size, sample_per_part, trans_std), eps=1e-4)) def example_dconv(): input = torch.randn(2, 64, 128, 128).cuda() # wrap all things (offset and mask) in DCN dcn = DCN(64, 64, kernel_size=(3, 3), stride=1, padding=1, deformable_groups=2).cuda() # print(dcn.weight.shape, input.shape) output = dcn(input) targert = output.new(*output.size()) targert.data.uniform_(-0.01, 0.01) error = (targert - output).mean() error.backward() print(output.shape) def example_dpooling(): input = torch.randn(2, 32, 64, 64).cuda() batch_inds = torch.randint(2, (20, 1)).cuda().float() x = torch.randint(256, (20, 1)).cuda().float() y = torch.randint(256, (20, 1)).cuda().float() w = torch.randint(64, (20, 1)).cuda().float() h = torch.randint(64, (20, 1)).cuda().float() rois = torch.cat((batch_inds, x, y, x + w, y + h), dim=1) offset = torch.randn(20, 2, 7, 7).cuda() input.requires_grad = True offset.requires_grad = True # normal roi_align pooling = DCNv2Pooling(spatial_scale=1.0 / 4, pooled_size=7, output_dim=32, no_trans=True, group_size=1, trans_std=0.1).cuda() # deformable pooling dpooling = DCNv2Pooling(spatial_scale=1.0 / 4, pooled_size=7, output_dim=32, no_trans=False, group_size=1, trans_std=0.1).cuda() out = pooling(input, rois, offset) dout = dpooling(input, rois, offset) print(out.shape) print(dout.shape) target_out = out.new(*out.size()) target_out.data.uniform_(-0.01, 0.01) target_dout = dout.new(*dout.size()) target_dout.data.uniform_(-0.01, 0.01) e = (target_out - out).mean() e.backward() e = (target_dout - dout).mean() e.backward() def example_mdpooling(): input = torch.randn(2, 32, 64, 64).cuda() input.requires_grad = True batch_inds = torch.randint(2, (20, 1)).cuda().float() x = torch.randint(256, (20, 1)).cuda().float() y = torch.randint(256, (20, 1)).cuda().float() w = torch.randint(64, (20, 1)).cuda().float() h = torch.randint(64, (20, 1)).cuda().float() rois = torch.cat((batch_inds, x, y, x + w, y + h), dim=1) # mdformable pooling (V2) dpooling = DCNPooling(spatial_scale=1.0 / 4, pooled_size=7, output_dim=32, no_trans=False, group_size=1, trans_std=0.1, deform_fc_dim=1024).cuda() dout = dpooling(input, rois) target = dout.new(*dout.size()) target.data.uniform_(-0.1, 0.1) error = (target - dout).mean() error.backward() print(dout.shape) if __name__ == '__main__': example_dconv() example_dpooling() example_mdpooling() check_pooling_zero_offset() # zero offset check if inC == outC: check_zero_offset() check_gradient_dpooling() check_gradient_dconv() # """ # ****** Note: backward is not reentrant error may not be a serious problem, # ****** since the max error is less than 1e-7, # ****** Still looking for what trigger this problem # """
8,506
30.391144
81
py
houghnet
houghnet-master/src/lib/models/networks/DCNv2/setup.py
#!/usr/bin/env python import os import glob import torch from torch.utils.cpp_extension import CUDA_HOME from torch.utils.cpp_extension import CppExtension from torch.utils.cpp_extension import CUDAExtension from setuptools import find_packages from setuptools import setup requirements = ["torch", "torchvision"] def get_extensions(): this_dir = os.path.dirname(os.path.abspath(__file__)) extensions_dir = os.path.join(this_dir, "src") main_file = glob.glob(os.path.join(extensions_dir, "*.cpp")) source_cpu = glob.glob(os.path.join(extensions_dir, "cpu", "*.cpp")) source_cuda = glob.glob(os.path.join(extensions_dir, "cuda", "*.cu")) sources = main_file + source_cpu extension = CppExtension extra_compile_args = {"cxx": []} define_macros = [] if torch.cuda.is_available() and CUDA_HOME is not None: extension = CUDAExtension sources += source_cuda define_macros += [("WITH_CUDA", None)] extra_compile_args["nvcc"] = [ "-DCUDA_HAS_FP16=1", "-D__CUDA_NO_HALF_OPERATORS__", "-D__CUDA_NO_HALF_CONVERSIONS__", "-D__CUDA_NO_HALF2_OPERATORS__", ] else: raise NotImplementedError('Cuda is not availabel') sources = [os.path.join(extensions_dir, s) for s in sources] include_dirs = [extensions_dir] ext_modules = [ extension( "_ext", sources, include_dirs=include_dirs, define_macros=define_macros, extra_compile_args=extra_compile_args, ) ] return ext_modules setup( name="DCNv2", version="0.1", author="charlesshang", url="https://github.com/charlesshang/DCNv2", description="deformable convolutional networks", packages=find_packages(exclude=("configs", "tests",)), # install_requires=requirements, ext_modules=get_extensions(), cmdclass={"build_ext": torch.utils.cpp_extension.BuildExtension}, )
1,977
28.969697
73
py
houghnet
houghnet-master/src/lib/models/networks/DCNv2/dcn_v2.py
#!/usr/bin/env python from __future__ import absolute_import from __future__ import print_function from __future__ import division import math import torch from torch import nn from torch.autograd import Function from torch.nn.modules.utils import _pair from torch.autograd.function import once_differentiable import _ext as _backend class _DCNv2(Function): @staticmethod def forward(ctx, input, offset, mask, weight, bias, stride, padding, dilation, deformable_groups): ctx.stride = _pair(stride) ctx.padding = _pair(padding) ctx.dilation = _pair(dilation) ctx.kernel_size = _pair(weight.shape[2:4]) ctx.deformable_groups = deformable_groups output = _backend.dcn_v2_forward(input, weight, bias, offset, mask, ctx.kernel_size[0], ctx.kernel_size[1], ctx.stride[0], ctx.stride[1], ctx.padding[0], ctx.padding[1], ctx.dilation[0], ctx.dilation[1], ctx.deformable_groups) ctx.save_for_backward(input, offset, mask, weight, bias) return output @staticmethod @once_differentiable def backward(ctx, grad_output): input, offset, mask, weight, bias = ctx.saved_tensors grad_input, grad_offset, grad_mask, grad_weight, grad_bias = \ _backend.dcn_v2_backward(input, weight, bias, offset, mask, grad_output, ctx.kernel_size[0], ctx.kernel_size[1], ctx.stride[0], ctx.stride[1], ctx.padding[0], ctx.padding[1], ctx.dilation[0], ctx.dilation[1], ctx.deformable_groups) return grad_input, grad_offset, grad_mask, grad_weight, grad_bias,\ None, None, None, None, dcn_v2_conv = _DCNv2.apply class DCNv2(nn.Module): def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, deformable_groups=1): super(DCNv2, self).__init__() self.in_channels = in_channels self.out_channels = out_channels self.kernel_size = _pair(kernel_size) self.stride = _pair(stride) self.padding = _pair(padding) self.dilation = _pair(dilation) self.deformable_groups = deformable_groups self.weight = nn.Parameter(torch.Tensor( out_channels, in_channels, *self.kernel_size)) self.bias = nn.Parameter(torch.Tensor(out_channels)) self.reset_parameters() def reset_parameters(self): n = self.in_channels for k in self.kernel_size: n *= k stdv = 1. / math.sqrt(n) self.weight.data.uniform_(-stdv, stdv) self.bias.data.zero_() def forward(self, input, offset, mask): assert 2 * self.deformable_groups * self.kernel_size[0] * self.kernel_size[1] == \ offset.shape[1] assert self.deformable_groups * self.kernel_size[0] * self.kernel_size[1] == \ mask.shape[1] return dcn_v2_conv(input, offset, mask, self.weight, self.bias, self.stride, self.padding, self.dilation, self.deformable_groups) class DCN(DCNv2): def __init__(self, in_channels, out_channels, kernel_size, stride, padding, dilation=1, deformable_groups=1): super(DCN, self).__init__(in_channels, out_channels, kernel_size, stride, padding, dilation, deformable_groups) channels_ = self.deformable_groups * 3 * self.kernel_size[0] * self.kernel_size[1] self.conv_offset_mask = nn.Conv2d(self.in_channels, channels_, kernel_size=self.kernel_size, stride=self.stride, padding=self.padding, bias=True) self.init_offset() def init_offset(self): self.conv_offset_mask.weight.data.zero_() self.conv_offset_mask.bias.data.zero_() def forward(self, input): out = self.conv_offset_mask(input) o1, o2, mask = torch.chunk(out, 3, dim=1) offset = torch.cat((o1, o2), dim=1) mask = torch.sigmoid(mask) return dcn_v2_conv(input, offset, mask, self.weight, self.bias, self.stride, self.padding, self.dilation, self.deformable_groups) class _DCNv2Pooling(Function): @staticmethod def forward(ctx, input, rois, offset, spatial_scale, pooled_size, output_dim, no_trans, group_size=1, part_size=None, sample_per_part=4, trans_std=.0): ctx.spatial_scale = spatial_scale ctx.no_trans = int(no_trans) ctx.output_dim = output_dim ctx.group_size = group_size ctx.pooled_size = pooled_size ctx.part_size = pooled_size if part_size is None else part_size ctx.sample_per_part = sample_per_part ctx.trans_std = trans_std output, output_count = \ _backend.dcn_v2_psroi_pooling_forward(input, rois, offset, ctx.no_trans, ctx.spatial_scale, ctx.output_dim, ctx.group_size, ctx.pooled_size, ctx.part_size, ctx.sample_per_part, ctx.trans_std) ctx.save_for_backward(input, rois, offset, output_count) return output @staticmethod @once_differentiable def backward(ctx, grad_output): input, rois, offset, output_count = ctx.saved_tensors grad_input, grad_offset = \ _backend.dcn_v2_psroi_pooling_backward(grad_output, input, rois, offset, output_count, ctx.no_trans, ctx.spatial_scale, ctx.output_dim, ctx.group_size, ctx.pooled_size, ctx.part_size, ctx.sample_per_part, ctx.trans_std) return grad_input, None, grad_offset, \ None, None, None, None, None, None, None, None dcn_v2_pooling = _DCNv2Pooling.apply class DCNv2Pooling(nn.Module): def __init__(self, spatial_scale, pooled_size, output_dim, no_trans, group_size=1, part_size=None, sample_per_part=4, trans_std=.0): super(DCNv2Pooling, self).__init__() self.spatial_scale = spatial_scale self.pooled_size = pooled_size self.output_dim = output_dim self.no_trans = no_trans self.group_size = group_size self.part_size = pooled_size if part_size is None else part_size self.sample_per_part = sample_per_part self.trans_std = trans_std def forward(self, input, rois, offset): assert input.shape[1] == self.output_dim if self.no_trans: offset = input.new() return dcn_v2_pooling(input, rois, offset, self.spatial_scale, self.pooled_size, self.output_dim, self.no_trans, self.group_size, self.part_size, self.sample_per_part, self.trans_std) class DCNPooling(DCNv2Pooling): def __init__(self, spatial_scale, pooled_size, output_dim, no_trans, group_size=1, part_size=None, sample_per_part=4, trans_std=.0, deform_fc_dim=1024): super(DCNPooling, self).__init__(spatial_scale, pooled_size, output_dim, no_trans, group_size, part_size, sample_per_part, trans_std) self.deform_fc_dim = deform_fc_dim if not no_trans: self.offset_mask_fc = nn.Sequential( nn.Linear(self.pooled_size * self.pooled_size * self.output_dim, self.deform_fc_dim), nn.ReLU(inplace=True), nn.Linear(self.deform_fc_dim, self.deform_fc_dim), nn.ReLU(inplace=True), nn.Linear(self.deform_fc_dim, self.pooled_size * self.pooled_size * 3) ) self.offset_mask_fc[4].weight.data.zero_() self.offset_mask_fc[4].bias.data.zero_() def forward(self, input, rois): offset = input.new() if not self.no_trans: # do roi_align first n = rois.shape[0] roi = dcn_v2_pooling(input, rois, offset, self.spatial_scale, self.pooled_size, self.output_dim, True, # no trans self.group_size, self.part_size, self.sample_per_part, self.trans_std) # build mask and offset offset_mask = self.offset_mask_fc(roi.view(n, -1)) offset_mask = offset_mask.view( n, 3, self.pooled_size, self.pooled_size) o1, o2, mask = torch.chunk(offset_mask, 3, dim=1) offset = torch.cat((o1, o2), dim=1) mask = torch.sigmoid(mask) # do pooling with offset and mask return dcn_v2_pooling(input, rois, offset, self.spatial_scale, self.pooled_size, self.output_dim, self.no_trans, self.group_size, self.part_size, self.sample_per_part, self.trans_std) * mask # only roi_align return dcn_v2_pooling(input, rois, offset, self.spatial_scale, self.pooled_size, self.output_dim, self.no_trans, self.group_size, self.part_size, self.sample_per_part, self.trans_std)
12,081
38.743421
92
py
houghnet
houghnet-master/src/lib/trains/exdet.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import numpy as np import cv2 import sys import time from src.lib.utils.debugger import Debugger from src.lib.models.data_parallel import DataParallel from src.lib.models.losses import FocalLoss, RegL1Loss from src.lib.models.decode import agnex_ct_decode, exct_decode from src.lib.models.utils import _sigmoid from .base_trainer import BaseTrainer class ExdetLoss(torch.nn.Module): def __init__(self, opt): super(ExdetLoss, self).__init__() self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss() self.crit_reg = RegL1Loss() self.opt = opt self.parts = ['t', 'l', 'b', 'r', 'c'] def forward(self, outputs, batch): opt = self.opt hm_loss, reg_loss = 0, 0 for s in range(opt.num_stacks): output = outputs[s] for p in self.parts: tag = 'hm_{}'.format(p) output[tag] = _sigmoid(output[tag]) hm_loss += self.crit(output[tag], batch[tag]) / opt.num_stacks if p != 'c' and opt.reg_offset and opt.off_weight > 0: reg_loss += self.crit_reg(output['reg_{}'.format(p)], batch['reg_mask'], batch['ind_{}'.format(p)], batch['reg_{}'.format(p)]) / opt.num_stacks loss = opt.hm_weight * hm_loss + opt.off_weight * reg_loss loss_stats = {'loss': loss, 'off_loss': reg_loss, 'hm_loss': hm_loss} return loss, loss_stats class ExdetTrainer(BaseTrainer): def __init__(self, opt, model, optimizer=None): super(ExdetTrainer, self).__init__(opt, model, optimizer=optimizer) self.decode = agnex_ct_decode if opt.agnostic_ex else exct_decode def _get_losses(self, opt): loss_states = ['loss', 'hm_loss', 'off_loss'] loss = ExdetLoss(opt) return loss_states, loss def debug(self, batch, output, iter_id): opt = self.opt detections = self.decode(output['hm_t'], output['hm_l'], output['hm_b'], output['hm_r'], output['hm_c']).detach().cpu().numpy() detections[:, :, :4] *= opt.input_res / opt.output_res for i in range(1): debugger = Debugger( dataset=opt.dataset, ipynb=(opt.debug==3), theme=opt.debugger_theme) pred_hm = np.zeros((opt.input_res, opt.input_res, 3), dtype=np.uint8) gt_hm = np.zeros((opt.input_res, opt.input_res, 3), dtype=np.uint8) img = batch['input'][i].detach().cpu().numpy().transpose(1, 2, 0) img = ((img * self.opt.std + self.opt.mean) * 255.).astype(np.uint8) for p in self.parts: tag = 'hm_{}'.format(p) pred = debugger.gen_colormap(output[tag][i].detach().cpu().numpy()) gt = debugger.gen_colormap(batch[tag][i].detach().cpu().numpy()) if p != 'c': pred_hm = np.maximum(pred_hm, pred) gt_hm = np.maximum(gt_hm, gt) if p == 'c' or opt.debug > 2: debugger.add_blend_img(img, pred, 'pred_{}'.format(p)) debugger.add_blend_img(img, gt, 'gt_{}'.format(p)) debugger.add_blend_img(img, pred_hm, 'pred') debugger.add_blend_img(img, gt_hm, 'gt') debugger.add_img(img, img_id='out') for k in range(len(detections[i])): if detections[i, k, 4] > 0.1: debugger.add_coco_bbox(detections[i, k, :4], detections[i, k, -1], detections[i, k, 4], img_id='out') if opt.debug == 4: debugger.save_all_imgs(opt.debug_dir, prefix='{}'.format(iter_id)) else: debugger.show_all_imgs(pause=True)
3,645
41.395349
79
py
houghnet
houghnet-master/src/lib/trains/ctdet.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import numpy as np from src.lib.models.losses import FocalLoss from src.lib.models.losses import RegL1Loss, RegLoss, NormRegL1Loss, RegWeightedL1Loss from src.lib.models.decode import ctdet_decode from src.lib.models.utils import _sigmoid from src.lib.utils.debugger import Debugger from src.lib.utils.post_process import ctdet_post_process from src.lib.utils.oracle_utils import gen_oracle_map from .base_trainer import BaseTrainer class CtdetLoss(torch.nn.Module): def __init__(self, opt): super(CtdetLoss, self).__init__() self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss() self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \ RegLoss() if opt.reg_loss == 'sl1' else None self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \ NormRegL1Loss() if opt.norm_wh else \ RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg self.opt = opt def forward(self, outputs, batch): opt = self.opt hm_loss, wh_loss, off_loss = 0, 0, 0 for s in range(opt.num_stacks): output = outputs[s] if not opt.mse_loss: output['hm'] = _sigmoid(output['hm']) if opt.eval_oracle_hm: output['hm'] = batch['hm'] if opt.eval_oracle_wh: output['wh'] = torch.from_numpy(gen_oracle_map( batch['wh'].detach().cpu().numpy(), batch['ind'].detach().cpu().numpy(), output['wh'].shape[3], output['wh'].shape[2])).to(opt.device) if opt.eval_oracle_offset: output['reg'] = torch.from_numpy(gen_oracle_map( batch['reg'].detach().cpu().numpy(), batch['ind'].detach().cpu().numpy(), output['reg'].shape[3], output['reg'].shape[2])).to(opt.device) hm_loss += self.crit(output['hm'], batch['hm']) / opt.num_stacks if opt.wh_weight > 0: if opt.dense_wh: mask_weight = batch['dense_wh_mask'].sum() + 1e-4 wh_loss += ( self.crit_wh(output['wh'] * batch['dense_wh_mask'], batch['dense_wh'] * batch['dense_wh_mask']) / mask_weight) / opt.num_stacks elif opt.cat_spec_wh: wh_loss += self.crit_wh( output['wh'], batch['cat_spec_mask'], batch['ind'], batch['cat_spec_wh']) / opt.num_stacks else: wh_loss += self.crit_reg( output['wh'], batch['reg_mask'], batch['ind'], batch['wh']) / opt.num_stacks if opt.reg_offset and opt.off_weight > 0: off_loss += self.crit_reg(output['reg'], batch['reg_mask'], batch['ind'], batch['reg']) / opt.num_stacks loss = opt.hm_weight * hm_loss + opt.wh_weight * wh_loss + \ opt.off_weight * off_loss loss_stats = {'loss': loss, 'hm_loss': hm_loss, 'wh_loss': wh_loss, 'off_loss': off_loss} return loss, loss_stats class CtdetTrainer(BaseTrainer): def __init__(self, opt, model, optimizer=None): super(CtdetTrainer, self).__init__(opt, model, optimizer=optimizer) def _get_losses(self, opt): loss_states = ['loss', 'hm_loss', 'wh_loss', 'off_loss'] loss = CtdetLoss(opt) return loss_states, loss def debug(self, batch, output, iter_id): opt = self.opt reg = output['reg'] if opt.reg_offset else None dets = ctdet_decode( output['hm'], output['wh'], reg=reg, cat_spec_wh=opt.cat_spec_wh, K=opt.K) dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets[:, :, :4] *= opt.down_ratio dets_gt = batch['meta']['gt_det'].numpy().reshape(1, -1, dets.shape[2]) dets_gt[:, :, :4] *= opt.down_ratio for i in range(1): debugger = Debugger( dataset=opt.dataset, ipynb=(opt.debug==3), theme=opt.debugger_theme) img = batch['input'][i].detach().cpu().numpy().transpose(1, 2, 0) img = np.clip((( img * opt.std + opt.mean) * 255.), 0, 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][i].detach().cpu().numpy()) gt = debugger.gen_colormap(batch['hm'][i].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm') debugger.add_blend_img(img, gt, 'gt_hm') debugger.add_img(img, img_id='out_pred') for k in range(len(dets[i])): if dets[i, k, 4] > opt.center_thresh: debugger.add_coco_bbox(dets[i, k, :4], dets[i, k, -1], dets[i, k, 4], img_id='out_pred') debugger.add_img(img, img_id='out_gt') for k in range(len(dets_gt[i])): if dets_gt[i, k, 4] > opt.center_thresh: debugger.add_coco_bbox(dets_gt[i, k, :4], dets_gt[i, k, -1], dets_gt[i, k, 4], img_id='out_gt') if opt.debug == 4: debugger.save_all_imgs(opt.debug_dir, prefix='{}'.format(iter_id)) else: debugger.show_all_imgs(pause=True) def save_result(self, output, batch, results): reg = output['reg'] if self.opt.reg_offset else None dets = ctdet_decode( output['hm'], output['wh'], reg=reg, cat_spec_wh=self.opt.cat_spec_wh, K=self.opt.K) dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets_out = ctdet_post_process( dets.copy(), batch['meta']['c'].cpu().numpy(), batch['meta']['s'].cpu().numpy(), output['hm'].shape[2], output['hm'].shape[3], output['hm'].shape[1]) results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out[0]
5,574
41.234848
86
py
houghnet
houghnet-master/src/lib/trains/ddd.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import numpy as np from src.lib.models.losses import FocalLoss, L1Loss, BinRotLoss from src.lib.models.decode import ddd_decode from src.lib.models.utils import _sigmoid from src.lib.utils.debugger import Debugger from src.lib.utils.post_process import ddd_post_process from src.lib.utils.oracle_utils import gen_oracle_map from .base_trainer import BaseTrainer class DddLoss(torch.nn.Module): def __init__(self, opt): super(DddLoss, self).__init__() self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss() self.crit_reg = L1Loss() self.crit_rot = BinRotLoss() self.opt = opt def forward(self, outputs, batch): opt = self.opt hm_loss, dep_loss, rot_loss, dim_loss = 0, 0, 0, 0 wh_loss, off_loss = 0, 0 for s in range(opt.num_stacks): output = outputs[s] output['hm'] = _sigmoid(output['hm']) output['dep'] = 1. / (output['dep'].sigmoid() + 1e-6) - 1. if opt.eval_oracle_dep: output['dep'] = torch.from_numpy(gen_oracle_map( batch['dep'].detach().cpu().numpy(), batch['ind'].detach().cpu().numpy(), opt.output_w, opt.output_h)).to(opt.device) hm_loss += self.crit(output['hm'], batch['hm']) / opt.num_stacks if opt.dep_weight > 0: dep_loss += self.crit_reg(output['dep'], batch['reg_mask'], batch['ind'], batch['dep']) / opt.num_stacks if opt.dim_weight > 0: dim_loss += self.crit_reg(output['dim'], batch['reg_mask'], batch['ind'], batch['dim']) / opt.num_stacks if opt.rot_weight > 0: rot_loss += self.crit_rot(output['rot'], batch['rot_mask'], batch['ind'], batch['rotbin'], batch['rotres']) / opt.num_stacks if opt.reg_bbox and opt.wh_weight > 0: wh_loss += self.crit_reg(output['wh'], batch['rot_mask'], batch['ind'], batch['wh']) / opt.num_stacks if opt.reg_offset and opt.off_weight > 0: off_loss += self.crit_reg(output['reg'], batch['rot_mask'], batch['ind'], batch['reg']) / opt.num_stacks loss = opt.hm_weight * hm_loss + opt.dep_weight * dep_loss + \ opt.dim_weight * dim_loss + opt.rot_weight * rot_loss + \ opt.wh_weight * wh_loss + opt.off_weight * off_loss loss_stats = {'loss': loss, 'hm_loss': hm_loss, 'dep_loss': dep_loss, 'dim_loss': dim_loss, 'rot_loss': rot_loss, 'wh_loss': wh_loss, 'off_loss': off_loss} return loss, loss_stats class DddTrainer(BaseTrainer): def __init__(self, opt, model, optimizer=None): super(DddTrainer, self).__init__(opt, model, optimizer=optimizer) def _get_losses(self, opt): loss_states = ['loss', 'hm_loss', 'dep_loss', 'dim_loss', 'rot_loss', 'wh_loss', 'off_loss'] loss = DddLoss(opt) return loss_states, loss def debug(self, batch, output, iter_id): opt = self.opt wh = output['wh'] if opt.reg_bbox else None reg = output['reg'] if opt.reg_offset else None dets = ddd_decode(output['hm'], output['rot'], output['dep'], output['dim'], wh=wh, reg=reg, K=opt.K) # x, y, score, r1-r8, depth, dim1-dim3, cls dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) calib = batch['meta']['calib'].detach().numpy() # x, y, score, rot, depth, dim1, dim2, dim3 # if opt.dataset == 'gta': # dets[:, 12:15] /= 3 dets_pred = ddd_post_process( dets.copy(), batch['meta']['c'].detach().numpy(), batch['meta']['s'].detach().numpy(), calib, opt) dets_gt = ddd_post_process( batch['meta']['gt_det'].detach().numpy().copy(), batch['meta']['c'].detach().numpy(), batch['meta']['s'].detach().numpy(), calib, opt) #for i in range(input.size(0)): for i in range(1): debugger = Debugger(dataset=opt.dataset, ipynb=(opt.debug==3), theme=opt.debugger_theme) img = batch['input'][i].detach().cpu().numpy().transpose(1, 2, 0) img = ((img * self.opt.std + self.opt.mean) * 255.).astype(np.uint8) pred = debugger.gen_colormap( output['hm'][i].detach().cpu().numpy()) gt = debugger.gen_colormap(batch['hm'][i].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'hm_pred') debugger.add_blend_img(img, gt, 'hm_gt') # decode debugger.add_ct_detection( img, dets[i], show_box=opt.reg_bbox, center_thresh=opt.center_thresh, img_id='det_pred') debugger.add_ct_detection( img, batch['meta']['gt_det'][i].cpu().numpy().copy(), show_box=opt.reg_bbox, img_id='det_gt') debugger.add_3d_detection( batch['meta']['image_path'][i], dets_pred[i], calib[i], center_thresh=opt.center_thresh, img_id='add_pred') debugger.add_3d_detection( batch['meta']['image_path'][i], dets_gt[i], calib[i], center_thresh=opt.center_thresh, img_id='add_gt') # debugger.add_bird_view( # dets_pred[i], center_thresh=opt.center_thresh, img_id='bird_pred') # debugger.add_bird_view(dets_gt[i], img_id='bird_gt') debugger.add_bird_views( dets_pred[i], dets_gt[i], center_thresh=opt.center_thresh, img_id='bird_pred_gt') # debugger.add_blend_img(img, pred, 'out', white=True) debugger.compose_vis_add( batch['meta']['image_path'][i], dets_pred[i], calib[i], opt.center_thresh, pred, 'bird_pred_gt', img_id='out') # debugger.add_img(img, img_id='out') if opt.debug ==4: debugger.save_all_imgs(opt.debug_dir, prefix='{}'.format(iter_id)) else: debugger.show_all_imgs(pause=True) def save_result(self, output, batch, results): opt = self.opt wh = output['wh'] if opt.reg_bbox else None reg = output['reg'] if opt.reg_offset else None dets = ddd_decode(output['hm'], output['rot'], output['dep'], output['dim'], wh=wh, reg=reg, K=opt.K) # x, y, score, r1-r8, depth, dim1-dim3, cls dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) calib = batch['meta']['calib'].detach().numpy() # x, y, score, rot, depth, dim1, dim2, dim3 dets_pred = ddd_post_process( dets.copy(), batch['meta']['c'].detach().numpy(), batch['meta']['s'].detach().numpy(), calib, opt) img_id = batch['meta']['img_id'].detach().numpy()[0] results[img_id] = dets_pred[0] for j in range(1, opt.num_classes + 1): keep_inds = (results[img_id][j][:, -1] > opt.center_thresh) results[img_id][j] = results[img_id][j][keep_inds]
6,967
43.954839
80
py
houghnet
houghnet-master/src/lib/trains/multi_pose.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import numpy as np from src.lib.models.losses import FocalLoss, RegL1Loss, RegLoss, RegWeightedL1Loss from src.lib.models.decode import multi_pose_decode from src.lib.models.utils import _sigmoid, flip_tensor, flip_lr_off, flip_lr from src.lib.utils.debugger import Debugger from src.lib.utils.post_process import multi_pose_post_process from src.lib.utils.oracle_utils import gen_oracle_map from .base_trainer import BaseTrainer class MultiPoseLoss(torch.nn.Module): def __init__(self, opt): super(MultiPoseLoss, self).__init__() self.crit = FocalLoss() self.crit_hm_hp = torch.nn.MSELoss() if opt.mse_loss else FocalLoss() self.crit_kp = RegWeightedL1Loss() if not opt.dense_hp else \ torch.nn.L1Loss(reduction='sum') self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \ RegLoss() if opt.reg_loss == 'sl1' else None self.opt = opt def forward(self, outputs, batch): opt = self.opt hm_loss, wh_loss, off_loss = 0, 0, 0 hp_loss, off_loss, hm_hp_loss, hp_offset_loss = 0, 0, 0, 0 for s in range(opt.num_stacks): output = outputs[s] output['hm'] = _sigmoid(output['hm']) if opt.hm_hp and not opt.mse_loss: output['hm_hp'] = _sigmoid(output['hm_hp']) if opt.eval_oracle_hmhp: output['hm_hp'] = batch['hm_hp'] if opt.eval_oracle_hm: output['hm'] = batch['hm'] if opt.eval_oracle_kps: if opt.dense_hp: output['hps'] = batch['dense_hps'] else: output['hps'] = torch.from_numpy(gen_oracle_map( batch['hps'].detach().cpu().numpy(), batch['ind'].detach().cpu().numpy(), opt.output_res, opt.output_res)).to(opt.device) if opt.eval_oracle_hp_offset: output['hp_offset'] = torch.from_numpy(gen_oracle_map( batch['hp_offset'].detach().cpu().numpy(), batch['hp_ind'].detach().cpu().numpy(), opt.output_res, opt.output_res)).to(opt.device) hm_loss += self.crit(output['hm'], batch['hm']) / opt.num_stacks if opt.dense_hp: mask_weight = batch['dense_hps_mask'].sum() + 1e-4 hp_loss += (self.crit_kp(output['hps'] * batch['dense_hps_mask'], batch['dense_hps'] * batch['dense_hps_mask']) / mask_weight) / opt.num_stacks else: hp_loss += self.crit_kp(output['hps'], batch['hps_mask'], batch['ind'], batch['hps']) / opt.num_stacks if opt.wh_weight > 0: wh_loss += self.crit_reg(output['wh'], batch['reg_mask'], batch['ind'], batch['wh']) / opt.num_stacks if opt.reg_offset and opt.off_weight > 0: off_loss += self.crit_reg(output['reg'], batch['reg_mask'], batch['ind'], batch['reg']) / opt.num_stacks if opt.reg_hp_offset and opt.off_weight > 0: hp_offset_loss += self.crit_reg( output['hp_offset'], batch['hp_mask'], batch['hp_ind'], batch['hp_offset']) / opt.num_stacks if opt.hm_hp and opt.hm_hp_weight > 0: hm_hp_loss += self.crit_hm_hp( output['hm_hp'], batch['hm_hp']) / opt.num_stacks loss = opt.hm_weight * hm_loss + opt.wh_weight * wh_loss + \ opt.off_weight * off_loss + opt.hp_weight * hp_loss + \ opt.hm_hp_weight * hm_hp_loss + opt.off_weight * hp_offset_loss loss_stats = {'loss': loss, 'hm_loss': hm_loss, 'hp_loss': hp_loss, 'hm_hp_loss': hm_hp_loss, 'hp_offset_loss': hp_offset_loss, 'wh_loss': wh_loss, 'off_loss': off_loss} return loss, loss_stats class MultiPoseTrainer(BaseTrainer): def __init__(self, opt, model, optimizer=None): super(MultiPoseTrainer, self).__init__(opt, model, optimizer=optimizer) def _get_losses(self, opt): loss_states = ['loss', 'hm_loss', 'hp_loss', 'hm_hp_loss', 'hp_offset_loss', 'wh_loss', 'off_loss'] loss = MultiPoseLoss(opt) return loss_states, loss def debug(self, batch, output, iter_id): opt = self.opt reg = output['reg'] if opt.reg_offset else None hm_hp = output['hm_hp'] if opt.hm_hp else None hp_offset = output['hp_offset'] if opt.reg_hp_offset else None dets = multi_pose_decode( output['hm'], output['wh'], output['hps'], reg=reg, hm_hp=hm_hp, hp_offset=hp_offset, K=opt.K) dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets[:, :, :4] *= opt.input_res / opt.output_res dets[:, :, 5:39] *= opt.input_res / opt.output_res dets_gt = batch['meta']['gt_det'].numpy().reshape(1, -1, dets.shape[2]) dets_gt[:, :, :4] *= opt.input_res / opt.output_res dets_gt[:, :, 5:39] *= opt.input_res / opt.output_res for i in range(1): debugger = Debugger( dataset=opt.dataset, ipynb=(opt.debug==3), theme=opt.debugger_theme) img = batch['input'][i].detach().cpu().numpy().transpose(1, 2, 0) img = np.clip((( img * opt.std + opt.mean) * 255.), 0, 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][i].detach().cpu().numpy()) gt = debugger.gen_colormap(batch['hm'][i].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm') debugger.add_blend_img(img, gt, 'gt_hm') debugger.add_img(img, img_id='out_pred') for k in range(len(dets[i])): if dets[i, k, 4] > opt.center_thresh: debugger.add_coco_bbox(dets[i, k, :4], dets[i, k, -1], dets[i, k, 4], img_id='out_pred') debugger.add_coco_hp(dets[i, k, 5:39], img_id='out_pred') debugger.add_img(img, img_id='out_gt') for k in range(len(dets_gt[i])): if dets_gt[i, k, 4] > opt.center_thresh: debugger.add_coco_bbox(dets_gt[i, k, :4], dets_gt[i, k, -1], dets_gt[i, k, 4], img_id='out_gt') debugger.add_coco_hp(dets_gt[i, k, 5:39], img_id='out_gt') if opt.hm_hp: pred = debugger.gen_colormap_hp(output['hm_hp'][i].detach().cpu().numpy()) gt = debugger.gen_colormap_hp(batch['hm_hp'][i].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hmhp') debugger.add_blend_img(img, gt, 'gt_hmhp') if opt.debug == 4: debugger.save_all_imgs(opt.debug_dir, prefix='{}'.format(iter_id)) else: debugger.show_all_imgs(pause=True) def save_result(self, output, batch, results): reg = output['reg'] if self.opt.reg_offset else None hm_hp = output['hm_hp'] if self.opt.hm_hp else None hp_offset = output['hp_offset'] if self.opt.reg_hp_offset else None dets = multi_pose_decode( output['hm'], output['wh'], output['hps'], reg=reg, hm_hp=hm_hp, hp_offset=hp_offset, K=self.opt.K) dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets_out = multi_pose_post_process( dets.copy(), batch['meta']['c'].cpu().numpy(), batch['meta']['s'].cpu().numpy(), output['hm'].shape[2], output['hm'].shape[3]) results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out[0]
7,300
44.347826
82
py
houghnet
houghnet-master/src/lib/trains/base_trainer.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import time import torch from progress.bar import Bar from src.lib.models.data_parallel import DataParallel from src.lib.utils.utils import AverageMeter class ModleWithLoss(torch.nn.Module): def __init__(self, model, loss): super(ModleWithLoss, self).__init__() self.model = model self.loss = loss def forward(self, batch): outputs = self.model(batch['input']) loss, loss_stats = self.loss(outputs, batch) return outputs[-1], loss, loss_stats class BaseTrainer(object): def __init__( self, opt, model, optimizer=None): self.opt = opt self.optimizer = optimizer self.loss_stats, self.loss = self._get_losses(opt) self.model_with_loss = ModleWithLoss(model, self.loss) def set_device(self, gpus, chunk_sizes, device): if len(gpus) > 1: self.model_with_loss = DataParallel( self.model_with_loss, device_ids=gpus, chunk_sizes=chunk_sizes).to(device) else: self.model_with_loss = self.model_with_loss.to(device) for state in self.optimizer.state.values(): for k, v in state.items(): if isinstance(v, torch.Tensor): state[k] = v.to(device=device, non_blocking=True) def run_epoch(self, phase, epoch, data_loader): model_with_loss = self.model_with_loss if phase == 'train': model_with_loss.train() else: if len(self.opt.gpus) > 1: model_with_loss = self.model_with_loss.module model_with_loss.eval() torch.cuda.empty_cache() opt = self.opt results = {} data_time, batch_time = AverageMeter(), AverageMeter() avg_loss_stats = {l: AverageMeter() for l in self.loss_stats} num_iters = len(data_loader) if opt.num_iters < 0 else opt.num_iters bar = Bar('{}/{}'.format(opt.task, opt.exp_id), max=num_iters) end = time.time() for iter_id, batch in enumerate(data_loader): if iter_id >= num_iters: break data_time.update(time.time() - end) for k in batch: if k != 'meta': batch[k] = batch[k].to(device=opt.device, non_blocking=True) output, loss, loss_stats = model_with_loss(batch) loss = loss.mean() if phase == 'train': self.optimizer.zero_grad() loss.backward() self.optimizer.step() batch_time.update(time.time() - end) end = time.time() Bar.suffix = '{phase}: [{0}][{1}/{2}]|Tot: {total:} |ETA: {eta:} '.format( epoch, iter_id, num_iters, phase=phase, total=bar.elapsed_td, eta=bar.eta_td) for l in avg_loss_stats: avg_loss_stats[l].update( loss_stats[l].mean().item(), batch['input'].size(0)) Bar.suffix = Bar.suffix + '|{} {:.4f} '.format(l, avg_loss_stats[l].avg) if not opt.hide_data_time: Bar.suffix = Bar.suffix + '|Data {dt.val:.3f}s({dt.avg:.3f}s) ' \ '|Net {bt.avg:.3f}s'.format(dt=data_time, bt=batch_time) if opt.print_iter > 0: if iter_id % opt.print_iter == 0: print('{}/{}| {}'.format(opt.task, opt.exp_id, Bar.suffix)) else: bar.next() if opt.debug > 0: self.debug(batch, output, iter_id) if opt.test: self.save_result(output, batch, results) del output, loss, loss_stats bar.finish() ret = {k: v.avg for k, v in avg_loss_stats.items()} ret['time'] = bar.elapsed_td.total_seconds() / 60. return ret, results def debug(self, batch, output, iter_id): raise NotImplementedError def save_result(self, output, batch, results): raise NotImplementedError def _get_losses(self, opt): raise NotImplementedError def val(self, epoch, data_loader): return self.run_epoch('val', epoch, data_loader) def train(self, epoch, data_loader): return self.run_epoch('train', epoch, data_loader)
3,929
32.02521
80
py
houghnet
houghnet-master/src/lib/trains/ctseg.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch import numpy as np from src.lib.models.losses import FocalLoss,SegLoss from src.lib.models.losses import RegL1Loss, RegLoss, NormRegL1Loss, RegWeightedL1Loss from src.lib.models.decode import ctdet_decode from src.lib.models.utils import _sigmoid from src.lib.utils.debugger import Debugger from src.lib.utils.post_process import ctdet_post_process from src.lib.utils.oracle_utils import gen_oracle_map from .base_trainer import BaseTrainer class CtsegLoss(torch.nn.Module): def __init__(self, opt): super(CtsegLoss, self).__init__() self.crit = torch.nn.MSELoss() if opt.mse_loss else FocalLoss() self.crit_reg = RegL1Loss() if opt.reg_loss == 'l1' else \ RegLoss() if opt.reg_loss == 'sl1' else None self.crit_wh = torch.nn.L1Loss(reduction='sum') if opt.dense_wh else \ NormRegL1Loss() if opt.norm_wh else \ RegWeightedL1Loss() if opt.cat_spec_wh else self.crit_reg self.crit_mask = SegLoss(opt.seg_feat_channel) self.opt = opt def forward(self, outputs, batch): opt = self.opt hm_loss, wh_loss, off_loss, mask_loss = 0, 0, 0,0 for s in range(opt.num_stacks): output = outputs[s] if not opt.mse_loss: output['hm'] = _sigmoid(output['hm']) if opt.eval_oracle_hm: output['hm'] = batch['hm'] if opt.eval_oracle_wh: output['wh'] = torch.from_numpy(gen_oracle_map( batch['wh'].detach().cpu().numpy(), batch['ind'].detach().cpu().numpy(), output['wh'].shape[3], output['wh'].shape[2])).to(opt.device) if opt.eval_oracle_offset: output['reg'] = torch.from_numpy(gen_oracle_map( batch['reg'].detach().cpu().numpy(), batch['ind'].detach().cpu().numpy(), output['reg'].shape[3], output['reg'].shape[2])).to(opt.device) hm_loss += self.crit(output['hm'], batch['hm']) / opt.num_stacks if opt.wh_weight > 0: if opt.dense_wh: mask_weight = batch['dense_wh_mask'].sum() + 1e-4 wh_loss += ( self.crit_wh(output['wh'] * batch['dense_wh_mask'], batch['dense_wh'] * batch['dense_wh_mask']) / mask_weight) / opt.num_stacks elif opt.cat_spec_wh: wh_loss += self.crit_wh( output['wh'], batch['cat_spec_mask'], batch['ind'], batch['cat_spec_wh']) / opt.num_stacks else: wh_loss += self.crit_reg( output['wh'], batch['reg_mask'], batch['ind'], batch['wh']) / opt.num_stacks if opt.reg_offset and opt.off_weight > 0: off_loss += self.crit_reg(output['reg'], batch['reg_mask'], batch['ind'], batch['reg']) / opt.num_stacks mask_loss+=self.crit_mask(output['saliency'], output['shape'], batch['gtboxes'], batch['reg_mask'], batch['ind'], batch['instance_mask'], output['hm'], batch['cat_spec_mask']) loss = opt.hm_weight * hm_loss + opt.wh_weight * wh_loss + \ opt.off_weight * off_loss + opt.seg_weight * mask_loss loss_stats = {'loss': loss, 'hm_loss': hm_loss, 'wh_loss': wh_loss, 'off_loss': off_loss,"mask_loss":mask_loss} return loss, loss_stats class CtsegTrainer(BaseTrainer): def __init__(self, opt, model, optimizer=None): super(CtsegTrainer, self).__init__(opt, model, optimizer=optimizer) def _get_losses(self, opt): loss_states = ['loss', 'hm_loss', 'wh_loss', 'off_loss','mask_loss'] loss = CtsegLoss(opt) return loss_states, loss def debug(self, batch, output, iter_id): opt = self.opt reg = output['reg'] if opt.reg_offset else None dets = ctdet_decode( output['hm'], output['wh'], reg=reg, cat_spec_wh=opt.cat_spec_wh, K=opt.K) dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets[:, :, :4] *= opt.down_ratio dets_gt = batch['meta']['gt_det'].numpy().reshape(1, -1, dets.shape[2]) dets_gt[:, :, :4] *= opt.down_ratio for i in range(1): debugger = Debugger( dataset=opt.dataset, ipynb=(opt.debug == 3), theme=opt.debugger_theme) img = batch['input'][i].detach().cpu().numpy().transpose(1, 2, 0) img = np.clip((( img * opt.std + opt.mean) * 255.), 0, 255).astype(np.uint8) pred = debugger.gen_colormap(output['hm'][i].detach().cpu().numpy()) gt = debugger.gen_colormap(batch['hm'][i].detach().cpu().numpy()) debugger.add_blend_img(img, pred, 'pred_hm') debugger.add_blend_img(img, gt, 'gt_hm') debugger.add_img(img, img_id='out_pred') for k in range(len(dets[i])): if dets[i, k, 4] > opt.center_thresh: debugger.add_coco_bbox(dets[i, k, :4], dets[i, k, -1], dets[i, k, 4], img_id='out_pred') debugger.add_img(img, img_id='out_gt') for k in range(len(dets_gt[i])): if dets_gt[i, k, 4] > opt.center_thresh: debugger.add_coco_bbox(dets_gt[i, k, :4], dets_gt[i, k, -1], dets_gt[i, k, 4], img_id='out_gt') if opt.debug == 4: debugger.save_all_imgs(opt.debug_dir, prefix='{}'.format(iter_id)) else: debugger.show_all_imgs(pause=True) def save_result(self, output, batch, results): reg = output['reg'] if self.opt.reg_offset else None dets = ctdet_decode( output['hm'], output['wh'], reg=reg, cat_spec_wh=self.opt.cat_spec_wh, K=self.opt.K) dets = dets.detach().cpu().numpy().reshape(1, -1, dets.shape[2]) dets_out = ctdet_post_process( dets.copy(), batch['meta']['c'].cpu().numpy(), batch['meta']['s'].cpu().numpy(), output['hm'].shape[2], output['hm'].shape[3], output['hm'].shape[1]) results[batch['meta']['img_id'].cpu().numpy()[0]] = dets_out[0]
6,685
47.100719
97
py
houghnet
houghnet-master/src/lib/datasets/sample/exdet.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch.utils.data as data import pycocotools.coco as coco import numpy as np import torch import json import cv2 import os from utils.image import flip, color_aug from utils.image import get_affine_transform, affine_transform from utils.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian import pycocotools.coco as coco import math class EXDetDataset(data.Dataset): def _coco_box_to_bbox(self, box): bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32) return bbox def _get_border(self, border, size): i = 1 while size - border // i <= border // i: i *= 2 return border // i def __getitem__(self, index): img_id = self.images[index] img_info = self.coco.loadImgs(ids=[img_id])[0] img_path = os.path.join(self.img_dir, img_info['file_name']) img = cv2.imread(img_path) height, width = img.shape[0], img.shape[1] c = np.array([img.shape[1] / 2., img.shape[0] / 2.]) s = max(img.shape[0], img.shape[1]) * 1.0 flipped = False if self.split == 'train': if not self.opt.not_rand_crop: s = s * np.random.choice(np.arange(0.6, 1.4, 0.1)) w_border = self._get_border(128, img.shape[1]) h_border = self._get_border(128, img.shape[0]) c[0] = np.random.randint(low=w_border, high=img.shape[1] - w_border) c[1] = np.random.randint(low=h_border, high=img.shape[0] - h_border) else: sf = self.opt.scale cf = self.opt.shift s = s * np.clip(np.random.randn()*sf + 1, 1 - sf, 1 + sf) c[0] += img.shape[1] * np.clip(np.random.randn()*cf, -2*cf, 2*cf) c[1] += img.shape[0] * np.clip(np.random.randn()*cf, -2*cf, 2*cf) if np.random.random() < self.opt.flip: flipped = True img = img[:, ::-1, :] trans_input = get_affine_transform( c, s, 0, [self.opt.input_res, self.opt.input_res]) inp = cv2.warpAffine(img, trans_input, (self.opt.input_res, self.opt.input_res), flags=cv2.INTER_LINEAR) inp = (inp.astype(np.float32) / 255.) if self.split == 'train' and not self.opt.no_color_aug: color_aug(self._data_rng, inp, self._eig_val, self._eig_vec) inp = (inp - self.mean) / self.std inp = inp.transpose(2, 0, 1) output_res = self.opt.output_res num_classes = self.opt.num_classes trans_output = get_affine_transform(c, s, 0, [output_res, output_res]) num_hm = 1 if self.opt.agnostic_ex else num_classes hm_t = np.zeros((num_hm, output_res, output_res), dtype=np.float32) hm_l = np.zeros((num_hm, output_res, output_res), dtype=np.float32) hm_b = np.zeros((num_hm, output_res, output_res), dtype=np.float32) hm_r = np.zeros((num_hm, output_res, output_res), dtype=np.float32) hm_c = np.zeros((num_classes, output_res, output_res), dtype=np.float32) reg_t = np.zeros((self.max_objs, 2), dtype=np.float32) reg_l = np.zeros((self.max_objs, 2), dtype=np.float32) reg_b = np.zeros((self.max_objs, 2), dtype=np.float32) reg_r = np.zeros((self.max_objs, 2), dtype=np.float32) ind_t = np.zeros((self.max_objs), dtype=np.int64) ind_l = np.zeros((self.max_objs), dtype=np.int64) ind_b = np.zeros((self.max_objs), dtype=np.int64) ind_r = np.zeros((self.max_objs), dtype=np.int64) reg_mask = np.zeros((self.max_objs), dtype=np.uint8) ann_ids = self.coco.getAnnIds(imgIds=[img_id]) anns = self.coco.loadAnns(ids=ann_ids) num_objs = min(len(anns), self.max_objs) draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \ draw_umich_gaussian for k in range(num_objs): ann = anns[k] # bbox = self._coco_box_to_bbox(ann['bbox']) # tlbr pts = np.array(ann['extreme_points'], dtype=np.float32).reshape(4, 2) # cls_id = int(self.cat_ids[ann['category_id']] - 1) # bug cls_id = int(self.cat_ids[ann['category_id']]) hm_id = 0 if self.opt.agnostic_ex else cls_id if flipped: pts[:, 0] = width - pts[:, 0] - 1 pts[1], pts[3] = pts[3].copy(), pts[1].copy() for j in range(4): pts[j] = affine_transform(pts[j], trans_output) pts = np.clip(pts, 0, self.opt.output_res - 1) h, w = pts[2, 1] - pts[0, 1], pts[3, 0] - pts[1, 0] if h > 0 and w > 0: radius = gaussian_radius((math.ceil(h), math.ceil(w))) radius = max(0, int(radius)) pt_int = pts.astype(np.int32) draw_gaussian(hm_t[hm_id], pt_int[0], radius) draw_gaussian(hm_l[hm_id], pt_int[1], radius) draw_gaussian(hm_b[hm_id], pt_int[2], radius) draw_gaussian(hm_r[hm_id], pt_int[3], radius) reg_t[k] = pts[0] - pt_int[0] reg_l[k] = pts[1] - pt_int[1] reg_b[k] = pts[2] - pt_int[2] reg_r[k] = pts[3] - pt_int[3] ind_t[k] = pt_int[0, 1] * output_res + pt_int[0, 0] ind_l[k] = pt_int[1, 1] * output_res + pt_int[1, 0] ind_b[k] = pt_int[2, 1] * output_res + pt_int[2, 0] ind_r[k] = pt_int[3, 1] * output_res + pt_int[3, 0] ct = [int((pts[3, 0] + pts[1, 0]) / 2), int((pts[0, 1] + pts[2, 1]) / 2)] draw_gaussian(hm_c[cls_id], ct, radius) reg_mask[k] = 1 ret = {'input': inp, 'hm_t': hm_t, 'hm_l': hm_l, 'hm_b': hm_b, 'hm_r': hm_r, 'hm_c': hm_c} if self.opt.reg_offset: ret.update({'reg_mask': reg_mask, 'reg_t': reg_t, 'reg_l': reg_l, 'reg_b': reg_b, 'reg_r': reg_r, 'ind_t': ind_t, 'ind_l': ind_l, 'ind_b': ind_b, 'ind_r': ind_r}) return ret
5,722
40.773723
81
py
houghnet
houghnet-master/src/lib/datasets/sample/ctdet.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch.utils.data as data import numpy as np import torch import json import cv2 import os from src.lib.utils.image import flip, color_aug from src.lib.utils.image import get_affine_transform, affine_transform from src.lib.utils.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian from src.lib.utils.image import draw_dense_reg import math class CTDetDataset(data.Dataset): def _coco_box_to_bbox(self, box): bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32) return bbox def _get_border(self, border, size): i = 1 while size - border // i <= border // i: i *= 2 return border // i def __getitem__(self, index): img_id = self.images[index] file_name = self.coco.loadImgs(ids=[img_id])[0]['file_name'] img_path = os.path.join(self.img_dir, file_name) ann_ids = self.coco.getAnnIds(imgIds=[img_id]) anns = self.coco.loadAnns(ids=ann_ids) num_objs = min(len(anns), self.max_objs) img = cv2.imread(img_path) height, width = img.shape[0], img.shape[1] c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32) if self.opt.keep_res: input_h = (height | self.opt.pad) + 1 input_w = (width | self.opt.pad) + 1 s = np.array([input_w, input_h], dtype=np.float32) else: s = max(img.shape[0], img.shape[1]) * 1.0 input_h, input_w = self.opt.input_h, self.opt.input_w flipped = False if self.split == 'train': if not self.opt.not_rand_crop: s = s * np.random.choice(np.arange(0.6, 1.4, 0.1)) w_border = self._get_border(128, img.shape[1]) h_border = self._get_border(128, img.shape[0]) c[0] = np.random.randint(low=w_border, high=img.shape[1] - w_border) c[1] = np.random.randint(low=h_border, high=img.shape[0] - h_border) else: sf = self.opt.scale cf = self.opt.shift c[0] += s * np.clip(np.random.randn()*cf, -2*cf, 2*cf) c[1] += s * np.clip(np.random.randn()*cf, -2*cf, 2*cf) s = s * np.clip(np.random.randn()*sf + 1, 1 - sf, 1 + sf) if np.random.random() < self.opt.flip: flipped = True img = img[:, ::-1, :] c[0] = width - c[0] - 1 trans_input = get_affine_transform( c, s, 0, [input_w, input_h]) inp = cv2.warpAffine(img, trans_input, (input_w, input_h), flags=cv2.INTER_LINEAR) inp = (inp.astype(np.float32) / 255.) if self.split == 'train' and not self.opt.no_color_aug: color_aug(self._data_rng, inp, self._eig_val, self._eig_vec) inp = (inp - self.mean) / self.std inp = inp.transpose(2, 0, 1) output_h = input_h // self.opt.down_ratio output_w = input_w // self.opt.down_ratio num_classes = self.num_classes trans_output = get_affine_transform(c, s, 0, [output_w, output_h]) hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32) wh = np.zeros((self.max_objs, 2), dtype=np.float32) dense_wh = np.zeros((2, output_h, output_w), dtype=np.float32) reg = np.zeros((self.max_objs, 2), dtype=np.float32) ind = np.zeros((self.max_objs), dtype=np.int64) reg_mask = np.zeros((self.max_objs), dtype=np.uint8) cat_spec_wh = np.zeros((self.max_objs, num_classes * 2), dtype=np.float32) cat_spec_mask = np.zeros((self.max_objs, num_classes * 2), dtype=np.uint8) draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \ draw_umich_gaussian gt_det = [] for k in range(num_objs): ann = anns[k] bbox = self._coco_box_to_bbox(ann['bbox']) cls_id = int(self.cat_ids[ann['category_id']]) if flipped: bbox[[0, 2]] = width - bbox[[2, 0]] - 1 bbox[:2] = affine_transform(bbox[:2], trans_output) bbox[2:] = affine_transform(bbox[2:], trans_output) bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, output_w - 1) bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, output_h - 1) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] if h > 0 and w > 0: radius = gaussian_radius((math.ceil(h), math.ceil(w))) radius = max(0, int(radius)) radius = self.opt.hm_gauss if self.opt.mse_loss else radius ct = np.array( [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) ct_int = ct.astype(np.int32) draw_gaussian(hm[cls_id], ct_int, radius) wh[k] = 1. * w, 1. * h ind[k] = ct_int[1] * output_w + ct_int[0] reg[k] = ct - ct_int reg_mask[k] = 1 cat_spec_wh[k, cls_id * 2: cls_id * 2 + 2] = wh[k] cat_spec_mask[k, cls_id * 2: cls_id * 2 + 2] = 1 if self.opt.dense_wh: draw_dense_reg(dense_wh, hm.max(axis=0), ct_int, wh[k], radius) gt_det.append([ct[0] - w / 2, ct[1] - h / 2, ct[0] + w / 2, ct[1] + h / 2, 1, cls_id]) ret = {'input': inp, 'hm': hm, 'reg_mask': reg_mask, 'ind': ind, 'wh': wh} if self.opt.dense_wh: hm_a = hm.max(axis=0, keepdims=True) dense_wh_mask = np.concatenate([hm_a, hm_a], axis=0) ret.update({'dense_wh': dense_wh, 'dense_wh_mask': dense_wh_mask}) del ret['wh'] elif self.opt.cat_spec_wh: ret.update({'cat_spec_wh': cat_spec_wh, 'cat_spec_mask': cat_spec_mask}) del ret['wh'] if self.opt.reg_offset: ret.update({'reg': reg}) if self.opt.debug > 0 or not self.split == 'train': gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \ np.zeros((1, 6), dtype=np.float32) meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id} ret['meta'] = meta return ret
5,835
39.248276
88
py
houghnet
houghnet-master/src/lib/datasets/sample/ddd.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch.utils.data as data import pycocotools.coco as coco import numpy as np import torch import json import cv2 import os import math from src.lib.utils.image import flip, color_aug from src.lib.utils.image import get_affine_transform, affine_transform from src.lib.utils.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian import pycocotools.coco as coco class DddDataset(data.Dataset): def _coco_box_to_bbox(self, box): bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32) return bbox def _convert_alpha(self, alpha): return math.radians(alpha + 45) if self.alpha_in_degree else alpha def __getitem__(self, index): img_id = self.images[index] img_info = self.coco.loadImgs(ids=[img_id])[0] img_path = os.path.join(self.img_dir, img_info['file_name']) img = cv2.imread(img_path) if 'calib' in img_info: calib = np.array(img_info['calib'], dtype=np.float32) else: calib = self.calib height, width = img.shape[0], img.shape[1] c = np.array([img.shape[1] / 2., img.shape[0] / 2.]) if self.opt.keep_res: s = np.array([self.opt.input_w, self.opt.input_h], dtype=np.int32) else: s = np.array([width, height], dtype=np.int32) aug = False if self.split == 'train' and np.random.random() < self.opt.aug_ddd: aug = True sf = self.opt.scale cf = self.opt.shift s = s * np.clip(np.random.randn()*sf + 1, 1 - sf, 1 + sf) c[0] += img.shape[1] * np.clip(np.random.randn()*cf, -2*cf, 2*cf) c[1] += img.shape[0] * np.clip(np.random.randn()*cf, -2*cf, 2*cf) trans_input = get_affine_transform( c, s, 0, [self.opt.input_w, self.opt.input_h]) inp = cv2.warpAffine(img, trans_input, (self.opt.input_w, self.opt.input_h), flags=cv2.INTER_LINEAR) inp = (inp.astype(np.float32) / 255.) # if self.split == 'train' and not self.opt.no_color_aug: # color_aug(self._data_rng, inp, self._eig_val, self._eig_vec) inp = (inp - self.mean) / self.std inp = inp.transpose(2, 0, 1) num_classes = self.opt.num_classes trans_output = get_affine_transform( c, s, 0, [self.opt.output_w, self.opt.output_h]) hm = np.zeros( (num_classes, self.opt.output_h, self.opt.output_w), dtype=np.float32) wh = np.zeros((self.max_objs, 2), dtype=np.float32) reg = np.zeros((self.max_objs, 2), dtype=np.float32) dep = np.zeros((self.max_objs, 1), dtype=np.float32) rotbin = np.zeros((self.max_objs, 2), dtype=np.int64) rotres = np.zeros((self.max_objs, 2), dtype=np.float32) dim = np.zeros((self.max_objs, 3), dtype=np.float32) ind = np.zeros((self.max_objs), dtype=np.int64) reg_mask = np.zeros((self.max_objs), dtype=np.uint8) rot_mask = np.zeros((self.max_objs), dtype=np.uint8) ann_ids = self.coco.getAnnIds(imgIds=[img_id]) anns = self.coco.loadAnns(ids=ann_ids) num_objs = min(len(anns), self.max_objs) draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \ draw_umich_gaussian gt_det = [] for k in range(num_objs): ann = anns[k] bbox = self._coco_box_to_bbox(ann['bbox']) cls_id = int(self.cat_ids[ann['category_id']]) if cls_id <= -99: continue # if flipped: # bbox[[0, 2]] = width - bbox[[2, 0]] - 1 bbox[:2] = affine_transform(bbox[:2], trans_output) bbox[2:] = affine_transform(bbox[2:], trans_output) bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, self.opt.output_w - 1) bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, self.opt.output_h - 1) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] if h > 0 and w > 0: radius = gaussian_radius((h, w)) radius = max(0, int(radius)) ct = np.array( [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) ct_int = ct.astype(np.int32) if cls_id < 0: ignore_id = [_ for _ in range(num_classes)] \ if cls_id == - 1 else [- cls_id - 2] if self.opt.rect_mask: hm[ignore_id, int(bbox[1]): int(bbox[3]) + 1, int(bbox[0]): int(bbox[2]) + 1] = 0.9999 else: for cc in ignore_id: draw_gaussian(hm[cc], ct, radius) hm[ignore_id, ct_int[1], ct_int[0]] = 0.9999 continue draw_gaussian(hm[cls_id], ct, radius) wh[k] = 1. * w, 1. * h gt_det.append([ct[0], ct[1], 1] + \ self._alpha_to_8(self._convert_alpha(ann['alpha'])) + \ [ann['depth']] + (np.array(ann['dim']) / 1).tolist() + [cls_id]) if self.opt.reg_bbox: gt_det[-1] = gt_det[-1][:-1] + [w, h] + [gt_det[-1][-1]] # if (not self.opt.car_only) or cls_id == 1: # Only estimate ADD for cars !!! if 1: alpha = self._convert_alpha(ann['alpha']) # print('img_id cls_id alpha rot_y', img_path, cls_id, alpha, ann['rotation_y']) if alpha < np.pi / 6. or alpha > 5 * np.pi / 6.: rotbin[k, 0] = 1 rotres[k, 0] = alpha - (-0.5 * np.pi) if alpha > -np.pi / 6. or alpha < -5 * np.pi / 6.: rotbin[k, 1] = 1 rotres[k, 1] = alpha - (0.5 * np.pi) dep[k] = ann['depth'] dim[k] = ann['dim'] # print(' cat dim', cls_id, dim[k]) ind[k] = ct_int[1] * self.opt.output_w + ct_int[0] reg[k] = ct - ct_int reg_mask[k] = 1 if not aug else 0 rot_mask[k] = 1 # print('gt_det', gt_det) # print('') ret = {'input': inp, 'hm': hm, 'dep': dep, 'dim': dim, 'ind': ind, 'rotbin': rotbin, 'rotres': rotres, 'reg_mask': reg_mask, 'rot_mask': rot_mask} if self.opt.reg_bbox: ret.update({'wh': wh}) if self.opt.reg_offset: ret.update({'reg': reg}) if self.opt.debug > 0 or not ('train' in self.split): gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \ np.zeros((1, 18), dtype=np.float32) meta = {'c': c, 's': s, 'gt_det': gt_det, 'calib': calib, 'image_path': img_path, 'img_id': img_id} ret['meta'] = meta return ret def _alpha_to_8(self, alpha): # return [alpha, 0, 0, 0, 0, 0, 0, 0] ret = [0, 0, 0, 1, 0, 0, 0, 1] if alpha < np.pi / 6. or alpha > 5 * np.pi / 6.: r = alpha - (-0.5 * np.pi) ret[1] = 1 ret[2], ret[3] = np.sin(r), np.cos(r) if alpha > -np.pi / 6. or alpha < -5 * np.pi / 6.: r = alpha - (0.5 * np.pi) ret[5] = 1 ret[6], ret[7] = np.sin(r), np.cos(r) return ret
6,825
38.918129
90
py
houghnet
houghnet-master/src/lib/datasets/sample/multi_pose.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch.utils.data as data import numpy as np import torch import json import cv2 import os from utils.image import flip, color_aug from utils.image import get_affine_transform, affine_transform from utils.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian from utils.image import draw_dense_reg import math class MultiPoseDataset(data.Dataset): def _coco_box_to_bbox(self, box): bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32) return bbox def _get_border(self, border, size): i = 1 while size - border // i <= border // i: i *= 2 return border // i def __getitem__(self, index): img_id = self.images[index] file_name = self.coco.loadImgs(ids=[img_id])[0]['file_name'] img_path = os.path.join(self.img_dir, file_name) ann_ids = self.coco.getAnnIds(imgIds=[img_id]) anns = self.coco.loadAnns(ids=ann_ids) num_objs = min(len(anns), self.max_objs) img = cv2.imread(img_path) height, width = img.shape[0], img.shape[1] c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32) s = max(img.shape[0], img.shape[1]) * 1.0 rot = 0 flipped = False if self.split == 'train': if not self.opt.not_rand_crop: s = s * np.random.choice(np.arange(0.6, 1.4, 0.1)) w_border = self._get_border(128, img.shape[1]) h_border = self._get_border(128, img.shape[0]) c[0] = np.random.randint(low=w_border, high=img.shape[1] - w_border) c[1] = np.random.randint(low=h_border, high=img.shape[0] - h_border) else: sf = self.opt.scale cf = self.opt.shift c[0] += s * np.clip(np.random.randn()*cf, -2*cf, 2*cf) c[1] += s * np.clip(np.random.randn()*cf, -2*cf, 2*cf) s = s * np.clip(np.random.randn()*sf + 1, 1 - sf, 1 + sf) if np.random.random() < self.opt.aug_rot: rf = self.opt.rotate rot = np.clip(np.random.randn()*rf, -rf*2, rf*2) if np.random.random() < self.opt.flip: flipped = True img = img[:, ::-1, :] c[0] = width - c[0] - 1 trans_input = get_affine_transform( c, s, rot, [self.opt.input_res, self.opt.input_res]) inp = cv2.warpAffine(img, trans_input, (self.opt.input_res, self.opt.input_res), flags=cv2.INTER_LINEAR) inp = (inp.astype(np.float32) / 255.) if self.split == 'train' and not self.opt.no_color_aug: color_aug(self._data_rng, inp, self._eig_val, self._eig_vec) inp = (inp - self.mean) / self.std inp = inp.transpose(2, 0, 1) output_res = self.opt.output_res num_joints = self.num_joints trans_output_rot = get_affine_transform(c, s, rot, [output_res, output_res]) trans_output = get_affine_transform(c, s, 0, [output_res, output_res]) hm = np.zeros((self.num_classes, output_res, output_res), dtype=np.float32) hm_hp = np.zeros((num_joints, output_res, output_res), dtype=np.float32) dense_kps = np.zeros((num_joints, 2, output_res, output_res), dtype=np.float32) dense_kps_mask = np.zeros((num_joints, output_res, output_res), dtype=np.float32) wh = np.zeros((self.max_objs, 2), dtype=np.float32) kps = np.zeros((self.max_objs, num_joints * 2), dtype=np.float32) reg = np.zeros((self.max_objs, 2), dtype=np.float32) ind = np.zeros((self.max_objs), dtype=np.int64) reg_mask = np.zeros((self.max_objs), dtype=np.uint8) kps_mask = np.zeros((self.max_objs, self.num_joints * 2), dtype=np.uint8) hp_offset = np.zeros((self.max_objs * num_joints, 2), dtype=np.float32) hp_ind = np.zeros((self.max_objs * num_joints), dtype=np.int64) hp_mask = np.zeros((self.max_objs * num_joints), dtype=np.int64) draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \ draw_umich_gaussian gt_det = [] for k in range(num_objs): ann = anns[k] bbox = self._coco_box_to_bbox(ann['bbox']) cls_id = int(ann['category_id']) - 1 pts = np.array(ann['keypoints'], np.float32).reshape(num_joints, 3) if flipped: bbox[[0, 2]] = width - bbox[[2, 0]] - 1 pts[:, 0] = width - pts[:, 0] - 1 for e in self.flip_idx: pts[e[0]], pts[e[1]] = pts[e[1]].copy(), pts[e[0]].copy() bbox[:2] = affine_transform(bbox[:2], trans_output) bbox[2:] = affine_transform(bbox[2:], trans_output) bbox = np.clip(bbox, 0, output_res - 1) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] if (h > 0 and w > 0) or (rot != 0): radius = gaussian_radius((math.ceil(h), math.ceil(w))) radius = self.opt.hm_gauss if self.opt.mse_loss else max(0, int(radius)) ct = np.array( [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) ct_int = ct.astype(np.int32) wh[k] = 1. * w, 1. * h ind[k] = ct_int[1] * output_res + ct_int[0] reg[k] = ct - ct_int reg_mask[k] = 1 num_kpts = pts[:, 2].sum() if num_kpts == 0: hm[cls_id, ct_int[1], ct_int[0]] = 0.9999 reg_mask[k] = 0 hp_radius = gaussian_radius((math.ceil(h), math.ceil(w))) hp_radius = self.opt.hm_gauss \ if self.opt.mse_loss else max(0, int(hp_radius)) for j in range(num_joints): if pts[j, 2] > 0: pts[j, :2] = affine_transform(pts[j, :2], trans_output_rot) if pts[j, 0] >= 0 and pts[j, 0] < output_res and \ pts[j, 1] >= 0 and pts[j, 1] < output_res: kps[k, j * 2: j * 2 + 2] = pts[j, :2] - ct_int kps_mask[k, j * 2: j * 2 + 2] = 1 pt_int = pts[j, :2].astype(np.int32) hp_offset[k * num_joints + j] = pts[j, :2] - pt_int hp_ind[k * num_joints + j] = pt_int[1] * output_res + pt_int[0] hp_mask[k * num_joints + j] = 1 if self.opt.dense_hp: # must be before draw center hm gaussian draw_dense_reg(dense_kps[j], hm[cls_id], ct_int, pts[j, :2] - ct_int, radius, is_offset=True) draw_gaussian(dense_kps_mask[j], ct_int, radius) draw_gaussian(hm_hp[j], pt_int, hp_radius) draw_gaussian(hm[cls_id], ct_int, radius) gt_det.append([ct[0] - w / 2, ct[1] - h / 2, ct[0] + w / 2, ct[1] + h / 2, 1] + pts[:, :2].reshape(num_joints * 2).tolist() + [cls_id]) if rot != 0: hm = hm * 0 + 0.9999 reg_mask *= 0 kps_mask *= 0 ret = {'input': inp, 'hm': hm, 'reg_mask': reg_mask, 'ind': ind, 'wh': wh, 'hps': kps, 'hps_mask': kps_mask} if self.opt.dense_hp: dense_kps = dense_kps.reshape(num_joints * 2, output_res, output_res) dense_kps_mask = dense_kps_mask.reshape( num_joints, 1, output_res, output_res) dense_kps_mask = np.concatenate([dense_kps_mask, dense_kps_mask], axis=1) dense_kps_mask = dense_kps_mask.reshape( num_joints * 2, output_res, output_res) ret.update({'dense_hps': dense_kps, 'dense_hps_mask': dense_kps_mask}) del ret['hps'], ret['hps_mask'] if self.opt.reg_offset: ret.update({'reg': reg}) if self.opt.hm_hp: ret.update({'hm_hp': hm_hp}) if self.opt.reg_hp_offset: ret.update({'hp_offset': hp_offset, 'hp_ind': hp_ind, 'hp_mask': hp_mask}) if self.opt.debug > 0 or not self.split == 'train': gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \ np.zeros((1, 40), dtype=np.float32) meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id} ret['meta'] = meta return ret
7,913
42.01087
81
py
houghnet
houghnet-master/src/lib/datasets/sample/ctseg.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch.utils.data as data import numpy as np import torch import json import cv2 import os from src.lib.utils.image import flip, color_aug from src.lib.utils.image import get_affine_transform, affine_transform from src.lib.utils.image import gaussian_radius, draw_umich_gaussian, draw_msra_gaussian from src.lib.utils.image import draw_dense_reg import math class CTSegDataset(data.Dataset): def _coco_box_to_bbox(self, box): bbox = np.array([box[0], box[1], box[0] + box[2], box[1] + box[3]], dtype=np.float32) return bbox def _get_border(self, border, size): i = 1 while size - border // i <= border // i: i *= 2 return border // i def __getitem__(self, index): img_id = self.images[index] file_name = self.coco.loadImgs(ids=[img_id])[0]['file_name'] img_path = os.path.join(self.img_dir, file_name) ann_ids = self.coco.getAnnIds(imgIds=[img_id]) anns = self.coco.loadAnns(ids=ann_ids) num_objs = min(len(anns), self.max_objs) img = cv2.imread(img_path) height, width = img.shape[0], img.shape[1] c = np.array([img.shape[1] / 2., img.shape[0] / 2.], dtype=np.float32) if self.opt.keep_res: input_h = (height | self.opt.pad) + 1 input_w = (width | self.opt.pad) + 1 s = np.array([input_w, input_h], dtype=np.float32) else: s = max(img.shape[0], img.shape[1]) * 1.0 input_h, input_w = self.opt.input_h, self.opt.input_w flipped = False if self.split == 'train': if not self.opt.not_rand_crop: s = s * np.random.choice(np.arange(0.6, 1.4, 0.1)) w_border = self._get_border(128, img.shape[1]) h_border = self._get_border(128, img.shape[0]) c[0] = np.random.randint(low=w_border, high=img.shape[1] - w_border) c[1] = np.random.randint(low=h_border, high=img.shape[0] - h_border) else: sf = self.opt.scale cf = self.opt.shift c[0] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf) c[1] += s * np.clip(np.random.randn() * cf, -2 * cf, 2 * cf) s = s * np.clip(np.random.randn() * sf + 1, 1 - sf, 1 + sf) if np.random.random() < self.opt.flip: flipped = True img = img[:, ::-1, :] c[0] = width - c[0] - 1 trans_input = get_affine_transform( c, s, 0, [input_w, input_h]) inp = cv2.warpAffine(img, trans_input, (input_w, input_h), flags=cv2.INTER_LINEAR) inp = (inp.astype(np.float32) / 255.) if self.split == 'train' and not self.opt.no_color_aug: color_aug(self._data_rng, inp, self._eig_val, self._eig_vec) inp = (inp - self.mean) / self.std inp = inp.transpose(2, 0, 1) output_h = input_h // self.opt.down_ratio output_w = input_w // self.opt.down_ratio num_classes = self.num_classes trans_output = get_affine_transform(c, s, 0, [output_w, output_h]) hm = np.zeros((num_classes, output_h, output_w), dtype=np.float32) wh = np.zeros((self.max_objs, 2), dtype=np.float32) gtboxes = np.zeros((self.max_objs, 4), dtype=np.float32) dense_wh = np.zeros((2, output_h, output_w), dtype=np.float32) reg = np.zeros((self.max_objs, 2), dtype=np.float32) ind = np.zeros((self.max_objs), dtype=np.int64) reg_mask = np.zeros((self.max_objs), dtype=np.uint8) cat_spec_wh = np.zeros((self.max_objs, num_classes * 2), dtype=np.float32) cat_spec_mask = np.zeros((self.max_objs, num_classes), dtype=np.uint8) instance_masks = np.zeros((self.max_objs, output_h,output_w),dtype=np.float32) draw_gaussian = draw_msra_gaussian if self.opt.mse_loss else \ draw_umich_gaussian gt_det = [] for k in range(num_objs): ann = anns[k] instance_mask = self.coco.annToMask(ann) bbox = self._coco_box_to_bbox(ann['bbox']) cls_id = int(self.cat_ids[ann['category_id']]) if flipped: bbox[[0, 2]] = width - bbox[[2, 0]] - 1 instance_mask = instance_mask[:, ::-1] bbox[:2] = affine_transform(bbox[:2], trans_output) bbox[2:] = affine_transform(bbox[2:], trans_output) bbox[[0, 2]] = np.clip(bbox[[0, 2]], 0, output_w - 1) bbox[[1, 3]] = np.clip(bbox[[1, 3]], 0, output_h - 1) instance_mask= cv2.warpAffine(instance_mask, trans_output, (output_w, output_h), flags=cv2.INTER_LINEAR) instance_mask = instance_mask.astype(np.float32) h, w = bbox[3] - bbox[1], bbox[2] - bbox[0] if h > 0 and w > 0: radius = gaussian_radius((math.ceil(h), math.ceil(w))) radius = max(0, int(radius)) radius = self.opt.hm_gauss if self.opt.mse_loss else radius ct = np.array( [(bbox[0] + bbox[2]) / 2, (bbox[1] + bbox[3]) / 2], dtype=np.float32) ct_int = ct.astype(np.int32) draw_gaussian(hm[cls_id], ct_int, radius) gtboxes[k] = bbox wh[k] = 1. * w, 1. * h ind[k] = ct_int[1] * output_w + ct_int[0] reg[k] = ct - ct_int reg_mask[k] = 1 cat_spec_wh[k, cls_id * 2: cls_id * 2 + 2] = wh[k] cat_spec_mask[k, cls_id] = 1 instance_masks[k] = instance_mask if self.opt.dense_wh: draw_dense_reg(dense_wh, hm.max(axis=0), ct_int, wh[k], radius) gt_det.append([ct[0] - w / 2, ct[1] - h / 2, ct[0] + w / 2, ct[1] + h / 2, 1, cls_id]) ret = {'input': inp, 'hm': hm, 'reg_mask': reg_mask, 'ind': ind, 'wh': wh, "instance_mask":instance_masks, 'gtboxes':gtboxes, 'cat_spec_mask': cat_spec_mask} if self.opt.dense_wh: hm_a = hm.max(axis=0, keepdims=True) dense_wh_mask = np.concatenate([hm_a, hm_a], axis=0) ret.update({'dense_wh': dense_wh, 'dense_wh_mask': dense_wh_mask}) del ret['wh'] elif self.opt.cat_spec_wh: ret.update({'cat_spec_wh': cat_spec_wh, 'cat_spec_mask': cat_spec_mask}) del ret['wh'] if self.opt.reg_offset: ret.update({'reg': reg}) if self.opt.debug > 0 or not self.split == 'train': gt_det = np.array(gt_det, dtype=np.float32) if len(gt_det) > 0 else \ np.zeros((1, 6), dtype=np.float32) meta = {'c': c, 's': s, 'gt_det': gt_det, 'img_id': img_id} ret['meta'] = meta return ret
7,112
43.735849
97
py
houghnet
houghnet-master/src/lib/datasets/dataset/kitti.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch.utils.data as data import pycocotools.coco as coco import numpy as np import torch import json import cv2 import os import math import torch.utils.data as data class KITTI(data.Dataset): num_classes = 3 default_resolution = [384, 1280] mean = np.array([0.485, 0.456, 0.406], np.float32).reshape(1, 1, 3) std = np.array([0.229, 0.224, 0.225], np.float32).reshape(1, 1, 3) def __init__(self, opt, split): super(KITTI, self).__init__() self.data_dir = os.path.join(opt.data_dir, 'kitti') self.img_dir = os.path.join(self.data_dir, 'images', 'trainval') if opt.trainval: split = 'trainval' if split == 'train' else 'test' self.img_dir = os.path.join(self.data_dir, 'images', split) self.annot_path = os.path.join( self.data_dir, 'annotations', 'kitti_{}.json').format(split) else: self.annot_path = os.path.join(self.data_dir, 'annotations', 'kitti_{}_{}.json').format(opt.kitti_split, split) self.max_objs = 50 self.class_name = [ '__background__', 'Pedestrian', 'Car', 'Cyclist'] self.cat_ids = {1:0, 2:1, 3:2, 4:-3, 5:-3, 6:-2, 7:-99, 8:-99, 9:-1} self._data_rng = np.random.RandomState(123) self._eig_val = np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32) self._eig_vec = np.array([ [-0.58752847, -0.69563484, 0.41340352], [-0.5832747, 0.00994535, -0.81221408], [-0.56089297, 0.71832671, 0.41158938] ], dtype=np.float32) self.split = split self.opt = opt self.alpha_in_degree = False print('==> initializing kitti {}, {} data.'.format(opt.kitti_split, split)) self.coco = coco.COCO(self.annot_path) self.images = self.coco.getImgIds() self.num_samples = len(self.images) print('Loaded {} {} samples'.format(split, self.num_samples)) def __len__(self): return self.num_samples def _to_float(self, x): return float("{:.2f}".format(x)) def convert_eval_format(self, all_bboxes): pass def save_results(self, results, save_dir): results_dir = os.path.join(save_dir, 'results') if not os.path.exists(results_dir): os.mkdir(results_dir) for img_id in results.keys(): out_path = os.path.join(results_dir, '{:06d}.txt'.format(img_id)) f = open(out_path, 'w') for cls_ind in results[img_id]: for j in range(len(results[img_id][cls_ind])): class_name = self.class_name[cls_ind] f.write('{} 0.0 0'.format(class_name)) for i in range(len(results[img_id][cls_ind][j])): f.write(' {:.2f}'.format(results[img_id][cls_ind][j][i])) f.write('\n') f.close() def run_eval(self, results, save_dir): self.save_results(results, save_dir) os.system('./src//tools/kitti_eval/evaluate_object_3d_offline ' + \ './data/kitti/training/label_2 ' + \ '{}/results/'.format(save_dir))
3,060
33.011111
79
py
houghnet
houghnet-master/src/lib/datasets/dataset/coco_hp.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import pycocotools.coco as coco from pycocotools.cocoeval import COCOeval import numpy as np import json import os import torch.utils.data as data class COCOHP(data.Dataset): num_classes = 1 num_joints = 17 default_resolution = [512, 512] mean = np.array([0.40789654, 0.44719302, 0.47026115], dtype=np.float32).reshape(1, 1, 3) std = np.array([0.28863828, 0.27408164, 0.27809835], dtype=np.float32).reshape(1, 1, 3) flip_idx = [[1, 2], [3, 4], [5, 6], [7, 8], [9, 10], [11, 12], [13, 14], [15, 16]] def __init__(self, opt, split): super(COCOHP, self).__init__() self.edges = [[0, 1], [0, 2], [1, 3], [2, 4], [4, 6], [3, 5], [5, 6], [5, 7], [7, 9], [6, 8], [8, 10], [6, 12], [5, 11], [11, 12], [12, 14], [14, 16], [11, 13], [13, 15]] self.acc_idxs = [1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16] self.data_dir = os.path.join(opt.data_dir) self.img_dir = os.path.join(self.data_dir, 'images', '{}2017'.format(split)) if split == 'test': self.annot_path = os.path.join( self.data_dir, 'annotations', 'image_info_test-dev2017.json').format(split) else: self.annot_path = os.path.join( self.data_dir, 'annotations', 'person_keypoints_{}2017.json').format(split) self.max_objs = 32 self._data_rng = np.random.RandomState(123) self._eig_val = np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32) self._eig_vec = np.array([ [-0.58752847, -0.69563484, 0.41340352], [-0.5832747, 0.00994535, -0.81221408], [-0.56089297, 0.71832671, 0.41158938] ], dtype=np.float32) self.split = split self.opt = opt print('==> initializing coco 2017 {} data.'.format(split)) self.coco = coco.COCO(self.annot_path) image_ids = self.coco.getImgIds() if split == 'train': self.images = [] for img_id in image_ids: idxs = self.coco.getAnnIds(imgIds=[img_id]) if len(idxs) > 0: self.images.append(img_id) else: self.images = image_ids self.num_samples = len(self.images) print('Loaded {} {} samples'.format(split, self.num_samples)) def _to_float(self, x): return float("{:.2f}".format(x)) def convert_eval_format(self, all_bboxes): # import pdb; pdb.set_trace() detections = [] for image_id in all_bboxes: for cls_ind in all_bboxes[image_id]: category_id = 1 for dets in all_bboxes[image_id][cls_ind]: bbox = dets[:4] bbox[2] -= bbox[0] bbox[3] -= bbox[1] score = dets[4] bbox_out = list(map(self._to_float, bbox)) keypoints = np.concatenate([ np.array(dets[5:39], dtype=np.float32).reshape(-1, 2), np.ones((17, 1), dtype=np.float32)], axis=1).reshape(51).tolist() keypoints = list(map(self._to_float, keypoints)) detection = { "image_id": int(image_id), "category_id": int(category_id), "bbox": bbox_out, "score": float("{:.2f}".format(score)), "keypoints": keypoints } detections.append(detection) return detections def __len__(self): return self.num_samples def save_results(self, results, save_dir): json.dump(self.convert_eval_format(results), open('{}/results.json'.format(save_dir), 'w')) def run_eval(self, results, save_dir): # result_json = os.path.join(opt.save_dir, "results.json") # detections = convert_eval_format(all_boxes) # json.dump(detections, open(result_json, "w")) self.save_results(results, save_dir) coco_dets = self.coco.loadRes('{}/results.json'.format(save_dir)) coco_eval = COCOeval(self.coco, coco_dets, "keypoints") coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize() coco_eval = COCOeval(self.coco, coco_dets, "bbox") coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize()
4,244
34.375
80
py
houghnet
houghnet-master/src/lib/datasets/dataset/pascal.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import pycocotools.coco as coco import numpy as np import torch import json import os import torch.utils.data as data class PascalVOC(data.Dataset): num_classes = 20 default_resolution = [384, 384] mean = np.array([0.485, 0.456, 0.406], dtype=np.float32).reshape(1, 1, 3) std = np.array([0.229, 0.224, 0.225], dtype=np.float32).reshape(1, 1, 3) def __init__(self, opt, split): super(PascalVOC, self).__init__() self.data_dir = os.path.join(opt.data_dir, 'voc') self.img_dir = os.path.join(self.data_dir, 'images') _ann_name = {'train': 'trainval0712', 'val': 'test2007'} self.annot_path = os.path.join( self.data_dir, 'annotations', 'pascal_{}.json').format(_ann_name[split]) self.max_objs = 50 self.class_name = ['__background__', "aeroplane", "bicycle", "bird", "boat", "bottle", "bus", "car", "cat", "chair", "cow", "diningtable", "dog", "horse", "motorbike", "person", "pottedplant", "sheep", "sofa", "train", "tvmonitor"] self._valid_ids = np.arange(1, 21, dtype=np.int32) self.cat_ids = {v: i for i, v in enumerate(self._valid_ids)} self._data_rng = np.random.RandomState(123) self._eig_val = np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32) self._eig_vec = np.array([ [-0.58752847, -0.69563484, 0.41340352], [-0.5832747, 0.00994535, -0.81221408], [-0.56089297, 0.71832671, 0.41158938] ], dtype=np.float32) self.split = split self.opt = opt print('==> initializing pascal {} data.'.format(_ann_name[split])) self.coco = coco.COCO(self.annot_path) self.images = sorted(self.coco.getImgIds()) self.num_samples = len(self.images) print('Loaded {} {} samples'.format(split, self.num_samples)) def _to_float(self, x): return float("{:.2f}".format(x)) def convert_eval_format(self, all_bboxes): detections = [[[] for __ in range(self.num_samples)] \ for _ in range(self.num_classes + 1)] for i in range(self.num_samples): img_id = self.images[i] for j in range(1, self.num_classes + 1): if isinstance(all_bboxes[img_id][j], np.ndarray): detections[j][i] = all_bboxes[img_id][j].tolist() else: detections[j][i] = all_bboxes[img_id][j] return detections def __len__(self): return self.num_samples def save_results(self, results, save_dir): json.dump(self.convert_eval_format(results), open('{}/results.json'.format(save_dir), 'w')) def run_eval(self, results, save_dir): # result_json = os.path.join(save_dir, "results.json") # detections = self.convert_eval_format(results) # json.dump(detections, open(result_json, "w")) self.save_results(results, save_dir) os.system('python tools/reval.py ' + \ '{}/results.json'.format(save_dir))
3,032
35.542169
80
py
houghnet
houghnet-master/src/lib/datasets/dataset/coco_seg.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import pycocotools.coco as coco from pycocotools.cocoeval import COCOeval import numpy as np import json import os import torch.utils.data as data class COCOSEG(data.Dataset): num_classes = 80 default_resolution = [512, 512] mean = np.array([0.40789654, 0.44719302, 0.47026115], dtype=np.float32).reshape(1, 1, 3) std = np.array([0.28863828, 0.27408164, 0.27809835], dtype=np.float32).reshape(1, 1, 3) def __init__(self, opt, split): super(COCOSEG, self).__init__() self.data_dir = os.path.join(opt.data_dir, opt.coco_dir) # self.img_dir = os.path.join(self.data_dir, '{}2017'.format(split)) self.img_dir = os.path.join(self.data_dir + '/images', '{}2017'.format(split)) if split == 'test': self.annot_path = os.path.join( self.data_dir, 'annotations', 'image_info_test-dev2017.json').format(split) else: if opt.task == 'exdet': self.annot_path = os.path.join( self.data_dir, 'annotations', 'instances_extreme_{}2017.json').format(split) else: self.annot_path = os.path.join( self.data_dir, 'annotations', 'instances_{}2017.json').format(split) self.max_objs = 70 self.class_name = [ '__background__', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush'] self._valid_ids = [ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] self.cat_ids = {v: i for i, v in enumerate(self._valid_ids)} self.voc_color = [(v // 32 * 64 + 64, (v // 8) % 4 * 64, v % 8 * 32) \ for v in range(1, self.num_classes + 1)] self._data_rng = np.random.RandomState(123) self._eig_val = np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32) self._eig_vec = np.array([ [-0.58752847, -0.69563484, 0.41340352], [-0.5832747, 0.00994535, -0.81221408], [-0.56089297, 0.71832671, 0.41158938] ], dtype=np.float32) # self.mean = np.array([0.485, 0.456, 0.406], np.float32).reshape(1, 1, 3) # self.std = np.array([0.229, 0.224, 0.225], np.float32).reshape(1, 1, 3) self.split = split self.opt = opt print('==> initializing coco 2017 {} data.'.format(split)) self.coco = coco.COCO(self.annot_path) self.images = self.coco.getImgIds() self.num_samples = len(self.images) print('Loaded {} {} samples'.format(split, self.num_samples)) def _to_float(self, x): return float("{:.2f}".format(x)) def convert_eval_format(self, all_bboxes): # import pdb; pdb.set_trace() detections = [] for image_id in all_bboxes: for cls_ind in all_bboxes[image_id]: category_id = self._valid_ids[cls_ind - 1] if type(all_bboxes[image_id][cls_ind]) == dict: for id in range(len(all_bboxes[image_id][cls_ind]['boxs'])): bbox = all_bboxes[image_id][cls_ind]['boxs'][id] mask = all_bboxes[image_id][cls_ind]['pred_mask'][id] bbox[2] -= bbox[0] bbox[3] -= bbox[1] score = bbox[4] bbox_out = list(map(self._to_float, bbox[0:4])) detection = { "image_id": int(image_id), "category_id": int(category_id), "bbox": bbox_out, "score": float("{:.2f}".format(score)), "segmentation": mask } detections.append(detection) else: for bbox in all_bboxes[image_id][cls_ind]: bbox[2] -= bbox[0] bbox[3] -= bbox[1] score = bbox[4] bbox_out = list(map(self._to_float, bbox[0:4])) detection = { "image_id": int(image_id), "category_id": int(category_id), "bbox": bbox_out, "score": float("{:.2f}".format(score)) } if len(bbox) > 5: extreme_points = list(map(self._to_float, bbox[5:13])) detection["extreme_points"] = extreme_points detections.append(detection) return detections def __len__(self): return self.num_samples def save_results(self, results, save_dir): json.dump(self.convert_eval_format(results), open('{}/results.json'.format(save_dir), 'w')) def run_eval(self, results, save_dir): detections = self.convert_eval_format(results) coco_dets = self.coco.loadRes(detections) coco_eval = COCOeval(self.coco, coco_dets, "bbox") coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize() coco_eval = COCOeval(self.coco, coco_dets, "segm") coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize()
6,004
38.768212
82
py
houghnet
houghnet-master/src/lib/datasets/dataset/coco.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import pycocotools.coco as coco from pycocotools.cocoeval import COCOeval import numpy as np import json import os import torch.utils.data as data class COCO(data.Dataset): num_classes = 80 default_resolution = [512, 512] mean = np.array([0.40789654, 0.44719302, 0.47026115], dtype=np.float32).reshape(1, 1, 3) std = np.array([0.28863828, 0.27408164, 0.27809835], dtype=np.float32).reshape(1, 1, 3) def __init__(self, opt, split): super(COCO, self).__init__() self.data_dir = os.path.join(opt.data_dir, opt.coco_dir) self.img_dir = os.path.join(self.data_dir + '/images', '{}2017'.format(split)) if split == 'test': self.annot_path = os.path.join( self.data_dir, 'annotations', 'image_info_test-dev2017.json').format(split) else: if opt.task == 'exdet': self.annot_path = os.path.join( self.data_dir, 'annotations', 'instances_extreme_{}2017.json').format(split) else: if opt.minicoco and split=="train": self.annot_path = os.path.join( self.data_dir, 'annotations', 'instances_mini{}2017.json').format(split) else: self.annot_path = os.path.join( self.data_dir, 'annotations', 'instances_{}2017.json').format(split) self.max_objs = 128 self.class_name = [ '__background__', 'person', 'bicycle', 'car', 'motorcycle', 'airplane', 'bus', 'train', 'truck', 'boat', 'traffic light', 'fire hydrant', 'stop sign', 'parking meter', 'bench', 'bird', 'cat', 'dog', 'horse', 'sheep', 'cow', 'elephant', 'bear', 'zebra', 'giraffe', 'backpack', 'umbrella', 'handbag', 'tie', 'suitcase', 'frisbee', 'skis', 'snowboard', 'sports ball', 'kite', 'baseball bat', 'baseball glove', 'skateboard', 'surfboard', 'tennis racket', 'bottle', 'wine glass', 'cup', 'fork', 'knife', 'spoon', 'bowl', 'banana', 'apple', 'sandwich', 'orange', 'broccoli', 'carrot', 'hot dog', 'pizza', 'donut', 'cake', 'chair', 'couch', 'potted plant', 'bed', 'dining table', 'toilet', 'tv', 'laptop', 'mouse', 'remote', 'keyboard', 'cell phone', 'microwave', 'oven', 'toaster', 'sink', 'refrigerator', 'book', 'clock', 'vase', 'scissors', 'teddy bear', 'hair drier', 'toothbrush'] self._valid_ids = [ 1, 2, 3, 4, 5, 6, 7, 8, 9, 10, 11, 13, 14, 15, 16, 17, 18, 19, 20, 21, 22, 23, 24, 25, 27, 28, 31, 32, 33, 34, 35, 36, 37, 38, 39, 40, 41, 42, 43, 44, 46, 47, 48, 49, 50, 51, 52, 53, 54, 55, 56, 57, 58, 59, 60, 61, 62, 63, 64, 65, 67, 70, 72, 73, 74, 75, 76, 77, 78, 79, 80, 81, 82, 84, 85, 86, 87, 88, 89, 90] self.cat_ids = {v: i for i, v in enumerate(self._valid_ids)} self.voc_color = [(v // 32 * 64 + 64, (v // 8) % 4 * 64, v % 8 * 32) \ for v in range(1, self.num_classes + 1)] self._data_rng = np.random.RandomState(123) self._eig_val = np.array([0.2141788, 0.01817699, 0.00341571], dtype=np.float32) self._eig_vec = np.array([ [-0.58752847, -0.69563484, 0.41340352], [-0.5832747, 0.00994535, -0.81221408], [-0.56089297, 0.71832671, 0.41158938] ], dtype=np.float32) # self.mean = np.array([0.485, 0.456, 0.406], np.float32).reshape(1, 1, 3) # self.std = np.array([0.229, 0.224, 0.225], np.float32).reshape(1, 1, 3) self.split = split self.opt = opt print('==> initializing coco 2017 {} data.'.format(split)) self.coco = coco.COCO(self.annot_path) self.images = self.coco.getImgIds() self.num_samples = len(self.images) print('Loaded {} {} samples'.format(split, self.num_samples)) def _to_float(self, x): return float("{:.2f}".format(x)) def convert_eval_format(self, all_bboxes): # import pdb; pdb.set_trace() detections = [] for image_id in all_bboxes: for cls_ind in all_bboxes[image_id]: category_id = self._valid_ids[cls_ind - 1] for bbox in all_bboxes[image_id][cls_ind]: bbox[2] -= bbox[0] bbox[3] -= bbox[1] score = bbox[4] bbox_out = list(map(self._to_float, bbox[0:4])) detection = { "image_id": int(image_id), "category_id": int(category_id), "bbox": bbox_out, "score": float("{:.2f}".format(score)) } if len(bbox) > 5: extreme_points = list(map(self._to_float, bbox[5:13])) detection["extreme_points"] = extreme_points detections.append(detection) return detections def __len__(self): return self.num_samples def save_results(self, results, save_dir): json.dump(self.convert_eval_format(results), open('{}/results.json'.format(save_dir), 'w')) def run_eval(self, results, save_dir): # result_json = os.path.join(save_dir, "results.json") # detections = self.convert_eval_format(results) # json.dump(detections, open(result_json, "w")) self.save_results(results, save_dir) coco_dets = self.coco.loadRes('{}/results.json'.format(save_dir)) coco_eval = COCOeval(self.coco, coco_dets, "bbox") coco_eval.evaluate() coco_eval.accumulate() coco_eval.summarize()
5,426
39.2
82
py
houghnet
houghnet-master/src/lib/utils/image.py
# ------------------------------------------------------------------------------ # Copyright (c) Microsoft # Licensed under the MIT License. # Written by Bin Xiao (Bin.Xiao@microsoft.com) # Modified by Xingyi Zhou # ------------------------------------------------------------------------------ from __future__ import absolute_import from __future__ import division from __future__ import print_function import numpy as np import cv2 import random import torch PI = np.pi def flip(img): return img[:, :, ::-1].copy() def transform_preds(coords, center, scale, output_size): target_coords = np.zeros(coords.shape) trans = get_affine_transform(center, scale, 0, output_size, inv=1) for p in range(coords.shape[0]): target_coords[p, 0:2] = affine_transform(coords[p, 0:2], trans) return target_coords def get_affine_transform(center, scale, rot, output_size, shift=np.array([0, 0], dtype=np.float32), inv=0): if not isinstance(scale, np.ndarray) and not isinstance(scale, list): scale = np.array([scale, scale], dtype=np.float32) scale_tmp = scale src_w = scale_tmp[0] dst_w = output_size[0] dst_h = output_size[1] rot_rad = np.pi * rot / 180 src_dir = get_dir([0, src_w * -0.5], rot_rad) dst_dir = np.array([0, dst_w * -0.5], np.float32) src = np.zeros((3, 2), dtype=np.float32) dst = np.zeros((3, 2), dtype=np.float32) src[0, :] = center + scale_tmp * shift src[1, :] = center + src_dir + scale_tmp * shift dst[0, :] = [dst_w * 0.5, dst_h * 0.5] dst[1, :] = np.array([dst_w * 0.5, dst_h * 0.5], np.float32) + dst_dir src[2:, :] = get_3rd_point(src[0, :], src[1, :]) dst[2:, :] = get_3rd_point(dst[0, :], dst[1, :]) if inv: trans = cv2.getAffineTransform(np.float32(dst), np.float32(src)) else: trans = cv2.getAffineTransform(np.float32(src), np.float32(dst)) return trans def affine_transform(pt, t): new_pt = np.array([pt[0], pt[1], 1.], dtype=np.float32).T new_pt = np.dot(t, new_pt) return new_pt[:2] def get_3rd_point(a, b): direct = a - b return b + np.array([-direct[1], direct[0]], dtype=np.float32) def get_dir(src_point, rot_rad): sn, cs = np.sin(rot_rad), np.cos(rot_rad) src_result = [0, 0] src_result[0] = src_point[0] * cs - src_point[1] * sn src_result[1] = src_point[0] * sn + src_point[1] * cs return src_result def crop(img, center, scale, output_size, rot=0): trans = get_affine_transform(center, scale, rot, output_size) dst_img = cv2.warpAffine(img, trans, (int(output_size[0]), int(output_size[1])), flags=cv2.INTER_LINEAR) return dst_img def gaussian_radius(det_size, min_overlap=0.7): height, width = det_size a1 = 1 b1 = (height + width) c1 = width * height * (1 - min_overlap) / (1 + min_overlap) sq1 = np.sqrt(b1 ** 2 - 4 * a1 * c1) r1 = (b1 + sq1) / 2 a2 = 4 b2 = 2 * (height + width) c2 = (1 - min_overlap) * width * height sq2 = np.sqrt(b2 ** 2 - 4 * a2 * c2) r2 = (b2 + sq2) / 2 a3 = 4 * min_overlap b3 = -2 * min_overlap * (height + width) c3 = (min_overlap - 1) * width * height sq3 = np.sqrt(b3 ** 2 - 4 * a3 * c3) r3 = (b3 + sq3) / 2 return min(r1, r2, r3) def gaussian2D(shape, sigma=1): m, n = [(ss - 1.) / 2. for ss in shape] y, x = np.ogrid[-m:m+1,-n:n+1] h = np.exp(-(x * x + y * y) / (2 * sigma * sigma)) h[h < np.finfo(h.dtype).eps * h.max()] = 0 return h def draw_umich_gaussian(heatmap, center, radius, k=1): diameter = 2 * radius + 1 gaussian = gaussian2D((diameter, diameter), sigma=diameter / 6) x, y = int(center[0]), int(center[1]) height, width = heatmap.shape[0:2] left, right = min(x, radius), min(width - x, radius + 1) top, bottom = min(y, radius), min(height - y, radius + 1) masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right] masked_gaussian = gaussian[radius - top:radius + bottom, radius - left:radius + right] if min(masked_gaussian.shape) > 0 and min(masked_heatmap.shape) > 0: # TODO debug np.maximum(masked_heatmap, masked_gaussian * k, out=masked_heatmap) return heatmap def draw_dense_reg(regmap, heatmap, center, value, radius, is_offset=False): diameter = 2 * radius + 1 gaussian = gaussian2D((diameter, diameter), sigma=diameter / 6) value = np.array(value, dtype=np.float32).reshape(-1, 1, 1) dim = value.shape[0] reg = np.ones((dim, diameter*2+1, diameter*2+1), dtype=np.float32) * value if is_offset and dim == 2: delta = np.arange(diameter*2+1) - radius reg[0] = reg[0] - delta.reshape(1, -1) reg[1] = reg[1] - delta.reshape(-1, 1) x, y = int(center[0]), int(center[1]) height, width = heatmap.shape[0:2] left, right = min(x, radius), min(width - x, radius + 1) top, bottom = min(y, radius), min(height - y, radius + 1) masked_heatmap = heatmap[y - top:y + bottom, x - left:x + right] masked_regmap = regmap[:, y - top:y + bottom, x - left:x + right] masked_gaussian = gaussian[radius - top:radius + bottom, radius - left:radius + right] masked_reg = reg[:, radius - top:radius + bottom, radius - left:radius + right] if min(masked_gaussian.shape) > 0 and min(masked_heatmap.shape) > 0: # TODO debug idx = (masked_gaussian >= masked_heatmap).reshape( 1, masked_gaussian.shape[0], masked_gaussian.shape[1]) masked_regmap = (1-idx) * masked_regmap + idx * masked_reg regmap[:, y - top:y + bottom, x - left:x + right] = masked_regmap return regmap def draw_msra_gaussian(heatmap, center, sigma): tmp_size = sigma * 3 mu_x = int(center[0] + 0.5) mu_y = int(center[1] + 0.5) w, h = heatmap.shape[0], heatmap.shape[1] ul = [int(mu_x - tmp_size), int(mu_y - tmp_size)] br = [int(mu_x + tmp_size + 1), int(mu_y + tmp_size + 1)] if ul[0] >= h or ul[1] >= w or br[0] < 0 or br[1] < 0: return heatmap size = 2 * tmp_size + 1 x = np.arange(0, size, 1, np.float32) y = x[:, np.newaxis] x0 = y0 = size // 2 g = np.exp(- ((x - x0) ** 2 + (y - y0) ** 2) / (2 * sigma ** 2)) g_x = max(0, -ul[0]), min(br[0], h) - ul[0] g_y = max(0, -ul[1]), min(br[1], w) - ul[1] img_x = max(0, ul[0]), min(br[0], h) img_y = max(0, ul[1]), min(br[1], w) heatmap[img_y[0]:img_y[1], img_x[0]:img_x[1]] = np.maximum( heatmap[img_y[0]:img_y[1], img_x[0]:img_x[1]], g[g_y[0]:g_y[1], g_x[0]:g_x[1]]) return heatmap def grayscale(image): return cv2.cvtColor(image, cv2.COLOR_BGR2GRAY) def lighting_(data_rng, image, alphastd, eigval, eigvec): alpha = data_rng.normal(scale=alphastd, size=(3, )) image += np.dot(eigvec, eigval * alpha) def blend_(alpha, image1, image2): image1 *= alpha image2 *= (1 - alpha) image1 += image2 def saturation_(data_rng, image, gs, gs_mean, var): alpha = 1. + data_rng.uniform(low=-var, high=var) blend_(alpha, image, gs[:, :, None]) def brightness_(data_rng, image, gs, gs_mean, var): alpha = 1. + data_rng.uniform(low=-var, high=var) image *= alpha def contrast_(data_rng, image, gs, gs_mean, var): alpha = 1. + data_rng.uniform(low=-var, high=var) blend_(alpha, image, gs_mean) def color_aug(data_rng, image, eig_val, eig_vec): functions = [brightness_, contrast_, saturation_] random.shuffle(functions) gs = grayscale(image) gs_mean = gs.mean() for f in functions: f(data_rng, image, gs, gs_mean, 0.4) lighting_(data_rng, image, 0.1, eig_val, eig_vec)
7,720
31.305439
88
py
houghnet
houghnet-master/src/lib/utils/utils.py
from __future__ import absolute_import from __future__ import division from __future__ import print_function import torch class AverageMeter(object): """Computes and stores the average and current value""" def __init__(self): self.reset() def reset(self): self.val = 0 self.avg = 0 self.sum = 0 self.count = 0 def update(self, val, n=1): self.val = val self.sum += val * n self.count += n if self.count > 0: self.avg = self.sum / self.count
542
22.608696
59
py
GaitForeMer
GaitForeMer-main/training/transformer_model_fn.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implments the model function for the POTR model.""" import numpy as np import os import sys import argparse import json import time # from potr.data.Gait17JointsDataset import Gait17JointsDataset import torch import torch.nn as nn import torch.optim as optim from torch.utils.tensorboard import SummaryWriter from sklearn.metrics import classification_report from numpyencoder import NumpyEncoder thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import training.seq2seq_model_fn as seq2seq_model_fn import models.PoseTransformer as PoseTransformer import models.PoseEncoderDecoder as PoseEncoderDecoder import data.NTURGDDataset as NTURGDDataset import data.GaitJointsDataset as GaitJointsDataset import utils.utils as utils _DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') _WEIGHT_DECAY = 0.00001 _NSEEDS = 8 class POTRModelFn(seq2seq_model_fn.ModelFn): def __init__(self, params, train_dataset_fn, eval_dataset_fn, pose_encoder_fn=None, pose_decoder_fn=None): super(POTRModelFn, self).__init__( params, train_dataset_fn, eval_dataset_fn, pose_encoder_fn, pose_decoder_fn) self._loss_fn = self.layerwise_loss_fn self.task = params['task'] if self.task == 'downstream': weights = torch.tensor([9., 28., 13., 4.]) weights = weights / weights.sum() # turn into percentage weights = 1.0 / weights # inverse weights = weights / weights.sum() self._loss_weights = weights.to(_DEVICE) self._weighted_ce_loss = nn.CrossEntropyLoss(weight=self._loss_weights) print('Using a weighted CE loss for gait impairment score prediction.') else: print('Using a standard CE loss for activity prediction.') def smooth_l1(self, decoder_pred, decoder_gt): l1loss = nn.SmoothL1Loss(reduction='mean') return l1loss(decoder_pred, decoder_gt) def loss_l1(self, decoder_pred, decoder_gt): return nn.L1Loss(reduction='mean')(decoder_pred, decoder_gt) def loss_activity(self, logits, class_gt): """Computes entropy loss from logits between predictions and class.""" if self.task == 'downstream': return self._weighted_ce_loss(logits, class_gt) else: return nn.functional.cross_entropy(logits, class_gt, reduction='mean') def compute_class_loss(self, class_logits, class_gt): """Computes the class loss for each of the decoder layers predictions or memory.""" class_loss = 0.0 for l in range(len(class_logits)): class_loss += self.loss_activity(class_logits[l], class_gt) return class_loss/len(class_logits) def select_loss_fn(self): if self._params['loss_fn'] == 'mse': return self.loss_mse elif self._params['loss_fn'] == 'smoothl1': return self.smooth_l1 elif self._params['loss_fn'] == 'l1': return self.loss_l1 else: raise ValueError('Unknown loss name {}.'.format(self._params['loss_fn'])) def layerwise_loss_fn(self, decoder_pred, decoder_gt, class_logits=None, class_gt=None): """Computes layerwise loss between predictions and ground truth.""" pose_loss = 0.0 loss_fn = self.select_loss_fn() for l in range(len(decoder_pred)): pose_loss += loss_fn(decoder_pred[l], decoder_gt) pose_loss = pose_loss/len(decoder_pred) if class_logits is not None: return pose_loss, self.compute_class_loss(class_logits, class_gt) return pose_loss, None def init_model(self, pose_encoder_fn=None, pose_decoder_fn=None): self._model = PoseTransformer.model_factory( self._params, pose_encoder_fn, pose_decoder_fn ) def select_optimizer(self): optimizer = optim.AdamW( self._model.parameters(), lr=self._params['learning_rate'], betas=(0.9, 0.999), weight_decay=_WEIGHT_DECAY ) return optimizer def dataset_factory(params, fold, model_prefix): if params['dataset'] == 'ntu_rgbd': return NTURGDDataset.dataset_factory(params) elif params['dataset'] == 'pd_gait': return GaitJointsDataset.dataset_factory(params, fold) else: raise ValueError('Unknown dataset {}'.format(params['dataset'])) def single_vote(pred): """ Get majority vote of predicted classes for the clips in one video. :param preds: list of predicted class for each clip of one video :return: majority vote of predicted class for one video """ p = np.array(pred) counts = np.bincount(p) max_count = 0 max_index = 0 for i in range(len(counts)): if max_count < counts[i]: max_index = i max_count = counts[i] return max_index def save_json(filename, attributes, names): """ Save training parameters and evaluation results to json file. :param filename: save filename :param attributes: attributes to save :param names: name of attributes to save in json file """ with open(filename, "w", encoding="utf8") as outfile: d = {} for i in range(len(attributes)): name = names[i] attribute = attributes[i] d[name] = attribute json.dump(d, outfile, indent=4, cls=NumpyEncoder) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--model_prefix', type=str, default='') parser.add_argument('--batch_size', type=int, default=16) parser.add_argument('--data_path', type=str) parser.add_argument('--learning_rate', type=float, default=1e-5) parser.add_argument('--max_epochs', type=int, default=500) parser.add_argument('--steps_per_epoch', type=int, default=200) parser.add_argument('--action', nargs='*', type=str, default=None) parser.add_argument('--use_one_hot', action='store_true') parser.add_argument('--init_fn', type=str, default='xavier_init') parser.add_argument('--include_last_obs', action='store_true') parser.add_argument('--task', type=str, default='downstream', choices=['pretext', 'downstream']) parser.add_argument('--downstream_strategy', default='both_then_class', choices=['both', 'class', 'both_then_class']) # pose transformers related parameters parser.add_argument('--model_dim', type=int, default=256) parser.add_argument('--num_encoder_layers', type=int, default=4) parser.add_argument('--num_decoder_layers', type=int, default=4) parser.add_argument('--num_heads', type=int, default=4) parser.add_argument('--dim_ffn', type=int, default=2048) parser.add_argument('--dropout', type=float, default=0.3) parser.add_argument('--source_seq_len', type=int, default=50) parser.add_argument('--target_seq_len', type=int, default=25) parser.add_argument('--max_gradient_norm', type=float, default=0.1) parser.add_argument('--lr_step_size',type=int, default=400) parser.add_argument('--learning_rate_fn',type=str, default='step') parser.add_argument('--warmup_epochs', type=int, default=100) parser.add_argument('--pose_format', type=str, default='rotmat') parser.add_argument('--remove_low_std', action='store_true') parser.add_argument('--remove_global_trans', action='store_true') parser.add_argument('--loss_fn', type=str, default='l1') parser.add_argument('--pad_decoder_inputs', action='store_true') parser.add_argument('--pad_decoder_inputs_mean', action='store_true') parser.add_argument('--use_wao_amass_joints', action='store_true') parser.add_argument('--non_autoregressive', action='store_true') parser.add_argument('--pre_normalization', action='store_true') parser.add_argument('--use_query_embedding', action='store_true') parser.add_argument('--predict_activity', action='store_true') parser.add_argument('--use_memory', action='store_true') parser.add_argument('--query_selection',action='store_true') parser.add_argument('--activity_weight', type=float, default=1.0) parser.add_argument('--pose_embedding_type', type=str, default='gcn_enc') parser.add_argument('--encoder_ckpt', type=str, default=None) parser.add_argument('--dataset', type=str, default='h36m_v2') parser.add_argument('--skip_rate', type=int, default=5) parser.add_argument('--eval_num_seeds', type=int, default=_NSEEDS) parser.add_argument('--copy_method', type=str, default=None) parser.add_argument('--finetuning_ckpt', type=str, default=None) parser.add_argument('--pos_enc_alpha', type=float, default=10) parser.add_argument('--pos_enc_beta', type=float, default=500) args = parser.parse_args() params = vars(args) if params['task'] == 'downstream': num_folds = 54 else: num_folds = 1 total_preds = [] total_gts = [] preds_votes = [] preds_probs = [] all_folds = range(1, 55) for fold in all_folds: print(f'Fold {fold} out of {num_folds}') utils.create_dir_tree(params['model_prefix']) # moving this up because dataset mean and std stored under it train_dataset_fn, eval_dataset_fn = dataset_factory(params, fold, params['model_prefix']) params['input_dim'] = train_dataset_fn.dataset._data_dim params['pose_dim'] = train_dataset_fn.dataset._pose_dim pose_encoder_fn, pose_decoder_fn = \ PoseEncoderDecoder.select_pose_encoder_decoder_fn(params) config_path = os.path.join(params['model_prefix'], 'config', 'config.json') with open(config_path, 'w') as file_: json.dump(params, file_, indent=4) model_fn = POTRModelFn( params, train_dataset_fn, eval_dataset_fn, pose_encoder_fn, pose_decoder_fn ) if params['task'] == 'downstream': predictions, gts, pred_probs = model_fn.train() print('predicitons:', predictions) # save predicted classes preds_votes.append(predictions.tolist()) # save predicted probabilities preds_probs.append(pred_probs.tolist()) # save final predictions and true labels if np.shape(gts)[0] == 1: # only 1 clip pred = int(predictions) else: pred = single_vote(predictions) gt = gts[0] total_preds.append(pred) total_gts.append(int(gt)) del model_fn, pose_encoder_fn, pose_decoder_fn attributes = [preds_votes, total_preds, preds_probs, total_gts] names = ['predicted_classes', 'predicted_final_classes', 'prediction_list', 'true_labels'] jsonfilename = os.path.join(params['model_prefix'], 'results.json') save_json(jsonfilename, attributes, names) else: model_fn.train() if params['task'] == 'downstream': print(classification_report(total_gts, total_preds))
11,607
37.059016
119
py
GaitForeMer
GaitForeMer-main/training/seq2seq_model_fn.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implements a model function estimator for training, evaluation and predict. Take and adapted from the code presented in [4] [1] https://github.com/asheshjain399/RNNexp/issues/6#issuecomment-249404882 [2] https://github.com/asheshjain399/RNNexp/blob/srnn/structural_rnn/CRFProblems/H3.6m/dataParser/Utils/motionGenerationError.m#L40-L54 [3] https://github.com/asheshjain399/RNNexp/issues/6#issuecomment-247769197 [4] https://arxiv.org/pdf/1705.02445.pdf """ import sys import numpy as np import json import sys import os import argparse import time from abc import abstractmethod import tqdm import torch import torch.optim as optim from torch.utils.tensorboard import SummaryWriter import torch.nn.functional as F # import wandb thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils import utils.WarmUpScheduler as warm_up_scheduler import visualize.viz as viz import models.seq2seq_model as seq2seq_model _DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') # min threshold for mean average precision in metters # Set to 10 cm _MAP_TRESH = 0.10 class ModelFn(object): """Implements the model functionalities: training, evaliation and prediction.""" def __init__( self, params, train_dataset_fn=None, eval_dataset_fn=None, pose_encoder_fn=None, pose_decoder_fn=None): """Initialization of model function.""" self._params = params self._train_dataset_fn = train_dataset_fn self._eval_dataset_fn = eval_dataset_fn self._visualize = False thisname = self.__class__.__name__ # self._norm_stats = train_dataset_fn.dataset._norm_stats self._norm_stats = None self._ms_range = [80, 160, 320, 400, 560, 1000] self.init_model(pose_encoder_fn, pose_decoder_fn) self._loss_fn = self.loss_mse self._model.to(_DEVICE) self._optimizer_fn = self.select_optimizer() self.select_lr_fn() self.finetune_init() self._lr_db_curve = [] lr_type = 'stepwise' if self._params['learning_rate_fn'] == 'beatles' \ else 'epochwise' self._params['lr_schedule_type'] = lr_type self.evaluate_fn = self.evaluate_nturgbd self._writer = SummaryWriter( os.path.join(self._params['model_prefix'], 'tf_logs')) self._time_range_eval = [] m_params = filter(lambda p: p.requires_grad, self._model.parameters()) nparams = sum([np.prod(p.size()) for p in m_params]) #print arguments # print('[INFO] ({}) This module has {} parameters!'.format(thisname, nparams)) # print('[INFO] ({}) Intializing ModelFn with params'.format(thisname)) # for k,v in self._params.items(): # print('[INFO] ({}) {}: {}'.format(thisname, k, v)) def finetune_init(self): if self._params['finetuning_ckpt'] is not None: print('[INFO] (finetune_model) Finetuning from:', self._params['finetuning_ckpt']) # edits made here to exclude activity prediction head model_state_dict = torch.load(self._params['finetuning_ckpt'], map_location=_DEVICE) if 'gait' in self._params['dataset']: # exclude prediction head del model_state_dict['_action_head.0.weight'] del model_state_dict['_action_head.0.bias'] self._model.load_state_dict(model_state_dict, strict=False) else: self._model.load_state_dict(model_state_dict) def select_lr_fn(self): """Calls the selection of learning rate function.""" self._lr_scheduler = self.get_lr_fn() lr_fn = self._params['learning_rate_fn'] if self._params['warmup_epochs'] > 0 and lr_fn != 'beatles': self._lr_scheduler = warm_up_scheduler.GradualWarmupScheduler( self._optimizer_fn, multiplier=1, total_epoch=self._params['warmup_epochs'], after_scheduler=self._lr_scheduler ) def get_lr_fn(self): """Creates the function to be used to generate the learning rate.""" if self._params['learning_rate_fn'] == 'step': return torch.optim.lr_scheduler.StepLR( self._optimizer_fn, step_size=self._params['lr_step_size'], gamma=0.1 ) elif self._params['learning_rate_fn'] == 'exponential': return torch.optim.lr_scheduler.ExponentialLR( self._optimizer_fn, gamma=0.95 ) elif self._params['learning_rate_fn'] == 'linear': # sets learning rate by multipliying initial learning rate times a function lr0, T = self._params['learning_rate'], self._params['max_epochs'] lrT = lr0*0.5 m = (lrT - 1) / T lambda_fn = lambda epoch: m*epoch + 1.0 return torch.optim.lr_scheduler.LambdaLR( self._optimizer_fn, lr_lambda=lambda_fn ) elif self._params['learning_rate_fn'] == 'beatles': # D^(-0.5)*min(i^(-0.5), i*warmup_steps^(-1.5)) D = float(self._params['model_dim']) warmup = self._params['warmup_epochs'] lambda_fn = lambda e: (D**(-0.5))*min((e+1.0)**(-0.5), (e+1.0)*warmup**(-1.5)) return torch.optim.lr_scheduler.LambdaLR( self._optimizer_fn, lr_lambda=lambda_fn ) else: raise ValueError('Unknown learning rate function: {}'.format( self._params['learning_rate_fn'])) @abstractmethod def init_model(self, pose_encoder_fn, pose_decoder_fn): pass @abstractmethod def select_optimizer(self): pass def loss_mse(self, decoder_pred, decoder_gt): """Computes the L2 loss between predictions and ground truth.""" step_loss = (decoder_pred - decoder_gt)**2 step_loss = step_loss.mean() return step_loss @abstractmethod def compute_loss(self, inputs=None, target=None, preds=None, class_logits=None, class_gt=None): return self._loss_fn(preds, target, class_logits, class_gt) def print_logs(self, step_loss, current_step, pose_loss, activity_loss, selection_loss): selection_logs = '' if self._params['query_selection']: selection_logs = 'selection loss {:.4f}'.format(selection_loss) if self._params['predict_activity']: print("[INFO] global {:06d}; step {:04d}; pose_loss {:4f} - class_loss {:4f}; step_loss: {:.4f}; lr: {:.2e} {:s}".\ format(self._global_step, current_step, pose_loss, activity_loss, step_loss, self._params['learning_rate'], selection_logs) ) else: print("[INFO] global {3:06d}; step {0:04d}; step_loss: {1:.4f}; lr: {2:.2e} {4:s}".\ format(current_step, step_loss, self._params['learning_rate'], self._global_step, selection_logs) ) def compute_selection_loss(self, inputs, target, cols_softmax=False): """Compute the query entry selection loss. Args: inputs: [batch_size, src_len, tgt_len] target: [batch_size, src_len, tgt_len] """ axis_ = 2 if cols_softmax else 1 target = F.softmax(-target, dim=axis_) return torch.nn.MSELoss(reduction='mean')(inputs, target) def train_one_epoch(self, epoch): """Trains for a number of steps before evaluation.""" epoch_loss = 0 act_loss = 0 sel_loss = 0 N = len(self._train_dataset_fn) for current_step, sample in enumerate(self._train_dataset_fn): self._optimizer_fn.zero_grad() for k in sample.keys(): if k == 'actions' or k == 'decoder_outputs_euler' or k=='action_str': continue sample[k] = sample[k].to(_DEVICE) decoder_pred = self._model( sample['encoder_inputs'], sample['decoder_inputs']) selection_loss = 0 if self._params['query_selection']: prob_mat = decoder_pred[-1][-1] selection_loss = self.compute_selection_loss( inputs=prob_mat, target=sample['src_tgt_distance'] ) sel_loss += selection_loss pred_class, gt_class = None, None if self._params['predict_activity']: gt_class = sample['action_ids'] # one label for the sequence pred_class = decoder_pred[1] pose_loss, activity_loss = self.compute_loss( inputs=sample['encoder_inputs'], target=sample['decoder_outputs'], preds=decoder_pred[0], class_logits=pred_class, class_gt=gt_class ) step_loss = pose_loss + selection_loss if self._params['predict_activity']: if self._params['task'] == 'pretext': step_loss += self._params['activity_weight']*activity_loss else: if self._params['downstream_strategy'] == 'both': step_loss += self._params['activity_weight']*activity_loss elif self._params['downstream_strategy'] == 'class': step_loss = activity_loss elif self._params['downstream_strategy'] == 'both_then_class': if epoch >= 50: step_loss = activity_loss else: step_loss += self._params['activity_weight']*activity_loss act_loss += activity_loss epoch_loss += step_loss.item() step_loss.backward() if self._params['max_gradient_norm'] is not None: torch.nn.utils.clip_grad_norm_( self._model.parameters(), self._params['max_gradient_norm']) self._optimizer_fn.step() if current_step % 10 == 0: step_loss = step_loss.cpu().data.numpy() # self.print_logs(step_loss, current_step, pose_loss, activity_loss, # selection_loss) self.update_learning_rate(self._global_step, mode='stepwise') self._global_step += 1 if self._params['query_selection']: self._scalars['train_selectioin_loss'] = sel_loss/N if self._params['predict_activity']: return epoch_loss/N, act_loss/N return epoch_loss/N def train(self): """Main training loop.""" self._params['learning_rate'] = self._lr_scheduler.get_lr()[0] self._global_step = 1 thisname = self.__class__.__name__ # wandb.init(name='training', project='GaitForeMer') for e in range(self._params['max_epochs']): self._scalars = {} self._model.train() start_time = time.time() epoch_loss = self.train_one_epoch(e) act_log = '' if self._params['predict_activity']: act_loss = epoch_loss[1] epoch_loss = epoch_loss[0] act_log = '; activity_loss: {}'.format(act_loss) self._scalars['act_loss_train'] = act_loss self._scalars['epoch_loss'] = epoch_loss print("epoch {0:04d}; epoch_loss: {1:.4f}".format(e, epoch_loss)+act_log) self.flush_extras(e, 'train') _time = time.time() - start_time self._model.eval() eval_loss = self.evaluate_fn(e, _time) act_log = '' if self._params['predict_activity']: self._scalars['act_loss_eval'] = eval_loss[1] self._scalars['accuracy'] = eval_loss[2] act_log = '; act_eval_loss {}; accuracy {}'.format(eval_loss[1], eval_loss[2]) eval_activity_loss = eval_loss[1] eval_accuracy = eval_loss[2] # eval_loss = eval_loss[0] self._scalars['eval_loss'] = eval_loss[0] print("[INFO] ({}) Epoch {:04d}; eval_loss: {:.4f}; lr: {:.2e}".format( thisname, e, eval_loss[0], self._params['learning_rate'])+act_log) self.write_summary(e) # wandb_logs = {"train loss": epoch_loss, "train activity loss": act_loss, "eval loss": eval_loss, "eval activity loss": eval_activity_loss, "eval accuracy": eval_accuracy} # wandb.log(wandb_logs) model_path = os.path.join( self._params['model_prefix'], 'models', 'ckpt_epoch_%04d.pt'%e) if (e+1)%100 == 0: torch.save(self._model.state_dict(), model_path) self.update_learning_rate(e, mode='epochwise') self.flush_extras(e, 'eval') # return predictions and real ones predictions = eval_loss[3] gt = eval_loss[4] pred_probs = eval_loss[5] return predictions, gt, pred_probs # save the last one # model_path = os.path.join( # self._params['model_prefix'], 'models', 'ckpt_epoch_%04d.pt'%e) # torch.save(self._model.state_dict(). model_path) # self.flush_curves() def write_summary(self, epoch): # for action_, ms_errors_ in ms_eval_loss.items(): self._writer.add_scalars( 'loss/recon_loss', {'train':self._scalars['epoch_loss'], 'eval': self._scalars['eval_loss']}, epoch ) # write scalars for H36M dataset prediction style action_ = self._train_dataset_fn.dataset._monitor_action if 'ms_eval_loss' in self._scalars.keys(): range_len = len(self._scalars['ms_eval_loss'][action_]) # range_len = len(self._ms_range) ms_dict = {str(self._ms_range[i]): self._scalars['ms_eval_loss'][action_][i] for i in range(range_len)} ms_e = np.concatenate([np.array(v).reshape(1,range_len) for k,v in self._scalars['ms_eval_loss'].items()], axis=0) self._writer.add_scalars('ms_loss/eval-'+action_, ms_dict, epoch) ms_e = np.mean(ms_e, axis=0) # (n_actions) self._time_range_eval.append(np.expand_dims(ms_e, axis=0)) # (1, n_actions) all_ms = {str(self._ms_range[i]): ms_e[i] for i in range(len(ms_e))} self._writer.add_scalars('ms_loss/eval-all', all_ms, epoch) self._writer.add_scalar('MSRE/msre_eval', self._scalars['msre'], epoch) self._writer.add_scalars('time_range/eval', {'short-term':np.mean(ms_e[:4]), 'long-term':np.mean(ms_e)}, epoch) if self._params['predict_activity']: self._writer.add_scalars( 'loss/class_loss', {'train': self._scalars['act_loss_train'], 'eval': self._scalars['act_loss_eval']}, epoch ) self._writer.add_scalar('class/accuracy', self._scalars['accuracy'], epoch) if self._params['query_selection']: self._writer.add_scalars( 'selection/query_selection', {'eval': self._scalars['eval_selection_loss'], 'train': self._scalars['train_selectioin_loss']}, epoch ) if 'mAP' in self._scalars.keys(): self._writer.add_scalar('mAP/mAP', self._scalars['mAP'], epoch) if 'MPJPE' in self._scalars.keys(): self._writer.add_scalar('MPJPE/MPJPE', self._scalars['MPJPE'], epoch) def print_range_summary(self, action, mean_mean_errors): mean_eval_error = [] # Pretty print of the results for 80, 160, 320, 400, 560 and 1000 ms print("{0: <16} |".format(action), end="") for ms in [1,3,7,9,13,24]: if self._params['target_seq_len'] >= ms + 1: print(" {0:.3f} |".format(mean_mean_errors[ms]), end="") mean_eval_error.append(mean_mean_errors[ms]) else: print(" n/a |", end="") print() return mean_eval_error def print_table_header(self): print() print("{0: <16} |".format("milliseconds"), end="") for ms in [80, 160, 320, 400, 560, 1000]: print(" {0:5d} |".format(ms), end="") print() def flush_curves(self): path_ = os.path.join(self._params['model_prefix'], 'loss_info') os.makedirs(path_, exist_ok=True) path_ = os.path.join(path_, 'eval_time_range.npy') np.save(path_, np.concatenate(self._time_range_eval, axis=0)) path_ = os.path.join(path_, 'lr_schedule.npy') np.save(path_, np.array(self._lr_db_curve)) def update_learning_rate(self, epoch_step, mode='stepwise'): """Update learning rate handler updating only when the mode matches.""" if self._params['lr_schedule_type'] == mode: self._lr_scheduler.step(epoch_step) self._writer.add_scalar( 'learning_rate/lr', self._params['learning_rate'], epoch_step) self._lr_db_curve.append([self._params['learning_rate'], epoch_step]) # self._params['learning_rate'] = self._lr_scheduler.get_last_lr()[0] self._params['learning_rate'] = self._lr_scheduler.get_lr()[0] @abstractmethod def flush_extras(self, epoch, phase): pass def compute_class_accurracy_sequence(self, class_logits, class_gt): # softmax on last dimension and get max on last dimension class_pred = torch.argmax(class_logits.softmax(-1), -1) accuracy = (class_pred == class_gt).float().sum() accuracy = accuracy / class_logits.size()[0] return accuracy.item() def compute_class_accurracy_instance(self, class_logits, class_gt): # softmax on last dimension and get max on last dimension tar_seq_len = self._params['target_seq_len'] class_pred = torch.argmax(class_logits.softmax(-1), -1) accuracy = (class_pred == class_gt).float().sum() accuracy = accuracy / (class_logits.size()[0]*tar_seq_len) return accuracy.item() def validation_srnn_ms(self, sample, decoder_pred): # the data was flatened from a sequence of size # [n_actions, n_seeds, target_length, pose_size] n_actions = len(self._params['action_subset']) seq_shape = (n_actions, self._params['eval_num_seeds'], self._params['target_seq_len'], self._params['pose_dim']) srnn_gts_euler = sample['decoder_outputs_euler'] decoder_pred_ = decoder_pred.cpu().numpy() decoder_pred_ = decoder_pred_.reshape(seq_shape) do_remove = self._params['remove_low_std'] mean_eval_error_dict = {} self.print_table_header() eval_ms_mean = [] for ai, action in enumerate(sample['actions']): action = action[0] decoder_pred = decoder_pred_[ai, :, :, :] if self._params['dataset'] == 'h36m': # seq_len x n_seeds x pose_dim decoder_pred = decoder_pred.transpose([1, 0, 2]) # a list or a vector of length n_seeds # each entry of: shape seq_len x complete_pose_dim (H36M == 99) srnn_pred_euler = self._eval_dataset_fn.dataset.post_process_to_euler(decoder_pred) # n_seeds x seq_len mean_errors = np.zeros((self._params['eval_num_seeds'], self._params['target_seq_len'])) # Training is done in exponential map or rotation matrix or quaternion # but the error is reported in Euler angles, as in previous work [3,4,5] for i in np.arange(self._params['eval_num_seeds']): # seq_len x complete_pose_dim (H36M==99) eulerchannels_pred = srnn_pred_euler[i] # n_seeds x seq_len x complete_pose_dim (H36M==96) action_gt = srnn_gts_euler[action] # seq_len x complete_pose_dim (H36M==96) gt_i = np.copy(action_gt.squeeze()[i].numpy()) # Only remove global rotation. Global translation was removed before gt_i[:, 0:3] = 0 # here [2,4,5] remove data based on the std of the batch THIS IS WEIRD! # (seq_len, 96) - (seq_len, 96) idx_to_use = np.where(np.std(gt_i, 0) > 1e-4)[0] euc_error = np.power(gt_i[:,idx_to_use] - eulerchannels_pred[:,idx_to_use], 2) # shape: seq_len euc_error = np.sum(euc_error, 1) # shape: seq_len euc_error = np.sqrt(euc_error) mean_errors[i,:] = euc_error # This is simply the mean error over the eval_num_seeds examples # with shape [eval_num_seeds] mean_mean_errors = np.mean(mean_errors, 0) mean_eval_error_dict[action] = self.print_range_summary(action, mean_mean_errors) return mean_eval_error_dict @torch.no_grad() def evaluate_nturgbd(self, current_step, dummy_entry=None): eval_loss = 0.0 mAP_all = 0.0 class_loss = 0.0 mean_accuracy = 0.0 N = len(self._eval_dataset_fn) gt_class_ = [] pred_class_ = [] num_joints = self._params['pose_dim'] // 3 TP = np.zeros((num_joints,)) FN = np.zeros((num_joints,)) MPJPE = np.zeros((num_joints,)) for (i, sample) in tqdm.tqdm(enumerate(self._eval_dataset_fn)): for k in sample.keys(): if k=='action_str': continue sample[k] = sample[k].to(_DEVICE) decoder_pred = self._model( sample['encoder_inputs'], sample['decoder_inputs']) pred_class, gt_class = None, None if self._params['predict_activity']: gt_class = sample['action_ids'] # one label for the sequence pred_class = decoder_pred[1] decoder_pred = decoder_pred[0] gt_class_.append(gt_class.item()) pred_class_.append(pred_class[-1].cpu().numpy()) pose_loss, activity_loss = self.compute_loss( inputs=sample['encoder_inputs'], target=sample['decoder_outputs'], preds=decoder_pred, class_logits=pred_class, class_gt=gt_class ) # Can save predicted outputs for visualization here # if i == 2: # predicted_pose = decoder_pred[-1].squeeze().reshape(20, 17, 3).cpu().numpy() # input_pose = sample['encoder_inputs'].squeeze().reshape(39, 17, 3).cpu().numpy() # gt_pose = sample['decoder_outputs'].squeeze().reshape(20, 17, 3).cpu().numpy() # np.save('output_poses/v37_pred.npy', predicted_pose) # np.save('output_poses/v37_gt.npy', gt_pose) # np.save('output_poses/v37_input.npy', input_pose) # # break eval_loss+= pose_loss eval_loss /= N if self._params['predict_activity']: class_loss /= N pred_class_ = torch.squeeze(torch.from_numpy(np.stack(pred_class_))) gt_class_ = torch.from_numpy(np.array(gt_class_)) # print(pred_class_.size(), gt_class_.size()) accuracy = self.compute_class_accurracy_sequence(pred_class_, gt_class_) return (eval_loss, class_loss, accuracy, torch.argmax(pred_class_.softmax(-1), -1), gt_class_, pred_class_.softmax(-1)) # return (eval_loss, class_loss, accuracy, torch.argmax(pred_class_.softmax(-1), -1), gt_class_) # return (eval_loss, class_loss, accuracy) return eval_loss def compute_mean_average_precision(self, prediction, target): pred = np.squeeze(prediction) tgt = np.squeeze(target) T, D = pred.shape pred = self._eval_dataset_fn.dataset.unormalize_sequence(pred) tgt = self._eval_dataset_fn.dataset.unormalize_sequence(tgt) # num_frames x num_joints x 3 pred = pred.reshape((T, -1, 3)) tgt = tgt.reshape((T, -1, 3)) # compute the norm for the last axis: (x,y,z) coordinates # num_frames x num_joints TP = np.linalg.norm(pred-tgt, axis=-1) <= _MAP_TRESH TP = TP.astype(int) FN = np.logical_not(TP).astype(int) # num_joints TP = np.sum(TP, axis=0) FN = np.sum(FN, axis=0) # compute recall for each joint recall = TP / (TP+FN) # average over joints mAP = np.mean(recall) return mAP, TP, FN def compute_MPJPE(self, prediction, target): pred = np.squeeze(prediction) tgt = np.squeeze(target) T, D = pred.shape pred = self._eval_dataset_fn.dataset.unormalize_sequence(pred) tgt = self._eval_dataset_fn.dataset.unormalize_sequence(tgt) # num_frames x num_joints x 3 pred = pred.reshape((T, -1, 3)) tgt = tgt.reshape((T, -1, 3)) # compute the norm for the last axis: (x,y,z) coordinates # num_frames x num_joints norm = np.linalg.norm(pred-tgt, axis=-1) # num_joints MPJPE = np.mean(norm, axis=0) return MPJPE
24,065
36.024615
179
py
GaitForeMer
GaitForeMer-main/training/pose_classifier_fn.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """ [1] https://arxiv.org/abs/1312.6114 """ import torch import torch.nn as nn from torch.utils.tensorboard import SummaryWriter import torch.optim as optim import numpy as np import os import sys import argparse import tqdm thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import models.PoseActionClassifier as ActionClass import data.H36MDatasetPose as H36MDataset import utils.utils as utils _DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') class PoseActionFn(object): def __init__(self, params, train_dataset, val_dataset=None): self._params = params self._train_dataset = train_dataset self._val_dataset = val_dataset self._writer = SummaryWriter( os.path.join(self._params['model_prefix'], 'tf_logs')) self.load_constants() self.init_model() self.select_optimizer() thisname = self.__class__.__name__ self._lr_scheduler = utils.get_lr_fn(self._params, self._optimizer_fn) for k, v in self._params.items(): print('[INFO] ({}) {}: {}'.format(thisname, k, v)) def load_constants(self): self._params['use_one_hot'] = False self._params['parent'], self._params['offset'], \ self._params['rot_ind'], self._params['exp_map_ind'] = \ utils.load_constants(self._params['data_path']) def init_model(self): self._model = ActionClass.ActionClassifier( dim=self._params['model_dim'], n_classes=len(self._params['action_subset']) ) self._model.to(_DEVICE) n_params = filter(lambda p: p.requires_grad, self._model.parameters()) n_params = sum([np.prod(p.size()) for p in n_params]) print('++++++++ Total Parameters:', n_params) def select_optimizer(self): self._optimizer_fn = optim.Adam( self._model.parameters(), lr=self._params['learning_rate'] ) def compute_accuracy(self, class_logits, class_gt): class_pred = torch.argmax(class_logits.softmax(-1), -1) accuracy = (class_pred == class_gt).float().sum() accuracy = accuracy / (class_logits.size()[0]) return accuracy def forward_model(self, sample): pose_gt = sample['pose'].to(_DEVICE) class_gt = sample['action'].to(_DEVICE) class_logits = self._model(pose_gt) loss = nn.functional.cross_entropy(class_logits, class_gt, reduction='mean') accuracy = self.compute_accuracy(class_logits, class_gt) return loss, accuracy def train_one_epoch(self, epoch): epoch_loss, epoch_accuracy = 0, 0 N = len(self._train_dataset) self._model.train() for i, sample in enumerate(self._train_dataset): self._optimizer_fn.zero_grad() loss, accuracy = self.forward_model(sample) if i%1000 == 0: print('[INFO] epoch: {:04d}; it: {:04d} loss: {:.4f}; acc: {:.4f}'.format( epoch, i, loss, accuracy)) loss.backward() self._optimizer_fn.step() epoch_loss += loss epoch_accuracy += accuracy return epoch_loss/N, epoch_accuracy/N @torch.no_grad() def validation(self, epoch): epoch_loss, epoch_accuracy = 0, 0 N = len(self._val_dataset) self._model.eval() for i, sample in tqdm.tqdm(enumerate(self._val_dataset)): loss, accuracy = self.forward_model(sample) epoch_loss += loss epoch_accuracy += accuracy return epoch_loss/N, epoch_accuracy/N def train(self): thisname = self.__class__.__name__ self._params['learning_rate'] = self._lr_scheduler.get_last_lr()[0] for e in range(self._params['max_epochs']): self._model.train() epoch_loss, epoch_accuracy = self.train_one_epoch(e) val_loss, val_accuracy = self.validation(e) # save models model_path = os.path.join( self._params['model_prefix'], 'models', 'ckpt_epoch_%04d.pt'%e) torch.save(self._model.state_dict(), model_path) # verbose and write the scalars print('[INFO] Epoch: {:04d}; epoch_loss: {:.4f}; epoch_accuracy: {:.4f}; val_loss: {:.4f}; val_accuracy: {:.4f}; lr: {:2.2e}'.format( e, epoch_loss, epoch_accuracy, val_loss, val_accuracy, self._params['learning_rate'])) self._writer.add_scalars( 'loss/loss', {'train': epoch_loss, 'val': val_loss}, e) self._writer.add_scalars( 'accurracy/accurracy', {'train': epoch_accuracy, 'val': val_accuracy}, e) self._writer.add_scalar( 'learning_rate/lr', self._params['learning_rate'], e) self._lr_scheduler.step(e) self._params['learning_rate'] = self._lr_scheduler.get_last_lr()[0] if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--data_path', type=str, default=None) parser.add_argument('--action', type=str, nargs='*', default=None) parser.add_argument('--batch_size', type=int, default=128) parser.add_argument('--pose_format', type=str, default='expmap') parser.add_argument('--remove_low_std', action='store_true') parser.add_argument('--model_dim', type=int, default=128) parser.add_argument('--max_epochs', type=int, default=500) parser.add_argument('--model_prefix', type=str, default=None) parser.add_argument('--learning_rate', type=float, default=1e-3) parser.add_argument('--learning_rate_fn', type=str, default='linear') args = parser.parse_args() params = vars(args) if 'all' in args.action: args.action = H36MDataset._ALL_ACTIONS params['action_subset'] = args.action dataset_t = H36MDataset.H36MDataset(params, mode='train') dataset_v = H36MDataset.H36MDataset( params, mode='eval', norm_stats=dataset_t._norm_stats) train_dataset_fn= torch.utils.data.DataLoader( dataset_t, batch_size=params['batch_size'], shuffle=True, num_workers=4, collate_fn=H36MDataset.collate_fn, drop_last=True ) val_dataset_fn = torch.utils.data.DataLoader( dataset_v, batch_size=1, shuffle=True, num_workers=1, collate_fn=H36MDataset.collate_fn, drop_last=True ) params['input_dim'] = train_dataset_fn.dataset._data_dim params['pose_dim'] = train_dataset_fn.dataset._pose_dim vae_trainer = PoseActionFn(params, train_dataset_fn, val_dataset_fn) vae_trainer.train()
7,534
34.21028
139
py
GaitForeMer
GaitForeMer-main/models/Conv1DEncoder.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Model of 1D convolutions for encoding pose sequences.""" import numpy as np import os import sys thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import torch import torch.nn as nn class Pose1DEncoder(nn.Module): def __init__(self, input_channels=3, output_channels=128, n_joints=21): super(Pose1DEncoder, self).__init__() self._input_channels = input_channels self._output_channels = output_channels self._n_joints = n_joints self.init_model() def init_model(self): self._model = nn.Sequential( nn.Conv1d(in_channels=self._input_channels, out_channels=32, kernel_size=7), nn.BatchNorm1d(32), nn.ReLU(True), nn.Conv1d(in_channels=32, out_channels=32, kernel_size=3), nn.BatchNorm1d(32), nn.ReLU(True), nn.Conv1d(in_channels=32, out_channels=64, kernel_size=3), nn.BatchNorm1d(64), nn.ReLU(True), nn.Conv1d(in_channels=64, out_channels=64, kernel_size=3), nn.BatchNorm1d(64), nn.ReLU(True), nn.Conv1d(in_channels=64, out_channels=128, kernel_size=3), nn.BatchNorm1d(128), nn.ReLU(True), nn.Conv1d(in_channels=128, out_channels=128, kernel_size=3), nn.BatchNorm1d(128), nn.ReLU(True), nn.Conv1d(in_channels=128, out_channels=self._output_channels, kernel_size=3), nn.BatchNorm1d(self._output_channels), nn.ReLU(True), nn.Conv1d(in_channels=self._output_channels, out_channels=self._output_channels, kernel_size=3) ) def forward(self, x): """ Args: x: [batch_size, seq_len, skeleton_dim]. """ # inputs to model is [batch_size, channels, n_joints] # transform the batch to [batch_size*seq_len, dof, n_joints] bs, seq_len, dim = x.size() dof = dim//self._n_joints x = x.view(bs*seq_len, dof, self._n_joints) # [batch_size*seq_len, dof, n_joints] x = self._model(x) # [batch_size, seq_len, output_channels] x = x.view(bs, seq_len, self._output_channels) return x class Pose1DTemporalEncoder(nn.Module): def __init__(self, input_channels, output_channels): super(Pose1DTemporalEncoder, self).__init__() self._input_channels = input_channels self._output_channels = output_channels self.init_model() def init_model(self): self._model = nn.Sequential( nn.Conv1d( in_channels=self._input_channels, out_channels=32, kernel_size=3, padding=1), nn.BatchNorm1d(32), nn.ReLU(True), nn.Conv1d(in_channels=32, out_channels=32, kernel_size=3, padding=1), nn.BatchNorm1d(32), nn.ReLU(True), nn.Conv1d(in_channels=32, out_channels=64, kernel_size=3, padding=1), nn.BatchNorm1d(64), nn.ReLU(True), nn.Conv1d(in_channels=64, out_channels=64, kernel_size=3, padding=1), nn.BatchNorm1d(64), nn.ReLU(True), nn.Conv1d(in_channels=64, out_channels=128, kernel_size=3, padding=1), nn.BatchNorm1d(128), nn.ReLU(True), nn.Conv1d(in_channels=128, out_channels=128, kernel_size=3, padding=1), nn.BatchNorm1d(128), nn.ReLU(True), nn.Conv1d(in_channels=128, out_channels=self._output_channels, kernel_size=3, padding=1), nn.BatchNorm1d(self._output_channels), nn.ReLU(True), nn.Conv1d(in_channels=self._output_channels, out_channels=self._output_channels, kernel_size=3, padding=1) ) def forward(self, x): """ Args: x: [batch_size, seq_len, skeleton_dim]. """ # batch_size, skeleton_dim, seq_len x = torch.transpose(x, 1,2) x = self._model(x) # batch_size, seq_len, skeleton_dim x = torch.transpose(x, 1, 2) return x if __name__ == '__main__': dof = 9 output_channels = 128 n_joints = 21 seq_len = 49 model = Pose1DTemporalEncoder(input_channels=dof*n_joints, output_channels=output_channels) inputs = torch.FloatTensor(10, seq_len, dof*n_joints) X = model(inputs) print(X.size()) # model = Pose1DEncoder(input_channels=dof, output_channels=output_channels) # inputs = torch.FloatTensor(10, seq_len, dof*n_joints) # X = model(inputs) # print(X.size())
5,262
32.954839
114
py
GaitForeMer
GaitForeMer-main/models/Transformer.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implementation of the Transformer for sequence-to-sequence decoding. Implementation of the transformer for sequence to sequence prediction as in [1] and [2]. [1] https://arxiv.org/pdf/1706.03762.pdf [2] https://arxiv.org/pdf/2005.12872.pdf """ import numpy as np import os import sys import copy import torch import torch.nn as nn import torch.nn.functional as F from scipy.optimize import linear_sum_assignment thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils import utils.PositionEncodings as PositionEncodings import models.TransformerEncoder as Encoder import models.TransformerDecoder as Decoder class Transformer(nn.Module): def __init__(self, num_encoder_layers=6, num_decoder_layers=6, model_dim=256, num_heads=8, dim_ffn=2048, dropout=0.1, init_fn=utils.normal_init_, use_query_embedding=False, pre_normalization=False, query_selection=False, target_seq_len=25): """Implements the Transformer model for sequence-to-sequence modeling.""" super(Transformer, self).__init__() self._model_dim = model_dim self._num_heads = num_heads self._dim_ffn = dim_ffn self._dropout = dropout self._use_query_embedding = use_query_embedding self._query_selection = query_selection self._tgt_seq_len = target_seq_len self._encoder = Encoder.TransformerEncoder( num_layers=num_encoder_layers, model_dim=model_dim, num_heads=num_heads, dim_ffn=dim_ffn, dropout=dropout, init_fn=init_fn, pre_normalization=pre_normalization ) self._decoder = Decoder.TransformerDecoder( num_layers=num_decoder_layers, model_dim=model_dim, num_heads=num_heads, dim_ffn=dim_ffn, dropout=dropout, init_fn=init_fn, use_query_embedding=use_query_embedding, pre_normalization=pre_normalization ) if self._query_selection: self._position_predictor = nn.Linear(self._model_dim, self._tgt_seq_len) def process_index_selection(self, self_attn, one_to_one_selection=False): """Selection of query elments using position predictor from encoder memory. After prediction a maximum assignement problem is solved to get indices for each element in the query sequence. Args: self_attn: Encoder memory with shape [src_len, batch_size, model_dim] Returns: A tuple with two list of i and j matrix entries of m """ batch_size = self_attn.size()[1] # batch_size x src_seq_len x model_dim in_pos = torch.transpose(self_attn, 0, 1) # predict the matrix of similitudes # batch_size x src_seq_len x tgt_seq_len prob_matrix = self._position_predictor(in_pos) # apply softmax on the similutes to get probabilities on positions # batch_size x src_seq_len x tgt_seq_len if one_to_one_selection: soft_matrix = F.softmax(prob_matrix, dim=2) # predict assignments in a one to one fashion maximizing the sum of probs indices = [linear_sum_assignment(soft_matrix[i].cpu().detach(), maximize=True) for i in range(batch_size) ] else: # perform softmax by rows to have many targets to one input assignements soft_matrix = F.softmax(prob_matrix, dim=1) indices_rows = torch.argmax(soft_matrix, 1) indices = [(indices_rows[i], list(range(prob_matrix.size()[2]))) for i in range(batch_size) ] return indices, soft_matrix def forward(self, source_seq, target_seq, encoder_position_encodings=None, decoder_position_encodings=None, query_embedding=None, mask_target_padding=None, mask_look_ahead=None, get_attn_weights=False, query_selection_fn=None, fold=None, eval_step=None): if self._use_query_embedding: bs = source_seq.size()[1] query_embedding = query_embedding.unsqueeze(1).repeat(1, bs, 1) decoder_position_encodings = encoder_position_encodings memory, enc_weights = self._encoder(source_seq, encoder_position_encodings) # Save encoder outputs # if fold is not None: # encoder_output_dir = 'encoder_outputs' # if not os.path.exists(f'{encoder_output_dir}f{fold}/'): # os.makedirs(f'{encoder_output_dir}f{fold}/') # outpath = f'{encoder_output_dir}f{fold}/{eval_step}.npy' # encoder_output = memory.detach().cpu().numpy() # np.save(outpath, encoder_output) tgt_plain = None # perform selection from input sequence if self._query_selection: indices, prob_matrix = self.process_index_selection(memory) tgt_plain, target_seq = query_selection_fn(indices) out_attn, out_weights = self._decoder( target_seq, memory, decoder_position_encodings, query_embedding=query_embedding, mask_target_padding=mask_target_padding, mask_look_ahead=mask_look_ahead, get_attn_weights=get_attn_weights ) out_weights_ = None enc_weights_ = None prob_matrix_ = None if get_attn_weights: out_weights_, enc_weights_ = out_weights, enc_weights if self._query_selection: prob_matrix_ = prob_matrix return out_attn, memory, out_weights_, enc_weights_, (tgt_plain, prob_matrix_)
6,581
33.103627
85
py
GaitForeMer
GaitForeMer-main/models/PoseTransformer.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implementation of the Transformer for sequence-to-sequence decoding. Implementation of the transformer for sequence to sequence prediction as in [1] and [2]. [1] https://arxiv.org/pdf/1706.03762.pdf [2] https://arxiv.org/pdf/2005.12872.pdf """ import numpy as np import os import sys import copy import torch import torch.nn as nn import torch.nn.functional as F thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils import utils.PositionEncodings as PositionEncodings import models.TransformerEncoder as Encoder import models.TransformerDecoder as Decoder from models.Transformer import Transformer _SOURCE_LENGTH = 110 _TARGET_LENGTH = 55 _POSE_DIM = 54 _PAD_LENGTH = _SOURCE_LENGTH class PoseTransformer(nn.Module): """Implements the sequence-to-sequence Transformer .model for pose prediction.""" def __init__(self, pose_dim=_POSE_DIM, model_dim=256, num_encoder_layers=6, num_decoder_layers=6, num_heads=8, dim_ffn=2048, dropout=0.1, target_seq_length=_TARGET_LENGTH, source_seq_length=_SOURCE_LENGTH, input_dim=None, init_fn=utils.xavier_init_, non_autoregressive=False, use_query_embedding=False, pre_normalization=False, predict_activity=False, use_memory=False, num_activities=None, pose_embedding=None, pose_decoder=None, copy_method='uniform_scan', query_selection=False, pos_encoding_params=(10000, 1)): """Initialization of pose transformers.""" super(PoseTransformer, self).__init__() self._target_seq_length = target_seq_length self._source_seq_length = source_seq_length self._pose_dim = pose_dim self._input_dim = pose_dim if input_dim is None else input_dim self._model_dim = model_dim self._use_query_embedding = use_query_embedding self._predict_activity = predict_activity self._num_activities = num_activities self._num_decoder_layers = num_decoder_layers self._mlp_dim = model_dim self._non_autoregressive = non_autoregressive self._pose_embedding = pose_embedding self._pose_decoder = pose_decoder self._query_selection = query_selection thisname = self.__class__.__name__ self._copy_method = copy_method self._pos_encoding_params = pos_encoding_params self._transformer = Transformer( num_encoder_layers=num_encoder_layers, num_decoder_layers=num_decoder_layers, model_dim=model_dim, num_heads=num_heads, dim_ffn=dim_ffn, dropout=dropout, init_fn=init_fn, use_query_embedding=use_query_embedding, pre_normalization=pre_normalization, query_selection=query_selection, target_seq_len=target_seq_length ) t_params = filter(lambda p: p.requires_grad, self._transformer.parameters()) nparams = sum([np.prod(p.size()) for p in t_params]) print('[INFO] ({}) Transformer has {} parameters!'.format(thisname, nparams)) self._pos_encoder = PositionEncodings.PositionEncodings1D( num_pos_feats=self._model_dim, temperature=self._pos_encoding_params[0], alpha=self._pos_encoding_params[1] ) self._pos_decoder = PositionEncodings.PositionEncodings1D( num_pos_feats=self._model_dim, temperature=self._pos_encoding_params[0], alpha=self._pos_encoding_params[1] ) # self.init_pose_encoder_decoders(init_fn) self._use_class_token = False self.init_position_encodings() self.init_query_embedding() if self._use_class_token: self.init_class_token() if self._predict_activity: self._action_head_size = self._model_dim if self._use_class_token \ else self._model_dim*(self._source_seq_length-1) self._action_head = nn.Sequential( nn.Linear(self._action_head_size, self._num_activities), ) def init_query_embedding(self): """Initialization of query sequence embedding.""" self._query_embed = nn.Embedding(self._target_seq_length, self._model_dim) print('[INFO] ({}) Init query embedding!'.format(self.__class__.__name__)) nn.init.xavier_uniform_(self._query_embed.weight.data) # self._query_embed.weight.data.normal_(0.0, 0.004) def init_class_token(self): token = torch.FloatTensor(1, self._model_dim) print('[INFO] ({}) Init class token!'.format(self.__class__.__name__)) self._class_token = nn.Parameter(token, requires_grad=True) nn.init.xavier_uniform_(self._class_token.data) def init_position_encodings(self): src_len = self._source_seq_length-1 # when using a token we need an extra element in the sequence if self._use_class_token: src_len = src_len + 1 encoder_pos_encodings = self._pos_encoder(src_len).view( src_len, 1, self._model_dim) decoder_pos_encodings = self._pos_decoder(self._target_seq_length).view( self._target_seq_length, 1, self._model_dim) mask_look_ahead = torch.from_numpy( utils.create_look_ahead_mask( self._target_seq_length, self._non_autoregressive)) self._encoder_pos_encodings = nn.Parameter( encoder_pos_encodings, requires_grad=False) self._decoder_pos_encodings = nn.Parameter( decoder_pos_encodings, requires_grad=False) self._mask_look_ahead = nn.Parameter( mask_look_ahead, requires_grad=False) def forward(self, input_pose_seq, target_pose_seq=None, mask_target_padding=None, get_attn_weights=False, fold=None, eval_step=None): """Performs the forward pass of the pose transformers. Args: input_pose_seq: Shape [batch_size, src_sequence_length, dim_pose]. target_pose_seq: Shape [batch_size, tgt_sequence_length, dim_pose]. Returns: A tensor of the predicted sequence with shape [batch_size, tgt_sequence_length, dim_pose]. """ if self.training: return self.forward_training( input_pose_seq, target_pose_seq, mask_target_padding, get_attn_weights) # no fold here, only want to save eval # eval forward for non auto regressive type of model if self._non_autoregressive: return self.forward_training( input_pose_seq, target_pose_seq, mask_target_padding, get_attn_weights, fold=fold, eval_step=eval_step) return self.forward_autoregressive( input_pose_seq, target_pose_seq, mask_target_padding, get_attn_weights) def handle_class_token(self, input_pose_seq): """ Args: input_pose_seq: [src_len, batch_size, model_dim] """ # concatenate extra token for activity prediction as an extra # element of the input sequence # specialized token is not a skeleton _, B, _ = input_pose_seq.size() token = self._class_token.squeeze().repeat(1, B, 1) input_pose_seq = torch.cat([token, input_pose_seq], axis=0) return input_pose_seq def handle_copy_query(self, indices, input_pose_seq_): """Handles the way queries are generated copying items from the inputs. Args: indices: A list of tuples len `batch_size`. Each tuple contains has the form (input_list, target_list) where input_list contains indices of elements in the input to be copy to elements in the target specified by target_list. input_pose_seq_: Source skeleton sequence [batch_size, src_len, pose_dim]. Returns: A tuple with first elements the decoder input skeletons with shape [tgt_len, batch_size, skeleton_dim], and the skeleton embeddings of the input sequence with shape [tgt_len, batch_size, pose_dim]. """ batch_size = input_pose_seq_.size()[0] decoder_inputs = torch.FloatTensor( batch_size, self._target_seq_length, self._pose_dim ).to(self._decoder_pos_encodings.device) for i in range(batch_size): for j in range(self._target_seq_length): src_idx, tgt_idx = indices[i][0][j], indices[i][1][j] decoder_inputs[i, tgt_idx] = input_pose_seq_[i, src_idx] dec_inputs_encode = self._pose_embedding(decoder_inputs) return torch.transpose(decoder_inputs, 0, 1), \ torch.transpose(dec_inputs_encode, 0, 1) def forward_training(self, input_pose_seq_, target_pose_seq_, mask_target_padding, get_attn_weights=False, fold=None, eval_step=None): """Compute forward pass for training and non recursive inference. Args: input_pose_seq_: Source sequence [batch_size, src_len, skeleton_dim]. target_pose_seq_: Query target sequence [batch_size, tgt_len, skeleton_dim]. mask_target_padding: Mask for target masking with ones where elements belong to the padding elements of shape [batch_size, tgt_len, skeleton_dim]. get_attn_weights: Boolean to indicate if attention weights should be returned. Returns: """ # 1) Encode the sequence with given pose encoder # [batch_size, sequence_length, model_dim] input_pose_seq = input_pose_seq_ target_pose_seq = target_pose_seq_ if self._pose_embedding is not None: input_pose_seq = self._pose_embedding(input_pose_seq) target_pose_seq = self._pose_embedding(target_pose_seq) # 2) compute the look-ahead mask and the positional encodings # [sequence_length, batch_size, model_dim] input_pose_seq = torch.transpose(input_pose_seq, 0, 1) target_pose_seq = torch.transpose(target_pose_seq, 0, 1) def query_copy_fn(indices): return self.handle_copy_query(indices, input_pose_seq_) # concatenate extra token for activity prediction as an extr element of the # input sequence, i.e. specialized token is not a skeleton if self._use_class_token: input_pose_seq = self.handle_class_token(input_pose_seq) # 3) compute the attention weights using the transformer # [target_sequence_length, batch_size, model_dim] attn_output, memory, attn_weights, enc_weights, mat = self._transformer( input_pose_seq, target_pose_seq, query_embedding=self._query_embed.weight, encoder_position_encodings=self._encoder_pos_encodings, decoder_position_encodings=self._decoder_pos_encodings, mask_look_ahead=self._mask_look_ahead, mask_target_padding=mask_target_padding, get_attn_weights=get_attn_weights, query_selection_fn=query_copy_fn, fold=fold, eval_step=eval_step ) end = self._input_dim if self._input_dim == self._pose_dim else self._pose_dim out_sequence = [] target_pose_seq_ = mat[0] if self._query_selection else \ torch.transpose(target_pose_seq_, 0, 1) # 4) decode sequence with pose decoder. The decoding process is time # independent. It means non-autoregressive or parallel decoding. # [batch_size, target_sequence_length, pose_dim] for l in range(self._num_decoder_layers): # [target_seq_length*batch_size, pose_dim] out_sequence_ = self._pose_decoder( attn_output[l].view(-1, self._model_dim)) # [target_seq_length, batch_size, pose_dim] out_sequence_ = out_sequence_.view( self._target_seq_length, -1, self._pose_dim) # apply residual connection between target query and predicted pose # [tgt_seq_len, batch_size, pose_dim] out_sequence_ = out_sequence_ + target_pose_seq_[:, :, 0:end] # [batch_size, tgt_seq_len, pose_dim] out_sequence_ = torch.transpose(out_sequence_, 0, 1) out_sequence.append(out_sequence_) if self._predict_activity: out_class = self.predict_activity(attn_output, memory) return out_sequence, out_class, attn_weights, enc_weights, mat return out_sequence, attn_weights, enc_weights, mat def predict_activity(self, attn_output, memory): """Performs activity prediction either from memory or class token. attn_output: Encoder memory. Shape [src_seq_len, batch_size, model_dim]. """ # [batch_size, src_len, model_dim] in_act = torch.transpose(memory, 0, 1) # use a single specialized token for predicting activity # the specialized token is in the first element of the sequence if self._use_class_token: # [batch_size, model_dim] token = in_act[:, 0] actions = self._action_head(token) return [actions] # use all the input sequence attention to predict activity # [batch_size, src_len*model_dim] in_act = torch.reshape(in_act, (-1, self._action_head_size)) actions = self._action_head(in_act) return [actions] #out_class = [] #for l in range(self._num_decoder_layers): # in_act = torch.transpose(attn_output[l], 0, 1) # in_act = torch.reshape(in_act, (-1, self._action_head_size)) # actions = self._action_head(in_act) # out_class.append(actions) #return out_class def forward_autoregressive(self, input_pose_seq, target_pose_seq=None, mask_target_padding=None, get_attn_weights=False): """Compute forward pass for auto-regressive inferece in test time.""" thisdevice = self._encoder_pos_encodings.device # the first query pose is the first in the target prev_target = target_pose_seq[:, 0, :] # 1) Enconde using the pose embeding if self._pose_embedding is not None: input_pose_seq = self._pose_embedding(input_pose_seq) target_pose_seq = self._pose_embedding(target_pose_seq) # [batch_size, 1, model_dim] target_seq = target_pose_seq[:, 0:1, :] # 2) compute the look-ahead mask and the positional encodings # [sequence_length, batch_size, model_dim] input_pose_seq = torch.transpose(input_pose_seq, 0, 1) target_seq = torch.transpose(target_seq, 0, 1) # concatenate extra token for activity prediction as an extra if self._use_class_token: input_pose_seq = self.handle_class_token(input_pose_seq) # 3) use auto recursion to compute the predicted set of tokens memory, enc_attn_weights = self._transformer._encoder( input_pose_seq, self._encoder_pos_encodings) # get only the first In teory it should only be one target pose at testing batch_size = memory.size()[1] out_pred_seq = torch.FloatTensor( batch_size, self._target_seq_length, self._pose_dim).to(thisdevice) for t in range(self._target_seq_length): position_encodings = self._pos_decoder(t+1).view( t+1, 1, self._model_dim).to(thisdevice) mask_look_ahead = torch.from_numpy( utils.create_look_ahead_mask(t+1)).to(thisdevice) # a list of length n_decoder_layers with elements of # shape [t, batch_size, model_dim] out_attn, out_weights = self._transformer._decoder( target_seq, memory, position_encodings, mask_look_ahead=mask_look_ahead ) # get only the last predicted token decode it to get the pose and # then encode the pose. shape [1*batch_size, pose_dim] # for 8 seeds of evaluation (batch_size) pred_pose = self._pose_decoder( out_attn[-1][t:(t+1), :, :].view(-1, self._model_dim)) # apply residual between last target pose and recently generated pose if self._pose_dim == self._input_dim: pred_pose = pred_pose + prev_target else: prev_target[:, 0:self._pose_dim] = pred_pose + prev_target[:,0:self._pose_dim] pred_pose = prev_target prev_target = pred_pose out_pred_seq[:, t, :] = pred_pose.view(-1, self._input_dim)[:, 0:self._pose_dim] if self._pose_embedding is not None: pose_code = self._pose_embedding(pred_pose) # [1, batch_size, model_dim] pose_code = pose_code.view(-1, batch_size, self._model_dim) # [t+1, batch_size, model_dim] target_seq = torch.cat([target_seq, pose_code], axis=0) # 1) the last attention output contains all the necessary sequence; or # 2) Use all the memory to predict if self._predict_activity: actions = self.predict_activity(out_attn, memory) if self._predict_activity: return [out_pred_seq], [actions[-1]], None, None return [out_pred_seq] def model_factory(params, pose_embedding_fn, pose_decoder_fn): init_fn = utils.normal_init_ \ if params['init_fn'] == 'normal_init' else utils.xavier_init_ return PoseTransformer( pose_dim=params['pose_dim'], input_dim=params['input_dim'], model_dim=params['model_dim'], num_encoder_layers=params['num_encoder_layers'], num_decoder_layers=params['num_decoder_layers'], num_heads=params['num_heads'], dim_ffn=params['dim_ffn'], dropout=params['dropout'], target_seq_length=params['target_seq_len'], source_seq_length=params['source_seq_len'], init_fn=init_fn, non_autoregressive=params['non_autoregressive'], use_query_embedding=params['use_query_embedding'], pre_normalization=params['pre_normalization'], predict_activity=params['predict_activity'], num_activities=params['num_activities'], use_memory=params['use_memory'], pose_embedding=pose_embedding_fn(params), pose_decoder=pose_decoder_fn(params), query_selection=params['query_selection'], pos_encoding_params=(params['pos_enc_beta'], params['pos_enc_alpha']) ) if __name__ == '__main__': transformer = PoseTransformer(model_dim=_POSE_DIM, num_heads=6) transformer.eval() batch_size = 8 model_dim = 256 tgt_seq = torch.FloatTensor(batch_size, _TARGET_LENGTH, _POSE_DIM).fill_(1) src_seq = torch.FloatTensor(batch_size, _SOURCE_LENGTH-1, _POSE_DIM).fill_(1) outputs = transformer(src_seq, tgt_seq) print(outputs[-1].size())
19,154
38.658385
120
py
GaitForeMer
GaitForeMer-main/models/seq2seq_model.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Sequence to sequence model for human motion prediction. The model has been implemented according to [1] and adapted from its pytorch version [2]. The reimplementation has the purpose of reducing clutter in code and for learing purposes. [1] [2] """ import torch import numpy as np from torch import nn import torch.nn.functional as F class Seq2SeqModel(nn.Module): """Sequence to sequence model.""" def __init__( self, architecture='tied', source_seq_len=50, target_seq_len=25, rnn_size=1024, # hidden recurrent layer size num_layers=1, max_gradient_norm=5, batch_size=16, learning_rate=0.005, learning_rate_decay_factor=0.95, loss_to_use='sampling_based', number_of_actions=1, one_hot=True, residual_velocities=False, dropout=0.0, dtype=torch.float32, device=None): """ Args: architecture: [basic, tied] whether to tie the decoder and decoder. source_seq_len: lenght of the input sequence. target_seq_len: lenght of the target sequence. rnn_size: number of units in the rnn. num_layers: number of rnns to stack. max_gradient_norm: gradients will be clipped to maximally this norm. batch_size: the size of the batches used during training; the model construction is independent of batch_size, so it can be changed after initialization if this is convenient, e.g., for decoding. learning_rate: learning rate to start with. learning_rate_decay_factor: decay learning rate by this much when needed. loss_to_use: [supervised, sampling_based]. Whether to use ground truth in each timestep to compute the loss after decoding, or to feed back the prediction from the previous time-step. number_of_actions: number of classes we have. one_hot: whether to use one_hot encoding during train/test (sup models). residual_velocities: whether to use a residual connection that models velocities. dtype: the data type to use to store internal variables. """ super(Seq2SeqModel, self).__init__() self.HUMAN_SIZE = 54 self.input_size = self.HUMAN_SIZE + number_of_actions if one_hot else self.HUMAN_SIZE print( "One hot is ", one_hot ) print( "Input size is %d" % self.input_size ) # Summary writers for train and test runs self.source_seq_len = source_seq_len self.target_seq_len = target_seq_len self.rnn_size = rnn_size self.batch_size = batch_size self.dropout = dropout # === Create the RNN that will keep the state === print('rnn_size = {0}'.format( rnn_size )) self.cell = torch.nn.GRUCell(self.input_size, self.rnn_size) # self.cell2 = torch.nn.GRUCell(self.rnn_size, self.rnn_size) self.fc1 = nn.Linear(self.rnn_size, self.input_size) def forward(self, encoder_inputs, decoder_inputs): def loop_function(prev, i): return prev batchsize = encoder_inputs.size()[0] encoder_inputs = torch.transpose(encoder_inputs, 0, 1) decoder_inputs = torch.transpose(decoder_inputs, 0, 1) state = torch.zeros(batchsize, self.rnn_size).\ to(encoder_inputs.get_device()) # state2 = torch.zeros(batchsize, self.rnn_size) # if use_cuda: # state = state.cuda() # # state2 = state2.cuda() for i in range(self.source_seq_len-1): state = self.cell(encoder_inputs[i], state) # state2 = self.cell2(state, state2) state = F.dropout(state, self.dropout, training=self.training) # if use_cuda: # state = state.cuda() ## state2 = state2.cuda() outputs = [] prev = None for i, inp in enumerate(decoder_inputs): # loop function is trained as in auto regressive if loop_function is not None and prev is not None: inp = loop_function(prev, i) inp = inp.detach() state = self.cell(inp, state) # state2 = self.cell2(state, state2) # output = inp + self.fc1(state2) # state = F.dropout(state, self.dropout, training=self.training) output = inp + self.fc1(F.dropout(state, self.dropout, training=self.training)) outputs.append(output.view([1, batchsize, self.input_size])) if loop_function is not None: prev = output # return outputs, state outputs = torch.cat(outputs, 0) return torch.transpose(outputs, 0, 1)
5,425
34.697368
89
py
GaitForeMer
GaitForeMer-main/models/PoseEncoderDecoder.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Definition of pose encoder and encoder embeddings and model factory.""" import numpy as np import os import sys import torch import torch.nn as nn thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils import models.PoseGCN as GCN import models.Conv1DEncoder as Conv1DEncoder def pose_encoder_mlp(params): # These two encoders should be experimented with a graph NN and # a prior based pose decoder using also the graph init_fn = utils.normal_init_ \ if params['init_fn'] == 'normal_init' else utils.xavier_init_ pose_embedding = nn.Sequential( nn.Linear(params['input_dim'], params['model_dim']), nn.Dropout(0.1) ) utils.weight_init(pose_embedding, init_fn_=init_fn) return pose_embedding def pose_decoder_mlp(params): init_fn = utils.normal_init_ \ if params['init_fn'] == 'normal_init' else utils.xavier_init_ pose_decoder = nn.Linear(params['model_dim'], params['pose_dim']) utils.weight_init(pose_decoder, init_fn_=init_fn) return pose_decoder def pose_decoder_gcn(params): decoder = GCN.PoseGCN( input_features=params['model_dim'], output_features = 9 if params['pose_format'] == 'rotmat' else 3, model_dim=params['model_dim'], output_nodes=params['n_joints'], p_dropout=params['dropout'], num_stage=1 ) return decoder def pose_encoder_gcn(params): encoder = GCN.SimpleEncoder( n_nodes=params['n_joints'], input_features=9 if params['pose_format'] == 'rotmat' else 3, #n_nodes=params['pose_dim'], #input_features=1, model_dim=params['model_dim'], p_dropout=params['dropout'] ) return encoder def pose_encoder_conv1d(params): encoder = Conv1DEncoder.Pose1DEncoder( input_channels=9 if params['pose_format'] == 'rotmat' else 3, output_channels=params['model_dim'], n_joints=params['n_joints'] ) return encoder def pose_encoder_conv1dtemporal(params): dof = 9 if params['pose_format'] == 'rotmat' else 3 encoder = Conv1DEncoder.Pose1DTemporalEncoder( input_channels=dof*params['n_joints'], output_channels=params['model_dim'] ) return encoder def select_pose_encoder_decoder_fn(params): if params['pose_embedding_type'].lower() == 'simple': return pose_encoder_mlp, pose_decoder_mlp if params['pose_embedding_type'].lower() == 'conv1d_enc': return pose_encoder_conv1d, pose_decoder_mlp if params['pose_embedding_type'].lower() == 'convtemp_enc': return pose_encoder_conv1dtemporal, pose_decoder_mlp if params['pose_embedding_type'].lower() == 'gcn_dec': return pose_encoder_mlp, pose_decoder_gcn if params['pose_embedding_type'].lower() == 'gcn_enc': return pose_encoder_gcn, pose_decoder_mlp if params['pose_embedding_type'].lower() == 'gcn_full': return pose_encoder_gcn, pose_decoder_gcn elif params['pose_embedding_type'].lower() == 'vae': return pose_encoder_vae, pose_decoder_mlp else: raise ValueError('Unknown pose embedding {}'.format(params['pose_embedding_type']))
4,120
32.504065
87
py
GaitForeMer
GaitForeMer-main/models/TransformerEncoder.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implementation of Transformer encoder and encoder layer with self attention. Implementation of the encoder layer as in [1] and [2] for sequence to sequence modeling. [1] https://arxiv.org/pdf/1706.03762.pdf [2] https://arxiv.org/pdf/2005.12872.pdf """ import numpy as np import sys import os import torch import torch.nn as nn thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils class EncoderLayer(nn.Module): """Implements the transformer encoder Layer.""" def __init__(self, model_dim=256, num_heads=8, dim_ffn=2048, dropout=0.1, init_fn=utils.normal_init_, pre_normalization=False): """Encoder layer initialization. Args: model_dim: num_heads: dim_ffn: dropout: """ super(EncoderLayer, self).__init__() self._model_dim = model_dim self._num_heads = num_heads self._dim_ffn = dim_ffn self._dropout = dropout self._pre_normalization = pre_normalization self._self_attn = nn.MultiheadAttention(model_dim, num_heads, dropout) self._relu = nn.ReLU() self._dropout_layer = nn.Dropout(self._dropout) self._linear1 = nn.Linear(model_dim, self._dim_ffn) self._linear2 = nn.Linear(self._dim_ffn, self._model_dim) self._norm1 = nn.LayerNorm(model_dim, eps=1e-5) self._norm2 = nn.LayerNorm(model_dim, eps=1e-5) utils.weight_init(self._linear1, init_fn_=init_fn) utils.weight_init(self._linear2, init_fn_=init_fn) def forward(self, source_seq, pos_encodings): """Computes forward pass according. Args: source_seq: [sequence_length, batch_size, model_dim]. pos_encodings: [sequence_length, model_dim]. Returns: Tensor of shape [sequence_length, batch_size, model_dim]. """ if self._pre_normalization: return self.forward_pre(source_seq, pos_encodings) return self.forward_post(source_seq, pos_encodings) def forward_post(self, source_seq, pos_encodings): """Computes decoder layer forward pass with pre normalization. Args: source_seq: [sequence_length, batch_size, model_dim]. pos_encodings: [sequence_length, model_dim]. Returns: Tensor of shape [sequence_length, batch_size, model_dim]. """ # add positional encodings to the input sequence # for self attention query is the same as key query = source_seq + pos_encodings key = query value = source_seq attn_output, attn_weights = self._self_attn( query, key, value, need_weights=True ) norm_attn = self._dropout_layer(attn_output) + source_seq norm_attn = self._norm1(norm_attn) output = self._linear1(norm_attn) output = self._relu(output) output = self._dropout_layer(output) output = self._linear2(output) output = self._dropout_layer(output) + norm_attn output = self._norm2(output) return output, attn_weights def forward_pre(self, source_seq_, pos_encodings): """Computes decoder layer forward pass with pre normalization. Args: source_seq: [sequence_length, batch_size, model_dim]. pos_encodings: [sequence_length, model_dim]. Returns: Tensor of shape [sequence_length, batch_size, model_dim]. """ # add positional encodings to the input sequence # for self attention query is the same as key source_seq = self._norm1(source_seq_) query = source_seq + pos_encodings key = query value = source_seq attn_output, attn_weights = self._self_attn( query, key, value, need_weights=True ) norm_attn_ = self._dropout_layer(attn_output) + source_seq_ norm_attn = self._norm2(norm_attn_) output = self._linear1(norm_attn) output = self._relu(output) output = self._dropout_layer(output) output = self._linear2(output) output = self._dropout_layer(output) + norm_attn_ return output, attn_weights class TransformerEncoder(nn.Module): def __init__(self, num_layers=6, model_dim=256, num_heads=8, dim_ffn=2048, dropout=0.1, init_fn=utils.normal_init_, pre_normalization=False): super(TransformerEncoder, self).__init__() """Transforme encoder initialization.""" self._num_layers = num_layers self._model_dim = model_dim self._num_heads = num_heads self._dim_ffn = dim_ffn self._dropout = dropout # self._norm = norm self._pre_normalization = pre_normalization self._encoder_stack = self.init_encoder_stack(init_fn) def init_encoder_stack(self, init_fn): """Create the stack of encoder layers.""" stack = nn.ModuleList() for s in range(self._num_layers): layer = EncoderLayer( model_dim=self._model_dim, num_heads=self._num_heads, dim_ffn=self._dim_ffn, dropout=self._dropout, init_fn=init_fn, pre_normalization=self._pre_normalization ) stack.append(layer) return stack def forward(self, input_sequence, pos_encodings): """Computes decoder forward pass. Args: source_seq: [sequence_length, batch_size, model_dim]. pos_encodings: [sequence_length, model_dim]. Returns: Tensor of shape [sequence_length, batch_size, model_dim]. """ outputs = input_sequence for l in range(self._num_layers): outputs, attn_weights = self._encoder_stack[l](outputs, pos_encodings) # if self._norm: # outputs = self._norm(outputs) return outputs, attn_weights if __name__ == '__main__': device = torch.device('cuda' if torch.cuda.is_available() else 'cpu') seq_length = 50 pos_encodings = torch.FloatTensor(seq_length, 1, 256).uniform_(0,1) seq = torch.FloatTensor(seq_length, 8, 256).fill_(1.0) pos_encodings = pos_encodings.to(device) seq = seq.to(device) encoder = TransformerEncoder(num_layers=6) encoder.to(device) encoder.eval() print(encoder(seq, pos_encodings).size())
7,169
28.628099
79
py
GaitForeMer
GaitForeMer-main/models/TransformerDecoder.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implementation of Transformer decoder and decoder layer with self attention. Implementation of the decoder layer as in [1] and [2] for sequence to sequence modeling. [1] https://arxiv.org/pdf/1706.03762.pdf [2] https://arxiv.org/pdf/2005.12872.pdf """ import numpy as np import sys import os import torch import torch.nn as nn thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils class DecoderLayer(nn.Module): """Implements the transformer decoder Layer.""" def __init__(self, model_dim=256, num_heads=8, dim_ffn=2048, dropout=0.1, init_fn=utils.normal_init_, pre_normalization=False, use_query_embedding=False): """Decoder layer initialization. Args: model_dim: num_heads: dim_ffn: dropout: """ super(DecoderLayer, self).__init__() self._model_dim = model_dim self._num_heads = num_heads self._dim_ffn = dim_ffn self._dropout = dropout self._pre_normalization = pre_normalization self._use_query_embedding = use_query_embedding self._self_attn = nn.MultiheadAttention( model_dim, num_heads, dropout=dropout ) self._multihead_attn = nn.MultiheadAttention( model_dim, num_heads, dropout=dropout ) # the so-called point-wise network self._linear1 = nn.Linear(model_dim, dim_ffn) self._linear2 = nn.Linear(dim_ffn, model_dim) self._relu = nn.ReLU() self._norm1 = nn.LayerNorm(model_dim) self._norm2 = nn.LayerNorm(model_dim) self._norm3 = nn.LayerNorm(model_dim) self._dropout1 = nn.Dropout(dropout) self._dropout2 = nn.Dropout(dropout) self._dropout3 = nn.Dropout(dropout) self._dropout4 = nn.Dropout(dropout) utils.weight_init(self._linear1, init_fn_=init_fn) utils.weight_init(self._linear2, init_fn_=init_fn) self._forward_fn = self.forward_pre if pre_normalization else self.forward_post def forward(self, target_seq, memory, pos_encodings, query_embedding=None, mask_look_ahead=None, mask_target_padding=None): """Forward pass of the layer. Args: target_seq: [target_seq_length, batch_size, model_dim] memory: [source_seq_length, batch_size, model_dim] mask_look_ahead: [] mask_target_padding: """ return self._forward_fn( target_seq, memory, pos_encodings, query_embedding=query_embedding, mask_look_ahead=mask_look_ahead, mask_target_padding=mask_target_padding ) def handle_query_embedding(self, sequence, embedding): """Handle """ if self._use_query_embedding: return sequence + embedding return sequence def forward_post(self, target_seq, memory, pos_encodings, query_embedding=None, mask_look_ahead=None, mask_target_padding=None): """Forward pass of the layer with post normalization. Args: target_seq: [target_seq_length, batch_size, model_dim] memory: [source_seq_length, batch_size, model_dim] mask_look_ahead: [] mask_target_padding: """ # 1) Compute self attention with current sequence of inferred tokens # query is the same as key for self attention # [batch_size, seq_length, model_dim] if self._use_query_embedding: q = k = v = target_seq + query_embedding else: q = k = v = target_seq + pos_encodings self_attn, self_attn_weights = self._self_attn( query=q, key=k, value=v, #target_seq, attn_mask=mask_look_ahead, key_padding_mask=mask_target_padding ) self_attn = self._dropout1(self_attn) out_self_attn = self._norm1(self_attn + target_seq) # 2) Attend the encoder's memory given the comptued self attention # [batch_size, seq_length, model_dim] attn, attn_weights = self._multihead_attn( query=self.handle_query_embedding(out_self_attn, query_embedding), key=self.handle_query_embedding(memory, pos_encodings), value=memory) attn = self._dropout2(attn) out_attn = self._norm2(attn + out_self_attn) # 3) Compute pointwise embeding by expanding and projecting + dropout ffn_output = self._linear1(out_attn) ffn_output = self._relu(ffn_output) ffn_output = self._dropout4(ffn_output) ffn_output = self._linear2(ffn_output) # 4) Compute residual connection as final output ffn_output = self._dropout3(ffn_output) outputs = self._norm3(ffn_output + out_attn) return outputs, self_attn_weights, attn_weights def forward_pre(self, target_seq_, memory, pos_encodings, query_embedding=None, mask_look_ahead=None, mask_target_padding=None): """Forward pass of the layer with pre normalization. Args: target_seq: [target_seq_length, batch_size, model_dim] memory: [source_seq_length, batch_size, model_dim] mask_look_ahead: [] mask_target_padding: """ target_seq = self._norm1(target_seq_) # 1) Compute self attention with current sequence of inferred tokens # query is the same as key for self attention # [batch_size, seq_length, model_dim] if self._use_query_embedding: # in case of using only the query embedding follow DETR [2] which drops # values to zero and uses only the query embeddings q = k = target_seq + query_embedding v = target_seq else: q = k = v = target_seq + pos_encodings self_attn, self_attn_weights = self._self_attn( query=q, key=k, value=v, attn_mask=mask_look_ahead, key_padding_mask=mask_target_padding ) self_attn = self._dropout1(self_attn) out_self_attn = self._norm2(self_attn + target_seq_) # 2) Attend the encoder's memory given the comptued self attention # [batch_size, seq_length, model_dim] attn, attn_weights = self._multihead_attn( query=self.handle_query_embedding(out_self_attn, query_embedding), key=self.handle_query_embedding(memory, pos_encodings), value=memory) attn = self._dropout2(attn) out_attn = self._norm3(attn + out_self_attn) # 3) Compute pointwise embeding by expanding and projecting + dropout ffn_output = self._linear1(out_attn) ffn_output = self._relu(ffn_output) ffn_output = self._dropout4(ffn_output) ffn_output = self._linear2(ffn_output) # 4) Compute residual connection as final output ffn_output = self._dropout3(ffn_output) return ffn_output, self_attn_weights, attn_weights class TransformerDecoder(nn.Module): """Transformer decoder module.""" def __init__(self, num_layers=6, model_dim=256, num_heads=8, dim_ffn=2048, dropout=0.1, init_fn=utils.normal_init_, pre_normalization=False, use_query_embedding=False): super(TransformerDecoder, self).__init__() self._model_dim = model_dim self._num_heads = num_heads self._dim_ffn = dim_ffn self._dropout = dropout self._num_layers = num_layers self._use_query_embedding = use_query_embedding self._pre_normalization = pre_normalization self._decoder_stack = self.init_decoder_stack(init_fn) def init_decoder_stack(self, init_fn): stack = nn.ModuleList() for s in range(self._num_layers): layer = DecoderLayer( model_dim=self._model_dim, num_heads=self._num_heads, dim_ffn=self._dim_ffn, dropout=self._dropout, init_fn=init_fn, pre_normalization=self._pre_normalization, use_query_embedding=self._use_query_embedding ) stack.append(layer) return stack def forward(self, target_seq, memory, pos_encodings, query_embedding=None, mask_target_padding=None, mask_look_ahead=None, get_attn_weights=False): """Computes forward pass of decoder. Args: target_seq: [target_sequence_length, batch_size, model_dim]. memory: [source_sequence_length, batch_size, model_dim]. pos_encodings: [target_seq_length, model_dim]. mask_look_ahead: [target_seq_length, model_dim]. Returns: A tensor with the decoded attention with shape [target_sequence_length, batch_size, model_dim]. """ seq_length = target_seq.size()[0] output_list = [] attn_weights_list = [] if get_attn_weights else None outputs = torch.zeros_like(target_seq) if self._use_query_embedding else target_seq for l in range(self._num_layers): outputs, self_attn_weights, attn_weights = self._decoder_stack[l]( outputs, memory, pos_encodings=pos_encodings, query_embedding=query_embedding, mask_target_padding=mask_target_padding, mask_look_ahead=mask_look_ahead ) if get_attn_weights: attn_weights_list.append(attn_weights) output_list.append(outputs) return output_list, attn_weights_list if __name__ == '__main__': thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils seq_length = 55 batch_size = 8 model_dim = 256 tgt_seq = torch.FloatTensor(seq_length, batch_size, model_dim).fill_(1) memory = torch.FloatTensor(seq_length, batch_size, model_dim).uniform_(0, 1) mask_look_ahead = utils.create_look_ahead_mask(seq_length) mask_look_ahead = torch.from_numpy(mask_look_ahead) encodings = torch.FloatTensor(seq_length, 1, model_dim).uniform_(0,1) decoder = TransformerDecoder() outputs = decoder(tgt_seq, memory, encodings, mask_look_ahead=mask_look_ahead) print(outputs.size())
11,059
31.818991
87
py
GaitForeMer
GaitForeMer-main/models/PoseGCN.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Graph Convolutional Neural Network implementation. Code adapted from [1]. [1] https://github.com/wei-mao-2019/HisRepItself [2] https://github.com/tkipf/gcn/blob/92600c39797c2bfb61a508e52b88fb554df30177/gcn/layers.py#L132 """ import os import sys import torch.nn as nn import torch from torch.nn.parameter import Parameter import math import numpy as np thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import utils.utils as utils class GraphConvolution(nn.Module): """Implements graph convolutions.""" def __init__(self, in_features, out_features, output_nodes=48, bias=False): """Constructor. The graph convolutions can be defined as \sigma(AxHxW), where A is the adjacency matrix, H is the feature representation from previous layer and W is the wegith of the current layer. The dimensions of such martices A\in R^{NxN}, H\in R^{NxM} and W\in R^{MxO} where - N is the number of nodes - M is the number of input features per node - O is the number of output features per node Args: in_features: Number of input features per node. out_features: Number of output features per node. output_nodes: Number of nodes in the graph. """ super(GraphConvolution, self).__init__() self.in_features = in_features self.out_features = out_features self._output_nodes = output_nodes # W\in R^{MxO} self.weight = Parameter(torch.FloatTensor(in_features, out_features)) # A\in R^{NxN} self.att = Parameter(torch.FloatTensor(output_nodes, output_nodes)) if bias: self.bias = Parameter(torch.FloatTensor(out_features)) else: self.register_parameter('bias', None) self.reset_parameters() def reset_parameters(self): stdv = 1. / math.sqrt(self.weight.size(1)) self.weight.data.uniform_(-stdv, stdv) self.att.data.uniform_(-stdv, stdv) if self.bias is not None: self.bias.data.uniform_(-stdv, stdv) def forward(self, x): """Forward pass. Args: x: [batch_size, n_nodes, input_features] Returns: Feature representation computed from inputs. Shape is [batch_size, n_nodes, output_features]. """ # [batch_size, input_dim, output_features] # HxW = {NxM}x{MxO} = {NxO} support = torch.matmul(x, self.weight) # [batch_size, n_nodes, output_features] # = {NxN}x{NxO} = {NxO} output = torch.matmul(self.att, support) if self.bias is not None: return output + self.bias else: return output def __repr__(self): return self.__class__.__name__ + ' (' \ + str(self.in_features) + ' -> ' \ + str(self.out_features) + ')' class GC_Block(nn.Module): """Residual block with graph convolutions. The implementation uses the same number of input features for outputs. """ def __init__(self, in_features, p_dropout, output_nodes=48, bias=False): """Constructor. Args: in_features: Number of input and output features. p_dropout: Dropout used in the layers. output_nodes: Number of output nodes in the graph. """ super(GC_Block, self).__init__() self.in_features = in_features self.out_features = in_features self.gc1 = GraphConvolution( in_features, in_features, output_nodes=output_nodes, bias=bias ) self.bn1 = nn.BatchNorm1d(output_nodes * in_features) self.gc2 = GraphConvolution( in_features, in_features, output_nodes=output_nodes, bias=bias ) self.bn2 = nn.BatchNorm1d(output_nodes * in_features) self.do = nn.Dropout(p_dropout) self.act_f = nn.Tanh() def forward(self, x): """Forward pass of the residual module""" y = self.gc1(x) b, n, f = y.shape y = self.bn1(y.view(b, -1)).view(b, n, f) y = self.act_f(y) y = self.do(y) y = self.gc2(y) b, n, f = y.shape y = self.bn2(y.view(b, -1)).view(b, n, f) y = self.act_f(y) y = self.do(y) return y + x def __repr__(self): return self.__class__.__name__ + ' (' \ + str(self.in_features) + ' -> ' \ + str(self.out_features) + ')' class PoseGCN(nn.Module): def __init__(self, input_features=128, output_features=3, model_dim=128, output_nodes=21, p_dropout=0.1, num_stage=1): """Constructor. Args: input_feature: num of input feature of the graph nodes. model_dim: num of hidden features of the generated embeddings. p_dropout: dropout probability num_stage: number of residual blocks in the network. output_nodes: number of nodes in graph. """ super(PoseGCN, self).__init__() self.num_stage = num_stage self._n_nodes = output_nodes self._model_dim = model_dim self._output_features = output_features self._hidden_dim = 512 self._front = nn.Sequential( nn.Linear(model_dim, output_nodes*self._hidden_dim), nn.Dropout(p_dropout) ) utils.weight_init(self._front, init_fn_=utils.xavier_init_) self.gc1 = GraphConvolution( self._hidden_dim, self._hidden_dim, output_nodes=output_nodes ) self.bn1 = nn.BatchNorm1d(output_nodes * self._hidden_dim) self.gcbs = [] for i in range(num_stage): self.gcbs.append(GC_Block( self._hidden_dim, p_dropout=p_dropout, output_nodes=output_nodes) ) self.gcbs = nn.ModuleList(self.gcbs) self.gc7 = GraphConvolution( self._hidden_dim, output_features, output_nodes=output_nodes ) self.do = nn.Dropout(p_dropout) self.act_f = nn.Tanh() gcn_params = filter(lambda p: p.requires_grad, self.parameters()) nparams = sum([np.prod(p.size()) for p in gcn_params]) print('[INFO] ({}) GCN has {} params!'.format(self.__class__.__name__, nparams)) def preprocess(self, x): if len(x.size()) < 3: _, D = x.size() # seq_len, batch_size, input_dim x = x.view(self._seq_len, -1, D) # [batch_size, seq_len, input_dim] x = torch.transpose(x, 0, 1) # [batch_size, input_dim, seq_len] x = torch.transpose(x, 1, 2) return x return x def postprocess(self, y): """Flattents the input tensor. Args: y: Input tensor of shape [batch_size, n_nodes, output_features]. """ y = y.view(-1, self._n_nodes*self._output_features) return y def forward(self, x): """Forward pass of network. Args: x: [batch_size, model_dim]. """ # [batch_size, model_dim*n_nodes] x = self._front(x) x = x.view(-1, self._n_nodes, self._hidden_dim) # [batch_size, n_joints, model_dim] y = self.gc1(x) b, n, f = y.shape y = self.bn1(y.view(b, -1)).view(b, n, f) y = self.act_f(y) y = self.do(y) for i in range(self.num_stage): y = self.gcbs[i](y) # [batch_size, n_joints, output_features] y = self.gc7(y) # y = y + x # [seq_len*batch_size, input_dim] y = self.postprocess(y) return y class SimpleEncoder(nn.Module): def __init__(self, n_nodes=63, input_features=1, model_dim=128, p_dropout=0.1): """Constructor. Args: input_dim: Dimension of the input vector. This will be equivalent to the number of nodes in the graph, each node with 1 feature each. model_dim: Dimension of the output vector to produce. p_dropout: Dropout to be applied for regularization. """ super(SimpleEncoder, self).__init__() #The graph convolutions can be defined as \sigma(AxHxW), where A is the #A\in R^{NxN} x H\in R^{NxM} x W\in R ^{MxO} self._input_features = input_features self._output_nodes = n_nodes self._hidden_dim = 512 self._model_dim = model_dim self._num_stage = 1 print('[INFO] ({}) Hidden dimension: {}!'.format( self.__class__.__name__, self._hidden_dim)) self.gc1 = GraphConvolution( in_features=self._input_features, out_features=self._hidden_dim, output_nodes=self._output_nodes ) self.bn1 = nn.BatchNorm1d(self._output_nodes*self._hidden_dim) self.gc2 = GraphConvolution( in_features=self._hidden_dim, out_features=model_dim, output_nodes=self._output_nodes ) self.gcbs = [] for i in range(self._num_stage): self.gcbs.append(GC_Block( self._hidden_dim, p_dropout=p_dropout, output_nodes=self._output_nodes) ) self.gcbs = nn.ModuleList(self.gcbs) self.do = nn.Dropout(p_dropout) self.act_f = nn.Tanh() self._back = nn.Sequential( nn.Linear(model_dim*self._output_nodes, model_dim), nn.Dropout(p_dropout) ) utils.weight_init(self._back, init_fn_=utils.xavier_init_) gcn_params = filter(lambda p: p.requires_grad, self.parameters()) nparams = sum([np.prod(p.size()) for p in gcn_params]) print('[INFO] ({}) GCN has {} params!'.format(self.__class__.__name__, nparams)) def forward(self, x): """Forward pass of network. Args: x: [batch_size, n_poses, pose_dim/input_dim]. """ B, S, D = x.size() # [batch_size, n_joints, model_dim] y = self.gc1(x.view(-1, self._output_nodes, self._input_features)) b, n, f = y.shape y = self.bn1(y.view(b, -1)).view(b, n, f) y = self.act_f(y) y = self.do(y) for i in range(self._num_stage): y = self.gcbs[i](y) # [batch_size, n_joints, model_dim] y = self.gc2(y) # [batch_size, model_dim] y = self._back(y.view(-1, self._model_dim*self._output_nodes)) # [batch_size, n_poses, model_dim] y = y.view(B, S, self._model_dim) return y def test_decoder(): seq_len = 25 input_size = 63 model_dim = 128 dropout = 0.3 n_stages = 2 output_nodes = 21 joint_dof = 1 n_joints = model_dim layer = GraphConvolution( in_features=joint_dof, out_features=model_dim, output_nodes=n_joints ) X = torch.FloatTensor(10, n_joints, joint_dof) print(layer(X).size()) gcn = PoseGCN( input_features=model_dim, output_features=3, model_dim=model_dim, output_nodes=output_nodes, p_dropout=0.1, num_stage=2 ) X = torch.FloatTensor(10*seq_len, model_dim) print(gcn(X).size()) def test_encoder(): input_size = 63 model_dim = 128 dropout = 0.3 n_stages = 2 output_nodes = 21 dof = 9 encoder = SimpleEncoder( n_nodes=output_nodes, model_dim=model_dim, input_features=dof, p_dropout=0.1 ) X = torch.FloatTensor(10, 25, output_nodes*dof) print(encoder(X).size()) if __name__ == '__main__': #test_decoder() test_encoder()
11,945
26.976581
97
py
GaitForeMer
GaitForeMer-main/models/potr_fn.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Model function to deploy POTR models for visualization and generation.""" import numpy as np import os import sys import argparse import json import time import cv2 from matplotlib import image from matplotlib import pyplot as plt from PIL import Image, ImageDraw from sklearn.metrics import confusion_matrix from sklearn.metrics import accuracy_score import torch import torch.nn as nn import torch.optim as optim from torch.utils.tensorboard import SummaryWriter thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import training.seq2seq_model_fn as seq2seq_model_fn import models.PoseTransformer as PoseTransformer import models.PoseEncoderDecoder as PoseEncoderDecoder import data.H36MDataset_v2 as H36MDataset_v2 import data.AMASSDataset as AMASSDataset import utils.utils as utils import radam.radam as radam import training.transformer_model_fn as tr_fn import tqdm # _DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') _DEVICE = torch.device('cpu') def plot_conf_mat(matrix): import matplotlib.pyplot as plt import matplotlib fig, ax = plt.subplots(figsize=(30,30)) #im = ax.imshow(matrix, cmap='Wistia') im = ax.imshow(matrix, cmap='Blues') action_labels = ['A%02d'%i for i in range(1, 61, 1)] ax.set_xticks(np.arange(len(action_labels))) ax.set_yticks(np.arange(len(action_labels))) ax.set_xticklabels(action_labels, fontdict={'fontsize':15})#, rotation=90) ax.set_yticklabels(action_labels, fontdict={'fontsize':15}) ax.tick_params(top=True, bottom=False, labeltop=True, labelbottom=False) for i in range(len(action_labels)): for j in range(len(action_labels)): # color= "w" if round(matrix[i, j],2) < nmax else "black" text = ax.text(j, i, round(matrix[i, j], 2), ha="center", va="center", color="black", fontsize=10) plt.ylabel("") plt.xlabel("") # ax.set_title("Small plot") fig.tight_layout() #plt.show() plt.savefig('confusion_matrix.png') plt.close() def crop_image(img): size = max(img.shape[0], img.shape[1]) h = int(size*0.30) w = int(size*0.30) cy = img.shape[0]//2 cx = img.shape[1]//2 crop = img[cy-h//2:cy+h//2, cx-w//2:cx+w//2] return crop def visualize_h36mdataset(): parser = argparse.ArgumentParser() parser.add_argument('--config_file', type=str) parser.add_argument('--model_file', type=str) parser.add_argument('--data_path', type=str, default=None) args = parser.parse_args() params = json.load(open(args.config_file)) if args.data_path is not None: params['data_path'] = args.data_path args.data_path = params['data_path'] train_dataset_fn, eval_dataset_fn = tr_fn.dataset_factory(params) pose_encoder_fn, pose_decoder_fn = \ PoseEncoderDecoder.select_pose_encoder_decoder_fn(params) for k,v in params.items(): print('[INFO] (POTRFn@main) {}: {}'.format(k, v)) # ids of most common actions in H36M actions = [('walking', 12), ('eating', 2), ('smoking', 9), ('discussion', 1), ('directions', 0)] with torch.no_grad(): for i in range(len(actions)): action, acidx = actions[i] sample = next(iter(eval_dataset_fn)) enc_inputs = sample['encoder_inputs'].to(_DEVICE) dec_inputs = sample['decoder_inputs'].to(_DEVICE) gts = np.squeeze(sample['decoder_outputs'].cpu().numpy())[8*acidx:8*acidx+8] ins = np.squeeze(sample['encoder_inputs'].cpu().numpy())[8*acidx:8*acidx+8] ins = eval_dataset_fn.dataset.unnormalize_pad_data_to_expmap(ins) H36MDataset_v2.visualize_sequence( ins[0:1], args.data_path, prefix='skeletons/%s/gt_in'%action, colors=['gray', 'gray']) #print(gts.shape) gts = eval_dataset_fn.dataset.unnormalize_pad_data_to_expmap(gts) H36MDataset_v2.visualize_sequence( gts[0:1], args.data_path, prefix='skeletons/%s/gt'%action, colors=['gray', 'gray']) enc_inputs = torch.squeeze(enc_inputs) dec_inputs = torch.squeeze(dec_inputs) model = PoseTransformer.model_factory( params, pose_encoder_fn, pose_decoder_fn ) model.load_state_dict(torch.load(args.model_file, map_location=_DEVICE)) model.to(_DEVICE) model.eval() prediction, attn_weights, memory = model( enc_inputs, dec_inputs, get_attn_weights=True ) prediction = prediction[-1][8*acidx:8*acidx+8].cpu().numpy()[0:1] preds = eval_dataset_fn.dataset.unnormalize_pad_data_to_expmap(prediction) H36MDataset_v2.visualize_sequence(preds, args.data_path, prefix='skeletons/%s/pred'%action, colors=['red', 'red']) def compute_mean_average_precision(prediction, target, dataset_fn): pred = prediction.cpu().numpy().squeeze() tgt = target.cpu().numpy().squeeze() T, D = pred.shape pred = dataset_fn.dataset.unormalize_sequence(pred) tgt = dataset_fn.dataset.unormalize_sequence(tgt) pred = pred.reshape((T, -1, 3)) tgt = tgt.reshape((T, -1, 3)) mAP, _, _, (TP, FN) = utils.compute_mean_average_precision( pred, tgt, seq2seq_model_fn._MAP_TRESH, per_frame=True ) return mAP, TP, FN def compute_mpjpe(prediction, target, dataset_fn): pred = prediction.cpu().numpy().squeeze() tgt = target.cpu().numpy().squeeze() T, D = pred.shape pred = dataset_fn.dataset.unormalize_sequence(pred) tgt = dataset_fn.dataset.unormalize_sequence(tgt) pred = pred.reshape((T, -1, 3)) tgt = tgt.reshape((T, -1, 3)) # seq_len x n_joints norm = np.squeeze(np.linalg.norm(pred-tgt, axis=-1)) return norm def compute_test_mAP_nturgbd(): parser = argparse.ArgumentParser() parser.add_argument('--config_file', type=str) parser.add_argument('--model_file', type=str) parser.add_argument('--data_path', type=str, default=None) args = parser.parse_args() params = json.load(open(args.config_file)) if args.data_path is not None: params['data_path'] = args.data_path args.data_path = params['data_path'] params['test_phase'] = True _, test_dataset_fn = tr_fn.dataset_factory(params) pose_encoder_fn, pose_decoder_fn = \ PoseEncoderDecoder.select_pose_encoder_decoder_fn(params) for k,v in params.items(): print('[INFO] (POTRFn@main) {}: {}'.format(k, v)) model = PoseTransformer.model_factory( params, pose_encoder_fn, pose_decoder_fn ) model.load_state_dict(torch.load(args.model_file, map_location=_DEVICE)) model.to(_DEVICE) model.eval() FN = np.zeros((params['target_seq_len'],), dtype=np.float32) TP = np.zeros((params['target_seq_len'],), dtype=np.float32) FN_joint = np.zeros((params['n_joints'],), dtype=np.float32) TP_joint = np.zeros((params['n_joints'],), dtype=np.float32) MPJPE = np.zeros((params['n_joints'],), dtype=np.float32) pred_activity = [] gt_activity = [] with torch.no_grad(): print('Running testing...') for n, sample in tqdm.tqdm(enumerate(test_dataset_fn)): enc_inputs = sample['encoder_inputs'].to(_DEVICE) dec_inputs = sample['decoder_inputs'].to(_DEVICE) gts = sample['decoder_outputs'].to(_DEVICE) # ins = np.squeeze(sample['encoder_inputs'].cpu().numpy()) outputs = model( enc_inputs, dec_inputs, get_attn_weights=True ) if params['predict_activity']: a_ids = sample['action_ids'] prediction, out_logits, attn_weights, memory = outputs out_class = torch.argmax(out_logits[-1].softmax(-1), -1) else: prediction, attn_weights, memory = outputs mAP, TP_, FN_ = compute_mean_average_precision(prediction[-1], gts, test_dataset_fn) MPJPE_ = compute_mpjpe(prediction[-1], gts, test_dataset_fn) # reduce by frame to get per joint MPJPE MPJPE = MPJPE + np.sum(MPJPE_, axis=0) # reduce by frame to get per joint AP TP_joint = TP_joint + np.sum(TP_, axis=0) FN_joint = FN_joint + np.sum(FN_, axis=0) # reduce by joint to get per frame AP TP_ = np.sum(TP_, axis=-1) FN_ = np.sum(FN_, axis=-1) TP = TP + TP_ FN = FN + FN_ # print(n, ':', prediction[-1].size(), out_class.item(), a_ids.item(), mAP, TP.shape, FN.shape) if params['predict_activity']: pred_activity.append(out_class.item()) gt_activity.append(a_ids.item()) #accurracy = (np.array(gt_activity) == np.array(pred_activity)).astype(np.float32).sum() #accurracy = accurracy / len(gt_activity) accurracy = -1 if params['predict_activity']: accurracy = accuracy_score(gt_activity, pred_activity, normalize='true') conf_matrix = confusion_matrix(gt_activity, pred_activity, normalize='true') plot_conf_mat(conf_matrix) AP = TP / (TP+FN) AP_joints = TP_joint / (TP_joint + FN_joint) MPJPE = MPJPE / (n*params['target_seq_len']) print('[INFO] The mAP per joint\n', np.around(AP_joints, 2)) print('[INFO] The MPJPE\n', np.around(MPJPE,4)*100.0) print('[INFO] The accuracy: {} mAP: {}'.format( round(accurracy, 2), round(np.mean(AP), 2))) ms_range = [0.08, 0.160, 0.320, 0.400, 0.5, 0.66] FPS = 30.0 ms_map = [] for ms in ms_range: nf = int(round(ms*FPS)) ms_map.append(np.mean(AP[0:nf])) print() print("{0: <16} |".format("milliseconds"), end="") for ms in ms_range: print(" {0:5d} |".format(int(ms*1000)), end="") print() print("{0: <16} |".format("global mAP"), end="") for mAP in ms_map: print(" {0:.3f} |".format(mAP), end="") print() def visualize_attn_weights(): parser = argparse.ArgumentParser() parser.add_argument('--config_file', type=str) parser.add_argument('--model_file', type=str) parser.add_argument('--data_path', type=str, default=None) args = parser.parse_args() params = json.load(open(args.config_file)) if args.data_path is not None: params['data_path'] = args.data_path args.data_path = params['data_path'] train_dataset_fn, eval_dataset_fn = tr_fn.dataset_factory(params) pose_encoder_fn, pose_decoder_fn = \ PoseEncoderDecoder.select_pose_encoder_decoder_fn(params) model = PoseTransformer.model_factory( params, pose_encoder_fn, pose_decoder_fn ) model.load_state_dict(torch.load(args.model_file, map_location=_DEVICE)) model.to(_DEVICE) model.eval() for k,v in params.items(): print('[INFO] (POTRFn@main) {}: {}'.format(k, v)) # ids of most common actions in H36M actions = [('walking', 12)] #[('walking', 12), ('eating', 2), ('smoking', 9), # ('discussion', 1), ('directions', 0)] with torch.no_grad(): for i in range(len(actions)): action, acidx = actions[i] sample = next(iter(eval_dataset_fn)) enc_inputs = sample['encoder_inputs'].to(_DEVICE) dec_inputs = sample['decoder_inputs'].to(_DEVICE) enc_inputs = torch.squeeze(enc_inputs) dec_inputs = torch.squeeze(dec_inputs) prediction, attn_weights, enc_weights = model( enc_inputs, dec_inputs, get_attn_weights=True ) attn_weights= attn_weights[-1][8*acidx:8*acidx+8].cpu().numpy()[0:1] attn_weights = np.squeeze(attn_weights) print(attn_weights.shape) path = 'skeletons/%s'%action in_imgs_ = [crop_image(cv2.imread(os.path.join(path, x)) ) for x in os.listdir(path) if 'gt_in' in x] in_imgs = [in_imgs_[i] for i in range(0, len(in_imgs_), 2)] pred_imgs = [crop_image(cv2.imread(os.path.join(path, x))) for x in os.listdir(path) if 'pred_0' in x] the_shape = in_imgs[0].shape cx = the_shape[1]//2 cy = the_shape[0]//2 in_imgs = np.concatenate(in_imgs, axis=1) pred_imgs = np.concatenate(pred_imgs, axis=1) #cv2.imshow('In IMG', in_imgs) #cv2.imshow('pred IMG', pred_imgs) #cv2.waitKey() spaces_between = 5 print(in_imgs.shape, pred_imgs.shape, the_shape) canvas = np.ones( (in_imgs.shape[0]*spaces_between, in_imgs.shape[1], 3), dtype=in_imgs.dtype)*255 canvas[0:the_shape[0], :] = in_imgs canvas[the_shape[0]*(spaces_between-1):, 0:pred_imgs.shape[1]] = pred_imgs #cx_pred = cx + the_shape[1]*(spaces_between-1) - cx//2 cy_pred = cy + the_shape[0]*(spaces_between-1) - cy//3*2 print(attn_weights.min(), attn_weights.max()) mean = attn_weights.mean() #plt.imshow(canvas, origin='lower') pil_canvas = Image.fromarray(canvas) d_canvas = ImageDraw.Draw(pil_canvas) for pred_idx in range(attn_weights.shape[0]): # cy_pred = cy + pred_idx*the_shape[0] cx_pred = cx + pred_idx*the_shape[1] #cv2.circle(canvas, (cx_pred, cy_pred), 5, [0,255,0], -1) for ii, in_idx in enumerate(range(0, attn_weights.shape[1], 2)): # cy_in = cy + ii*the_shape[0] cx_in = cx + ii*the_shape[1] this_weight = attn_weights[pred_idx, in_idx] if this_weight > mean: #d_canvas.line([(cx+cx//2, cy_in), (cx_pred, cy_pred)], fill=(255,0,0, 25), width=this_weight/mean) d_canvas.line([(cx_in, cy+cy//3*2), (cx_pred, cy_pred)], fill=(255,0,0, 25), width=this_weight/mean) name = 'the_canvas.png' #cv2.imwrite('the_canvas.jpg', canvas) # plt.show() #plt.imsave(name, canvas) pil_canvas.save(name) print(pil_canvas.info) fig, ax = plt.subplots(figsize=(20,10)) ax.matshow(attn_weights) plt.ylabel("") plt.xlabel("") fig.tight_layout() #plt.show() name = 'attn_map.png' plt.savefig(name) plt.close() if __name__ == '__main__': # visualize_h36mdataset() visualize_attn_weights() #compute_test_mAP_nturgbd()
14,622
30.245726
110
py
GaitForeMer
GaitForeMer-main/utils/WarmUpScheduler.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Warm up scheduler implementation. Proposed in 'Accurate, Large Minibatch SGD: Training ImageNet in 1 Hour'. Adapted from https://github.com/ildoonet/pytorch-gradual-warmup-lr/blob/master/warmup_scheduler/scheduler.py """ from torch.optim.lr_scheduler import _LRScheduler from torch.optim.lr_scheduler import ReduceLROnPlateau class GradualWarmupScheduler(_LRScheduler): """ Gradually warm-up(increasing) learning rate in optimizer.""" def __init__(self, optimizer, multiplier, total_epoch, after_scheduler=None): """Constructor. Args: optimizer (Optimizer): Wrapped optimizer. multiplier: target learning rate = base lr * multiplier if multiplier > 1.0. if multiplier = 1.0, lr starts from 0 and ends up with the base_lr. total_epoch: target learning rate is reached at total_epoch, gradually after_scheduler: after target_epoch, use this scheduler (eg. ReduceLROnPlateau) """ self.multiplier = multiplier if self.multiplier < 1.: raise ValueError('multiplier should be greater than or equal to 1.') self.total_epoch = total_epoch self.after_scheduler = after_scheduler self.finished = False super(GradualWarmupScheduler, self).__init__(optimizer) def get_lr(self): if self.last_epoch > self.total_epoch: if self.after_scheduler: if not self.finished: self.after_scheduler.base_lrs = [ base_lr * self.multiplier for base_lr in self.base_lrs] self.finished = True # return self.after_scheduler.get_last_lr() return self.after_scheduler.get_lr() return [base_lr * self.multiplier for base_lr in self.base_lrs] if self.multiplier == 1.0: return [base_lr * (float(self.last_epoch) / self.total_epoch) for base_lr in self.base_lrs] else: return [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in self.base_lrs] def step_ReduceLROnPlateau(self, metrics, epoch=None): if epoch is None: epoch = self.last_epoch + 1 # ReduceLROnPlateau is called at the end of epoch, whereas others # are called at beginning self.last_epoch = epoch if epoch != 0 else 1 if self.last_epoch <= self.total_epoch: warmup_lr = [base_lr * ((self.multiplier - 1.) * self.last_epoch / self.total_epoch + 1.) for base_lr in self.base_lrs] for param_group, lr in zip(self.optimizer.param_groups, warmup_lr): param_group['lr'] = lr else: if epoch is None: self.after_scheduler.step(metrics, None) else: self.after_scheduler.step(metrics, epoch - self.total_epoch) def step(self, epoch=None, metrics=None): if type(self.after_scheduler) != ReduceLROnPlateau: if self.finished and self.after_scheduler: if epoch is None: self.after_scheduler.step(None) else: self.after_scheduler.step(epoch - self.total_epoch) # self._last_lr = self.after_scheduler.get_last_lr() self._last_lr = self.after_scheduler.get_lr() else: return super(GradualWarmupScheduler, self).step(epoch) else: self.step_ReduceLROnPlateau(metrics, epoch)
4,280
39.771429
108
py
GaitForeMer
GaitForeMer-main/utils/utils.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Set of utility functions.""" import torch import numpy as np import copy import json import os import cv2 import torch.nn as nn def expmap_to_euler(action_sequence): rotmats = expmap_to_rotmat(action_sequence) eulers = rotmat_to_euler(rotmats) return eulers def expmap_to_rotmat(action_sequence): """Convert exponential maps to rotmats. Args: action_sequence: [n_samples, n_joints, 3] Returns: Rotation matrices for exponenital maps [n_samples, n_joints, 9]. """ n_samples, n_joints, _ = action_sequence.shape expmap = np.reshape(action_sequence, [n_samples*n_joints, 1, 3]) # first three values are positions, so technically it's meaningless to convert them, # but we do it anyway because later we discard this values anywho rotmats = np.zeros([n_samples*n_joints, 3, 3]) for i in range(rotmats.shape[0]): rotmats[i] = cv2.Rodrigues(expmap[i])[0] rotmats = np.reshape(rotmats, [n_samples, n_joints, 3*3]) return rotmats def rotmat_to_expmap(action_sequence): """Convert rotmats to expmap. Args: action_sequence: [n_samples, n_joints, 9] Returns: Rotation exponenital maps [n_samples, n_joints, 3]. """ n_samples, n_joints, _ = action_sequence.shape rotmats = np.reshape(action_sequence, [n_samples*n_joints, 3, 3]) # first three values are positions, so technically it's meaningless to convert them, # but we do it anyway because later we discard this values anywho expmaps = np.zeros([n_samples*n_joints, 3, 1]) for i in range(rotmats.shape[0]): expmaps[i] = cv2.Rodrigues(rotmats[i])[0] expmaps = np.reshape(expmaps, [n_samples, n_joints, 3]) return expmaps def rotmat_to_euler(action_sequence): """Convert exponential maps to rotmats. Args: action_sequence: [n_samples, n_joints, 9] Returns: Euler angles for rotation maps given [n_samples, n_joints, 3]. """ n_samples, n_joints, _ = action_sequence.shape rotmats = np.reshape(action_sequence, [n_samples*n_joints, 3, 3]) eulers = np.zeros([n_samples*n_joints, 3]) for i in range(eulers.shape[0]): eulers[i] = rotmat2euler(rotmats[i]) eulers = np.reshape(eulers, [n_samples, n_joints, 3]) return eulers def rotmat2euler(R): """Converts a rotation matrix to Euler angles. Matlab port to python for evaluation purposes https://github.com/asheshjain399/RNNexp/blob/srnn/structural_rnn/CRFProblems/H3.6m/mhmublv/Motion/RotMat2Euler.m#L1 Args: R: a 3x3 rotation matrix Returns: eul: a 3x1 Euler angle representation of R """ if R[0,2] >= 1 or R[0,2] <= -1: # special case values are out of bounds for arcsinc E3 = 0 # set arbitrarily dlta = np.arctan2( R[0,1], R[0,2] ); if R[0,2] == -1: E2 = np.pi/2; E1 = E3 + dlta; else: E2 = -np.pi/2; E1 = -E3 + dlta; else: E2 = -np.arcsin(R[0,2]) E1 = np.arctan2(R[1,2]/np.cos(E2), R[2,2]/np.cos(E2) ) E3 = np.arctan2(R[0,1]/np.cos(E2), R[0,0]/np.cos(E2) ) eul = np.array([E1, E2, E3]); return eul def load_constants(data_path): offset = json.load(open(os.path.join(data_path, 'offset.json'))) parent = json.load(open(os.path.join(data_path, 'parent.json'))) rot_ind = json.load(open(os.path.join(data_path, 'rot_ind.json'))) parent = np.array(parent)-1 offset = np.array(offset).reshape(-1, 3) exp_map_ind = np.split(np.arange(4, 100)-1, 32) return parent, offset, rot_ind, exp_map_ind def compute_forward_kinematics(angles, parent, offset, rotInd, expmapInd): """Computes forward kinematics from angles to 3d points. Convert joint angles and bone lenghts into the 3d points of a person. Based on expmap2xyz.m, available at https://github.com/asheshjain399/RNNexp/blob/7fc5a53292dc0f232867beb66c3a9ef845d705cb/structural_rnn/CRFProblems/H3.6m/mhmublv/Motion/exp2xyz.m Args angles: 99-long vector with 3d position and 3d joint angles in expmap format parent: 32-long vector with parent-child relationships in the kinematic tree offset: 96-long vector with bone lenghts rotInd: 32-long list with indices into angles expmapInd: 32-long list with indices into expmap angles Returns xyz: 32x3 3d points that represent a person in 3d space """ assert len(angles) == 99, 'Incorrect number of angles.' # Structure that indicates parents for each joint njoints = 32 xyzStruct = [dict() for x in range(njoints)] for i in np.arange(njoints): if not rotInd[i] : # If the list is empty xangle, yangle, zangle = 0, 0, 0 else: xangle = angles[rotInd[i][0]-1] yangle = angles[rotInd[i][1]-1] zangle = angles[rotInd[i][2]-1] r = angles[expmapInd[i]] thisRotation = expmap2rotmat(r) thisPosition = np.array([xangle, yangle, zangle]) if parent[i] == -1: # Root node xyzStruct[i]['rotation'] = thisRotation xyzStruct[i]['xyz'] = np.reshape(offset[i,:], (1,3)) + thisPosition else: xyzStruct[i]['xyz'] = (offset[i,:] + thisPosition).dot( xyzStruct[parent[i]]['rotation']) + xyzStruct[parent[i]]['xyz'] xyzStruct[i]['rotation'] = thisRotation.dot( xyzStruct[parent[i]]['rotation']) xyz = [xyzStruct[i]['xyz'] for i in range(njoints)] xyz = np.array(xyz).squeeze() xyz = xyz[:,[0,2,1]] # xyz = xyz[:,[2,0,1]] return np.reshape( xyz, [-1] ) def revert_coordinate_space(channels, R0, T0): """Arrange poses to a canonical form to face the camera. Bring a series of poses to a canonical form so they are facing the camera when they start. Adapted from https://github.com/asheshjain399/RNNexp/blob/7fc5a53292dc0f232867beb66c3a9ef845d705cb/structural_rnn/CRFProblems/H3.6m/dataParser/Utils/revertCoordinateSpace.m Args: channels: n-by-99 matrix of poses R0: 3x3 rotation for the first frame T0: 1x3 position for the first frame Returns: channels_rec: The passed poses, but the first has T0 and R0, and the rest of the sequence is modified accordingly. """ n, d = channels.shape channels_rec = copy.copy(channels) R_prev = R0 T_prev = T0 rootRotInd = np.arange(3,6) for ii in range(n): R_diff = expmap2rotmat(channels[ii, rootRotInd]) R = R_diff.dot(R_prev) channels_rec[ii, rootRotInd] = rotmat2expmap(R) T = T_prev + (R_prev.T).dot(np.reshape(channels[ii,:3],[3,1])).reshape(-1) channels_rec[ii,:3] = T T_prev = T R_prev = R return channels_rec def rotmat2quat(R): """Converts a rotation matrix to a quaternion. Matlab port to python for evaluation purposes https://github.com/asheshjain399/RNNexp/blob/srnn/structural_rnn/CRFProblems/H3.6m/mhmublv/Motion/rotmat2quat.m#L4 Args: R: 3x3 rotation matrix Returns: q: 1x4 quaternion """ rotdiff = R - R.T; r = np.zeros(3) r[0] = -rotdiff[1,2] r[1] = rotdiff[0,2] r[2] = -rotdiff[0,1] sintheta = np.linalg.norm(r) / 2; r0 = np.divide(r, np.linalg.norm(r) + np.finfo(np.float32).eps ); costheta = (np.trace(R)-1) / 2; theta = np.arctan2( sintheta, costheta ); q = np.zeros(4) q[0] = np.cos(theta/2) q[1:] = r0*np.sin(theta/2) return q def quat2expmap(q): """Convert quaternions to an exponential map. Matlab port to python for evaluation purposes https://github.com/asheshjain399/RNNexp/blob/srnn/structural_rnn/CRFProblems/H3.6m/mhmublv/Motion/quat2expmap.m#L1 Args: q: 1x4 quaternion Returns: r: 1x3 exponential map Raises: ValueError if the l2 norm of the quaternion is not close to 1 """ if (np.abs(np.linalg.norm(q)-1)>1e-3): raise(ValueError, "quat2expmap: input quaternion is not norm 1") sinhalftheta = np.linalg.norm(q[1:]) coshalftheta = q[0] r0 = np.divide( q[1:], (np.linalg.norm(q[1:]) + np.finfo(np.float32).eps)); theta = 2 * np.arctan2( sinhalftheta, coshalftheta ) theta = np.mod( theta + 2*np.pi, 2*np.pi ) if theta > np.pi: theta = 2 * np.pi - theta r0 = -r0 r = r0 * theta return r def rotmat2expmap(R): return quat2expmap( rotmat2quat(R) ) def expmap2rotmat(r): """Converts an exponential map (axis angle number) to rotation matrix. Converts an exponential map angle to a rotation matrix Matlab port to python for evaluation purposes. This is also called Rodrigues' formula and can be found also implemented in opencv as cv2.Rodrigues. https://github.com/asheshjain399/RNNexp/blob/srnn/structural_rnn/CRFProblems/H3.6m/mhmublv/Motion/expmap2rotmat.m Args: r: 1x3 exponential map Returns: R: 3x3 rotation matrix """ theta = np.linalg.norm( r ) r0 = np.divide( r, theta + np.finfo(np.float32).eps ) r0x = np.array([0, -r0[2], r0[1], 0, 0, -r0[0], 0, 0, 0]).reshape(3,3) r0x = r0x - r0x.T R = np.eye(3,3) + np.sin(theta)*r0x + (1-np.cos(theta))*(r0x).dot(r0x); return R def revert_output_format( poses, data_mean, data_std, dim_to_ignore, actions, use_one_hot): """Transforms pose predictions to a more interpretable format. Converts the output of the neural network to a format that is more easy to manipulate for, e.g. conversion to other format or visualization Args: poses: Sequence of pose predictions. A list with (seq_length) entries, each with a (batch_size, dim) output Returns: poses_out: List of tensors each of size (batch_size, seq_length, dim). """ seq_len = len(poses) if seq_len == 0: return [] batch_size, dim = poses[0].shape poses_out = np.concatenate(poses) poses_out = np.reshape(poses_out, (seq_len, batch_size, dim)) poses_out = np.transpose(poses_out, [1, 0, 2]) poses_out_list = [] for i in range(poses_out.shape[0]): poses_out_list.append( unnormalize_data(poses_out[i, :, :], data_mean, data_std, dim_to_ignore, actions, use_one_hot)) return poses_out_list def unnormalize_data( normalizedData, data_mean, data_std, dimensions_to_ignore=None, actions=[], use_one_hot=False): """ https://github.com/asheshjain399/RNNexp/blob/srnn/structural_rnn/CRFProblems/H3.6m/generateMotionData.py#L12 Args normalizedData: nxd matrix with normalized data data_mean: vector of mean used to normalize the data data_std: vector of standard deviation used to normalize the data dimensions_to_ignore: vector with dimensions not used by the model actions: list of strings with the encoded actions use_one_hot: whether the data comes with one-hot encoding Returns origData: data originally used to """ T = normalizedData.shape[0] D = data_mean.shape[0] origData = np.zeros((T, D), dtype=np.float32) dimensions_to_ignore = [] if dimensions_to_ignore==None else dimensions_to_ignore dimensions_to_use = [i for i in range(D) if i not in dimensions_to_ignore] dimensions_to_use = np.array(dimensions_to_use) #print('Size of the normalized data', normalizedData.shape) #print('Size of the mean data', data_mean.shape[0]) #print('Lenght of the dimensions to use', len(dimensions_to_use)) if use_one_hot: origData[:, dimensions_to_use] = normalizedData[:, :-len(actions)] else: # print('++++++++++++++++++++',origData.shape, normalizedData.shape, len(dimensions_to_use)) origData[:, dimensions_to_use] = normalizedData # potentially ineficient, but only done once per experiment stdMat = data_std.reshape((1, D)) stdMat = np.repeat(stdMat, T, axis=0) meanMat = data_mean.reshape((1, D)) meanMat = np.repeat(meanMat, T, axis=0) origData = np.multiply(origData, stdMat) + meanMat return origData def get_srnn_gts( actions, model, test_set, data_mean, data_std, dim_to_ignore, one_hot, to_euler=True): """ Get the ground truths for srnn's sequences, and convert to Euler angles. (the error is always computed in Euler angles). Args actions: a list of actions to get ground truths for. model: training model we are using (we only use the "get_batch" method). test_set: dictionary with normalized training data. data_mean: d-long vector with the mean of the training data. data_std: d-long vector with the standard deviation of the training data. dim_to_ignore: dimensions that we are not using to train/predict. one_hot: whether the data comes with one-hot encoding indicating action. to_euler: whether to convert the angles to Euler format or keep thm in exponential map Returns srnn_gts_euler: a dictionary where the keys are actions, and the values are the ground_truth, denormalized expected outputs of srnns's seeds. """ srnn_gts_euler = {} for action in actions: srnn_gt_euler = [] _, _, srnn_expmap = model.get_batch_srnn( test_set, action ) # expmap -> rotmat -> euler for i in np.arange( srnn_expmap.shape[0] ): denormed = data_utils.unNormalizeData(srnn_expmap[i,:,:], data_mean, data_std, dim_to_ignore, actions, one_hot ) if to_euler: for j in np.arange( denormed.shape[0] ): for k in np.arange(3,97,3): denormed[j,k:k+3] = data_utils.rotmat2euler( data_utils.expmap2rotmat( denormed[j,k:k+3] )) srnn_gt_euler.append( denormed ); # Put back in the dictionary srnn_gts_euler[action] = srnn_gt_euler return srnn_gts_euler def normal_init_(layer, mean_, sd_, bias, norm_bias=True): """Intialization of layers with normal distribution with mean and bias""" classname = layer.__class__.__name__ # Only use the convolutional layers of the module #if (classname.find('Conv') != -1 ) or (classname.find('Linear')!=-1): if classname.find('Linear') != -1: print('[INFO] (normal_init) Initializing layer {}'.format(classname)) layer.weight.data.normal_(mean_, sd_) if norm_bias: layer.bias.data.normal_(bias, 0.05) else: layer.bias.data.fill_(bias) def weight_init( module, mean_=0, sd_=0.004, bias=0.0, norm_bias=False, init_fn_=normal_init_): """Initialization of layers with normal distribution""" moduleclass = module.__class__.__name__ try: for layer in module: if layer.__class__.__name__ == 'Sequential': for l in layer: init_fn_(l, mean_, sd_, bias, norm_bias) else: init_fn_(layer, mean_, sd_, bias, norm_bias) except TypeError: init_fn_(module, mean_, sd_, bias, norm_bias) def xavier_init_(layer, mean_, sd_, bias, norm_bias=True): classname = layer.__class__.__name__ if classname.find('Linear')!=-1: print('[INFO] (xavier_init) Initializing layer {}'.format(classname)) nn.init.xavier_uniform_(layer.weight.data) # nninit.xavier_normal(layer.bias.data) if norm_bias: layer.bias.data.normal_(0, 0.05) else: layer.bias.data.zero_() def create_dir_tree(base_dir): dir_tree = ['models', 'tf_logs', 'config', 'std_log'] for dir_ in dir_tree: os.makedirs(os.path.join(base_dir, dir_), exist_ok=True) def create_look_ahead_mask(seq_length, is_nonautoregressive=False): """Generates a binary mask to prevent to use future context in a sequence.""" if is_nonautoregressive: return np.zeros((seq_length, seq_length), dtype=np.float32) x = np.ones((seq_length, seq_length), dtype=np.float32) mask = np.triu(x, 1).astype(np.float32) return mask # (seq_len, seq_len) def pose_expmap2rotmat(input_pose): """Convert exponential map pose format to rotation matrix pose format.""" pose_rotmat = [] for j in np.arange(input_pose.shape[0]): rot_mat = [expmap2rotmat(input_pose[j, k:k+3]) for k in range(3, 97, 3)] pose_rotmat.append(np.stack(rot_mat).flatten()) pose_rotmat = np.stack(pose_rotmat) return pose_rotmat def expmap23d_sequence(sequence, norm_stats, params): viz_poses = revert_output_format( [sequence], norm_stats['mean'], norm_stats['std'], norm_stats['dim_to_ignore'], params['action_subset'], params['use_one_hot']) nframes = sequence.shape[0] expmap = revert_coordinate_space( viz_poses[0], np.eye(3), np.zeros(3)) xyz_data = np.zeros((nframes, 96)) for i in range(nframes): xyz_data[i, :] = compute_forward_kinematics( expmap[i, :], params['parent'], params['offset'], params['rot_ind'], params['exp_map_ind'] ) return xyz_data def get_lr_fn(params, optimizer_fn): """Creates the function to be used to generate the learning rate.""" if params['learning_rate_fn'] == 'step': return torch.optim.lr_scheduler.StepLR( optimizer_fn, step_size=params['lr_step_size'], gamma=0.1 ) elif params['learning_rate_fn'] == 'exponential': return torch.optim.lr_scheduler.ExponentialLR( optimizer_fn, gamma=0.95 ) elif params['learning_rate_fn'] == 'linear': # sets learning rate by multipliying initial learning rate times a function lr0, T = params['learning_rate'], params['max_epochs'] lrT = lr0*0.1 m = (lrT - 1) / T lambda_fn = lambda epoch: m*epoch + 1.0 return torch.optim.lr_scheduler.LambdaLR( optimizer_fn, lr_lambda=lambda_fn ) elif params['learning_rate_fn'] == 'beatles': # D^(-0.5)*min(i^(-0.5), i*warmup_steps^(-1.5)) D = float(params['model_dim']) warmup = params['warmup_epochs'] lambda_fn = lambda e: (D**(-0.5))*min((e+1.0)**(-0.5), (e+1.0)*warmup**(-1.5)) return torch.optim.lr_scheduler.LambdaLR( optimizer_fn, lr_lambda=lambda_fn ) else: raise ValueError('Unknown learning rate function: {}'.format( params['learning_rate_fn'])) def compute_mean_average_precision(prediction, target, threshold, per_frame=False): """ Args: prediction: unormalized sequece of shape [seq_len, num_joints, 3] target: unormalized sequence of shape [seq_len, num_joints, 3] threshold: float """ # compute the norm for the last axis: (x,y,z) coordinates # [num_frames x num_joints] TP = np.linalg.norm(prediction-target, axis=-1) <= threshold TP_ = TP.astype(int) FN_ = np.logical_not(TP).astype(int) # [num_joints] TP = np.sum(TP_, axis=0) FN = np.sum(FN_, axis=0) # compute recall for each joint recall = TP / (TP+FN) # average over joints mAP = np.mean(recall) if per_frame: return mAP, TP, FN, (TP_, FN_) return mAP, TP, FN
19,278
30.708882
161
py
GaitForeMer
GaitForeMer-main/utils/visualize_attention_weights.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Visualization of the attention weights.""" import torch import torch.nn as nn import numpy as np import argparse import sys import os import tqdm import json from sklearn.metrics import confusion_matrix thispath = os.path.dirname(os.path.abspath(__file__)) sys.path.insert(0, thispath+"/../") import data.H36MDataset_v2 as H36M_v2 import models.PoseTransformer as PoseTransformer import models.PoseEncoderDecoder as PoseEncoderDecoder import matplotlib.pyplot as plt import matplotlib def plot_conf_mat(matrix, action_labels): fig, ax = plt.subplots() im = ax.imshow(matrix, cmap='Wistia') ax.set_xticks(np.arange(len(action_labels))) ax.set_yticks(np.arange(len(action_labels))) ax.set_xticklabels(action_labels, fontdict={'fontsize':10}, rotation=90) ax.set_yticklabels(action_labels, fontdict={'fontsize':10}) # cbar_kw={} # if set_colorbar: # cbar = ax.figure.colorbar(im, ax=ax, **cbar_kw) # cbar.ax.set_ylabel("", rotation=-90, va="bottom") # nmax= np.max(matrix)/2. ax.tick_params(top=True, bottom=False, labeltop=True, labelbottom=False) for i in range(len(action_labels)): for j in range(len(action_labels)): # color= "w" if round(matrix[i, j],2) < nmax else "black" text = ax.text(j, i, round(matrix[i, j], 2), ha="center", va="center", color="black", fontsize=5) plt.ylabel("") plt.xlabel("") # ax.set_title("Small plot") fig.tight_layout() plt.show() #plt.savefig(name) plt.close() the_keys = [(5, 'directions', 1), (5, 'directions', 2), (5, 'discussion', 1), (5, 'discussion', 2), (5, 'eating', 1), (5, 'eating', 2), (5, 'greeting', 1), (5, 'greeting', 2), (5, 'phoning', 1), (5, 'phoning', 2), (5, 'posing', 1), (5, 'posing', 2), (5, 'purchases', 1), (5, 'purchases', 2), (5, 'sitting', 1), (5, 'sitting', 2), (5, 'sittingdown', 1), (5, 'sittingdown', 2), (5, 'smoking', 1), (5, 'smoking', 2), (5, 'takingphoto', 1), (5, 'takingphoto', 2), (5, 'waiting', 1), (5, 'waiting', 2), (5, 'walking', 1), (5, 'walking', 2), (5, 'walkingdog', 1), (5, 'walkingdog', 2), (5, 'walkingtogether', 1), (5, 'walkingtogether', 2) ] def get_windows( data, source_seq_len, target_seq_len, pad_decoder_inputs, input_size, n_windows): N, _ = data.shape src_seq_len = source_seq_len - 1 encoder_inputs_ = [] decoder_inputs_ = [] decoder_outputs_ = [] start_frame = 0 for n in range(n_windows): encoder_inputs = np.zeros((src_seq_len, input_size), dtype=np.float32) decoder_inputs = np.zeros((target_seq_len, input_size), dtype=np.float32) decoder_outputs = np.zeros((target_seq_len, input_size), dtype=np.float32) # total_frames x n_joints*joint_dim total_frames = source_seq_len + target_seq_len data_sel = data[start_frame:(start_frame+total_frames), :] encoder_inputs[:, 0:input_size] = data_sel[0:src_seq_len,:] decoder_inputs[:, 0:input_size] = data_sel[src_seq_len:src_seq_len+target_seq_len, :] decoder_outputs[:, 0:input_size] = data_sel[source_seq_len:, 0:input_size] if pad_decoder_inputs: query = decoder_inputs[0:1, :] decoder_inputs = np.repeat(query, target_seq_len, axis=0) encoder_inputs_.append(encoder_inputs) decoder_inputs_.append(decoder_inputs) decoder_outputs_.append(decoder_outputs) start_frame = start_frame + src_seq_len return ( torch.from_numpy(np.stack(encoder_inputs_)), torch.from_numpy(np.stack(decoder_inputs_)), torch.from_numpy(np.stack(decoder_outputs_)) ) if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--params_json', type=str, default=None) parser.add_argument('--model', type=str, default= None) args = parser.parse_args() _DEVICE = torch.device('cuda' if torch.cuda.is_available() else 'cpu') params = json.load(open(args.params_json)) train_dataset_fn, eval_dataset_fn = H36M_v2.dataset_factory(params) pose_encoder_fn, pose_decoder_fn = \ PoseEncoderDecoder.select_pose_encoder_decoder_fn(params) potr = PoseTransformer.model_factory( params, pose_encoder_fn, pose_decoder_fn) potr.load_state_dict(torch.load(args.model, map_location=_DEVICE)) potr.to(_DEVICE) potr.eval() all_pred, all_gt = [], [] n_windows = 8 the_keys_ = [the_keys[i] for i in range(1, len(the_keys), 2)] with torch.no_grad(): for i in range(len(the_keys_)): entry_key = the_keys_[i] # (5, 'walking', 1) data = eval_dataset_fn.dataset._data[entry_key] encoder_inputs, decoder_inputs, decoder_outputs = get_windows( data, params['source_seq_len'], params['target_seq_len'], params['pad_decoder_inputs'], params['input_dim'], n_windows ) pred_sequence, attn_weights, enc_weights= potr( encoder_inputs.to(_DEVICE), decoder_inputs.to(_DEVICE), get_attn_weights=True ) enc_weights = enc_weights.cpu().numpy() attn_weights = attn_weights[-1].cpu().numpy() attn_weights = [attn_weights[j] for j in range(n_windows)] mat = np.concatenate(attn_weights, axis=-1) mat = np.concatenate([enc_weights[j] for j in range(n_windows)], axis=-1) print(enc_weights.shape) fig, ax = plt.subplots(figsize=(20,10)) ax.matshow(mat) plt.ylabel("") plt.xlabel("") fig.tight_layout() #plt.show() name = 'vis_attn/%s_.png'%(entry_key[1]) plt.savefig(name) plt.close()
6,648
33.630208
632
py
GaitForeMer
GaitForeMer-main/utils/PositionEncodings.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Implementation of the 2D positional encodings used in [1]. Position encodings gives a signature to each pixel in the image by a set of sine frequecies computed with a 2D sine function. [1] https://arxiv.org/abs/2005.12872 [2] https://arxiv.org/pdf/1706.03762.pdf """ import numpy as np import math import torch from torch import nn class PositionEncodings2D(object): """Implementation of 2d masked position encodings as a NN layer. This is a more general version of the position embedding, very similar to the one used by the Attention is all you need paper, but generalized to work on images as used in [1]. """ def __init__( self, num_pos_feats=64, temperature=10000, normalize=False, scale=None): """Constructs position embeding layer. Args: num_pos_feats: An integer for the depth of the encoding signature per pixel for each axis `x` and `y`. temperature: Value of the exponential temperature. normalize: Bool indicating if the encodings shuld be normalized by number of pixels in each image row. scale: Use for scaling factor. Normally None is used for 2*pi scaling. """ super().__init__() self._num_pos_feats = num_pos_feats self._temperature = temperature self._normalize = normalize if scale is not None and normalize is False: raise ValueError("normalize should be True if scale is passed") if scale is None: scale = 2 * math.pi self._scale = scale def __call__(self, mask): """Generates the positional encoding given image boolean mask. Args: mask: Boolean tensor of shape [batch_size, width, height] with ones in pixels that belong to the padding and zero in valid pixels. Returns: Sine position encodings. Shape [batch_size, num_pos_feats*2, width, height] """ # the positional encodings are generated for valid pixels hence # we need to take the negation of the boolean mask not_mask = ~mask y_embed = not_mask.cumsum(1, dtype=torch.float32) x_embed = not_mask.cumsum(2, dtype=torch.float32) if self._normalize: eps = 1e-6 y_embed = y_embed / (y_embed[:, -1:, :] + eps) * self._scale x_embed = x_embed / (x_embed[:, :, -1:] + eps) * self._scale dim_t = torch.arange( self._num_pos_feats, dtype=torch.float32) dim_t = self._temperature ** (2 * (dim_t // 2) / self._num_pos_feats) pos_x = x_embed[:, :, :, None] / dim_t pos_y = y_embed[:, :, :, None] / dim_t pos_x = torch.stack((pos_x[:, :, :, 0::2].sin(), pos_x[:, :, :, 1::2].cos()), dim=4).flatten(3) pos_y = torch.stack((pos_y[:, :, :, 0::2].sin(), pos_y[:, :, :, 1::2].cos()), dim=4).flatten(3) pos = torch.cat((pos_y, pos_x), dim=3).permute(0, 3, 1, 2) return pos class PositionEncodings1D(object): """Positional encodings for `1D` sequences. Implements the following equations: PE_{(pos, 2i)} = sin(pos/10000^{2i/d_model}) PE_{(pos, 2i+1)} = cos(pos/10000^{2i/d_model}) Where d_model is the number of positional features. Also known as the depth of the positional encodings. These are the positional encodings proposed in [2]. """ def __init__(self, num_pos_feats=512, temperature=10000, alpha=1): self._num_pos_feats = num_pos_feats self._temperature = temperature self._alpha = alpha def __call__(self, seq_length): angle_rads = self.get_angles( np.arange(seq_length)[:, np.newaxis], np.arange(self._num_pos_feats)[np.newaxis, :] ) # apply sin to even indices in the array; 2i angle_rads[:, 0::2] = np.sin(angle_rads[:, 0::2]) # apply cos to odd indices in the array; 2i+1 angle_rads[:, 1::2] = np.cos(angle_rads[:, 1::2]) pos_encoding = angle_rads[np.newaxis, ...] pos_encoding = pos_encoding.astype(np.float32) return torch.from_numpy(pos_encoding) def get_angles(self, pos, i): angle_rates = 1 / np.power( self._temperature, (2 * (i//2)) / np.float32(self._num_pos_feats)) return self._alpha*pos * angle_rates def visualize_2d_encodings(): import cv2 import numpy as np import matplotlib.pyplot as pplt # Create a mask where pixels are all valid mask = torch.BoolTensor(1, 32, 32).fill_(False) # position encodigns with a signature of depth per pixel # the efective pixel signature is num_pos_feats*2 (128 for each axis) pos_encodings_gen = PositionEncodings2D(num_pos_feats=128, normalize=True) encodings = pos_encodings_gen(mask).numpy() print('Shape of encodings', encodings.shape) # visualize the first frequency channel for x and y y_encodings = encodings[0,0, :, :] x_encodings = encodings[0,128, : ,:] pplt.matshow(x_encodings, cmap=pplt.get_cmap('jet')) pplt.matshow(y_encodings, cmap=pplt.get_cmap('jet')) pplt.show() def visualize_1d_encodings(): import matplotlib.pyplot as plt pos_encoder_gen = PositionEncodings1D() pos_encoding = pos_encoder_gen(50).numpy() print(pos_encoding.shape) plt.pcolormesh(pos_encoding[0], cmap='RdBu') plt.xlabel('Depth') plt.xlim((0, 512)) plt.ylabel('position in sequence') plt.colorbar() plt.show() if __name__ == "__main__": visualize_2d_encodings() # visualize_1d_encodings()
6,337
32.712766
81
py
GaitForeMer
GaitForeMer-main/data/GaitJointsDataset.py
import os import sys import numpy as np import torch import argparse import tqdm import pickle import random _TOTAL_ACTIONS = 4 # Mapping from 1-base of NTU to vibe 49 joints # hip, thorax, _MAJOR_JOINTS = [39, 41, 37, 43, 34, 35, 36, 33, 32, 31, 28, 29, 30, 27, 26, 25, 40] # 1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15, 17, 18, 19, 21 _NMAJOR_JOINTS = len(_MAJOR_JOINTS) _MIN_STD = 1e-4 _SPINE_ROOT = 0 # after only taking major joints (ie index in _MAJOR_JOINTS) def collate_fn(batch): """Collate function for data loaders.""" e_inp = torch.from_numpy(np.stack([e['encoder_inputs'] for e in batch])) d_inp = torch.from_numpy(np.stack([e['decoder_inputs'] for e in batch])) d_out = torch.from_numpy(np.stack([e['decoder_outputs'] for e in batch])) action_id = torch.from_numpy(np.stack([e['action_id'] for e in batch])) action = [e['action_str'] for e in batch] batch_ = { 'encoder_inputs': e_inp, 'decoder_inputs': d_inp, 'decoder_outputs': d_out, 'action_str': action, 'action_ids': action_id } return batch_ class GaitJointsDataset(torch.utils.data.Dataset): def __init__(self, params=None, mode='train', fold=1): super(GaitJointsDataset, self).__init__() self._params = params self._mode = mode thisname = self.__class__.__name__ self._monitor_action = 'normal' self._action_str = ['normal', 'slight', 'moderate', 'severe'] self.data_dir = self._params['data_path'] self.fold = fold self.load_data() def load_data(self): train_data = pickle.load(open(self.data_dir+"EPG_train_" + str(self.fold) + ".pkl", "rb")) test_data = pickle.load(open(self.data_dir+"EPG_test_" + str(self.fold) + ".pkl", "rb")) if self._mode == 'train': X_1, Y = self.data_generator(train_data, mode='train', fold_number=self.fold) else: X_1, Y = self.data_generator(test_data) self.X_1 = X_1 self.Y = Y self._action_str = ['none', 'mild', 'moderate', 'severe'] self._pose_dim = 3 * _NMAJOR_JOINTS self._data_dim = self._pose_dim def data_generator(self, T, mode='test', fold_number=1): X_1 = [] Y = [] # bootstrap_number = 3 # num_samples = 39 total_num_clips = 0 for i in range(len(T['pose'])): total_num_clips += 1 p = np.copy(T['pose'][i]) # print(np.shape(p)) y_label_index = T['label'][i] label = y_label_index X_1.append(p) Y.append(label) # can't stack X_1 because not all have equal frames Y = np.stack(Y) # For using a subset of the dataset (few-shot) # if mode == 'train': # sampling_dir = 'PATH/TO/BOOTSTRAP_SAMPLING_DIR' # all_clip_video_names = pickle.load(open(sampling_dir + "all_clip_video_names.pkl", "rb")) # clip_video_names = all_clip_video_names[fold_number - 1] # all_bootstrap_samples = pickle.load(open(sampling_dir + f'{num_samples}_samples/bootstrap_{bootstrap_number}_samples.pkl', 'rb')) # bootstrap_samples = all_bootstrap_samples[fold_number - 1] # mask_list = [1 if video_name in bootstrap_samples else 0 for video_name in clip_video_names] # train_indices = [train_idx for train_idx, mask_value in enumerate(mask_list) if mask_value == 1] # X_1 = [X_1[train_idx] for train_idx in train_indices] # Y = Y[train_indices] return X_1, Y def __len__(self): return len(self.Y) def __getitem__(self, idx): return self._get_item_train(idx) def _get_item_train(self, idx): """Get item for the training mode.""" x = self.X_1[idx] y = self.Y[idx] # adjust for mapping/subset of joints from vibe to ntu x = x[:,_MAJOR_JOINTS,:] action_id = y source_seq_len = self._params['source_seq_len'] target_seq_len = self._params['target_seq_len'] input_size = 3 * _NMAJOR_JOINTS # not sure if this is right pose_size = 3 * _NMAJOR_JOINTS # note sure if thiis is right total_frames = source_seq_len + target_seq_len src_seq_len = source_seq_len - 1 encoder_inputs = np.zeros((src_seq_len, input_size), dtype=np.float32) decoder_inputs = np.zeros((target_seq_len, input_size), dtype=np.float32) decoder_outputs = np.zeros((target_seq_len, pose_size), dtype=np.float32) # total_framesxn_joints*joint_dim N = np.shape(x)[0] x = x.reshape(N, -1) start_frame = np.random.randint(0, N - total_frames + 1) # original code did not change start frame between epochs start_frame = random.randint(0, N - total_frames) # high inclusive data_sel = x[start_frame:(start_frame + total_frames), :] encoder_inputs[:, 0:input_size] = data_sel[0:src_seq_len,:] decoder_inputs[:, 0:input_size] = \ data_sel[src_seq_len:src_seq_len+target_seq_len, :] decoder_outputs[:, 0:pose_size] = data_sel[source_seq_len:, 0:pose_size] if self._params['pad_decoder_inputs']: query = decoder_inputs[0:1, :] decoder_inputs = np.repeat(query, target_seq_len, axis=0) return { 'encoder_inputs': encoder_inputs, 'decoder_inputs': decoder_inputs, 'decoder_outputs': decoder_outputs, 'action_id': action_id, 'action_str': self._action_str[action_id], } def dataset_factory(params, fold): """Defines the datasets that will be used for training and validation.""" params['num_activities'] = _TOTAL_ACTIONS params['virtual_dataset_size'] = params['steps_per_epoch']*params['batch_size'] params['n_joints'] = _NMAJOR_JOINTS eval_mode = 'test' if 'test_phase' in params.keys() else 'eval' if eval_mode == 'test': train_dataset_fn = None else: train_dataset = GaitJointsDataset(params, mode='train', fold=fold) train_dataset_fn = torch.utils.data.DataLoader( train_dataset, batch_size=params['batch_size'], shuffle=True, num_workers=4, collate_fn=collate_fn, ) eval_dataset = GaitJointsDataset( params, mode=eval_mode, fold=fold, ) eval_dataset_fn = torch.utils.data.DataLoader( eval_dataset, batch_size=1, shuffle=False, num_workers=1, collate_fn=collate_fn, ) return train_dataset_fn, eval_dataset_fn
6,232
31.128866
137
py
GaitForeMer
GaitForeMer-main/data/NTURGDDataset.py
############################################################################### # Pose Transformers (POTR): Human Motion Prediction with Non-Autoregressive # Transformers # # Copyright (c) 2021 Idiap Research Institute, http://www.idiap.ch/ # Written by # Angel Martinez <angel.martinez@idiap.ch>, # # This file is part of # POTR: Human Motion Prediction with Non-Autoregressive Transformers # # POTR is free software: you can redistribute it and/or modify # it under the terms of the GNU General Public License version 3 as # published by the Free Software Foundation. # # POTR is distributed in the hope that it will be useful, # but WITHOUT ANY WARRANTY; without even the implied warranty of # MERCHANTABILITY or FITNESS FOR A PARTICULAR PURPOSE. See the # GNU General Public License for more details. # # You should have received a copy of the GNU General Public License # along with POTR. If not, see <http://www.gnu.org/licenses/>. ############################################################################### """Pytorch dataset of skeletons for the NTU-RGB+D [1] dataset. [1] http://rose1.ntu.edu.sg/Datasets/actionRecognition.asp [2] https://www.cv-foundation.org/openaccess/content_cvpr_2016/papers/Shahroudy_NTU_RGBD_A_CVPR_2016_paper.pdf """ import os import sys import numpy as np import torch import argparse import tqdm # tran subjects id can be found in [2] _TRAIN_SUBJECTS = [ 1, 2, 4, 5, 8, 9, 13, 14, 15,16, 17, 18, 19, 25, 27, 28, 31, 34, 35, 38 ] _TEST_SUBJECTS = [x for x in range(1, 40) if x not in _TRAIN_SUBJECTS] # the joints according to [2] in 1-base # 1-base of the spine 2-middle of the spine 3-neck 4-head 5-left shoulder # 6-left elbow 7-left wrist 8-left hand 9-right shoulder 10-right elbow # 11-right wrist 12-right hand 13-left hip 14-left knee 15-left ankle # 16-left foot 17-right hip 18-right knee 19-right ankle 20-right foot # 21-spine 22-tip of the left hand 23-left thumb 24-tip of the right # hand 25-right thumb # here set the joint indices in base 0 _MAJOR_JOINTS = [x-1 for x in [1, 2, 3, 4, 5, 6, 7, 9, 10, 11, 13, 14, 15, 17, 18, 19, 21] ] _NMAJOR_JOINTS = len(_MAJOR_JOINTS) _SPINE_ROOT = 0 _MIN_STD = 1e-4 # NTURGB+D contains 60 actions _TOTAL_ACTIONS = 60 _MIN_REQ_FRAMES = 65 def collate_fn(batch): """Collate function for data loaders.""" e_inp = torch.from_numpy(np.stack([e['encoder_inputs'] for e in batch])) d_inp = torch.from_numpy(np.stack([e['decoder_inputs'] for e in batch])) d_out = torch.from_numpy(np.stack([e['decoder_outputs'] for e in batch])) action_id = torch.from_numpy(np.stack([e['action_id'] for e in batch])) action = [e['action_str'] for e in batch] batch_ = { 'encoder_inputs': e_inp, 'decoder_inputs': d_inp, 'decoder_outputs': d_out, 'action_str': action, 'action_ids': action_id } return batch_ def load_action_labels(data_path): data_labels = [] with open(os.path.join(data_path, 'action_labels.txt')) as file_: for line in file_: data_labels.append(line.strip()) return data_labels def get_activity_from_file(path_file): # The pattern is SsssCcccPpppRrrrAaaa.skeleton pattern = path_file.split('/')[-1].split('.')[0] setup_id= int(pattern[1:4]) camera_id = int(pattern[5:8]) subject_id = int(pattern[9:12]) replication_id = int(pattern[13:16]) activity_id = int(pattern[17:]) return (setup_id, camera_id, subject_id, replication_id, activity_id) def select_fold_files(path_to_data, skip_files_path): all_files = [x for x in os.listdir(path_to_data) if x.endswith('skeleton')] with open(skip_files_path) as file_: skip_files = [line.strip() for line in file_] training_files = [] test_files = [] for path_file in all_files: if path_file.split('.')[0] in skip_files: print('Skiping file:', path_file) continue seq_info = get_activity_from_file(path_file) if seq_info[2] in _TRAIN_SUBJECTS: training_files.append(path_file) else: test_files.append(path_file) return training_files, test_files def save_fold_files(path_to_data, output_path, skip_files_path): training_files, test_files = select_fold_files(path_to_data, skip_files_path) val_idx = np.random.choice( len(training_files), int(len(training_files)*0.05), replace=False) training_files = [training_files[i] for i in range(len(training_files)) if i not in val_idx] val_files = [training_files[i] for i in range(len(training_files)) if i in val_idx] with open(os.path.join(output_path, 'training_files.txt'), 'w') as file_: for f in training_files: print(f, file=file_) with open(os.path.join(output_path, 'testing_files.txt'), 'w') as file_: for f in test_files: print(f, file=file_) with open(os.path.join(output_path, 'validation_files.txt'), 'w') as file_: for f in val_files: print(f, file=file_) def read_sequence_kinect_skeletons(path_file): """Reads the text file provided in the """ fid = open(path_file, 'r') seq_info = get_activity_from_file(path_file) # first line is the number of frames framecount = int(fid.readline().strip()) bodies = {} for i in range(framecount): bodycount = int(fid.readline().strip()) for b in range(bodycount): # traccking ID of the skeleton line = fid.readline().strip().split(' ') body_id = int(line[0]) arrayint = [int(x) for x in line[1:7]] lean = [float(x) for x in line[7:9]] tracking_state = int(line[-1]) #number of joints joint_count = int(fid.readline().strip()) joints = [] for j in range(joint_count): line = fid.readline().strip().split(' ') # 3D location of the joint joint_3d = [float(x) for x in line[0:3]] # 2D depth location of joints joint_2d_depth = [float(x) for x in line[3:5]] # 2D color location of joints joint_2d_color = [float(x) for x in line[5:7]] # orientation of joints (?) joint_orientation = [float(x) for x in line[7:11]] # tracking state joint_track_state = int(line[-1]) joints.append(joint_3d) if body_id in list(bodies.keys()): bodies[body_id].append(np.array(joints, dtype=np.float32)) else: bodies[body_id] = [np.array(joints, dtype=np.float32)] for k, v in bodies.items(): bodies[k] = np.stack(v) return bodies, seq_info def select_sequence_based_var(action_sequence_dict): """Selects the actor in sequence based on the sum of variance of X, Y, Z.""" larger_var = -1 selected_key = None for k, v in action_sequence_dict.items(): var = np.var(v, axis=-1) sum_var = np.sum(var) if sum_var > larger_var: larger_var = sum_var selected_key = k return action_sequence_dict[selected_key] class NTURGDDatasetSkeleton(torch.utils.data.Dataset): def __init__(self, params=None, mode='train'): super(NTURGDDatasetSkeleton, self).__init__() self._params = params self._mode = mode thisname = self.__class__.__name__ self._monitor_action = 'walking' for k, v in params.items(): print('[INFO] ({}) {}: {}'.format(thisname, k, v)) data_path = self._params['data_path'] self._action_str = load_action_labels(data_path) self._fold_file = '' if self._mode.lower() == 'train': self._fold_file = os.path.join(data_path, 'training_files.txt') elif self._mode.lower() == 'eval': self._fold_file = os.path.join(data_path, 'validation_files.txt') elif self._mode.lower() == 'test': self._fold_file = os.path.join(data_path, 'testing_files.txt') else: raise ValueError('Unknown launching mode: {}'.format(self._mode)) self.load_data() def read_fold_file(self, fold_file): files = [] with open(fold_file) as file_: for line in file_: files.append(line.strip()) return files def compute_norm_stats(self, data): self._norm_stats = {} mean = np.mean(data, axis=0) std = np.mean(data, axis=0) std[np.where(std<_MIN_STD)] = 1 self._norm_stats['mean'] = mean.ravel() self._norm_stats['std'] = std.ravel() def load_compute_norm_stats(self, data): mean_path = os.path.join(self._params['data_path'], 'mean.npy') std_path = os.path.join(self._params['data_path'], 'std.npy') thisname = self.__class__.__name__ self._norm_stats = {} if os.path.exists(mean_path): print('[INFO] ({}) Loading normalization stats!'.format(thisname)) self._norm_stats['mean'] = np.load(mean_path) self._norm_stats['std'] = np.load(std_path) elif self._mode == 'train': print('[INFO] ({}) Computing normalization stats!'.format(thisname)) self.compute_norm_stats(data) np.save(mean_path, self._norm_stats['mean']) np.save(std_path, self._norm_stats['std']) else: raise ValueError('Cant compute statistics in not training mode!') def normalize_data(self): for k in self._data.keys(): tmp_data = self._data[k] tmp_data = tmp_data - self._norm_stats['mean'] tmp_data = np.divide(tmp_data, self._norm_stats['std']) self._data[k] = tmp_data def load_data(self): seq_files = self.read_fold_file(self._fold_file) self._data = {} all_dataset = [] seq_lens = [] for sequence_file in tqdm.tqdm(seq_files): sequence_file = os.path.join(self._params['data_path'], 'nturgb+d_skeletons', sequence_file) # the sequence key contains # (setup_id, camera_id, subject_id, replication_id, activity_id) # sequence shape [num_frames, 25, 3] action_sequence, seq_key = read_sequence_kinect_skeletons(sequence_file) # added code, there are no actors in sequence if len(action_sequence) == 0: continue # FOR TESTING PURPOSES, EXIT LOADING CODE EARLY # if len(all_dataset) > 100: # break action_sequence = select_sequence_based_var(action_sequence) # sequence shape [num_frames, 16, 3] action_sequence = action_sequence[:, _MAJOR_JOINTS, :] # Only consider sequences with more than _MIN_REQ_FRAMES frames if action_sequence.shape[0]<_MIN_REQ_FRAMES: continue # center joints in the spine of the skeleton root_sequence = np.expand_dims(action_sequence[:, _SPINE_ROOT, :], axis=1) action_sequence = action_sequence - root_sequence T, N, D = action_sequence.shape seq_lens.append(T) # total_frames x n_joints*3 self._data[seq_key] = action_sequence.reshape((T, -1)) all_dataset.append(action_sequence) all_dataset = np.concatenate(all_dataset, axis=0) self.load_compute_norm_stats(all_dataset) self.normalize_data() self._pose_dim = self._norm_stats['std'].shape[-1] self._data_dim = self._pose_dim self._data_keys = list(self._data.keys()) thisname = self.__class__.__name__ print('[INFO] ({}) The min seq len for mode: {} is: {}'.format( thisname, self._mode, min(seq_lens))) print('[INFO] ({}) Pose dim: {} Data dim: {}'.format( thisname, self._pose_dim, self._data_dim)) def __len__(self): if self._mode == 'train': return max(len(self._data_keys), self._params['virtual_dataset_size']) return len(self._data_keys) def __getitem__(self, idx): return self._get_item_train(idx) def _get_item_train(self, idx): """Get item for the training mode.""" if self._mode == 'train': # idx = np.random.choice(len(self._data_keys), 1)[0] idx = np.random.choice(len(self._data_keys)) the_key = self._data_keys[idx] # the action id in the files come in 1 based index action_id = the_key[-1] - 1 source_seq_len = self._params['source_seq_len'] target_seq_len = self._params['target_seq_len'] input_size = self._pose_dim pose_size = self._pose_dim total_frames = source_seq_len + target_seq_len src_seq_len = source_seq_len - 1 encoder_inputs = np.zeros((src_seq_len, input_size), dtype=np.float32) decoder_inputs = np.zeros((target_seq_len, input_size), dtype=np.float32) decoder_outputs = np.zeros((target_seq_len, pose_size), dtype=np.float32) N, _ = self._data[the_key].shape start_frame = np.random.randint(0, N-total_frames) # total_framesxn_joints*joint_dim data_sel = self._data[the_key][start_frame:(start_frame+total_frames), :] encoder_inputs[:, 0:input_size] = data_sel[0:src_seq_len,:] decoder_inputs[:, 0:input_size] = \ data_sel[src_seq_len:src_seq_len+target_seq_len, :] decoder_outputs[:, 0:pose_size] = data_sel[source_seq_len:, 0:pose_size] if self._params['pad_decoder_inputs']: query = decoder_inputs[0:1, :] decoder_inputs = np.repeat(query, target_seq_len, axis=0) return { 'encoder_inputs': encoder_inputs, 'decoder_inputs': decoder_inputs, 'decoder_outputs': decoder_outputs, 'action_id': action_id, 'action_str': self._action_str[action_id], } def unormalize_sequence(self, action_sequence): sequence = action_sequence*self._norm_stats['std'] sequence = sequence + self._norm_stats['mean'] return sequence def dataset_factory(params): """Defines the datasets that will be used for training and validation.""" params['num_activities'] = _TOTAL_ACTIONS params['virtual_dataset_size'] = params['steps_per_epoch']*params['batch_size'] params['n_joints'] = _NMAJOR_JOINTS eval_mode = 'test' if 'test_phase' in params.keys() else 'eval' if eval_mode == 'test': train_dataset_fn = None else: train_dataset = NTURGDDatasetSkeleton(params, mode='train') train_dataset_fn = torch.utils.data.DataLoader( train_dataset, batch_size=params['batch_size'], shuffle=True, num_workers=4, collate_fn=collate_fn, drop_last=True ) eval_dataset = NTURGDDatasetSkeleton( params, mode=eval_mode ) eval_dataset_fn = torch.utils.data.DataLoader( eval_dataset, batch_size=1, shuffle=True, num_workers=1, drop_last=True, collate_fn=collate_fn, ) return train_dataset_fn, eval_dataset_fn if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--data_path', type=str, default=None) parser.add_argument('--pad_decoder_inputs', action='store_true') parser.add_argument('--source_seq_len', type=int, default=40) parser.add_argument('--target_seq_len', type=int, default=20) parser.add_argument('--batch_size', type=int, default=128) parser.add_argument('--steps_per_epoch', type=int, default=200) args = parser.parse_args() params = vars(args) print('Creating fold files') save_fold_files( 'data/nturgb+d_data/nturgb+d_skeletons', # path_to_data 'data/nturgb+d_data', # output_path 'data/nturgb+d_data/missing_skeletons.txt' # skip_files_path ) train_dataset_load, val_dataset_load = dataset_factory(params) for n, sample in enumerate(val_dataset_load): print(n, sample['encoder_inputs'].size(), sample['decoder_inputs'].size(), sample['decoder_outputs'].size(), sample['action_ids'].size())
15,284
32.084416
110
py
CPFN
CPFN-master/training_PatchSelection.py
# Importation of packages import os import sys import torch import argparse import numpy as np # Importing the Dataset file from Dataset import dataloaders # Importing the Network file from PointNet2 import pn2_network # Importing the Utils files from Utils import config_loader, training_utils, training_visualisation if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--config_file', help='YAML configuration file', default='Configs/config_patchSelec.yml') parser.add_argument('--lowres_dataset', help='Directory of the Lowres Dataset', default=os.path.expanduser('data/TraceParts_v2_LowRes/')) parser.add_argument('--highres_dataset', help='Directory of the Highres Dataset', default=os.path.expanduser('data/TraceParts_v2/')) parser.add_argument('--scale', help='Scale of the Primitives', type=float, default=0.05) parser.add_argument('--patchselec_weigths', help='Filename of the model weights to load', default='') args = parser.parse_args() # Loading the config file conf = config_loader.Patch_SelecConfig(args.config_file) # Selecting the visible GPUs visible_GPUs = conf.get_CUDA_visible_GPUs() device = torch.device('cuda') if visible_GPUs is not None: os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(visible_GPUs) # Training Parameters nb_epochs = conf.get_n_epochs() init_learning_rate = conf.get_init_learning_rate() val_interval = conf.get_val_interval() snapshot_interval = conf.get_snapshot_interval() # Training Dataset csv_path_train = os.path.join('Dataset', conf.get_train_data_file()) noisy_train = conf.get_train_data_first_n() first_n_train = conf.is_train_data_noisy() num_workers_train = conf.get_nb_train_workers() if not os.path.isdir(conf.get_weights_folder()): os.mkdir(conf.get_weights_folder()) # Validation Dataset csv_path_val = os.path.join('Dataset', conf.get_val_data_file()) noisy_val = conf.get_val_data_first_n() first_n_val = conf.is_val_data_noisy() num_workers_val = conf.get_nb_val_workers() # Launching the Network patchselec_weights_filename = 'patchselec_%s_module'%str(round(args.scale, 2)) patchselec_module = pn2_network.PointNet2(dim_input=3, dim_pos=3, output_sizes=[2]).to(device) if os.path.isfile(os.path.join(conf.get_weights_folder(), args.patchselec_weigths)): dict = torch.load(os.path.join(conf.get_weights_folder(), args.patchselec_weigths)) patchselec_module.load_state_dict(dict, strict=True) train_dataset = dataloaders.Dataset_PatchSelection(csv_path_train, args.lowres_dataset, args.highres_dataset, args.scale, n_points=8192, normalisation=True) train_datasampler = dataloaders.RandomSampler(data_source=train_dataset, seed=12345, identical_epochs=False) train_dataloader = torch.utils.data.DataLoader(train_dataset, sampler=train_datasampler, batch_size=conf.get_batch_size(), num_workers=conf.get_nb_train_workers(), pin_memory=True) val_dataset = dataloaders.Dataset_PatchSelection(csv_path_val, args.lowres_dataset, args.highres_dataset, args.scale, n_points=8192, normalisation=True) val_datasampler = dataloaders.RandomSampler(data_source=val_dataset, seed=12345, identical_epochs=False) val_dataloader = torch.utils.data.DataLoader(val_dataset, sampler=val_datasampler, batch_size=conf.get_batch_size(), num_workers=conf.get_nb_val_workers(), pin_memory=True) # Optimizer optimizer = torch.optim.Adam(patchselec_module.parameters(), lr=init_learning_rate) # Visualisation visualiser = training_visualisation.Visualiser(conf.get_visualisation_interval()) # Initialisation global_step = 0 best_loss = np.inf for epoch in range(nb_epochs): global_step, _ = training_utils.patch_selection_train_val_epoch(train_dataloader, patchselec_module, epoch, optimizer, global_step, visualiser, args, conf, device, network_mode='train') if (epoch % conf.get_val_interval() == 0) and (epoch > 0): with torch.no_grad(): _, loss = training_utils.patch_selection_train_val_epoch(val_dataloader, patchselec_module, epoch, optimizer, global_step, visualiser, args, conf, device, network_mode='val') if loss < best_loss: torch.save(patchselec_module.state_dict(), os.path.join(conf.get_weights_folder(), patchselec_weights_filename + '.pth')) best_loss = loss if (epoch % conf.get_snapshot_interval() == 0) and (epoch > 0): torch.save(patchselec_module.state_dict(), os.path.join(conf.get_weights_folder(), patchselec_weights_filename + '%d.pth'%epoch)) torch.save(patchselec_module.state_dict(), os.path.join(conf.get_weights_folder(), patchselec_weights_filename + '%d.pth' % epoch))
4,812
54.321839
193
py
CPFN
CPFN-master/evaluation_globalSPFN.py
# Importation of packages import os import sys import torch import argparse import numpy as np import pandas as pd # Importing the Dataset files from Dataset import dataloaders # Importing the Network files from SPFN import fitter_factory, metric_implementation, losses_implementation from PointNet2 import pn2_network # Importing Utils files from Utils import config_loader if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--config_file', help='YAML configuration file', default='Configs/config_globalSPFN.yml') parser.add_argument('--lowres_dataset', help='Directory of the Lowres Input Dataset', default=os.path.expanduser('data/TraceParts_v2_lowres/')) parser.add_argument('--highres_dataset', help='Directory of the Highres Input Dataset', default=os.path.expanduser('data/TraceParts_v2/')) parser.add_argument('--path_patches', help='Path to Sampled Patches h5 files', default=os.path.expanduser('data/TraceParts_v2_patches/')) parser.add_argument('--scale', help='Scale to select the smallest primitive', default=0.05, type=float) parser.add_argument('--output_folder', help='Directory of the output folder', default=os.path.expanduser('data/TraceParts_v2_globalspfn/')) parser.add_argument('--evaluation_set', help='Whether to evaluate on the train or test set', default='test') args = parser.parse_args() path_patches = os.path.join(args.path_patches, str(round(args.scale,2))) if not os.path.isdir(args.output_folder): os.mkdir(args.output_folder) # Loading the config file conf = config_loader.SPFNConfig(args.config_file) # Selecting the visible GPUs visible_GPUs = conf.get_CUDA_visible_GPUs() device = torch.device('cuda') if visible_GPUs is not None: os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(visible_GPUs) # Primtive Types and Numbers fitter_factory.register_primitives(conf.get_list_of_primitives()) n_registered_primitives = fitter_factory.get_n_registered_primitives() n_max_global_instances = conf.get_n_max_global_instances() # Test Dataset if args.evaluation_set == 'train': csv_path_test = os.path.join('Dataset', conf.get_train_data_file()) else: csv_path_test = os.path.join('Dataset', conf.get_test_data_file()) noisy_test = conf.is_test_data_noisy() first_n_test = conf.get_test_data_first_n() # Launching the Network spfn_module_filename = 'globalspfn_module.pth' spfn_module = pn2_network.PointNet2(dim_input=3, dim_pos=3, output_sizes=[3, n_registered_primitives, n_max_global_instances]).to(device) dict = torch.load(os.path.join(conf.get_weights_folder(), spfn_module_filename)) spfn_module.load_state_dict(dict, strict=True) spfn_module.eval() test_dataset = dataloaders.Dataset_GlobalSPFN(n_max_global_instances, csv_path_test, args.lowres_dataset, args.highres_dataset, path_patches, noisy_test, test=True, n_points=None, first_n=first_n_test, fixed_order=True) test_datasampler = dataloaders.Sampler(data_source=test_dataset) test_dataloader = torch.utils.data.DataLoader(test_dataset, sampler=test_datasampler, batch_size=1, num_workers=0, pin_memory=True) dataframe_results = pd.DataFrame(columns=['Filename', 'mIoU', 'Type', 'Normal', 'Axis', 'MeanRes', 'StdRes', 'SkCoverage0.01', 'SkCoverage0.02', 'PCoverage0.01', 'PCoverage0.02']) list_mIoU = [] for batch_id, data in enumerate(test_dataloader, 0): if batch_id%100==0: print('Iteration %d / %d' % (batch_id, len(test_dataloader))) P = data[0].type(torch.FloatTensor).to(device) X_gt = data[1].type(torch.FloatTensor).to(device) points_per_instance = data[2].type(torch.FloatTensor).to(device) I_gt = data[3].type(torch.LongTensor).to(device) T_gt = data[4].type(torch.LongTensor).to(device) plane_n_gt = data[5].type(torch.FloatTensor).to(device) cylinder_axis_gt = data[6].type(torch.FloatTensor).to(device) cone_axis_gt = data[7].type(torch.FloatTensor).to(device) patch_centers = data[8].type(torch.LongTensor).to(device) gt_parameters = {'plane_normal': plane_n_gt, 'cylinder_axis': cylinder_axis_gt, 'cone_axis': cone_axis_gt} glob_features = None loc_features = None if not os.path.isdir(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5',''))): os.mkdir(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5',''))) with torch.no_grad(): X, T, W, global_feat, local_feat = spfn_module(P, glob_features=glob_features, loc_features=loc_features) if args.evaluation_set == 'test': np.save(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5',''), 'local_feat_full.npy'), local_feat[0].cpu().numpy()) local_feat = local_feat[:,:,patch_centers[0]] X = X / torch.norm(X, dim=2, keepdim=True) W = torch.softmax(W, dim=2) with torch.no_grad(): W = metric_implementation.hard_W_encoding(W) matching_indices, mask = metric_implementation.hungarian_matching(W, I_gt) mask = mask.float() mIoU = metric_implementation.compute_segmentation_iou(W, I_gt, matching_indices, mask) if not os.path.isdir(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5', ''))): os.mkdir(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5', ''))) if args.evaluation_set == 'test': np.save(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5', ''), 'object_seg.npy'), W[0].cpu().numpy()) np.save(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5', ''), 'object_normals.npy'), X[0].cpu().numpy()) np.save(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5', ''), 'object_type.npy'), T[0].cpu().numpy()) mIoU, type_accuracy, normal_difference, axis_difference, mean_residual, std_residual, Sk_coverage, P_coverage, W, predicted_parameters, T = metric_implementation.compute_all_metrics(P, X, X_gt, W, I_gt, T, T_gt, points_per_instance, gt_parameters, list_epsilon=[0.01, 0.02], classes=conf.get_list_of_primitives()) list_mIoU.append(mIoU.item()) if batch_id%100==0: print('mIoU: ', np.mean(list_mIoU)) dataframe_results.loc[batch_id] = [test_dataset.hdf5_file_list[batch_id].replace('.h5',''), mIoU.item(), type_accuracy.item(), normal_difference.item(), axis_difference.item(), mean_residual.item(), std_residual.item(), Sk_coverage[0].item(), Sk_coverage[1].item(), P_coverage[0].item(), P_coverage[1].item()] np.save(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5',''), 'global_feat.npy'), global_feat[0,:,0].cpu().numpy()) np.save(os.path.join(args.output_folder, test_dataset.hdf5_file_list[batch_id].replace('.h5',''), 'local_feat.npy'), local_feat[0].cpu().numpy()) dataframe_results.to_csv(os.path.join(args.output_folder, 'Results.csv'))
7,235
63.607143
321
py
CPFN
CPFN-master/evaluation_baselineSPFN.py
# Importation of packages import os import sys import torch import argparse import numpy as np import pandas as pd # Importing the Dataset files from Dataset import dataloaders # Importing the Network files from SPFN import fitter_factory, metric_implementation, losses_implementation from PointNet2 import pn2_network # Importing Utils files from Utils import config_loader, merging_utils if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--config_file', help='YAML configuration file', default='Configs/config_localSPFN.yml') parser.add_argument('--lowres_dataset', help='Directory of the Lowres Input Dataset', default=os.path.expanduser('data/TraceParts_v2_LowRes/')) parser.add_argument('--highres_dataset', help='Directory of the Highres Input Dataset', default=os.path.expanduser('data/TraceParts_v2/')) parser.add_argument('--dir_spfn', help='Directory of the global SPFN output', default=os.path.expanduser('data/GlobalSPFN_Results/')) parser.add_argument('--dir_indices', help='Directory of the indices', default=os.path.expanduser('data/Heatmap/')) parser.add_argument('--output_folder', help='Directory of the output folder', default=os.path.expanduser('data/LocalSPFN_Results/')) parser.add_argument('--scale', help='Scale of the primitives', default=0.05) args = parser.parse_args() dir_indices = os.path.join(args.dir_indices, str(round(args.scale,2))) if not os.path.isdir(args.output_folder): os.mkdir(args.output_folder) # Loading the config file conf = config_loader.Local_SPFNConfig(args.config_file) # Selecting the visible GPUs visible_GPUs = conf.get_CUDA_visible_GPUs() device = torch.device('cuda') if visible_GPUs is not None: os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(visible_GPUs) # Primtive Types and Numbers fitter_factory.register_primitives(conf.get_list_of_primitives()) n_registered_primitives = fitter_factory.get_n_registered_primitives() n_max_global_instances = conf.get_n_max_global_instances() n_max_local_instances = conf.get_n_max_local_instances() # Test Dataset csv_path_test = os.path.join(args.lowres_dataset, conf.get_test_data_file()) noisy_test = conf.get_test_data_first_n() first_n_test = conf.is_test_data_noisy() test_dataset = dataloaders.Dataset_TestLocalSPFN(n_max_global_instances, n_max_local_instances, csv_path_test, args.dir_spfn, args.lowres_dataset, args.highres_dataset, dir_indices, noisy_test, first_n=first_n_test, fixed_order=True) test_datasampler = dataloaders.Sampler(data_source=test_dataset) test_dataloader = torch.utils.data.DataLoader(test_dataset, sampler=test_datasampler, batch_size=1, num_workers=0, pin_memory=True) dataframe_results = pd.DataFrame(columns=['Filename', 'mIoU', 'Type', 'Normal', 'Axis', 'MeanRes', 'StdRes', 'SkCoverage0.01', 'SkCoverage0.02', 'PCoverage0.01', 'PCoverage0.02']) cpt_df_stats = 0 dataframe_results_stats = pd.DataFrame(columns=['Filename', 'Primtive Id', 'Mask', 'Nb Points', 'mIoU']) list_mIoU = [] for batch_id, data in enumerate(test_dataloader, 0): if batch_id%100==0: print('Iteration %d / %d' % (batch_id, len(test_dataloader))) P = data[0].type(torch.FloatTensor).squeeze(0).to(device) nb_patches, num_points, _ = P.size() P_gt = data[2].type(torch.FloatTensor).squeeze(0).to(device) I_gt = data[3].type(torch.LongTensor).squeeze(0).to(device) T_gt = data[4].type(torch.LongTensor).squeeze(0).to(device) patch_indices = data[5].type(torch.LongTensor).squeeze(0).to(device) spfn_labels = data[6].type(torch.LongTensor).squeeze(0).to(device) num_global_points = spfn_labels.size(0) spfn_normals = data[7].type(torch.FloatTensor).squeeze(0).to(device) spfn_type = data[8].type(torch.FloatTensor).squeeze(0).to(device) glob_features = data[9].type(torch.FloatTensor).squeeze(0).to(device) loc_features = data[10].type(torch.FloatTensor).squeeze(0).to(device) P_global = data[11].type(torch.FloatTensor).squeeze(0).to(device) X_gt_global = data[12].type(torch.FloatTensor).squeeze(0).to(device) I_gt_global = data[13].type(torch.LongTensor).squeeze(0).to(device) plane_n_gt = data[14].type(torch.FloatTensor).to(device) cylinder_axis_gt = data[15].type(torch.FloatTensor).to(device) cone_axis_gt = data[16].type(torch.FloatTensor).to(device) gt_parameters = {'plane_normal': plane_n_gt, 'cylinder_axis': cylinder_axis_gt, 'cone_axis': cone_axis_gt} W_fusion = torch.eye(n_max_global_instances + 1).to(spfn_labels.device)[torch.argmax(spfn_labels, dim=1) + 1] W_fusion = W_fusion[:, 1:] X_global = spfn_normals T_global = spfn_type with torch.no_grad(): W_fusion = metric_implementation.hard_W_encoding(W_fusion.unsqueeze(0)) matching_indices_fusion, mask_fusion = metric_implementation.hungarian_matching(W_fusion, I_gt_global.unsqueeze(0)) mask_fusion = mask_fusion.float() mIoU_fusion = metric_implementation.compute_segmentation_iou(W_fusion, I_gt_global.unsqueeze(0), matching_indices_fusion, mask_fusion) mIoU_fusion_per_primitive = 1 - losses_implementation.compute_miou_loss(W_fusion, I_gt_global.unsqueeze(0), matching_indices_fusion)[0] _, unique_counts_primitives_fusion = np.unique(I_gt_global.cpu().numpy(), return_counts=True) for j in range(len(unique_counts_primitives_fusion)): dataframe_results_stats.loc[cpt_df_stats] = [test_dataset.hdf5_file_list[batch_id], j, mask_fusion[0, j].item(), unique_counts_primitives_fusion[j], mIoU_fusion_per_primitive[0, j].item()] cpt_df_stats += 1 # ADDED mIoU, type_accuracy, normal_difference, axis_difference, mean_residual, std_residual, Sk_coverage, P_coverage, W, predicted_parameters, T = metric_implementation.compute_all_metrics( P_global.unsqueeze(0), X_global.unsqueeze(0), X_gt_global.unsqueeze(0), W_fusion, I_gt_global.unsqueeze(0), T_global.unsqueeze(0), T_gt.unsqueeze(0), P_gt.unsqueeze(0), gt_parameters, list_epsilon=[0.01, 0.02], classes=['plane', 'sphere', 'cylinder', 'cone']) list_mIoU.append(mIoU.item()) if batch_id%100==0: print('mIoU: ', np.mean(list_mIoU)) dataframe_results.loc[batch_id] = [test_dataset.hdf5_file_list[batch_id], mIoU.item(), type_accuracy.item(), normal_difference.item(), axis_difference.item(), mean_residual.item(), std_residual.item(), Sk_coverage[0].item(), Sk_coverage[1].item(), P_coverage[0].item(), P_coverage[1].item()] dataframe_results.to_csv(os.path.join(args.output_folder, 'Results_baseline.csv'), index=False) dataframe_results_stats.to_csv(os.path.join(args.output_folder, 'Results_Stats_baseline.csv'), index=False)
7,131
60.482759
197
py
CPFN
CPFN-master/evaluation_PatchSelection.py
# Importation of packages import os import sys import h5py import torch import argparse import numpy as np # Importing the Dataset file from Dataset import dataloaders # Importing the Network file from PointNet2 import pn2_network # Importing the Utils files from Utils import config_loader, sampling_utils if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--config_file', help='YAML configuration file', default='Configs/config_patchSelec.yml') parser.add_argument('--lowres_dataset', help='Directory of the Lowres Dataset', default=os.path.expanduser('data/TraceParts_v2_lowres/')) parser.add_argument('--highres_dataset', help='Directory of the Highres Dataset', default=os.path.expanduser('data/TraceParts_v2/')) parser.add_argument('--heatmap_folder', help='Directory to save the heatmaps in', default=os.path.expanduser('data/TraceParts_v2_heatmaps/')) parser.add_argument('--scale', help='Scale of the Primitives', type=float, default=0.05) args = parser.parse_args() heatmap_folder = os.path.join(args.heatmap_folder, str(args.scale)) os.makedirs(heatmap_folder, exist_ok=True) # Loading the config file conf = config_loader.Patch_SelecConfig(args.config_file) # Selecting the visible GPUs visible_GPUs = conf.get_CUDA_visible_GPUs() device = torch.device('cuda') if visible_GPUs is not None: os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(visible_GPUs) # Test Dataset csv_path_test = os.path.join('Dataset', conf.get_test_data_file()) noisy_test = conf.get_test_data_first_n() first_n_test = conf.is_test_data_noisy() # Launching the Network if args.scale<1: patchselec_module_filename = 'patchselec_%s_module'%str(round(args.scale, 2)) + '.pth' patchselec_module = pn2_network.PointNet2(dim_input=3, dim_pos=3, output_sizes=[2]).to(device) dict = torch.load(os.path.join(conf.get_weights_folder(), patchselec_module_filename)) patchselec_module.load_state_dict(dict, strict=True) patchselec_module.eval() test_dataset = dataloaders.Dataset_PatchSelection(csv_path_test, args.lowres_dataset, args.highres_dataset, args.scale, n_points=8192, normalisation=True) test_datasampler = dataloaders.Sampler(data_source=test_dataset) test_dataloader = torch.utils.data.DataLoader(test_dataset, sampler=test_datasampler, batch_size=1, num_workers=0, pin_memory=True) # Initialisation if args.scale<1: confusion_matrix = np.zeros([2, 2]) for batch_id, data in enumerate(test_dataloader, 0): if batch_id%100==0: print('Iteration %d / %d' % (batch_id, len(test_dataloader))) # Computing the prediction points = data[0].type(torch.FloatTensor).to(device) batch_size_current, num_points, _ = points.size() output_labels = data[1].type(torch.LongTensor).to(device) shuffled_indices = data[2].type(torch.LongTensor).to(device) if args.scale<1: predicted_labels = patchselec_module(points)[0] predicted_labels = torch.argmax(predicted_labels, dim=2) else: predicted_labels = output_labels[0] if not os.path.isdir(os.path.join(heatmap_folder, test_dataset.hdf5_file_list_lowres[batch_id].split('/')[-1].replace('.h5',''))): os.mkdir(os.path.join(heatmap_folder, test_dataset.hdf5_file_list_lowres[batch_id].split('/')[-1].replace('.h5', ''))) # Computing the confusion matrix if args.scale<1: confusion_matrix[0, 0] += torch.sum((predicted_labels == 0) * (output_labels == 0)).item() confusion_matrix[0, 1] += torch.sum((predicted_labels == 0) * (output_labels == 1)).item() confusion_matrix[1, 0] += torch.sum((predicted_labels == 1) * (output_labels == 0)).item() confusion_matrix[1, 1] += torch.sum((predicted_labels == 1) * (output_labels == 1)).item() predicted_labels = torch.gather(predicted_labels[0], 0, shuffled_indices[0]) # Selecting the indices with h5py.File(os.path.join(args.highres_dataset, test_dataset.hdf5_file_list_lowres[batch_id].split('/')[-1]), 'r') as f: gt_points_hr = f['gt_points'][()] gt_labels_hr = f['gt_labels'][()] with h5py.File(os.path.join(os.path.join(args.lowres_dataset, test_dataset.hdf5_file_list_lowres[batch_id].split('/')[-1])), 'r') as f: gt_points_lr = f['gt_points'][()] gt_labels_lr = f['gt_labels'][()] pool_indices = np.where(predicted_labels.detach().cpu().numpy())[0] if len(pool_indices) > 0: patch_indices = sampling_utils.sample(gt_points_lr, gt_points_hr, pool_indices, max_number_patches=len(pool_indices)) np.save(os.path.join(heatmap_folder, test_dataset.hdf5_file_list_lowres[batch_id].split('/')[-1].replace('.h5', '_indices.npy')), patch_indices) if args.scale<1: confusion_matrix = confusion_matrix / np.sum(confusion_matrix) print('Confusion Matrix', confusion_matrix) np.save(os.path.join(heatmap_folder, 'confusion_matrix.npy'), confusion_matrix)
5,095
54.391304
158
py
CPFN
CPFN-master/evaluation_localSPFN.py
# Importation of packages import os import sys import torch import argparse import numpy as np import pandas as pd # Importing the Dataset files from Dataset import dataloaders # Importing the Network files from SPFN import fitter_factory, metric_implementation, losses_implementation from PointNet2 import pn2_network # Importing utils files from Utils import config_loader, merging_utils if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--config_file', help='YAML configuration file', default='Configs/config_localSPFN.yml') parser.add_argument('--lowres_dataset', help='Directory of the Lowres Input Dataset', default=os.path.expanduser('data/TraceParts_v2_lowres/')) parser.add_argument('--highres_dataset', help='Directory of the Highres Input Dataset', default=os.path.expanduser('data/TraceParts_v2/')) parser.add_argument('--dir_spfn', help='Directory of the global SPFN output', default=os.path.expanduser('data/TraceParts_v2_globalspfn/')) parser.add_argument('--dir_indices', help='Directory of the indices', default=os.path.expanduser('data/TraceParts_v2_heatmaps/')) parser.add_argument('--output_folder', help='Directory of the output folder', default=os.path.expanduser('data/TraceParts_v2_localspfn/')) parser.add_argument('--scale', help='Scale of the primitives', default=0.05, type=float) args = parser.parse_args() dir_indices = os.path.join(args.dir_indices, str(round(args.scale,2))) if not os.path.isdir(args.output_folder): os.mkdir(args.output_folder) # Loading the config file conf = config_loader.Local_SPFNConfig(args.config_file) # Selecting the visible GPUs visible_GPUs = conf.get_CUDA_visible_GPUs() device = torch.device('cuda') if visible_GPUs is not None: os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(visible_GPUs) # Primtive Types and Numbers fitter_factory.register_primitives(conf.get_list_of_primitives()) n_registered_primitives = fitter_factory.get_n_registered_primitives() n_max_global_instances = conf.get_n_max_global_instances() n_max_local_instances = conf.get_n_max_local_instances() # Test Dataset csv_path_test = os.path.join('Dataset', conf.get_test_data_file()) noisy_test = conf.get_test_data_first_n() first_n_test = conf.is_test_data_noisy() # Launching the Network spfn_module_filename = 'localspfn_%s_module.pth'%str(round(args.scale, 2)) spfn_module = pn2_network.PointNet2(dim_input=3, dim_pos=3, output_sizes=[3, n_registered_primitives, n_max_local_instances], use_glob_features=False, use_loc_features=False).to(device) dict = torch.load(os.path.join(conf.get_weights_folder(), spfn_module_filename)) spfn_module.load_state_dict(dict, strict=True) spfn_module.eval() test_dataset = dataloaders.Dataset_TestLocalSPFN(n_max_global_instances, n_max_local_instances, csv_path_test, args.dir_spfn, args.lowres_dataset, args.highres_dataset, dir_indices, noisy_test, first_n=first_n_test, fixed_order=True) test_datasampler = dataloaders.Sampler(data_source=test_dataset) test_dataloader = torch.utils.data.DataLoader(test_dataset, sampler=test_datasampler, batch_size=1, num_workers=0, pin_memory=True) dataframe_results = pd.DataFrame(columns=['Filename', 'mIoU', 'Type', 'Normal', 'Axis', 'MeanRes', 'StdRes', 'SkCoverage0.01', 'SkCoverage0.02', 'PCoverage0.01', 'PCoverage0.02']) cpt_df_stats = 0 dataframe_results_stats = pd.DataFrame(columns=['Filename', 'Primitive Id', 'Mask', 'Nb Points', 'mIoU']) list_mIoU = [] for batch_id, data in enumerate(test_dataloader, 0): if batch_id%100==0: print('Iteration %d / %d' % (batch_id, len(test_dataloader))) P = data[0].type(torch.FloatTensor).squeeze(0).to(device) nb_patches, num_points, _ = P.size() P_gt = data[2].type(torch.FloatTensor).squeeze(0).to(device) I_gt = data[3].type(torch.LongTensor).squeeze(0).to(device) T_gt = data[4].type(torch.LongTensor).squeeze(0).to(device) patch_indices = data[5].type(torch.LongTensor).squeeze(0).to(device) spfn_labels = data[6].type(torch.LongTensor).squeeze(0).to(device) num_global_points = spfn_labels.size(0) spfn_normals = data[7].type(torch.FloatTensor).squeeze(0).to(device) spfn_type = data[8].type(torch.FloatTensor).squeeze(0).to(device) glob_features = data[9].type(torch.FloatTensor).squeeze(0).to(device) loc_features = data[10].type(torch.FloatTensor).squeeze(0).to(device) P_global = data[11].type(torch.FloatTensor).squeeze(0).to(device) X_gt_global = data[12].type(torch.FloatTensor).squeeze(0).to(device) I_gt_global = data[13].type(torch.LongTensor).squeeze(0).to(device) plane_n_gt = data[14].type(torch.FloatTensor).to(device) cylinder_axis_gt = data[15].type(torch.FloatTensor).to(device) cone_axis_gt = data[16].type(torch.FloatTensor).to(device) gt_parameters = {'plane_normal': plane_n_gt, 'cylinder_axis': cylinder_axis_gt, 'cone_axis': cone_axis_gt} if nb_patches > 0: X, T, W, _, _ = spfn_module(P, glob_features=glob_features, loc_features=loc_features) X = X / torch.norm(X, dim=2, keepdim=True) W = torch.softmax(W, dim=2) with torch.no_grad(): W_fusion = W similarity_fusion = merging_utils.similarity_soft(spfn_labels, W_fusion, patch_indices) labels_fusion = merging_utils.run_heuristic_solver(similarity_fusion.cpu().numpy(), nb_patches, n_max_global_instances, n_max_local_instances) point2primitive_fusion = torch.zeros([num_global_points, nb_patches * n_max_local_instances + n_max_global_instances]).float().to(device) for b in range(nb_patches): point2primitive_fusion[patch_indices[b], b * n_max_local_instances:(b + 1) * n_max_local_instances] = W_fusion[b] point2primitive_fusion[:, (b+1)*n_max_local_instances:] = spfn_labels # Deleting the patch prediction for points within any patches flag = torch.sum(point2primitive_fusion[:,:(b+1)*n_max_local_instances], dim=1)>0 point2primitive_fusion[flag,(b+1)*n_max_local_instances:] = 0 W_fusion = merging_utils.get_point_final(point2primitive_fusion, torch.from_numpy(labels_fusion).to(device)) with torch.no_grad(): patch_indices = patch_indices.contiguous() X = X.contiguous() T = T.contiguous() # Normal estimation X_global = torch.zeros_like(X_gt_global) X_global = X_global.scatter_add_(0, patch_indices.view(-1).unsqueeze(1).expand(-1, 3), X.view(-1, 3)) empty_indices = torch.all(X_global==0, axis=1) X_global[empty_indices] = spfn_normals[empty_indices] X_global = torch.nn.functional.normalize(X_global, p=2, dim=1, eps=1e-12) # Type estimation T_gt_perpoint = torch.gather(T_gt, 0, I_gt_global) patch_indices = patch_indices.view(-1).unsqueeze(1).expand(-1, len(conf.get_list_of_primitives())) num = torch.zeros_like(T_gt_perpoint).float().unsqueeze(1).expand(-1, len(conf.get_list_of_primitives())) num = num.scatter_add(0, patch_indices, T.view(-1, len(conf.get_list_of_primitives()))) den = torch.zeros_like(T_gt_perpoint).float().unsqueeze(1).expand(-1, len(conf.get_list_of_primitives())) den = den.scatter_add(0, patch_indices, torch.ones_like(patch_indices).float()) T_global = num / den.clamp(min=1) T_global[empty_indices] = spfn_type[empty_indices] else: W_fusion = torch.eye(n_max_global_instances + 1).to(spfn_labels.device)[torch.argmax(spfn_labels, dim=1) + 1] W_fusion = W_fusion[:, 1:] X_global = spfn_normals T_global = spfn_type W_fusion = W_fusion[:,torch.sum(W_fusion, dim=0)>1] if W_fusion.shape[1] < n_max_global_instances: W_fusion = torch.cat((W_fusion, torch.zeros([W_fusion.shape[0], n_max_global_instances-W_fusion.shape[1]]).to(device)), dim=1) with torch.no_grad(): W_fusion = metric_implementation.hard_W_encoding(W_fusion.unsqueeze(0)) matching_indices_fusion, mask_fusion = metric_implementation.hungarian_matching(W_fusion, I_gt_global.unsqueeze(0)) mask_fusion = mask_fusion.float() mIoU_fusion = metric_implementation.compute_segmentation_iou(W_fusion, I_gt_global.unsqueeze(0), matching_indices_fusion, mask_fusion) mIoU_fusion_per_primitive = 1 - losses_implementation.compute_miou_loss(W_fusion, I_gt_global.unsqueeze(0), matching_indices_fusion)[0] _, unique_counts_primitives_fusion = np.unique(I_gt_global.cpu().numpy(), return_counts=True) for j in range(len(unique_counts_primitives_fusion)): dataframe_results_stats.loc[cpt_df_stats] = [test_dataset.hdf5_file_list[batch_id], j, mask_fusion[0, j].item(), unique_counts_primitives_fusion[j], mIoU_fusion_per_primitive[0, j].item()] cpt_df_stats += 1 with torch.no_grad(): mIoU, type_accuracy, normal_difference, axis_difference, mean_residual, std_residual, Sk_coverage, P_coverage, W, predicted_parameters, T = metric_implementation.compute_all_metrics( P_global.unsqueeze(0), X_global.unsqueeze(0), X_gt_global.unsqueeze(0), W_fusion, I_gt_global.unsqueeze(0), T_global.unsqueeze(0), T_gt.unsqueeze(0), P_gt.unsqueeze(0), gt_parameters, list_epsilon=[0.01, 0.02], classes=['plane', 'sphere', 'cylinder', 'cone']) list_mIoU.append(mIoU.item()) if batch_id%100==0: print('mIoU: ', np.mean(list_mIoU)) dataframe_results.loc[batch_id] = [test_dataset.hdf5_file_list[batch_id], mIoU.item(), type_accuracy.item(), normal_difference.item(), axis_difference.item(), mean_residual.item(), std_residual.item(), Sk_coverage[0].item(), Sk_coverage[1].item(), P_coverage[0].item(), P_coverage[1].item()] dataframe_results.to_csv(os.path.join(args.output_folder, 'Results.csv'), index=False) dataframe_results_stats.to_csv(os.path.join(args.output_folder, 'Results_Stats.csv'), index=False)
10,606
63.284848
197
py
CPFN
CPFN-master/training_SPFN.py
# Importation of packages import os import sys import torch import argparse import numpy as np # Importing the Dataset files from Dataset import dataloaders # Importing the Network files from SPFN import fitter_factory, losses_implementation from PointNet2 import pn2_network # Importing Utils files from Utils import config_loader, training_utils, training_visualisation if __name__ == '__main__': parser = argparse.ArgumentParser() parser.add_argument('--config_file', help='YAML configuration file', type=str, default='Configs/config_globalSPFN.yml') parser.add_argument('--lowres_dataset', help='Directory of the Input Dataset', type=str, default=os.path.expanduser('data/TraceParts_v2_lowres/')) parser.add_argument('--network', help='Network to train: GlobalSPFN, LocalSPFN', type=str, default='GlobalSPFN') parser.add_argument('--path_patches', help='Path to Sampled Patches h5 files', type=str, default=os.path.expanduser('data/TraceParts_v2_patches')) parser.add_argument('--scale', help='Scale to select the smallest primitive', type=float, default=0.05) parser.add_argument('--spfn_weigths', help='Filename of the model weights to load', type=str, default='') args = parser.parse_args() # Loading the config file assert (args.network in ['GlobalSPFN', 'LocalSPFN']) if args.network == 'GlobalSPFN': conf = config_loader.Global_SPFNConfig(args.config_file) elif args.network == 'LocalSPFN': conf = config_loader.Local_SPFNConfig(args.config_file) # Selecting the visible GPUs visible_GPUs = conf.get_CUDA_visible_GPUs() device = torch.device('cuda') if visible_GPUs is not None: os.environ["CUDA_VISIBLE_DEVICES"] = ','.join(visible_GPUs) # Primtive Types and Numbers fitter_factory.register_primitives(conf.get_list_of_primitives()) n_registered_primitives = fitter_factory.get_n_registered_primitives() n_max_global_instances = conf.get_n_max_global_instances() if args.network == 'LocalSPFN': n_max_local_instances = conf.get_n_max_local_instances() # Training Parameters nb_epochs = conf.get_n_epochs() init_learning_rate = conf.get_init_learning_rate() val_interval = conf.get_val_interval() snapshot_interval = conf.get_snapshot_interval() # Training Dataset csv_path_train = os.path.join('Dataset', conf.get_train_data_file()) noisy_train = conf.is_train_data_noisy() first_n_train = conf.get_train_data_first_n() num_workers_train = conf.get_nb_train_workers() path_patches = os.path.join(args.path_patches, str(round(args.scale, 2))) # Validation Dataset csv_path_val = os.path.join('Dataset', conf.get_val_data_file()) noisy_val = conf.is_val_data_noisy() first_n_val = conf.get_val_data_first_n() num_workers_val = conf.get_nb_val_workers() # Launching the Network if args.network == 'GlobalSPFN': spfn_weights_filename = 'globalspfn_module' spfn_module = pn2_network.PointNet2(dim_input=3, dim_pos=3, output_sizes=[3, n_registered_primitives, n_max_global_instances]).to(device) elif args.network == 'LocalSPFN': spfn_weights_filename = 'localspfn_%s_module'%str(round(args.scale, 2)) spfn_module = pn2_network.PointNet2(dim_input=3, dim_pos=3, output_sizes=[3, n_registered_primitives, n_max_local_instances]).to(device) if os.path.isfile(os.path.join(conf.get_weights_folder(), args.spfn_weigths)): dict = torch.load(os.path.join(conf.get_weights_folder(), args.spfn_weigths)) spfn_module.load_state_dict(dict, strict=True) # Loading the dataset if args.network == 'GlobalSPFN': train_dataset = dataloaders.Dataset_GlobalSPFN(n_max_global_instances, csv_path_train, args.lowres_dataset, None, None, noisy_train, n_points=8192, first_n=first_n_train, fixed_order=False) train_datasampler = dataloaders.RandomSampler(data_source=train_dataset, seed=12345, identical_epochs=False) train_dataloader = torch.utils.data.DataLoader(train_dataset, sampler=train_datasampler, batch_size=conf.get_batch_size(), num_workers=num_workers_train, pin_memory=True) val_dataset = dataloaders.Dataset_GlobalSPFN(n_max_global_instances, csv_path_val, args.lowres_dataset, None, None, noisy_val, n_points=8192, first_n=first_n_val, fixed_order=False) val_datasampler = dataloaders.RandomSampler(data_source=val_dataset, seed=12345, identical_epochs=False) val_dataloader = torch.utils.data.DataLoader(val_dataset, sampler=val_datasampler, batch_size=conf.get_batch_size(), num_workers=conf.get_nb_val_workers(), pin_memory=True) elif args.network == 'LocalSPFN': train_dataset = dataloaders.Dataset_TrainLocalSPFN(n_max_local_instances, csv_path_train, path_patches, noisy_train, first_n=first_n_train, fixed_order=False, lean=True) train_datasampler = dataloaders.RandomSampler(data_source=train_dataset, seed=12345, identical_epochs=False) train_dataloader = torch.utils.data.DataLoader(train_dataset, sampler=train_datasampler, batch_size=conf.get_batch_size(), num_workers=num_workers_train, pin_memory=True) val_dataset = dataloaders.Dataset_TrainLocalSPFN(n_max_local_instances, csv_path_val, path_patches, noisy_val, first_n=first_n_val, fixed_order=False, lean=True) val_datasampler = dataloaders.RandomSampler(data_source=val_dataset, seed=12345, identical_epochs=False) val_dataloader = torch.utils.data.DataLoader(val_dataset, sampler=val_datasampler, batch_size=conf.get_batch_size(), num_workers=conf.get_nb_val_workers(), pin_memory=True) # Optimizer optimizer = torch.optim.Adam(spfn_module.parameters(), lr=init_learning_rate) # Visualisation visualiser = training_visualisation.Visualiser(conf.get_visualisation_interval()) # Initialisation global_step = 0 old_learning_rate = init_learning_rate best_loss = np.inf for epoch in range(nb_epochs): global_step, _ = training_utils.spfn_train_val_epoch(train_dataloader, spfn_module, epoch, optimizer, global_step, visualiser, args, conf, device, network_mode='train') if (epoch % conf.get_val_interval() == 0) and (epoch > 0): with torch.no_grad(): _, loss = training_utils.spfn_train_val_epoch(val_dataloader, spfn_module, epoch, optimizer, global_step, visualiser, args, conf, device, network_mode='val') if loss < best_loss: torch.save(spfn_module.state_dict(), os.path.join(conf.get_weights_folder(), spfn_weights_filename + '.pth')) best_loss = loss if (epoch % conf.get_snapshot_interval() == 0) and (epoch > 0): torch.save(spfn_module.state_dict(), os.path.join(conf.get_weights_folder(), spfn_weights_filename + '%d.pth' % epoch)) torch.save(spfn_module.state_dict(), os.path.join(conf.get_weights_folder(), spfn_weights_filename + '%d.pth' % epoch))
6,960
59.530435
198
py
CPFN
CPFN-master/SPFN/sphere_fitter.py
# Importatiomn of packages import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from SPFN.primitives import Sphere from SPFN.geometry_utils import weighted_sphere_fitting, weighted_sphere_fitting_tensorflow def compute_parameters(P, W): batch_size, n_points, _ = P.size() _, _, n_max_primitives = W.size() P = P.unsqueeze(1).expand(batch_size, n_max_primitives, n_points, 3).contiguous() W = W.transpose(1, 2).contiguous() P = P.view(batch_size * n_max_primitives, n_points, 3) W = W.view(batch_size * n_max_primitives, n_points) center, radius_squared = weighted_sphere_fitting(P, W) center = center.view(batch_size, n_max_primitives, 3) radius_squared = radius_squared.view(batch_size, n_max_primitives) return center, radius_squared def compute_parameters_tensorflow(P, W): batch_size = tf.shape(P)[0] n_points = tf.shape(P)[1] n_max_primitives = tf.shape(W)[2] P = tf.tile(tf.expand_dims(P, axis=1), [1, n_max_primitives, 1, 1]) # BxKxNx3 W = tf.transpose(W, perm=[0, 2, 1]) # BxKxN P = tf.reshape(P, [batch_size * n_max_primitives, n_points, 3]) # BKxNx3 W = tf.reshape(W, [batch_size * n_max_primitives, n_points]) # BKxN center, radius_squared = weighted_sphere_fitting_tensorflow(P, W) center = tf.reshape(center, [batch_size, n_max_primitives, 3]) radius_squared = tf.reshape(radius_squared, [batch_size, n_max_primitives]) return center, radius_squared if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) center_torch, radius_squared_torch = compute_parameters(P_torch, W_torch) center_torch = center_torch.detach().cpu().numpy() radius_squared_torch = radius_squared_torch.detach().cpu().numpy() print('center_torch', center_torch) print('radius_squared_torch', radius_squared_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) center_tensorflow, radius_squared_tensorflow = compute_parameters_tensorflow(P_tensorflow, W_tensorflow) sess = tf.Session() center_tensorflow, radius_squared_tensorflow = sess.run([center_tensorflow, radius_squared_tensorflow]) print(np.abs(center_tensorflow-center_torch).max()) print(np.abs(radius_squared_tensorflow-radius_squared_torch).max()) def sqrt_safe(x): return torch.sqrt(torch.abs(x) + 1e-10) def compute_residue_single(center, radius_squared, p): return (sqrt_safe(torch.sum((p - center)**2, dim=-1)) - sqrt_safe(radius_squared))**2 def sqrt_safe_tensorflow(x): return tf.sqrt(tf.abs(x) + 1e-10) def compute_residue_single_tensorflow(center, radius_squared, p): return tf.square(sqrt_safe_tensorflow(tf.reduce_sum(tf.square(p - center), axis=-1)) - sqrt_safe_tensorflow(radius_squared)) if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) center = np.random.randn(batch_size, num_points, 3) radius_squared = np.random.rand(batch_size, num_points) p = np.random.rand(batch_size, num_points, 3) center_torch = torch.from_numpy(center).float().to(device) radius_squared_torch = torch.from_numpy(radius_squared).float().to(device) p_torch = torch.from_numpy(p).float().to(device) residue_loss_torch = compute_residue_single(center_torch, radius_squared_torch, p_torch) residue_loss_torch = residue_loss_torch.detach().cpu().numpy() print('residue_loss_torch', residue_loss_torch) # Debugging with Tensorflow center_tensorflow = tf.constant(center, dtype=tf.float32) radius_squared_tensorflow = tf.constant(radius_squared, dtype=tf.float32) p_tensorflow = tf.constant(p, dtype=tf.float32) residue_loss_torch_tensorflow = compute_residue_single_tensorflow(center_tensorflow, radius_squared_tensorflow, p_tensorflow) sess = tf.Session() residue_loss_torch_tensorflow = sess.run(residue_loss_torch_tensorflow) print(np.abs(residue_loss_torch_tensorflow - residue_loss_torch).max()) def create_primitive_from_dict(d): assert d['type'] == 'sphere' location = np.array([d['location_x'], d['location_y'], d['location_z']], dtype=float) radius = float(d['radius']) return Sphere(center=location, radius=radius) def extract_parameter_data_as_dict(primitives, n_max_primitives): return {} def extract_predicted_parameters_as_json(sphere_center, sphere_radius_squared, k): sphere = Sphere(sphere_center, np.sqrt(sphere_radius_squared)) return { 'type': 'sphere', 'center_x': float(sphere.center[0]), 'center_y': float(sphere.center[1]), 'center_z': float(sphere.center[2]), 'radius': float(sphere.radius), 'label': k, }
5,084
44.810811
129
py
CPFN
CPFN-master/SPFN/plane_fitter.py
# Importatiomn of packages import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from SPFN.primitives import Plane from SPFN.geometry_utils import weighted_plane_fitting, weighted_plane_fitting_tensorflow def compute_parameters(P, W): batch_size, n_points, _ = P.size() _, _, n_max_instances = W.size() W_reshaped = W.transpose(1, 2).contiguous().view(batch_size * n_max_instances, n_points) P_tiled = P.unsqueeze(1).expand(batch_size, n_max_instances, n_points, 3).contiguous().view(batch_size * n_max_instances, n_points, 3) n, c = weighted_plane_fitting(P_tiled, W_reshaped) # BKx3 n = n.view(batch_size, n_max_instances, 3) c = c.view(batch_size, n_max_instances) return n, c def compute_parameters_tensorflow(P, W): batch_size = tf.shape(P)[0] n_points = tf.shape(P)[1] n_max_instances = tf.shape(W)[2] W_reshaped = tf.reshape(tf.transpose(W, [0, 2, 1]), [batch_size * n_max_instances, n_points]) # BKxN P_tiled = tf.reshape(tf.tile(tf.expand_dims(P, axis=1), [1, n_max_instances, 1, 1]), [batch_size * n_max_instances, n_points, 3]) # BKxNx3, important there to match indices in W_reshaped!!! n, c = weighted_plane_fitting_tensorflow(P_tiled, W_reshaped) # BKx3 n = tf.reshape(n, [batch_size, n_max_instances, 3]) # BxKx3 c = tf.reshape(c, [batch_size, n_max_instances]) # BxK return n, c if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) n_torch, c_torch = compute_parameters(P_torch, W_torch) n_torch = n_torch.detach().cpu().numpy() c_torch = c_torch.detach().cpu().numpy() print('n_torch', n_torch) print('c_torch', c_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) n_tensorflow, c_tensorflow = compute_parameters_tensorflow(P_tensorflow, W_tensorflow) sess = tf.Session() n_tensorflow, c_tensorflow = sess.run([n_tensorflow, c_tensorflow]) print(np.minimum(np.abs(n_tensorflow - n_torch), np.abs(n_tensorflow + n_torch)).max()) print(np.minimum(np.abs(c_tensorflow - c_torch), np.abs(c_tensorflow + c_torch)).max()) def compute_residue_single(n, c, p): return (torch.sum(p * n, dim=-1) - c)**2 def compute_residue_single_tensorflow(n, c, p): # n: ...x3, c: ..., p: ...x3 return tf.square(tf.reduce_sum(p * n, axis=-1) - c) if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) n = np.random.randn(batch_size, num_points, 3) c = np.random.rand(batch_size, num_points) p = np.random.rand(batch_size, num_points, 3) n_torch = torch.from_numpy(n).float().to(device) c_torch = torch.from_numpy(c).float().to(device) p_torch = torch.from_numpy(p).float().to(device) residue_loss_torch = compute_residue_single(n_torch, c_torch, p_torch) residue_loss_torch = residue_loss_torch.detach().cpu().numpy() print('residue_loss_torch', residue_loss_torch) # Debugging with Tensorflow n_tensorflow = tf.constant(n, dtype=tf.float32) c_tensorflow = tf.constant(c, dtype=tf.float32) p_tensorflow = tf.constant(p, dtype=tf.float32) residue_loss_torch_tensorflow = compute_residue_single_tensorflow(n_tensorflow, c_tensorflow, p_tensorflow) sess = tf.Session() residue_loss_torch_tensorflow = sess.run(residue_loss_torch_tensorflow) print(np.abs(residue_loss_torch_tensorflow-residue_loss_torch).max()) def acos_safe(x): return torch.acos(torch.clamp(x, min=-1.0+1e-6, max=1.0-1e-6)) def compute_parameter_loss(predicted_n, gt_n, matching_indices, angle_diff): # predicted_axis: BxK1x3 # gt_axis: BXK2x3 # matching indices: BxK2 batch_size, nb_primitives, _ = gt_n.size() predicted_n = torch.gather(predicted_n, 1, matching_indices.unsqueeze(2).expand(batch_size, nb_primitives, 3)) dot_abs = torch.abs(torch.sum(predicted_n * gt_n, axis=2)) if angle_diff: return acos_safe(dot_abs) # BxK else: return 1.0 - dot_abs # BxK def batched_gather(data, indices, axis): # data - Bx...xKx..., axis is where dimension K is # indices - BxK # output[b, ..., k, ...] = in[b, ..., indices[b, k], ...] assert axis >= 1 ndims = data.get_shape().ndims # allow dynamic rank if axis > 1: # tranpose data to BxKx... perm = np.arange(ndims) perm[axis] = 1 perm[1] = axis data = tf.transpose(data, perm=perm) batch_size = tf.shape(data)[0] batch_nums = tf.tile(tf.expand_dims(tf.expand_dims(tf.range(batch_size), axis=1), axis=2), multiples=[1, tf.shape(indices)[1], 1]) # BxKx1 indices = tf.concat([batch_nums, tf.expand_dims(indices, axis=2)], axis=2) # BxKx2 gathered_data = tf.gather_nd(data, indices=indices) if axis > 1: gathered_data = tf.transpose(gathered_data, perm=perm) return gathered_data def acos_safe_tensorflow(x): return tf.math.acos(tf.clip_by_value(x, -1.0+1e-6, 1.0-1e-6)) def compute_parameter_loss_tensorflow(predicted_n, gt_n, matching_indices, angle_diff): n = batched_gather(predicted_n, matching_indices, axis=1) dot_abs = tf.abs(tf.reduce_sum(n * gt_n, axis=2)) if angle_diff: return acos_safe_tensorflow(dot_abs) # BxK else: return 1.0 - dot_abs # BxK if __name__ == '__main__': batch_size = 100 num_primitives1 = 15 num_primitives2 = 5 device = torch.device('cuda:0') np.random.seed(0) predicted_axis = np.random.randn(batch_size, num_primitives1, 3) gt_axis = np.random.randn(batch_size, num_primitives2, 3) matching_indices = np.random.randint(0, 15, (batch_size, num_primitives2)) angle_diff = True predicted_axis_torch = torch.from_numpy(predicted_axis).float().to(device) gt_axis_torch = torch.from_numpy(gt_axis).float().to(device) matching_indices_torch = torch.from_numpy(matching_indices).long().to(device) loss_torch = compute_parameter_loss(predicted_axis_torch, gt_axis_torch, matching_indices_torch, angle_diff) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow predicted_axis_tensorflow = tf.constant(predicted_axis, dtype=tf.float32) gt_axis_tensorflow = tf.constant(gt_axis, dtype=tf.float32) matching_indices_tensorflow = tf.constant(matching_indices, dtype=tf.int32) loss_tensorflow = compute_parameter_loss_tensorflow(predicted_axis_tensorflow, gt_axis_tensorflow, matching_indices_tensorflow, angle_diff) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) def create_primitive_from_dict(d): assert d['type'] == 'plane' location = np.array([d['location_x'], d['location_y'], d['location_z']], dtype=float) axis = np.array([d['axis_x'], d['axis_y'], d['axis_z']], dtype=float) return Plane(n=axis, c=np.dot(location, axis)) def extract_parameter_data_as_dict(primitives, n_max_instances): n = np.zeros(dtype=float, shape=[n_max_instances, 3]) for i, primitive in enumerate(primitives): if isinstance(primitive, Plane): n[i] = primitive.n return { 'plane_n_gt': n } def extract_predicted_parameters_as_json(plane_normal, plane_center, k): # This is only for a single plane plane = Plane(plane_normal, plane_center) json_info = { 'type': 'plane', 'center_x': float(plane.center[0]), 'center_y': float(plane.center[1]), 'center_z': float(plane.center[2]), 'normal_x': float(plane.n[0]), 'normal_y': float(plane.n[1]), 'normal_z': float(plane.n[2]), 'x_size': float(plane.x_range[1] - plane.x_range[0]), 'y_size': float(plane.y_range[1] - plane.y_range[0]), 'x_axis_x': float(plane.x_axis[0]), 'x_axis_y': float(plane.x_axis[1]), 'x_axis_z': float(plane.x_axis[2]), 'y_axis_x': float(plane.y_axis[0]), 'y_axis_y': float(plane.y_axis[1]), 'y_axis_z': float(plane.y_axis[2]), 'label': k, } return json_info
8,518
43.369792
194
py
CPFN
CPFN-master/SPFN/metric_implementation.py
# Importation of packages import torch import numpy as np from scipy.optimize import linear_sum_assignment from SPFN import plane_fitter, sphere_fitter, cylinder_fitter, cone_fitter from SPFN import losses_implementation def hungarian_matching(W_pred, I_gt): # This non-tf function does not backprob gradient, only output matching indices # W_pred - BxNxK # I_gt - BxN, may contain -1's # Output: matching_indices - BxK, where (b,k)th ground truth primitive is matched with (b, matching_indices[b, k]) # where only n_gt_labels entries on each row have meaning. The matching does not include gt background instance batch_size, n_points, n_max_labels = W_pred.size() matching_indices = torch.zeros([batch_size, n_max_labels], dtype=torch.long).to(W_pred.device) mask = torch.zeros([batch_size, n_max_labels], dtype=torch.bool).to(W_pred.device) for b in range(batch_size): # assuming I_gt[b] does not have gap n_gt_labels = torch.max(I_gt[b]).item() + 1 # this is K' W_gt = torch.eye(n_gt_labels+1).to(I_gt.device)[I_gt[b]] dot = torch.mm(W_gt.transpose(0,1), W_pred[b]) denominator = torch.sum(W_gt, dim=0).unsqueeze(1) + torch.sum(W_pred[b], dim=0).unsqueeze(0) - dot cost = dot / torch.clamp(denominator, min=1e-10, max=None) # K'xK cost = cost[:n_gt_labels, :] # remove last row, corresponding to matching gt background instance _, col_ind = linear_sum_assignment(-cost.detach().cpu().numpy()) # want max solution col_ind = torch.from_numpy(col_ind).long().to(matching_indices.device) matching_indices[b, :n_gt_labels] = col_ind mask[b, :n_gt_labels] = True return matching_indices, mask # Converting W to hard encoding def hard_W_encoding(W): # W - BxNxK _, _, num_labels = W.size() hardW = torch.eye(num_labels).to(W.device)[torch.argmax(W, dim=2)] return hardW if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) W_torch = torch.from_numpy(W).float().to(device) hardW = hard_W_encoding(W_torch) print('hardW', hardW.size()) # Getting the per instance type def get_instance_type(T, W): instance_type = torch.bmm(W.transpose(1,2), T) instance_type = torch.argmax(instance_type, dim=2) return instance_type if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 n_max_instances = 12 n_type = 4 device = torch.device('cuda:0') np.random.seed(0) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, n_type) T = T / np.linalg.norm(T, axis=2, keepdims=True) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) instance_type = get_instance_type(T_torch, W_torch) print('instance_type', instance_type.size()) def sqrt_safe(x): return torch.sqrt(torch.abs(x) + 1e-10) # Getting the residual loss def get_residual_loss(parameters, matching_indices, points_per_instance, T, classes=['plane','sphere','cylinder','cone']): batch_size, num_primitives, num_primitive_points, _ = points_per_instance.shape _, residue_per_point_array = losses_implementation.compute_residue_loss(parameters, matching_indices, points_per_instance, torch.gather(T, 1, matching_indices), classes=classes) residue_per_point_array = torch.gather(residue_per_point_array, 3, T.view(batch_size, num_primitives, 1, 1).expand(batch_size, num_primitives, num_primitive_points, 1)).squeeze(3) residual_loss = sqrt_safe(residue_per_point_array) return residual_loss if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) points_per_instance_torch = torch.from_numpy(points_per_instance).float().to(device) T_torch = torch.from_numpy(T).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) W_torch = hard_W_encoding(W_torch) parameters_torch = losses_implementation.compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch, _ = hungarian_matching(W_torch, I_gt_torch) T_torch = get_instance_type(T_torch, W_torch) residual_loss = get_residual_loss(parameters_torch, matching_indices_torch, points_per_instance_torch, T_torch, classes=['plane','sphere','cylinder','cone']) print('residual_loss', residual_loss.size()) # Arccos safe def acos_safe(x): return torch.acos(torch.clamp(x, min=-1.0+1e-6, max=1.0-1e-6)) # Segmentation mIoU def compute_segmentation_iou(W, I_gt, matching_indices, mask):# W - BxNxK mIoU = 1 - losses_implementation.compute_miou_loss(W, I_gt, matching_indices)[0] mIoU = torch.sum(mask * mIoU, dim=1) / torch.sum(mask, dim=1) return mIoU if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) W_torch = torch.from_numpy(W).float().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) W_torch = hard_W_encoding(W_torch) matching_indices_torch, mask_torch = hungarian_matching(W_torch, I_gt_torch) mIou = compute_segmentation_iou(W_torch, I_gt_torch, matching_indices_torch, mask_torch) print('mIou', mIou.size()) # Mean primitive type accuracy def compute_type_accuracy(T, T_gt, matching_indices, mask): T_reordered = torch.gather(T, 1, matching_indices) # BxNxK type_accuracy = torch.sum(mask*(T_reordered == T_gt), dim=1) / torch.sum(mask, dim=1) return type_accuracy if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) mask = np.random.randint(0, 2, (batch_size, n_max_instances)) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) mask_torch = torch.from_numpy(mask).float().to(device) W_torch = hard_W_encoding(W_torch) T_torch = get_instance_type(T_torch, W_torch) type_accuracy = compute_type_accuracy(T_torch, T_gt_torch, mask_torch) print('type_accuracy', type_accuracy.size()) # Mean point normal difference def compute_normal_difference(X, X_gt): normal_difference = torch.mean(acos_safe(torch.abs(torch.sum(X*X_gt, dim=2))), dim=1) return normal_difference if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) X_gt = np.random.randn(batch_size, num_points, 3) X_gt = X_gt / np.linalg.norm(X_gt, axis=2, keepdims=True) X_torch = torch.from_numpy(X).float().to(device) X_gt_torch = torch.from_numpy(X_gt).float().to(device) normal_difference = compute_normal_difference(X_gt_torch, X_gt_torch) print('normal_difference', normal_difference.size()) # Mean primitive axis difference def compute_axis_difference(predicted_parameters, gt_parameters, matching_indices, T, T_gt, mask, classes=['plane','sphere','cylinder','cone'], div_eps=1e-10): mask = mask * (T == T_gt).float() parameter_loss = losses_implementation.compute_parameter_loss(predicted_parameters, gt_parameters, matching_indices, T_gt, is_eval=True, classes=classes) axis_difference = torch.sum(mask * parameter_loss, dim=1) / torch.clamp(torch.sum(parameter_loss, dim=1), min=div_eps, max=None) return axis_difference if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) X_torch = torch.from_numpy(X).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to(device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy(gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} W_torch = hard_W_encoding(W_torch) predicted_parameters_torch = losses_implementation.compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch, mask_torch = hungarian_matching(W_torch, I_gt_torch) T_torch = get_instance_type(T_torch, W_torch) axis_difference = compute_axis_difference(predicted_parameters_torch, gt_parameters_torch, matching_indices_torch, T_torch, T_gt_torch, mask_torch, classes=['plane', 'sphere', 'cylinder', 'cone']) print('axis_difference', axis_difference.size()) # Mean/Std Sk residual def compute_meanstd_Sk_residual(residue_loss, mask): mean_residual = torch.sum(mask * torch.mean(residue_loss, dim=2), dim=1) / torch.sum(mask, dim=1) std_residual = torch.sum(mask * torch.std(residue_loss, dim=2), dim=1) / torch.sum(mask, dim=1) return mean_residual, std_residual if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) X_torch = torch.from_numpy(X).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to( device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy( gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} points_per_instance_torch = torch.from_numpy(points_per_instance).float().to(device) W_torch = hard_W_encoding(W_torch) predicted_parameters_torch = losses_implementation.compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch, mask_torch = hungarian_matching(W_torch, I_gt_torch) T_torch = get_instance_type(T_torch, W_torch) residue_loss_torch = get_residual_loss(predicted_parameters_torch, matching_indices_torch, points_per_instance_torch, T_torch, classes=['plane', 'sphere', 'cylinder', 'cone']) mean_residual, std_residual = compute_meanstd_Sk_residual(residue_loss_torch, mask_torch) print('Mean Sk Residual Loss: ', mean_residual) print('Std Sk Residual Loss: ', std_residual) # Sk coverage def compute_Sk_coverage(residue_loss, epsilon, mask): residue_loss = torch.mean((residue_loss < epsilon).float(), dim=2) Sk_coverage = torch.sum(mask * residue_loss, dim=1) / torch.sum(mask, dim=1) return Sk_coverage if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) epsilon = 0.01 P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) X_torch = torch.from_numpy(X).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to( device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy( gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} points_per_instance_torch = torch.from_numpy(points_per_instance).float().to(device) W_torch = hard_W_encoding(W_torch) predicted_parameters_torch = losses_implementation.compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch, mask_torch = hungarian_matching(W_torch, I_gt_torch) T_torch = get_instance_type(T_torch, W_torch) residue_loss_torch = get_residual_loss(predicted_parameters_torch, matching_indices_torch, points_per_instance_torch, T_torch, classes=['plane', 'sphere', 'cylinder', 'cone']) Sk_coverage = compute_Sk_coverage(residue_loss_torch, epsilon, mask_torch) print('Sk Coverage : ', Sk_coverage) # P coverage def compute_P_coverage(P, T, matching_indices, predicted_parameters, epsilon, classes=['plane', 'sphere', 'cylinder', 'cone']): batch_size, num_points, _ = P.size() _, num_primitives = T.size() residue_loss = get_residual_loss(predicted_parameters, matching_indices, P.unsqueeze(1).expand(batch_size, num_primitives, num_points, 3), torch.gather(T, 1, matching_indices), classes=classes) residue_loss, _ = torch.min(residue_loss, dim=1) P_coverage = torch.mean((residue_loss < epsilon).float(), dim=1) return P_coverage if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) epsilon = 0.01 P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) X_torch = torch.from_numpy(X).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to( device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy( gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} points_per_instance_torch = torch.from_numpy(points_per_instance).float().to(device) W_torch = hard_W_encoding(W_torch) predicted_parameters_torch = losses_implementation.compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch, mask_torch = hungarian_matching(W_torch, I_gt_torch) T_torch = get_instance_type(T_torch, W_torch) P_coverage = compute_P_coverage(P_torch, T_torch, matching_indices_torch, predicted_parameters_torch, classes=['plane', 'sphere', 'cylinder', 'cone']) print('P Coverage : ', P_coverage) def compute_all_metrics(P, X, X_gt, W, I_gt, T, T_gt, points_per_instance, gt_parameters, list_epsilon=[0.01, 0.02], classes=['plane', 'sphere', 'cylinder', 'cone']): W = hard_W_encoding(W) T = get_instance_type(T, W) diff = T.size(1) - T_gt.size(1) if diff>0: T_gt = torch.cat((T_gt, torch.zeros_like(T_gt[:, 0:1]).expand(-1, diff)), dim=1) elif diff < 0: W = torch.cat((W, torch.zeros_like(W[:,:,0:1]).expand(-1, -1, -diff)), dim=2) T = torch.cat((T, torch.zeros_like(T[:, 0:1]).expand(-1, -diff)), dim=1) matching_indices, mask = hungarian_matching(W, I_gt) mask = mask.float() mIoU = compute_segmentation_iou(W, I_gt, matching_indices, mask) type_accuracy = compute_type_accuracy(T, T_gt, matching_indices, mask) normal_difference = compute_normal_difference(X, X_gt) predicted_parameters = losses_implementation.compute_parameters(P, W, X) if diff > 0: gt_parameters['plane_normal'] = torch.cat((gt_parameters['plane_normal'], torch.zeros_like(gt_parameters['plane_normal'][:, 0:1]).expand(-1, diff, 3)), dim=1) gt_parameters['cylinder_axis'] = torch.cat((gt_parameters['cylinder_axis'], torch.zeros_like(gt_parameters['cylinder_axis'][:, 0:1]).expand(-1, diff, 3)), dim=1) gt_parameters['cone_axis'] = torch.cat((gt_parameters['cone_axis'], torch.zeros_like(gt_parameters['cone_axis'][:, 0:1]).expand(-1, diff, 3)), dim=1) points_per_instance = torch.cat((points_per_instance, torch.zeros_like(points_per_instance[:,0:1]).expand(-1, diff, 512, 3)), dim=1) axis_difference = compute_axis_difference(predicted_parameters, gt_parameters, matching_indices, T, T_gt, mask, classes=classes) residue_loss = get_residual_loss(predicted_parameters, matching_indices, points_per_instance, T_gt, classes=classes) mean_residual, std_residual = compute_meanstd_Sk_residual(residue_loss, mask) Sk_coverage = [] for epsilon in list_epsilon: Sk_coverage.append(compute_Sk_coverage(residue_loss, epsilon, mask)) P_coverage = [] for epsilon in list_epsilon: P_coverage.append(compute_P_coverage(P, T, matching_indices, predicted_parameters, epsilon, classes=classes)) return mIoU, type_accuracy, normal_difference, axis_difference, mean_residual, std_residual, Sk_coverage, P_coverage, W, predicted_parameters, T if __name__ == '__main__' and 1: batch_size = 8 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) W = W / np.linalg.norm(W, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T = T / np.linalg.norm(T, axis=2, keepdims=True) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) X_gt = np.random.randn(batch_size, num_points, 3) X_gt = X_gt / np.linalg.norm(X_gt, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) T_torch = torch.from_numpy(T).float().to(device) X_torch = torch.from_numpy(X).float().to(device) X_gt_torch = torch.from_numpy(X_gt).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to( device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy( gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} points_per_instance_torch = torch.from_numpy(points_per_instance).float().to(device) mIoU, type_accuracy, normal_difference, axis_difference, mean_residual, std_residual, Sk_coverage, P_coverage = compute_all_metrics(P_torch, X_torch, X_gt_torch, W_torch, I_gt_torch, T_torch, T_gt_torch, points_per_instance_torch, gt_parameters_torch, classes=['plane', 'sphere', 'cylinder', 'cone']) print('mIoU', mIoU.size()) print('type_accuracy', type_accuracy.size()) print('normal_difference', normal_difference.size()) print('axis_difference', axis_difference.size()) print('mean_residual', mean_residual.size()) print('std_residual', std_residual.size()) for i in range(len(Sk_coverage)): print('Sk_coverage_%d'%i, Sk_coverage[i].size()) for i in range(len(P_coverage)): print('P_coverage_%d'%i, P_coverage[i].size()) def creates_json(T, predicted_parameters): list_json = [] for i, type_id in enumerate(T): if type_id == 0: json = plane_fitter.extract_predicted_parameters_as_json(predicted_parameters['plane_normal'][0,i].cpu().numpy(), predicted_parameters['plane_center'][0,i].cpu().numpy(), i) elif type_id == 1: json = sphere_fitter.extract_predicted_parameters_as_json(predicted_parameters['sphere_center'][0,i].cpu().numpy(), predicted_parameters['sphere_radius_squared'][0,i].cpu().numpy(), i) elif type_id == 2: json = cylinder_fitter.extract_predicted_parameters_as_json(predicted_parameters['cylinder_center'][0,i].cpu().numpy(), predicted_parameters['cylinder_radius_squared'][0,i].cpu().numpy(), predicted_parameters['cylinder_axis'][0,i].cpu().numpy(), i) elif type_id == 3: json = cone_fitter.extract_predicted_parameters_as_json(predicted_parameters['cone_apex'][0,i].cpu().numpy(), predicted_parameters['cone_axis'][0,i].cpu().numpy(), predicted_parameters['cone_half_angle'][0,i].cpu().numpy(), i) list_json.append(json) return list_json
36,132
59.121464
304
py
CPFN
CPFN-master/SPFN/geometry_utils.py
# Importing packages import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from SPFN.differentiable_tls import solve_weighted_tls, solve_weighted_tls_tensorflow def compute_consistent_plane_frame(normal): # Input: normal is Bx3 # Returns: x_axis, y_axis, both of dimension Bx3 device = normal.get_device() batch_size, _ = normal.size() candidate_axes = [[1, 0, 0], [0, 1, 0], [0, 0, 1]] # Actually, 2 should be enough. This may still cause singularity TODO!!! y_axes = [] for tmp_axis in candidate_axes: torch_axis = torch.FloatTensor(tmp_axis).to(device).unsqueeze(0) y_axes.append(torch.cross(normal, torch_axis.expand(batch_size, 3))) y_axes = torch.stack(y_axes, dim=0) # QxBx3 y_axes_norm = torch.norm(y_axes, dim=2) # QxB # choose the axis with largest norm y_axes_chosen_idx = torch.argmax(y_axes_norm, dim=0) # B y_axes_chosen_idx = y_axes_chosen_idx.view(1, batch_size, 1).expand(1, batch_size, 3) # y_axes_chosen[b, :] = y_axes[y_axes_chosen_idx[b], b, :] y_axes = torch.gather(y_axes, 0, y_axes_chosen_idx).squeeze(0) y_axes = torch.nn.functional.normalize(y_axes, p=2, dim=1, eps=1e-12) x_axes = torch.cross(y_axes, normal) # Bx3 return x_axes, y_axes def compute_consistent_plane_frame_tensorflow(normal): # Input: normal is Bx3 # Returns: x_axis, y_axis, both of dimension Bx3 batch_size = tf.shape(normal)[0] candidate_axes = [[1, 0, 0], [0, 1, 0], [0, 0, 1]] # Actually, 2 should be enough. This may still cause singularity TODO!!! y_axes = [] for tmp_axis in candidate_axes: tf_axis = tf.tile(tf.expand_dims(tf.constant(dtype=tf.float32, value=tmp_axis), axis=0), [batch_size, 1]) # Bx3 y_axes.append(tf.cross(normal, tf_axis)) y_axes = tf.stack(y_axes, axis=0) # QxBx3 y_axes_norm = tf.norm(y_axes, axis=2) # QxB # choose the axis with largest norm y_axes_chosen_idx = tf.argmax(y_axes_norm, axis=0) # B # y_axes_chosen[b, :] = y_axes[y_axes_chosen_idx[b], b, :] indices_0 = tf.tile(tf.expand_dims(y_axes_chosen_idx, axis=1), [1, 3]) # Bx3 indices_1 = tf.tile(tf.expand_dims(tf.range(batch_size), axis=1), [1, 3]) # Bx3 indices_2 = tf.tile(tf.expand_dims(tf.range(3), axis=0), [batch_size, 1]) # Bx3 indices = tf.stack([tf.cast(indices_0, tf.int32), indices_1, indices_2], axis=2) # Bx3x3 y_axes = tf.gather_nd(y_axes, indices=indices) # Bx3 if tf.VERSION == '1.4.1': y_axes = tf.nn.l2_normalize(y_axes, dim=1) else: y_axes = tf.nn.l2_normalize(y_axes, axis=1) x_axes = tf.cross(y_axes, normal) # Bx3 return x_axes, y_axes if __name__ == '__main__': batch_size = 100 device = torch.device('cuda:0') np.random.seed(0) normal = np.random.randn(batch_size, 3) normal = normal / np.linalg.norm(normal, axis=1, keepdims=True) normal_torch = torch.from_numpy(normal).float().to(device) x_axes_torch, y_axes_torch = compute_consistent_plane_frame(normal_torch) x_axes_torch = x_axes_torch.detach().cpu().numpy() y_axes_torch = y_axes_torch.detach().cpu().numpy() print('x_axes_torch', x_axes_torch) print('y_axes_torch', y_axes_torch) # Debugging with Tensorflow normal_tensorflow = tf.constant(normal, dtype=tf.float32) x_axes_tensorflow, y_axes_tensorflow = compute_consistent_plane_frame_tensorflow(normal_tensorflow) sess = tf.Session() x_axes_tensorflow, y_axes_tensorflow = sess.run([x_axes_tensorflow, y_axes_tensorflow]) print(np.max(np.abs(x_axes_tensorflow-x_axes_torch))) def weighted_plane_fitting(P, W, division_eps=1e-10): # P - BxNx3 # W - BxN # Returns n, c, with n - Bx3, c - B WP = P * W.unsqueeze(2) # BxNx3 W_sum = torch.sum(W, dim=1, keepdim=True) # Bx1 P_weighted_mean = torch.sum(WP, dim=1) / torch.clamp(W_sum, min=division_eps, max=None) # Bx3 A = P - P_weighted_mean.unsqueeze(1) # BxNx3 n = solve_weighted_tls(A, W) # Bx3 c = torch.sum(n * P_weighted_mean, dim=1) return n, c def weighted_plane_fitting_tensorflow(P, W, division_eps=1e-10): # P - BxNx3 # W - BxN # Returns n, c, with n - Bx3, c - B WP = P * tf.expand_dims(W, axis=2) # BxNx3 W_sum = tf.reduce_sum(W, axis=1) # B P_weighted_mean = tf.reduce_sum(WP, axis=1) / tf.maximum(tf.expand_dims(W_sum, 1), division_eps) # Bx3 A = P - tf.expand_dims(P_weighted_mean, axis=1) # BxNx3 n = solve_weighted_tls_tensorflow(A, W) # Bx3 c = tf.reduce_sum(n * P_weighted_mean, axis=1) return n, c if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) n_torch, c_torch = weighted_plane_fitting(P_torch, W_torch) n_torch = n_torch.detach().cpu().numpy() c_torch = c_torch.detach().cpu().numpy() print('n_torch', n_torch) print('c_torch', c_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) n_tensorflow, c_tensorflow = weighted_plane_fitting_tensorflow(P_tensorflow, W_tensorflow) sess = tf.Session() n_tensorflow, c_tensorflow = sess.run([n_tensorflow, c_tensorflow]) print(np.minimum(np.abs(n_tensorflow - n_torch), np.abs(n_tensorflow + n_torch)).max()) print(np.minimum(np.abs(c_tensorflow - c_torch), np.abs(c_tensorflow + c_torch)).max()) def guarded_matrix_solve_ls(A, b, W, condition_number_cap=1e5, sqrt_eps=1e-10, ls_l2_regularizer=1e-8): # Solve weighted least square ||\sqrt(W)(Ax-b)||^2 # A - BxNxD # b - BxNx1 # W - BxN batch_size, _, dim = A.size() sqrt_W = torch.sqrt(torch.clamp(W, min=sqrt_eps, max=None)).unsqueeze(2) # BxN A = A * sqrt_W # BxNxD b = b * sqrt_W # BxNx1 # Compute singular value, trivializing the problem when condition number is too large AtA = torch.bmm(A.transpose(1,2), A) _, s, _ = torch.svd(AtA, compute_uv=False) # s will be BxD s = s.detach() mask = s[:,0] / s[:,-1] < condition_number_cap # B #import pdb; pdb.set_trace() #A = A * mask.float().view(batch_size, 1, 1) #x = torch.linalg.lstsq(A, b).solution AtA = AtA * mask.float().view(batch_size, 1, 1) + ls_l2_regularizer * torch.eye(dim).unsqueeze(0).to(A.device) # zero out badly conditioned data Atb = torch.bmm(A.transpose(1, 2) * mask.float().view(batch_size, 1, 1), b) x, _ = torch.solve(Atb, AtA) x = x.squeeze(2) return x # BxD if __name__ == '__main__': sqrt_eps = 1e-10 ls_l2_regularizer = 1e-8 batch_size = 100 num_points = 1024 dimension = 3 device = torch.device('cuda:0') np.random.seed(0) A = np.random.randn(batch_size, num_points, dimension) b = np.random.randn(batch_size, num_points, 1) W = np.random.rand(batch_size, num_points) A = torch.from_numpy(A).float().to(device) b = torch.from_numpy(b).float().to(device) W = torch.from_numpy(W).float().to(device) sqrt_W = torch.sqrt(torch.clamp(W, sqrt_eps)).unsqueeze(2) # BxN A = A * sqrt_W # BxNxD b = b * sqrt_W # BxNx1 AtA = torch.bmm(A.transpose(1, 2), A) mask = torch.zeros([batch_size]).float().to(A.device) # B AtA = AtA * mask.view(batch_size, 1, 1) + ls_l2_regularizer * torch.eye(dimension).unsqueeze(0).to(device) # zero out badly conditioned data Atb = torch.bmm(A.transpose(1, 2) * mask.view(batch_size, 1, 1), b) x = torch.solve(Atb, AtA) def guarded_matrix_solve_ls_tensorflow(A, b, W, condition_number_cap=1e5, sqrt_eps=1e-10, ls_l2_regularizer=1e-8): # Solve weighted least square ||\sqrt(W)(Ax-b)||^2 # A - BxNxD # b - BxNx1 # W - BxN sqrt_W = tf.sqrt(tf.maximum(W, sqrt_eps)) # BxN A *= tf.expand_dims(sqrt_W, axis=2) # BxNxD b *= tf.expand_dims(sqrt_W, axis=2) # BxNx1 # Compute singular value, trivializing the problem when condition number is too large AtA = tf.matmul(a=A, b=A, transpose_a=True) s, _, _ = [tf.stop_gradient(u) for u in tf.svd(AtA)] # s will be BxD mask = tf.less(s[:, 0] / s[:, -1], condition_number_cap) # B A *= tf.to_float(tf.expand_dims(tf.expand_dims(mask, axis=1), axis=2)) # zero out badly conditioned data x = tf.matrix_solve_ls(A, b, l2_regularizer=ls_l2_regularizer, fast=True) # BxDx1 return tf.squeeze(x, axis=2) # BxD if __name__ == '__main__': batch_size = 100 num_points = 1024 dimension = 3 device = torch.device('cuda:0') np.random.seed(0) A = np.random.randn(batch_size, num_points, dimension) b = np.random.randn(batch_size, num_points, 1) W = np.random.rand(batch_size, num_points) A_torch = torch.from_numpy(A).float().to(device) b_torch = torch.from_numpy(b).float().to(device) W_torch = torch.from_numpy(W).float().to(device) x_torch = guarded_matrix_solve_ls(A_torch, b_torch, W_torch) x_torch = x_torch.detach().cpu().numpy() print('x_torch', x_torch) # Debugging with Tensorflow A_tensorflow = tf.constant(A, dtype=tf.float32) b_tensorflow = tf.constant(b, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) x_tensorflow = guarded_matrix_solve_ls_tensorflow(A_tensorflow, b_tensorflow, W_tensorflow) sess = tf.Session() x_tensorflow = sess.run(x_tensorflow) print(np.max(np.abs(x_tensorflow - x_torch))) def weighted_sphere_fitting(P, W, division_eps=1e-10): # P - BxNxD # W - BxN W_sum = torch.sum(W, axis=1) # B WP_sqr_sum = torch.sum(W * torch.sum(P**2, axis=2), axis=1) # B P_sqr = torch.sum(P**2, axis=2) # BxN b = ((WP_sqr_sum / torch.clamp(W_sum, min=division_eps, max=None)).unsqueeze(1) - P_sqr).unsqueeze(2) # BxNx1 WP_sum = torch.sum(W.unsqueeze(2) * P, dim=1) # BxD A = 2 * ((WP_sum / torch.clamp(W_sum, min=division_eps, max=None).unsqueeze(1)).unsqueeze(1) - P) # BxNxD # Seek least norm solution to the least square center = guarded_matrix_solve_ls(A, b, W) # BxD W_P_minus_C_sqr_sum = P - center.unsqueeze(1) # BxNxD W_P_minus_C_sqr_sum = W * torch.sum(W_P_minus_C_sqr_sum**2, dim=2) # BxN r_sqr = torch.sum(W_P_minus_C_sqr_sum, dim=1) / torch.clamp(W_sum, min=division_eps, max=None) # B return center, r_sqr def weighted_sphere_fitting_tensorflow(P, W, division_eps=1e-10): # P - BxNxD # W - BxN W_sum = tf.reduce_sum(W, axis=1) # B WP_sqr_sum = tf.reduce_sum(W * tf.reduce_sum(tf.square(P), axis=2), axis=1) # B P_sqr = tf.reduce_sum(tf.square(P), axis=2) # BxN b = tf.expand_dims(tf.expand_dims(WP_sqr_sum / tf.maximum(W_sum, division_eps), axis=1) - P_sqr, axis=2) # BxNx1 WP_sum = tf.reduce_sum(tf.expand_dims(W, axis=2) * P, axis=1) # BxD A = 2 * (tf.expand_dims(WP_sum / tf.expand_dims(tf.maximum(W_sum, division_eps), axis=1), axis=1) - P) # BxNxD # Seek least norm solution to the least square center = guarded_matrix_solve_ls_tensorflow(A, b, W) # BxD W_P_minus_C_sqr_sum = P - tf.expand_dims(center, axis=1) # BxNxD W_P_minus_C_sqr_sum = W * tf.reduce_sum(tf.square(W_P_minus_C_sqr_sum), axis=2) # BxN r_sqr = tf.reduce_sum(W_P_minus_C_sqr_sum, axis=1) / tf.maximum(W_sum, division_eps) # B return center, r_sqr if __name__ == '__main__': batch_size = 100 num_points = 1024 dimension = 3 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, dimension) W = np.random.rand(batch_size, num_points) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) center_torch, r_sqr_torch = weighted_sphere_fitting(P_torch, W_torch) center_torch = center_torch.detach().cpu().numpy() r_sqr_torch = r_sqr_torch.detach().cpu().numpy() print('center_torch', center_torch) print('r_sqr_torch', r_sqr_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) center_tensorflow, r_sqr_tensorflow = weighted_sphere_fitting_tensorflow(P_tensorflow, W_tensorflow) sess = tf.Session() center_tensorflow, r_sqr_tensorflow = sess.run([center_tensorflow, r_sqr_tensorflow]) print(np.max(np.abs(center_tensorflow - center_torch))) print(np.max(np.abs(r_sqr_tensorflow - r_sqr_torch)))
12,490
46.494297
148
py
CPFN
CPFN-master/SPFN/differentiable_tls.py
# Importation of packages import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from torch.autograd import gradcheck def guard_one_over_matrix(M, min_abs_value=1e-10): _, row, _ = M.size() device = M.get_device() up = torch.triu(torch.clamp(M, min=min_abs_value, max=None), diagonal=0) low = torch.tril(torch.clamp(M, min=None, max=-min_abs_value), diagonal=0) M = up + low M = M + torch.eye(row).to(device) M = 1 / M M = M - torch.eye(row).to(device) return M def guard_one_over_matrix_tensorflow(M, min_abs_value=1e-10): up = tf.matrix_band_part(tf.maximum(min_abs_value, M), 0, -1) low = tf.matrix_band_part(tf.minimum(-min_abs_value, M), -1, 0) M = up + low M += tf.eye(tf.shape(M)[1]) M = 1 / M M -= tf.eye(tf.shape(M)[1]) return M if __name__ == '__main__': batch_size = 100 P = 5 device = torch.device('cuda:0') np.random.seed(0) M = np.random.randn(batch_size, P, P) M_torch = torch.from_numpy(M).float().to(device) M_torch = guard_one_over_matrix(M_torch) M_torch = M_torch.detach().cpu().numpy() print('M_torch', M_torch) # Debugging with Tensorflow M_tensorflow = tf.constant(M, dtype=tf.float32) M_tensorflow_ = guard_one_over_matrix_tensorflow(M_tensorflow) sess = tf.Session() M_tensorflow = sess.run(M_tensorflow_) print(np.max(np.abs(M_tensorflow - M_torch))) def compute_svd_K(s): # s should be BxP # res[b,i,j] = 1/(s[b,i]^2 - s[b,j]^2) if i != j, 0 otherwise # res will be BxPxP s = s**2 res = s.unsqueeze(2) - s.unsqueeze(1) # making absolute value in res is at least 1e-10 res = guard_one_over_matrix(res) return res def compute_svd_K_tensorflow(s): # s should be BxP # res[b,i,j] = 1/(s[b,i]^2 - s[b,j]^2) if i != j, 0 otherwise # res will be BxPxP s = tf.square(s) res = tf.expand_dims(s, 2) - tf.expand_dims(s, 1) # making absolute value in res is at least 1e-10 res = guard_one_over_matrix_tensorflow(res) return res if __name__ == '__main__': batch_size = 100 P = 5 device = torch.device('cuda:0') np.random.seed(0) s = np.random.randn(batch_size, P) s_torch = torch.from_numpy(s).float().to(device) res_torch = compute_svd_K(s_torch) res_torch = res_torch.detach().cpu().numpy() print('res_torch', res_torch) # Debugging with Tensorflow s_tensorflow = tf.constant(s, dtype=tf.float32) res_tensorflow = compute_svd_K_tensorflow(s_tensorflow) sess = tf.Session() res_tensorflow = sess.run(res_tensorflow) print(np.max(np.abs(res_tensorflow - res_torch))) def custom_svd_v_column_tensorflow(M, col_index=-1): # Must make sure M is finite. Otherwise cudaSolver might fail. assert_op = tf.Assert(tf.logical_not(tf.reduce_any(tf.logical_not(tf.is_finite(M)))), [M], summarize=10) with tf.control_dependencies([assert_op]): with tf.get_default_graph().gradient_override_map({'Svd': 'CustomSvd'}): s, u, v = tf.svd(M, name='Svd') # M = usv^T return v[:, :, col_index] def register_custom_svd_gradient_tensorflow(): tf.RegisterGradient('CustomSvd')(custom_gradient_svd_tensorflow) def custom_gradient_svd_tensorflow(op, grad_s, grad_u, grad_v): s, u, v = op.outputs # s - BxP # u - BxNxP, N >= P # v - BxPxP v_t = tf.transpose(v, [0, 2, 1]) K = compute_svd_K_tensorflow(s) inner = tf.transpose(K, [0, 2, 1]) * tf.matmul(v_t, grad_v) inner = (inner + tf.transpose(inner, [0, 2, 1])) / 2 # ignoring gradient coming from grad_s and grad_u for our purpose res = tf.matmul(u, tf.matmul(2 * tf.matmul(tf.matrix_diag(s), inner), v_t)) return res if __name__ == '__main__' and 1: register_custom_svd_gradient_tensorflow() batch_size = 100 P = 5 device = torch.device('cuda:0') np.random.seed(0) M = np.random.randn(batch_size, P, P) M_tensorflow = tf.constant(M, dtype=tf.float32) M_input = tf.placeholder(dtype=tf.float32, shape=[None, P, P]) with tf.get_default_graph().gradient_override_map({'Svd': 'CustomSvd'}): s, u, v = tf.svd(M_input, name='Svd') # M = usv^T with tf.Session() as sess: error = tf.test.compute_gradient_error(M_input, [batch_size, P, P], v, [batch_size, P, P]) print('Error: ', error) class Custom_svd_v_colum(torch.autograd.Function): @staticmethod def forward(ctx, M, col_index=-1): u, s, v = torch.svd(M, some=True) out = v[:,:,col_index] ctx.save_for_backward(u, s, v) ctx.col_index = col_index return out @staticmethod def backward(ctx, grad_out): u, s, v = ctx.saved_tensors col_index = ctx.col_index grad_v = torch.zeros_like(v) grad_v[:,:,col_index] = grad_out v_t = v.transpose(1, 2) K = compute_svd_K(s) inner = K.transpose(1,2) * torch.bmm(v_t, grad_v) inner = (inner + inner.transpose(1, 2)) / 2 # ignoring gradient coming from grad_s and grad_u for our purpose res = torch.bmm(u, torch.bmm(2 * torch.bmm(torch.diag_embed(s, offset=0, dim1=-2, dim2=-1), inner), v_t)) return res, None if __name__ == '__main__': batch_size = 100 P = 5 device = torch.device('cuda:0') np.random.seed(0) M = np.random.randn(batch_size, P, P) M_torch = torch.from_numpy(M).float().to(device) out_torch = Custom_svd_v_colum().apply(M_torch) out_torch = out_torch.detach().cpu().numpy() print('out_torch', out_torch) # Debugging with Tensorflow M_tensorflow = tf.constant(M, dtype=tf.float32) out_tensorflow = custom_svd_v_column_tensorflow(M_tensorflow) sess = tf.Session() out_tensorflow = sess.run(out_tensorflow) print(np.minimum(np.abs(out_tensorflow-out_torch), np.abs(out_tensorflow+out_torch)).max()) if __name__ == '__main__' and 1: batch_size = 4 P = 5 device = torch.device('cuda:0') np.random.seed(0) M = np.random.randn(batch_size, P, P) M_torch = torch.from_numpy(M).float().to(device) M_torch = torch.nn.Parameter(M_torch, requires_grad=True) try: custom_svd_v_colum = Custom_svd_v_colum.apply torch.autograd.gradcheck(custom_svd_v_colum, (M_torch, -1), raise_exception=True) print('Test on Custom_svd_v_colum: Success') except: print('Test on Custom_svd_v_colum: Failure') raise if __name__ == '__main__' and 1: register_custom_svd_gradient_tensorflow() batch_size = 100 P = 5 device = torch.device('cuda:0') np.random.seed(0) M = np.random.randn(batch_size, P, P) M_torch = torch.from_numpy(M).float().to(device) M_torch = torch.nn.Parameter(M_torch, requires_grad=True) out = Custom_svd_v_colum().apply(M_torch) out.backward(torch.ones_like(out)) M_grad_torch = M_torch.grad.detach().cpu().numpy() M_tensorflow = tf.constant(M, dtype=tf.float32) out = custom_svd_v_column_tensorflow(M_tensorflow) M_grad_tensorflow = tf.gradients(out, [M_tensorflow])[0] sess = tf.Session() M_grad_tensorflow = sess.run(M_grad_tensorflow) print(np.minimum(np.abs(M_grad_tensorflow - M_grad_torch), np.abs(M_grad_tensorflow + M_grad_torch)).max()) def solve_weighted_tls(A, W): # A - BxNx3 # W - BxN, positive weights # Find solution to min x^T A^T diag(W) A x = min ||\sqrt{diag(W)} A x||^2, subject to ||x|| = 1 batch_size, num_points, _ = A.size() A_p = A.unsqueeze(2) * A.unsqueeze(3) # BxNx3x3 W_p = W.view(batch_size, num_points, 1, 1) M = torch.sum(W_p * A_p, dim=1) # Bx3x3 x = Custom_svd_v_colum().apply(M) # Bx3 return x def solve_weighted_tls_tensorflow(A, W): # A - BxNx3 # W - BxN, positive weights # Find solution to min x^T A^T diag(W) A x = min ||\sqrt{diag(W)} A x||^2, subject to ||x|| = 1 A_p = tf.expand_dims(A, axis=2) * tf.expand_dims(A, axis=3) # BxNx3x3 W_p = tf.expand_dims(tf.expand_dims(W, axis=2), axis=3) # BxNx1x1 M = tf.reduce_sum(W_p * A_p, axis=1) # Bx3x3 x = custom_svd_v_column_tensorflow(M) # Bx3 return x if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) A = np.random.randn(batch_size, num_points, 3) W = np.random.randn(batch_size, num_points) A_torch = torch.from_numpy(A).float().to(device) W_torch = torch.from_numpy(W).float().to(device) x_torch = solve_weighted_tls(A_torch, W_torch) x_torch = x_torch.detach().cpu().numpy() print('x_torch', x_torch) # Debugging with Tensorflow A_tensorflow = tf.constant(A, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) x_tensorflow = solve_weighted_tls_tensorflow(A_tensorflow, W_tensorflow) sess = tf.Session() x_tensorflow = sess.run(x_tensorflow) print(np.minimum(np.abs(x_tensorflow-x_torch), np.abs(x_tensorflow+x_torch)).max())
8,960
36.493724
113
py
CPFN
CPFN-master/SPFN/cylinder_fitter.py
# Importation of pqckqges import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from SPFN.primitives import Cylinder from SPFN.differentiable_tls import solve_weighted_tls, solve_weighted_tls_tensorflow from SPFN.geometry_utils import compute_consistent_plane_frame, compute_consistent_plane_frame_tensorflow, weighted_sphere_fitting, weighted_sphere_fitting_tensorflow def compute_parameters(P, W, X): # First determine n as the solution to \min \sum W_i (X_i \cdot n)^2 batch_size, n_points, _ = P.size() _, _, n_max_primitives = W.size() W_reshaped = W.transpose(1,2).contiguous().view(batch_size * n_max_primitives, n_points) # BKxN X_reshaped = X.unsqueeze(1).expand(batch_size, n_max_primitives, n_points, 3).contiguous().view(batch_size * n_max_primitives, n_points, 3) n = solve_weighted_tls(X_reshaped, W_reshaped).view(batch_size, n_max_primitives, 3) # BxKx3 x_axes, y_axes = compute_consistent_plane_frame(n.view(batch_size * n_max_primitives, 3)) x_axes = x_axes.view(batch_size, n_max_primitives, 3) # BxKx3 y_axes = y_axes.view(batch_size, n_max_primitives, 3) # BxKx3 x_coord = torch.sum(P.unsqueeze(1) * x_axes.unsqueeze(2), dim=3) # BxKxN y_coord = torch.sum(P.unsqueeze(1) * y_axes.unsqueeze(2), dim=3) # BxKxN P_proj = torch.stack([x_coord, y_coord], dim=3) # BxKxNx2, 2D projection point P_proj_reshaped = P_proj.view(batch_size * n_max_primitives, n_points, 2) # BKxNx2 circle_center, circle_radius_squared = weighted_sphere_fitting(P_proj_reshaped, W_reshaped) circle_center = circle_center.view(batch_size, n_max_primitives, 2) # BxKx2 center = circle_center[:,:,0].unsqueeze(2) * x_axes + circle_center[:,:,1].unsqueeze(2) * y_axes # BxKx3 radius_square = circle_radius_squared.view(batch_size, n_max_primitives) # BxK return n, center, radius_square def compute_parameters_tensorflow(P, W, X): # First determine n as the solution to \min \sum W_i (X_i \cdot n)^2 batch_size = tf.shape(P)[0] n_points = tf.shape(P)[1] n_max_primitives = tf.shape(W)[2] W_reshaped = tf.reshape(tf.transpose(W, [0, 2, 1]), [batch_size * n_max_primitives, n_points]) # BKxN X_reshaped = tf.reshape(tf.tile(tf.expand_dims(X, axis=1), [1, n_max_primitives, 1, 1]), [batch_size * n_max_primitives, n_points, 3]) # BKxNx3 n = tf.reshape(solve_weighted_tls_tensorflow(X_reshaped, W_reshaped), [batch_size, n_max_primitives, 3]) # BxKx3 x_axes, y_axes = compute_consistent_plane_frame_tensorflow(tf.reshape(n, [batch_size * n_max_primitives, 3])) x_axes = tf.reshape(x_axes, [batch_size, n_max_primitives, 3]) # BxKx3 y_axes = tf.reshape(y_axes, [batch_size, n_max_primitives, 3]) # BxKx3 x_coord = tf.reduce_sum(tf.expand_dims(P, axis=1) * tf.expand_dims(x_axes, axis=2), axis=3) # BxKxN y_coord = tf.reduce_sum(tf.expand_dims(P, axis=1) * tf.expand_dims(y_axes, axis=2), axis=3) # BxKxN P_proj = tf.stack([x_coord, y_coord], axis=3) # BxKxNx2, 2D projection point P_proj_reshaped = tf.reshape(P_proj, [batch_size * n_max_primitives, n_points, 2]) # BKxNx2 circle_center, circle_radius_squared = weighted_sphere_fitting_tensorflow(P_proj_reshaped, W_reshaped) circle_center = tf.reshape(circle_center, [batch_size, n_max_primitives, 2]) # BxKx2 center = tf.expand_dims(circle_center[:, :, 0], axis=2) * x_axes + tf.expand_dims(circle_center[:, :, 1], axis=2) * y_axes # BxKx3 radius_square = tf.reshape(circle_radius_squared, [batch_size, n_max_primitives]) # BxK return n, center, radius_square if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) n_torch, center_torch, radius_square_torch = compute_parameters(P_torch, W_torch, X_torch) n_torch = n_torch.detach().cpu().numpy() center_torch = center_torch.detach().cpu().numpy() radius_square_torch = radius_square_torch.detach().cpu().numpy() print('n_torch', n_torch) print('center_torch', center_torch) print('radius_square_torch', radius_square_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) X_tensorflow = tf.constant(X, dtype=tf.float32) n_tensorflow, center_tensorflow, radius_square_tensorflow = compute_parameters_tensorflow(P_tensorflow, W_tensorflow, X_tensorflow) sess = tf.Session() n_tensorflow, center_tensorflow, radius_square_tensorflow = sess.run([n_tensorflow, center_tensorflow, radius_square_tensorflow]) print(np.minimum(np.abs(n_tensorflow - n_torch), np.abs(n_tensorflow + n_torch)).max()) print(np.abs(center_tensorflow - center_torch).max()) print(np.abs(radius_square_tensorflow - radius_square_torch).max()) def sqrt_safe(x): return torch.sqrt(torch.abs(x) + 1e-10) def compute_residue_single(axis, center, radius_squared, p): p_minus_c = p - center p_minus_c_sqr = torch.sum(p_minus_c**2, dim=-1) p_minus_c_dot_n = torch.sum(p_minus_c * axis, dim=-1) return (sqrt_safe(p_minus_c_sqr - p_minus_c_dot_n**2) - sqrt_safe(radius_squared))**2 def sqrt_safe_tensorflow(x): return tf.sqrt(tf.abs(x) + 1e-10) def compute_residue_single_tensorflow(axis, center, radius_squared, p): p_minus_c = p - center p_minus_c_sqr = tf.reduce_sum(tf.square(p_minus_c), axis=-1) p_minus_c_dot_n = tf.reduce_sum(p_minus_c * axis, axis=-1) return tf.square(sqrt_safe_tensorflow(p_minus_c_sqr - tf.square(p_minus_c_dot_n)) - sqrt_safe_tensorflow(radius_squared)) if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) axis = np.random.randn(batch_size, num_points, 3) center = np.random.randn(batch_size, num_points, 3) radius_squared = np.random.randn(batch_size, num_points) p = np.random.randn(batch_size, num_points, 3) axis_torch = torch.from_numpy(axis).float().to(device) center_torch = torch.from_numpy(center).float().to(device) radius_squared_torch = torch.from_numpy(radius_squared).float().to(device) p_torch = torch.from_numpy(p).float().to(device) loss_torch = compute_residue_single(axis_torch, center_torch, radius_squared_torch, p_torch) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow axis_tensorflow = tf.constant(axis, dtype=tf.float32) center_tensorflow = tf.constant(center, dtype=tf.float32) radius_squared_tensorflow = tf.constant(radius_squared, dtype=tf.float32) p_tensorflow = tf.constant(p, dtype=tf.float32) loss_tensorflow = compute_residue_single_tensorflow(axis_tensorflow, center_tensorflow, radius_squared_tensorflow, p_tensorflow) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) def acos_safe(x): return torch.acos(torch.clamp(x, min=-1.0+1e-6, max=1.0-1e-6)) def compute_parameter_loss(predicted_axis, gt_axis, matching_indices, angle_diff): # predicted_axis: BxK1x3 # gt_axis: BXK2x3 # matching indices: BxK2 batch_size, nb_primitives, _ = gt_axis.size() predicted_axis = torch.gather(predicted_axis, 1, matching_indices.unsqueeze(2).expand(batch_size, nb_primitives, 3)) dot_abs = torch.abs(torch.sum(predicted_axis * gt_axis, axis=2)) if angle_diff: return acos_safe(dot_abs) # BxK else: return 1.0 - dot_abs # BxK def batched_gather(data, indices, axis): # data - Bx...xKx..., axis is where dimension K is # indices - BxK # output[b, ..., k, ...] = in[b, ..., indices[b, k], ...] assert axis >= 1 ndims = data.get_shape().ndims # allow dynamic rank if axis > 1: # tranpose data to BxKx... perm = np.arange(ndims) perm[axis] = 1 perm[1] = axis data = tf.transpose(data, perm=perm) batch_size = tf.shape(data)[0] batch_nums = tf.tile(tf.expand_dims(tf.expand_dims(tf.range(batch_size), axis=1), axis=2), multiples=[1, tf.shape(indices)[1], 1]) # BxKx1 indices = tf.concat([batch_nums, tf.expand_dims(indices, axis=2)], axis=2) # BxKx2 gathered_data = tf.gather_nd(data, indices=indices) if axis > 1: gathered_data = tf.transpose(gathered_data, perm=perm) return gathered_data def acos_safe_tensorflow(x): return tf.math.acos(tf.clip_by_value(x, -1.0+1e-6, 1.0-1e-6)) def compute_parameter_loss_tensorflow(predicted_axis, gt_axis, matching_indices, angle_diff): n = batched_gather(predicted_axis, matching_indices, axis=1) dot_abs = tf.abs(tf.reduce_sum(n * gt_axis, axis=2)) if angle_diff: return acos_safe_tensorflow(dot_abs) # BxK else: return 1.0 - dot_abs # BxK if __name__ == '__main__': batch_size = 100 num_primitives1 = 15 num_primitives2 = 5 device = torch.device('cuda:0') np.random.seed(0) predicted_axis = np.random.randn(batch_size, num_primitives1, 3) gt_axis = np.random.randn(batch_size, num_primitives2, 3) matching_indices = np.random.randint(0, 15, (batch_size, num_primitives2)) angle_diff = True predicted_axis_torch = torch.from_numpy(predicted_axis).float().to(device) gt_axis_torch = torch.from_numpy(gt_axis).float().to(device) matching_indices_torch = torch.from_numpy(matching_indices).long().to(device) loss_torch = compute_parameter_loss(predicted_axis_torch, gt_axis_torch, matching_indices_torch, angle_diff) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow predicted_axis_tensorflow = tf.constant(predicted_axis, dtype=tf.float32) gt_axis_tensorflow = tf.constant(gt_axis, dtype=tf.float32) matching_indices_tensorflow = tf.constant(matching_indices, dtype=tf.int32) loss_tensorflow = compute_parameter_loss_tensorflow(predicted_axis_tensorflow, gt_axis_tensorflow, matching_indices_tensorflow, angle_diff) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) def create_primitive_from_dict(d): assert d['type'] == 'cylinder' location = np.array([d['location_x'], d['location_y'], d['location_z']], dtype=float) axis = np.array([d['axis_x'], d['axis_y'], d['axis_z']], dtype=float) radius = float(d['radius']) return Cylinder(center=location, radius=radius, axis=axis) def extract_parameter_data_as_dict(primitives, n_max_primitives): n = np.zeros(dtype=float, shape=[n_max_primitives, 3]) for i, primitive in enumerate(primitives): if isinstance(primitive, Cylinder): n[i] = primitive.axis return { 'cylinder_axis_gt': n } def extract_predicted_parameters_as_json(cylinder_center, cylinder_radius_squared, cylinder_axis, k): cylinder = Cylinder(cylinder_center, np.sqrt(cylinder_radius_squared), cylinder_axis, height=5) return { 'type': 'cylinder', 'center_x': float(cylinder.center[0]), 'center_y': float(cylinder.center[1]), 'center_z': float(cylinder.center[2]), 'radius': float(cylinder.radius), 'axis_x': float(cylinder.axis[0]), 'axis_y': float(cylinder.axis[1]), 'axis_z': float(cylinder.axis[2]), 'height': float(cylinder.height), 'label': k, }
11,804
51.234513
166
py
CPFN
CPFN-master/SPFN/losses_implementation.py
# Importation of packages import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from scipy.optimize import linear_sum_assignment from SPFN import plane_fitter, sphere_fitter, cylinder_fitter, cone_fitter # Segmentation Loss def hungarian_matching(W_pred, I_gt): # This non-tf function does not backprob gradient, only output matching indices # W_pred - BxNxK # I_gt - BxN, may contain -1's # Output: matching_indices - BxK, where (b,k)th ground truth primitive is matched with (b, matching_indices[b, k]) # where only n_gt_labels entries on each row have meaning. The matching does not include gt background instance batch_size, n_points, n_max_labels = W_pred.size() matching_indices = torch.zeros([batch_size, n_max_labels], dtype=torch.long).to(W_pred.device) for b in range(batch_size): # assuming I_gt[b] does not have gap n_gt_labels = torch.max(I_gt[b]).item() + 1 # this is K' W_gt = torch.eye(n_gt_labels+1).to(I_gt.device)[I_gt[b]] dot = torch.mm(W_gt.transpose(0,1), W_pred[b]) denominator = torch.sum(W_gt, dim=0).unsqueeze(1) + torch.sum(W_pred[b], dim=0).unsqueeze(0) - dot cost = dot / torch.clamp(denominator, min=1e-10, max=None) # K'xK cost = cost[:n_gt_labels, :] # remove last row, corresponding to matching gt background instance _, col_ind = linear_sum_assignment(-cost.detach().cpu().numpy()) # want max solution col_ind = torch.from_numpy(col_ind).long().to(matching_indices.device) matching_indices[b, :n_gt_labels] = col_ind return matching_indices def hungarian_matching_tensorflow(W_pred, I_gt): # This non-tf function does not backprob gradient, only output matching indices # W_pred - BxNxK # I_gt - BxN, may contain -1's # Output: matching_indices - BxK, where (b,k)th ground truth primitive is matched with (b, matching_indices[b, k]) # where only n_gt_labels entries on each row have meaning. The matching does not include gt background instance batch_size = I_gt.shape[0] n_points = I_gt.shape[1] n_max_labels = W_pred.shape[2] matching_indices = np.zeros([batch_size, n_max_labels], dtype=np.int32) for b in range(batch_size): # assuming I_gt[b] does not have gap n_gt_labels = np.max(I_gt[b]) + 1 # this is K' W_gt = np.zeros([n_points, n_gt_labels + 1]) # HACK: add an extra column to contain -1's W_gt[np.arange(n_points), I_gt[b]] = 1.0 # NxK' dot = np.sum(np.expand_dims(W_gt, axis=2) * np.expand_dims(W_pred[b], axis=1), axis=0) # K'xK denominator = np.expand_dims(np.sum(W_gt, axis=0), axis=1) + np.expand_dims(np.sum(W_pred[b], axis=0), axis=0) - dot cost = dot / np.maximum(denominator, 1e-10) # K'xK cost = cost[:n_gt_labels, :] # remove last row, corresponding to matching gt background instance _, col_ind = linear_sum_assignment(-cost) # want max solution matching_indices[b, :n_gt_labels] = col_ind return matching_indices if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) W_pred = np.random.rand(batch_size, num_points, n_max_instances) I_gt = np.random.randint(-1, n_max_instances, (batch_size, num_points)) W_pred = W_pred / np.linalg.norm(W_pred, axis=2, keepdims=True) W_pred_torch = torch.from_numpy(W_pred).float().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) matching_indices_torch = hungarian_matching(W_pred_torch, I_gt_torch) matching_indices_torch = matching_indices_torch.detach().cpu().numpy() print('matching_indices_torch', matching_indices_torch) # Debugging with Tensorflow W_pred_tensorflow = tf.constant(W_pred, dtype=tf.float32) I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) matching_indices_tensorflow = tf.py_func(hungarian_matching_tensorflow, [W_pred_tensorflow, I_gt_tensorflow], Tout=tf.int32) sess = tf.Session() matching_indices_tensorflow = sess.run(matching_indices_tensorflow) print(np.abs(matching_indices_torch - matching_indices_tensorflow).max()) def compute_miou_loss(W, I_gt, matching_indices, div_eps=1e-10): # W - BxNxK # I_gt - BxN batch_size, n_points, n_max_labels = W.size() _, n_labels = matching_indices.size() W_reordered = torch.gather(W, 2, matching_indices.unsqueeze(1).expand(batch_size, n_points, n_labels)) # BxNxK # notice in tf.one_hot, -1 will result in a zero row, which is what we want W_gt = torch.eye(n_labels+2).to(I_gt.device)[I_gt] W_gt = W_gt[:,:,:n_labels] dot = torch.sum(W_gt * W_reordered, axis=1) # BxK denominator = torch.sum(W_gt, dim=1) + torch.sum(W_reordered, dim=1) - dot mIoU = dot / (denominator + div_eps) # BxK return 1.0 - mIoU, 1 - dot / n_points def batched_gather_tensorflow(data, indices, axis): # data - Bx...xKx..., axis is where dimension K is # indices - BxK # output[b, ..., k, ...] = in[b, ..., indices[b, k], ...] assert axis >= 1 ndims = data.get_shape().ndims # allow dynamic rank if axis > 1: # tranpose data to BxKx... perm = np.arange(ndims) perm[axis] = 1 perm[1] = axis data = tf.transpose(data, perm=perm) batch_size = tf.shape(data)[0] batch_nums = tf.tile(tf.expand_dims(tf.expand_dims(tf.range(batch_size), axis=1), axis=2), multiples=[1, tf.shape(indices)[1], 1]) # BxKx1 indices = tf.concat([batch_nums, tf.expand_dims(indices, axis=2)], axis=2) # BxKx2 gathered_data = tf.gather_nd(data, indices=indices) if axis > 1: gathered_data = tf.transpose(gathered_data, perm=perm) return gathered_data def compute_miou_loss_tensorflow(W, I_gt, matching_indices): # W - BxNxK # I_gt - BxN W_reordered = batched_gather_tensorflow(W, indices=matching_indices, axis=2) # BxNxK depth = tf.shape(W)[2] # notice in tf.one_hot, -1 will result in a zero row, which is what we want W_gt = tf.one_hot(I_gt, depth=depth, dtype=tf.float32) # BxNxK dot = tf.reduce_sum(W_gt * W_reordered, axis=1) # BxK denominator = tf.reduce_sum(W_gt, axis=1) + tf.reduce_sum(W_reordered, axis=1) - dot mIoU = dot / (denominator + 1e-10) # BxK return 1.0 - mIoU if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) W = np.random.rand(batch_size, num_points, n_max_instances) I_gt = np.random.randint(-1, n_max_instances, (batch_size, num_points)) W = W / np.linalg.norm(W, axis=2, keepdims=True) W_torch = torch.from_numpy(W).float().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) with torch.no_grad(): matching_indices_torch = hungarian_matching(W_torch, I_gt_torch) loss_torch, _ = compute_miou_loss(W_torch, I_gt_torch, matching_indices_torch) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow W_tensorflow = tf.constant(W, dtype=tf.float32) I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) matching_indices_tensorflow = tf.stop_gradient(tf.py_func(hungarian_matching_tensorflow, [W_tensorflow, I_gt_tensorflow], Tout=tf.int32)) loss_tensorflow = compute_miou_loss_tensorflow(W_tensorflow, I_gt_tensorflow, matching_indices_tensorflow) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) # Normal Loss def acos_safe(x): return torch.acos(torch.clamp(x, min=-1.0+1e-6, max=1.0-1e-6)) def compute_normal_loss(normal, normal_gt, angle_diff): # normal, normal_gt: BxNx3 # Assume normals are unoriented dot_abs = torch.abs(torch.sum(normal * normal_gt, dim=2)) # BxN if angle_diff: return torch.mean(acos_safe(dot_abs), dim=1) else: return torch.mean(1.0 - dot_abs, dim=1) def acos_safe_tensorflow(x): return tf.math.acos(tf.clip_by_value(x, -1.0+1e-6, 1.0-1e-6)) def compute_normal_loss_tensorflow(normal, normal_gt, angle_diff): # normal, normal_gt: BxNx3 # Assume normals are unoriented dot_abs = tf.abs(tf.reduce_sum(normal * normal_gt, axis=2)) # BxN if angle_diff: return tf.reduce_mean(acos_safe_tensorflow(dot_abs), axis=1) else: return tf.reduce_mean(1.0 - dot_abs, axis=1) if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) normal = np.random.randn(batch_size, num_points, 3) normal_gt = np.random.randn(batch_size, num_points, 3) angle_diff = True normal_torch = torch.from_numpy(normal).float().to(device) normal_gt_torch = torch.from_numpy(normal_gt).float().to(device) loss_torch = compute_normal_loss(normal_torch, normal_gt_torch, angle_diff) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow normal_tensorflow = tf.constant(normal, dtype=tf.float32) normal_gt_tensorflow = tf.constant(normal_gt, dtype=tf.float32) loss_tensorflow = compute_normal_loss_tensorflow(normal_tensorflow, normal_gt_tensorflow, angle_diff) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) # Type Loss def compute_per_point_type_loss(per_point_type, I_gt, T_gt, is_eval): # For training, per_point_type is BxNxQ, where Q = n_registered_primitives # For test, per_point_type is BxN # I_gt - BxN, allow -1 # T_gt - BxK batch_size, n_points = I_gt.size() per_point_type_gt = torch.gather(T_gt, 1, torch.clamp(I_gt, min=0, max=None)) if is_eval: type_loss = 1.0 - (per_point_type == per_point_type_gt).float() else: type_loss = torch.nn.functional.cross_entropy(per_point_type.contiguous().view(batch_size*n_points, -1), per_point_type_gt.view(batch_size*n_points), reduction='none') # BxN type_loss = type_loss.view(batch_size, n_points) # do not add loss to background points in gt type_loss = torch.where(I_gt == -1, torch.zeros_like(type_loss), type_loss) return torch.sum(type_loss, dim=1) / (torch.sum((I_gt != -1).float(), dim=1).float()) # B def compute_per_point_type_loss_tensorflow(per_point_type, I_gt, T_gt, is_eval): # For training, per_point_type is BxNxQ, where Q = n_registered_primitives # For test, per_point_type is BxN # I_gt - BxN, allow -1 # T_gt - BxK batch_size = tf.shape(I_gt)[0] n_points = tf.shape(I_gt)[1] indices_0 = tf.tile(tf.expand_dims(tf.range(batch_size), axis=1), [1, n_points]) # BxN indices = tf.stack([indices_0, tf.maximum(0, I_gt)], axis=2) per_point_type_gt = tf.gather_nd(T_gt, indices=indices) # BxN if is_eval: type_loss = 1.0 - tf.to_float(tf.equal(per_point_type, per_point_type_gt)) else: type_loss = tf.nn.sparse_softmax_cross_entropy_with_logits(logits=per_point_type, labels=per_point_type_gt) # BxN # do not add loss to background points in gt type_loss = tf.where(tf.equal(I_gt, -1), tf.zeros_like(type_loss), type_loss) return tf.reduce_sum(type_loss, axis=1) / tf.to_float(tf.count_nonzero(tf.not_equal(I_gt, -1), axis=1)) # B if __name__ == '__main__': batch_size = 100 num_points = 1024 Q = 4 K = 10 device = torch.device('cuda:0') np.random.seed(0) per_point_type = np.random.randn(batch_size, num_points, Q) I_gt = np.random.randint(-1, K, (batch_size, num_points)) T_gt = np.random.randint(0, Q, (batch_size, K)) is_eval = False per_point_type_torch = torch.from_numpy(per_point_type).float().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) loss_torch = compute_per_point_type_loss(per_point_type_torch, I_gt_torch, T_gt_torch, is_eval) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow per_point_type_tensorflow = tf.constant(per_point_type, dtype=tf.float32) I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) T_gt_tensorflow = tf.constant(T_gt, dtype=tf.int32) loss_tensorflow = compute_per_point_type_loss_tensorflow(per_point_type_tensorflow, I_gt_tensorflow, T_gt_tensorflow, is_eval) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) def compute_parameters(P, W, X, classes=['plane','sphere','cylinder','cone']): parameters = {} for class_ in classes: if class_ == 'plane': plane_normal, plane_center = plane_fitter.compute_parameters(P, W) parameters['plane_normal'] = plane_normal parameters['plane_center'] = plane_center elif class_ == 'sphere': sphere_center, sphere_radius_squared = sphere_fitter.compute_parameters(P, W) parameters['sphere_center'] = sphere_center parameters['sphere_radius_squared'] = sphere_radius_squared elif class_ == 'cylinder': cylinder_axis, cylinder_center, cylinder_radius_squared = cylinder_fitter.compute_parameters(P, W, X) parameters['cylinder_axis'] = cylinder_axis parameters['cylinder_center'] = cylinder_center parameters['cylinder_radius_squared'] = cylinder_radius_squared elif class_ == 'cone': cone_apex, cone_axis, cone_half_angle = cone_fitter.compute_parameters(P, W, X) parameters['cone_apex'] = cone_apex parameters['cone_axis'] = cone_axis parameters['cone_half_angle'] = cone_half_angle else: raise NotImplementedError return parameters def compute_parameters_tensorflow(P, W, X, classes=['plane','sphere','cylinder','cone']): parameters = {} for class_ in classes: if class_ == 'plane': plane_normal, plane_center = plane_fitter.compute_parameters_tensorflow(P, W) parameters['plane_normal'] = plane_normal parameters['plane_center'] = plane_center elif class_ == 'sphere': sphere_center, sphere_radius_squared = sphere_fitter.compute_parameters_tensorflow(P, W) parameters['sphere_center'] = sphere_center parameters['sphere_radius_squared'] = sphere_radius_squared elif class_ == 'cylinder': cylinder_axis, cylinder_center, cylinder_radius_squared = cylinder_fitter.compute_parameters_tensorflow(P, W, X) parameters['cylinder_axis'] = cylinder_axis parameters['cylinder_center'] = cylinder_center parameters['cylinder_radius_squared'] = cylinder_radius_squared elif class_ == 'cone': cone_apex, cone_axis, cone_half_angle = cone_fitter.compute_parameters_tensorflow(P, W, X) parameters['cone_apex'] = cone_apex parameters['cone_axis'] = cone_axis parameters['cone_half_angle'] = cone_half_angle else: raise NotImplementedError return parameters if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) parameters = compute_parameters(P_torch, W_torch, X_torch) plane_normal_torch, plane_center_torch, sphere_center_torch, sphere_radius_squared_torch, cylinder_axis_torch, cylinder_center_torch, cylinder_radius_square_torch, cone_apex_torch, cone_axis_torch, cone_half_angle_torch = \ parameters['plane_normal'], parameters['plane_center'], parameters['sphere_center'], parameters['sphere_radius_squared'], parameters['cylinder_axis'], parameters['cylinder_center'], parameters['cylinder_radius_square'], parameters['cone_apex'] ,parameters['cone_axis'], parameters['cone_half_angle'] plane_normal_torch = plane_normal_torch.detach().cpu().numpy() plane_center_torch = plane_center_torch.detach().cpu().numpy() sphere_center_torch = sphere_center_torch.detach().cpu().numpy() sphere_radius_squared_torch = sphere_radius_squared_torch.detach().cpu().numpy() cylinder_axis_torch = cylinder_axis_torch.detach().cpu().numpy() cylinder_center_torch = cylinder_center_torch.detach().cpu().numpy() cylinder_radius_square_torch = cylinder_radius_square_torch.detach().cpu().numpy() cone_apex_torch = cone_apex_torch.detach().cpu().numpy() cone_axis_torch = cone_axis_torch.detach().cpu().numpy() cone_half_angle_torch = cone_half_angle_torch.detach().cpu().numpy() # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) X_tensorflow = tf.constant(X, dtype=tf.float32) parameters = compute_parameters_tensorflow(P_tensorflow, W_tensorflow, X_tensorflow) sess = tf.Session() plane_normal_tensorflow, plane_center_tensorflow, sphere_center_tensorflow, sphere_radius_squared_tensorflow, cylinder_axis_tensorflow, cylinder_center_tensorflow, cylinder_radius_square_tensorflow, cone_apex_tensorflow, cone_axis_tensorflow, cone_half_angle_tensorflow = \ sess.run([parameters['plane_normal'], parameters['plane_center'], parameters['sphere_center'], parameters['sphere_radius_squared'], parameters['cylinder_axis'], parameters['cylinder_center'], parameters['cylinder_radius_square'], parameters['cone_apex'] ,parameters['cone_axis'], parameters['cone_half_angle']]) print(np.minimum(np.abs(plane_normal_tensorflow - plane_normal_torch), np.abs(plane_normal_tensorflow + plane_normal_torch)).max()) print(np.minimum(np.abs(plane_center_tensorflow - plane_center_torch), np.abs(plane_center_tensorflow + plane_center_torch)).max()) print(np.abs(sphere_center_tensorflow - sphere_center_torch).max()) print(np.abs(sphere_radius_squared_tensorflow - sphere_radius_squared_torch).max()) print(np.minimum(np.abs(cylinder_axis_tensorflow - cylinder_axis_torch), np.abs(cylinder_axis_tensorflow + cylinder_axis_torch)).max()) print(np.abs(cylinder_center_tensorflow - cylinder_center_torch).max()) print(np.abs(cylinder_radius_square_tensorflow - cylinder_radius_square_torch).max()) print(np.abs(cone_apex_tensorflow - cone_apex_torch).max()) print(np.minimum(np.abs(cone_axis_tensorflow - cone_axis_torch), np.abs(cone_axis_tensorflow + cone_axis_torch)).max()) print(np.abs(cone_half_angle_tensorflow - cone_half_angle_torch).max()) # Residue Loss def compute_residue_loss(parameters, matching_indices, points_per_instance, T_gt, classes=['plane','sphere','cylinder','cone']): # parameters is a dictionary where each key represents a different parameter # points_per_instance of size BxKxN'x3 residue_losses = [] # a length T array of BxK tensors residue_per_point_array = [] # a length T array of BxKxN' tensors #residue_per_class = [] batch_size, n_labels = matching_indices.size() for class_ in classes: if class_ == 'plane': residue_per_point = plane_fitter.compute_residue_single(torch.gather(parameters['plane_normal'], 1, matching_indices.unsqueeze(2).expand(batch_size, n_labels, 3)).unsqueeze(2), torch.gather(parameters['plane_center'], 1, matching_indices).unsqueeze(2), points_per_instance) elif class_ == 'sphere': residue_per_point = sphere_fitter.compute_residue_single(torch.gather(parameters['sphere_center'], 1, matching_indices.unsqueeze(2).expand(batch_size, n_labels, 3)).unsqueeze(2), torch.gather(parameters['sphere_radius_squared'], 1, matching_indices.expand(batch_size, n_labels)).unsqueeze(2), points_per_instance) elif class_ == 'cylinder': residue_per_point = cylinder_fitter.compute_residue_single(torch.gather(parameters['cylinder_axis'], 1, matching_indices.unsqueeze(2).expand(batch_size, n_labels, 3)).unsqueeze(2), torch.gather(parameters['cylinder_center'], 1, matching_indices.unsqueeze(2).expand(batch_size, n_labels, 3)).unsqueeze(2), torch.gather(parameters['cylinder_radius_squared'], 1, matching_indices.expand(batch_size, n_labels)).unsqueeze(2), points_per_instance) elif class_ == 'cone': residue_per_point = cone_fitter.compute_residue_single(torch.gather(parameters['cone_apex'], 1, matching_indices.unsqueeze(2).expand(batch_size, n_labels, 3)).unsqueeze(2), torch.gather(parameters['cone_axis'], 1, matching_indices.unsqueeze(2).expand(batch_size, n_labels, 3)).unsqueeze(2), torch.gather(parameters['cone_half_angle'], 1, matching_indices.expand(batch_size, n_labels)).unsqueeze(2), points_per_instance) else: raise NotImplementedError #residue_per_class.append(residue_per_point) residue_per_point_array.append(residue_per_point) residue_losses.append(torch.mean(residue_per_point, dim=2)) residue_losses = torch.stack(residue_losses, dim=2) residue_loss = torch.gather(residue_losses, 2, T_gt.unsqueeze(2)).squeeze(2) residue_per_point_array = torch.stack(residue_per_point_array, dim=3) # BxKxN'xT return residue_loss, residue_per_point_array#, residue_per_class def aggregate_loss_from_stacked_tensorflow(loss_stacked, T_gt): # loss_stacked - BxKxT, T_gt - BxK # out[b, k] = loss_stacked[b, k, T_gt[b, k]] B = tf.shape(loss_stacked)[0] K = tf.shape(loss_stacked)[1] indices_0 = tf.tile(tf.expand_dims(tf.range(B), axis=1), multiples=[1, K]) # BxK indices_1 = tf.tile(tf.expand_dims(tf.range(K), axis=0), multiples=[B, 1]) # BxK indices = tf.stack([indices_0, indices_1, T_gt], axis=2) # BxKx3 return tf.gather_nd(loss_stacked, indices=indices) def compute_residue_loss_tensorflow(parameters, matching_indices, points_per_instance, T_gt, classes=['plane','sphere','cylinder','cone']): residue_losses = [] # a length T array of BxK tensors residue_per_point_array = [] # a length T array of BxKxN' tensors #residue_per_class = [] for class_ in classes: if class_ == 'plane': residue_per_point = plane_fitter.compute_residue_single_tensorflow(tf.expand_dims(batched_gather_tensorflow(parameters['plane_normal'], matching_indices, axis=1), axis=2), tf.expand_dims(batched_gather_tensorflow(parameters['plane_center'], matching_indices, axis=1), axis=2), points_per_instance) elif class_ == 'sphere': residue_per_point = sphere_fitter.compute_residue_single_tensorflow(tf.expand_dims(batched_gather_tensorflow(parameters['sphere_center'], matching_indices, axis=1), axis=2), tf.expand_dims(batched_gather_tensorflow(parameters['sphere_radius_squared'], matching_indices, axis=1), axis=2), points_per_instance) elif class_ == 'cylinder': residue_per_point = cylinder_fitter.compute_residue_single_tensorflow(tf.expand_dims(batched_gather_tensorflow(parameters['cylinder_axis'], matching_indices, axis=1), axis=2), tf.expand_dims(batched_gather_tensorflow(parameters['cylinder_center'], matching_indices, axis=1), axis=2), tf.expand_dims(batched_gather_tensorflow(parameters['cylinder_radius_squared'], matching_indices, axis=1), axis=2), points_per_instance) elif class_ == 'cone': residue_per_point = cone_fitter.compute_residue_single_tensorflow(tf.expand_dims(batched_gather_tensorflow(parameters['cone_apex'], matching_indices, axis=1), axis=2), tf.expand_dims(batched_gather_tensorflow(parameters['cone_axis'], matching_indices, axis=1), axis=2), tf.expand_dims(batched_gather_tensorflow(parameters['cone_half_angle'], matching_indices, axis=1), axis=2), points_per_instance) else: raise NotImplementedError #residue_per_class.append(residue_per_point) residue_per_point_array.append(residue_per_point) residue_losses.append(tf.reduce_mean(residue_per_point, axis=2)) residue_losses = tf.stack(residue_losses, axis=2) residue_per_point_array = tf.stack(residue_per_point_array, axis=3) # BxKxN'xT # Aggregate losses across fitters residue_loss = aggregate_loss_from_stacked_tensorflow(residue_losses, T_gt) # BxK return residue_loss, residue_per_point_array#, residue_per_class if __name__ == '__main__': batch_size = 100 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) points_per_instance_torch = torch.from_numpy(points_per_instance).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) parameters_torch = compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch = hungarian_matching(W_torch, I_gt_torch) residue_loss_torch, residue_per_point_array_torch, residue_per_class_torch = compute_residue_loss(parameters_torch, matching_indices_torch, points_per_instance_torch, T_gt_torch, classes=['plane', 'sphere', 'cylinder', 'cone']) residue_loss_torch = residue_loss_torch.detach().cpu().numpy() residue_per_point_array_torch = residue_per_point_array_torch.detach().cpu().numpy() residue_per_class_torch = [elt.detach().cpu().numpy() for elt in residue_per_class_torch] print('residue_loss_torch', residue_loss_torch) print('residue_per_point_array_torch', residue_per_point_array_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) X_tensorflow = tf.constant(X, dtype=tf.float32) points_per_instance_tensorflow = tf.constant(points_per_instance, dtype=tf.float32) T_gt_tensorflow = tf.constant(T_gt, dtype=tf.int32) I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) parameters_tensorflow = compute_parameters_tensorflow(P_tensorflow, W_tensorflow, X_tensorflow) matching_indices_tensorflow = tf.stop_gradient(tf.py_func(hungarian_matching_tensorflow, [W_tensorflow, I_gt_tensorflow], Tout=tf.int32)) residue_loss_tensorflow, residue_per_point_array_tensorflow, residue_per_class_tensorflow = compute_residue_loss_tensorflow(parameters_tensorflow, matching_indices_tensorflow, points_per_instance_tensorflow, T_gt_tensorflow, classes=['plane', 'sphere', 'cylinder', 'cone']) sess = tf.Session() residue_loss_tensorflow, residue_per_point_array_tensorflow, residue_per_class_tensorflow = sess.run([residue_loss_tensorflow, residue_per_point_array_tensorflow, residue_per_class_tensorflow]) print(np.abs(residue_loss_tensorflow - residue_loss_torch).max()) print(np.abs(residue_per_point_array_tensorflow - residue_per_point_array_torch).max()) for i, class_ in enumerate(['plane', 'sphere', 'cylinder', 'cone']): print(class_, np.abs(residue_per_class_tensorflow[i] - residue_per_class_torch[i]).max()) def compute_parameter_loss(predicted_parameters, gt_parameters, matching_indices, T_gt, is_eval=False, classes=['plane','sphere','cylinder','cone']): parameter_losses = [] # a length T array of BxK tensors batch_size, n_max_instances = predicted_parameters[list(predicted_parameters.keys())[0]].size()[0:2] for class_ in classes: if class_ == 'plane': parameter_loss = plane_fitter.compute_parameter_loss(predicted_parameters['plane_normal'], gt_parameters['plane_normal'], matching_indices, angle_diff=is_eval) elif class_ == 'sphere': parameter_loss = torch.zeros([batch_size, n_max_instances], dtype=torch.float).to(T_gt.device) elif class_ == 'cylinder': parameter_loss = cylinder_fitter.compute_parameter_loss(predicted_parameters['cylinder_axis'], gt_parameters['cylinder_axis'], matching_indices, angle_diff=is_eval) elif class_ == 'cone': parameter_loss = cone_fitter.compute_parameter_loss(predicted_parameters['cone_axis'], gt_parameters['cone_axis'], matching_indices, angle_diff=is_eval) else: raise NotImplementedError parameter_losses.append(parameter_loss) parameter_losses = torch.stack(parameter_losses, dim=2) parameter_loss = torch.gather(parameter_losses, 2, T_gt.unsqueeze(2)).squeeze(2) # BxK return parameter_loss def compute_parameter_loss_tensorflow(predicted_parameters, gt_parameters, matching_indices, T_gt, is_eval=False, classes=['plane','sphere','cylinder','cone']): parameter_losses = [] # a length T array of BxK tensors for class_ in classes: if class_ == 'plane': parameter_loss = plane_fitter.compute_parameter_loss_tensorflow(predicted_parameters['plane_normal'], gt_parameters['plane_normal'], matching_indices, angle_diff=is_eval) elif class_ == 'sphere': parameter_loss = tf.zeros(dtype=tf.float32, shape=[batch_size, n_max_instances]) elif class_ == 'cylinder': parameter_loss = cylinder_fitter.compute_parameter_loss_tensorflow(predicted_parameters['cylinder_axis'], gt_parameters['cylinder_axis'], matching_indices, angle_diff=is_eval) elif class_ == 'cone': parameter_loss = cone_fitter.compute_parameter_loss_tensorflow(predicted_parameters['cone_axis'], gt_parameters['cone_axis'], matching_indices, angle_diff=is_eval) else: raise NotImplementedError parameter_losses.append(parameter_loss) parameter_losses = tf.stack(parameter_losses, axis=2) parameter_loss = aggregate_loss_from_stacked_tensorflow(parameter_losses, T_gt) # BxK return parameter_loss if __name__ == '__main__': batch_size = 100 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to(device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy(gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} predicted_parameters_torch = compute_parameters(P_torch, W_torch, X_torch) matching_indices_torch = hungarian_matching(W_torch, I_gt_torch) parameter_loss_torch = compute_parameter_loss(predicted_parameters_torch, gt_parameters_torch, matching_indices_torch, T_gt_torch, is_eval=False, classes=['plane','sphere','cylinder','cone']) parameter_loss_torch = parameter_loss_torch.detach().cpu().numpy() print('parameter_loss_torch', parameter_loss_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) X_tensorflow = tf.constant(X, dtype=tf.float32) T_gt_tensorflow = tf.constant(T_gt, dtype=tf.int32) I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) gt_parameters_tensorflow = {'plane_normal': tf.constant(gt_parameters['plane_normal'], dtype=tf.float32), 'plane_center': tf.constant(gt_parameters['plane_center'], dtype=tf.float32), 'sphere_center': tf.constant(gt_parameters['sphere_center'], dtype=tf.float32), 'sphere_radius_squared': tf.constant(gt_parameters['sphere_radius_squared'], dtype=tf.float32), 'cylinder_axis': tf.constant(gt_parameters['cylinder_axis'], dtype=tf.float32), 'cylinder_center': tf.constant(gt_parameters['cylinder_center'], dtype=tf.float32), 'cylinder_radius_square': tf.constant(gt_parameters['cylinder_radius_square'], dtype=tf.float32), 'cone_apex': tf.constant(gt_parameters['cone_apex'], dtype=tf.float32), 'cone_axis': tf.constant(gt_parameters['cone_axis'], dtype=tf.float32), 'cone_half_angle': tf.constant(gt_parameters['cone_half_angle'], dtype=tf.float32)} predicted_parameters_tensorflow = compute_parameters_tensorflow(P_tensorflow, W_tensorflow, X_tensorflow) matching_indices_tensorflow = tf.stop_gradient(tf.py_func(hungarian_matching_tensorflow, [W_tensorflow, I_gt_tensorflow], Tout=tf.int32)) parameter_loss_tensorflow = compute_parameter_loss_tensorflow(predicted_parameters_tensorflow, gt_parameters_tensorflow, matching_indices_tensorflow, T_gt_tensorflow, is_eval=False, classes=['plane', 'sphere', 'cylinder', 'cone']) sess = tf.Session() parameter_loss_tensorflow = sess.run(parameter_loss_tensorflow) print(np.abs(parameter_loss_tensorflow - parameter_loss_torch).max()) def sequence_mask(lengths, maxlen=None): if maxlen is None: maxlen = lengths.max() row_vector = torch.arange(0, maxlen, 1).to(lengths.device) matrix = lengths.unsqueeze(dim=-1) mask = row_vector < matrix return mask def get_mask_gt(I_gt, n_max_instances): n_instances_gt = torch.max(I_gt, dim=1)[0] + 1 # only count known primitive type instances, as -1 will be ignored mask_gt = sequence_mask(n_instances_gt, maxlen=n_max_instances) return mask_gt def get_mask_gt_tensorflow(I_gt, n_max_instances): n_instances_gt = tf.reduce_max(I_gt, axis=1) + 1 # only count known primitive type instances, as -1 will be ignored mask_gt = tf.sequence_mask(n_instances_gt, maxlen=n_max_instances) return mask_gt if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) I_gt1 = np.random.randint(0, n_max_instances, (batch_size-batch_size//2, num_points)) I_gt2 = np.random.randint(0, n_max_instances//2, (batch_size//2, num_points)) I_gt = np.concatenate((I_gt1, I_gt2), axis=0) I_gt_torch = torch.from_numpy(I_gt).long().to(device) mask_gt_torch = get_mask_gt(I_gt_torch, n_max_instances) mask_gt_torch = mask_gt_torch.detach().cpu().numpy() print('mask_gt_torch', mask_gt_torch) # Debugging with Tensorflow I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) mask_gt_tensorflow = get_mask_gt_tensorflow(I_gt_tensorflow, n_max_instances) sess = tf.Session() mask_gt_tensorflow = sess.run(mask_gt_tensorflow) print(np.all(mask_gt_torch == mask_gt_tensorflow)) def reduce_mean_masked_instance(loss, mask_gt): # loss: BxK loss = torch.where(mask_gt, loss, torch.zeros_like(loss)) reduced_loss = torch.sum(loss, axis=1) # B denom = torch.sum(mask_gt.float(), dim=1) # B return torch.where(denom > 0, reduced_loss / denom, torch.zeros_like(reduced_loss)) # B def collect_losses(normal_loss, normal_loss_multiplier, type_loss, type_loss_multiplier, avg_miou_loss, miou_loss, miou_loss_multiplier, avg_residue_loss, residue_loss, residue_loss_multiplier, avg_parameter_loss, parameter_loss, parameter_loss_multiplier, total_loss_multiplier): total_loss = 0 # Normal Loss normal_loss_per_data = normal_loss total_normal_loss = torch.mean(normal_loss_per_data) if normal_loss_multiplier > 0: total_loss = total_loss + normal_loss_multiplier * total_normal_loss # Total loss type_loss_per_data = type_loss total_type_loss = torch.mean(type_loss_per_data) if type_loss_multiplier > 0: total_loss = total_loss + type_loss_multiplier * total_type_loss # mIoU Loss miou_loss_per_data = avg_miou_loss miou_loss_per_instance = miou_loss total_miou_loss = torch.mean(miou_loss_per_data) if miou_loss_multiplier > 0: total_loss = total_loss + miou_loss_multiplier * total_miou_loss # Residue Loss residue_loss_per_data = avg_residue_loss residue_loss_per_instance = residue_loss total_residue_loss = torch.mean(residue_loss_per_data) if residue_loss_multiplier > 0: total_loss = total_loss + residue_loss_multiplier * total_residue_loss # Paramerer Loss parameter_loss_per_data = avg_parameter_loss parameter_loss_per_instance = parameter_loss total_parameter_loss = torch.mean(parameter_loss_per_data) if parameter_loss_multiplier > 0: total_loss = total_loss + parameter_loss_multiplier * total_parameter_loss total_loss = total_loss * total_loss_multiplier return total_loss, total_normal_loss, total_type_loss, total_miou_loss, total_residue_loss, total_parameter_loss def compute_all_losses(P, W, I_gt, X, X_gt, T, T_gt, gt_parameters, points_per_instance, normal_loss_multiplier, type_loss_multiplier, miou_loss_multiplier, residue_loss_multiplier, parameter_loss_multiplier, total_loss_multiplier, is_eval, mode_seg='mIoU', classes=['plane','sphere','cylinder','cone']): assert(mode_seg in ['mIoU', 'intersection']) batch_size, _, n_max_instances = W.size() matching_indices = hungarian_matching(W, I_gt) if (residue_loss_multiplier>0) or (parameter_loss_multiplier>0): predicted_parameters = compute_parameters(P, W, X) mask_gt = get_mask_gt(I_gt, n_max_instances) if normal_loss_multiplier>0: normal_loss = compute_normal_loss(X, X_gt, angle_diff=is_eval) else: normal_loss = torch.zeros([batch_size, n_max_instances]).to(P.device) if type_loss_multiplier>0: type_loss = compute_per_point_type_loss(T, I_gt, T_gt, is_eval) else: type_loss = torch.zeros([batch_size, n_max_instances]).to(P.device) if (mode_seg == 'mIoU') and (miou_loss_multiplier>0): miou_loss, _ = compute_miou_loss(W, I_gt, matching_indices) avg_miou_loss = reduce_mean_masked_instance(miou_loss, mask_gt) elif (mode_seg == 'intersection') and (miou_loss_multiplier>0): _, miou_loss = compute_miou_loss(W, I_gt, matching_indices) avg_miou_loss = reduce_mean_masked_instance(miou_loss, mask_gt) else: miou_loss = torch.zeros([batch_size, n_max_instances]).to(P.device) avg_miou_loss = torch.zeros([batch_size]).to(P.device) if residue_loss_multiplier>0: residue_loss, residue_per_point_array = compute_residue_loss(predicted_parameters, matching_indices, points_per_instance, T_gt, classes=classes) avg_residue_loss = reduce_mean_masked_instance(residue_loss, mask_gt) else: residue_loss = torch.zeros([batch_size, n_max_instances]).to(P.device) avg_residue_loss = torch.zeros([batch_size]).to(P.device) if parameter_loss_multiplier>0: parameter_loss = compute_parameter_loss(predicted_parameters, gt_parameters, matching_indices, T_gt, is_eval, classes=classes) avg_parameter_loss = reduce_mean_masked_instance(parameter_loss, mask_gt) else: parameter_loss = torch.zeros([batch_size, n_max_instances]).to(P.device) avg_parameter_loss = torch.zeros([batch_size]).to(P.device) total_loss, total_normal_loss, total_type_loss, total_miou_loss, total_residue_loss, total_parameter_loss = \ collect_losses(normal_loss, normal_loss_multiplier, type_loss, type_loss_multiplier, avg_miou_loss, miou_loss, miou_loss_multiplier, avg_residue_loss, residue_loss, residue_loss_multiplier, avg_parameter_loss, parameter_loss, parameter_loss_multiplier, total_loss_multiplier) if (residue_loss_multiplier > 0) or (parameter_loss_multiplier > 0): return total_loss, total_normal_loss, total_type_loss, total_miou_loss, total_residue_loss, total_parameter_loss, predicted_parameters['plane_normal'], predicted_parameters['cylinder_axis'], predicted_parameters['cone_axis'] else: return total_loss, total_normal_loss, total_type_loss, total_miou_loss, total_residue_loss, total_parameter_loss, None, None, None def reduce_mean_masked_instance_tensorflow(loss, mask_gt): # loss: BxK loss = tf.where(mask_gt, loss, tf.zeros_like(loss)) reduced_loss = tf.reduce_sum(loss, axis=1) # B denom = tf.reduce_sum(tf.to_float(mask_gt), axis=1) # B return tf.where(denom > 0, reduced_loss / denom, tf.zeros_like(reduced_loss)) # B def collect_losses_tensorflow(normal_loss, normal_loss_multiplier, type_loss, type_loss_multiplier, avg_miou_loss, miou_loss, miou_loss_multiplier, avg_residue_loss, residue_loss, residue_loss_multiplier, avg_parameter_loss, parameter_loss, parameter_loss_multiplier, total_loss_multiplier): total_loss = tf.zeros(shape=[], dtype=tf.float32) normal_loss_per_data = normal_loss total_normal_loss = tf.reduce_mean(normal_loss_per_data) if normal_loss_multiplier > 0: total_loss += normal_loss_multiplier * total_normal_loss type_loss_per_data = type_loss total_type_loss = tf.reduce_mean(type_loss_per_data) if type_loss_multiplier > 0: total_loss += type_loss_multiplier * total_type_loss miou_loss_per_data = avg_miou_loss miou_loss_per_instance = miou_loss total_miou_loss = tf.reduce_mean(miou_loss_per_data) if miou_loss_multiplier > 0: total_loss += miou_loss_multiplier * total_miou_loss residue_loss_per_data = avg_residue_loss residue_loss_per_instance = residue_loss total_residue_loss = tf.reduce_mean(residue_loss_per_data) if residue_loss_multiplier > 0: total_loss += residue_loss_multiplier * total_residue_loss parameter_loss_per_data = avg_parameter_loss parameter_loss_per_instance = parameter_loss total_parameter_loss = tf.reduce_mean(parameter_loss_per_data) if parameter_loss_multiplier > 0: total_loss += parameter_loss_multiplier * total_parameter_loss total_loss *= total_loss_multiplier return total_loss def compute_all_losses_tensorflow(P, W, I_gt, X, X_gt, T, T_gt, gt_parameters, points_per_instance, normal_loss_multiplier, type_loss_multiplier, miou_loss_multiplier, residue_loss_multiplier, parameter_loss_multiplier, total_loss_multiplier, is_eval, classes=['plane','sphere','cylinder','cone']): b_max_instances = W.shape[2] matching_indices = tf.stop_gradient(tf.py_func(hungarian_matching_tensorflow, [W, I_gt], Tout=tf.int32)) predicted_parameters = compute_parameters_tensorflow(P, W, X) mask_gt = get_mask_gt_tensorflow(I_gt, n_max_instances) normal_loss = compute_normal_loss_tensorflow(X, X_gt, angle_diff=is_eval) type_loss = compute_per_point_type_loss_tensorflow(T, I_gt, T_gt, is_eval) miou_loss = compute_miou_loss_tensorflow(W, I_gt, matching_indices) avg_miou_loss = reduce_mean_masked_instance_tensorflow(miou_loss, mask_gt) residue_loss, residue_per_point_array = compute_residue_loss_tensorflow(predicted_parameters, matching_indices, points_per_instance, T_gt, classes=classes) avg_residue_loss = reduce_mean_masked_instance_tensorflow(residue_loss, mask_gt) parameter_loss = compute_parameter_loss_tensorflow(predicted_parameters, gt_parameters, matching_indices, T_gt, is_eval, classes=classes) avg_parameter_loss = reduce_mean_masked_instance_tensorflow(parameter_loss, mask_gt) total_loss = collect_losses_tensorflow(normal_loss, normal_loss_multiplier, type_loss, type_loss_multiplier, avg_miou_loss, miou_loss, miou_loss_multiplier, avg_residue_loss, residue_loss, residue_loss_multiplier, avg_parameter_loss, parameter_loss, parameter_loss_multiplier, total_loss_multiplier) return total_loss if __name__ == '__main__': batch_size = 100 num_points = 1024 num_points_instance = 512 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) X_gt = np.random.randn(batch_size, num_points, 3) X_gt = X_gt / np.linalg.norm(X_gt, axis=2, keepdims=True) T = np.random.rand(batch_size, num_points, 4) T_gt = np.random.randint(0, 4, (batch_size, n_max_instances)) I_gt = np.random.randint(0, n_max_instances, (batch_size, num_points)) plane_normal = np.random.randn(batch_size, n_max_instances, 3) plane_normal = plane_normal / np.linalg.norm(plane_normal, axis=2, keepdims=True) plane_center = np.random.randn(batch_size, n_max_instances) sphere_center = np.random.randn(batch_size, n_max_instances, 3) sphere_radius_squared = np.abs(np.random.randn(batch_size, n_max_instances)) cylinder_axis = np.random.randn(batch_size, n_max_instances, 3) cylinder_axis = cylinder_axis / np.linalg.norm(cylinder_axis, axis=2, keepdims=True) cylinder_center = np.random.randn(batch_size, n_max_instances, 3) cylinder_radius_square = np.abs(np.random.randn(batch_size, n_max_instances)) cone_apex = np.random.randn(batch_size, n_max_instances, 3) cone_axis = np.random.randn(batch_size, n_max_instances, 3) cone_half_angle = np.abs(np.random.randn(batch_size, n_max_instances)) points_per_instance = np.random.randn(batch_size, n_max_instances, num_points_instance, 3) normal_loss_multiplier = 1.0 type_loss_multiplier = 1.0 miou_loss_multiplier = 1.0 residue_loss_multiplier = 1.0 parameter_loss_multiplier = 1.0 total_loss_multiplier = 1.0 is_eval = False gt_parameters = {'plane_normal': plane_normal, 'plane_center': plane_center, 'sphere_center': sphere_center, 'sphere_radius_squared': sphere_radius_squared, 'cylinder_axis': cylinder_axis, 'cylinder_center': cylinder_center, 'cylinder_radius_square': cylinder_radius_square, 'cone_apex': cone_apex, 'cone_axis': cone_axis, 'cone_half_angle': cone_half_angle} P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) X_gt_torch = torch.from_numpy(X_gt).float().to(device) T_torch = torch.from_numpy(T).float().to(device) T_gt_torch = torch.from_numpy(T_gt).long().to(device) I_gt_torch = torch.from_numpy(I_gt).long().to(device) gt_parameters_torch = {'plane_normal': torch.from_numpy(gt_parameters['plane_normal']).float().to(device), 'plane_center': torch.from_numpy(gt_parameters['plane_center']).float().to(device), 'sphere_center': torch.from_numpy(gt_parameters['sphere_center']).float().to(device), 'sphere_radius_squared': torch.from_numpy(gt_parameters['sphere_radius_squared']).float().to(device), 'cylinder_axis': torch.from_numpy(gt_parameters['cylinder_axis']).float().to(device), 'cylinder_center': torch.from_numpy(gt_parameters['cylinder_center']).float().to(device), 'cylinder_radius_square': torch.from_numpy(gt_parameters['cylinder_radius_square']).float().to(device), 'cone_apex': torch.from_numpy(gt_parameters['cone_apex']).float().to(device), 'cone_axis': torch.from_numpy(gt_parameters['cone_axis']).float().to(device), 'cone_half_angle': torch.from_numpy(gt_parameters['cone_half_angle']).float().to(device)} points_per_instance_torch = torch.from_numpy(points_per_instance).long().to(device) total_loss_torch = compute_all_losses(P_torch, W_torch, I_gt_torch, X_torch, X_gt_torch, T_torch, T_gt_torch, gt_parameters_torch, points_per_instance_torch, normal_loss_multiplier, type_loss_multiplier, miou_loss_multiplier, residue_loss_multiplier, parameter_loss_multiplier, total_loss_multiplier, is_eval)[0] total_loss_torch = total_loss_torch.detach().cpu().numpy() print('total_loss_torch', total_loss_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) X_tensorflow = tf.constant(X, dtype=tf.float32) X_gt_tensorflow = tf.constant(X_gt, dtype=tf.float32) T_tensorflow = tf.constant(T, dtype=tf.float32) T_gt_tensorflow = tf.constant(T_gt, dtype=tf.int32) I_gt_tensorflow = tf.constant(I_gt, dtype=tf.int32) gt_parameters_tensorflow = {'plane_normal': tf.constant(gt_parameters['plane_normal'], dtype=tf.float32), 'plane_center': tf.constant(gt_parameters['plane_center'], dtype=tf.float32), 'sphere_center': tf.constant(gt_parameters['sphere_center'], dtype=tf.float32), 'sphere_radius_squared': tf.constant(gt_parameters['sphere_radius_squared'], dtype=tf.float32), 'cylinder_axis': tf.constant(gt_parameters['cylinder_axis'], dtype=tf.float32), 'cylinder_center': tf.constant(gt_parameters['cylinder_center'], dtype=tf.float32), 'cylinder_radius_square': tf.constant(gt_parameters['cylinder_radius_square'], dtype=tf.float32), 'cone_apex': tf.constant(gt_parameters['cone_apex'], dtype=tf.float32), 'cone_axis': tf.constant(gt_parameters['cone_axis'], dtype=tf.float32), 'cone_half_angle': tf.constant(gt_parameters['cone_half_angle'], dtype=tf.float32)} points_per_instance_tensorflow = tf.constant(points_per_instance, dtype=tf.float32) total_loss_tensorflow = compute_all_losses_tensorflow(P_tensorflow, W_tensorflow, I_gt_tensorflow, X_tensorflow, X_gt_tensorflow, T_tensorflow, T_gt_tensorflow, gt_parameters_tensorflow, points_per_instance_tensorflow, normal_loss_multiplier, type_loss_multiplier, miou_loss_multiplier, residue_loss_multiplier, parameter_loss_multiplier, total_loss_multiplier, is_eval) sess = tf.Session() total_loss_tensorflow = sess.run(total_loss_tensorflow) print(np.abs(total_loss_tensorflow - total_loss_torch).max())
55,650
62.819954
319
py
CPFN
CPFN-master/SPFN/cone_fitter.py
# Importation of packages import torch import numpy as np if __name__ == '__main__': import tensorflow as tf from SPFN.primitives import Cone from SPFN.geometry_utils import guarded_matrix_solve_ls, guarded_matrix_solve_ls_tensorflow, weighted_plane_fitting, weighted_plane_fitting_tensorflow def acos_safe(x): return torch.acos(torch.clamp(x, min=-1.0+1e-6, max=1.0-1e-6)) def compute_parameters(P, W, X, div_eps=1e-10): batch_size, n_points, _ = P.size() _, _, n_max_instances = W.size() W_reshaped = W.transpose(1,2).contiguous().view(batch_size * n_max_instances, n_points) # BKxN # A - BKxNx3 A = X.unsqueeze(1).expand(batch_size, n_max_instances, n_points, 3).contiguous().view(batch_size * n_max_instances, n_points, 3) # b - BKxNx1 b = torch.sum(P * X, dim=2).unsqueeze(1).expand(batch_size, n_max_instances, n_points).contiguous().view(batch_size * n_max_instances, n_points, 1) apex = guarded_matrix_solve_ls(A, b, W_reshaped).view(batch_size, n_max_instances, 3) # BxKx3 X_tiled = A # TODO: use P-apex instead of X for plane fitting plane_n, plane_c = weighted_plane_fitting(X_tiled, W_reshaped) axis = plane_n.view(batch_size, n_max_instances, 3) # BxKx3 P_minus_apex = P.unsqueeze(2) - apex.unsqueeze(1) # BxNxKx3 P_minus_apex_normalized = torch.nn.functional.normalize(P_minus_apex, p=2, dim=3, eps=1e-12) P_minus_apex_normalized_dot_axis = torch.sum(axis.unsqueeze(1) * P_minus_apex_normalized, dim=3) # BxNxK # flip direction of axis if wrong sgn_axis = torch.sign(torch.sum(W * P_minus_apex_normalized_dot_axis, dim=1)) # BxK sgn_axis = sgn_axis + (sgn_axis==0.0).float() # prevent sgn == 0 axis = axis * sgn_axis.unsqueeze(2) # BxKx3 tmp = W * acos_safe(torch.abs(P_minus_apex_normalized_dot_axis)) # BxNxK W_sum = torch.sum(W, dim=1) # BxK half_angle = torch.sum(tmp, dim=1) / (W_sum + div_eps) # BxK half_angle = torch.clamp(half_angle, min=1e-3, max=np.pi/2-1e-3) # angle cannot be too weird return apex, axis, half_angle def acos_safe_tensorflow(x): return tf.math.acos(tf.clip_by_value(x, -1.0+1e-6, 1.0-1e-6)) def compute_parameters_tensorflow(P, W, X): batch_size = tf.shape(P)[0] n_points = tf.shape(P)[1] n_max_instances = W.get_shape()[2] W_reshaped = tf.reshape(tf.transpose(W, [0, 2, 1]), [batch_size * n_max_instances, n_points]) # BKxN # A - BKxNx3 A = tf.reshape(tf.tile(tf.expand_dims(X, axis=1), [1, n_max_instances, 1, 1]), [batch_size * n_max_instances, n_points, 3]) # BKxNx3 # b - BKxNx1 b = tf.expand_dims(tf.reshape(tf.tile(tf.expand_dims(tf.reduce_sum(P * X, axis=2), axis=1), [1, n_max_instances, 1]), [batch_size * n_max_instances, n_points]), axis=2) apex = tf.reshape(guarded_matrix_solve_ls_tensorflow(A, b, W_reshaped), [batch_size, n_max_instances, 3]) # BxKx3 X_tiled = A # TODO: use P-apex instead of X for plane fitting plane_n, plane_c = weighted_plane_fitting_tensorflow(X_tiled, W_reshaped) axis = tf.reshape(plane_n, [batch_size, n_max_instances, 3]) # BxKx3 P_minus_apex_normalized = tf.nn.l2_normalize(tf.expand_dims(P, axis=2) - tf.expand_dims(apex, 1), axis=3) # BxNxKx3 P_minus_apex_normalized_dot_axis = tf.reduce_sum(tf.expand_dims(axis, axis=1) * P_minus_apex_normalized, axis=3) # BxNxK # flip direction of axis if wrong sgn_axis = tf.sign(tf.reduce_sum(W * P_minus_apex_normalized_dot_axis, axis=1)) # BxK sgn_axis += tf.to_float(tf.equal(sgn_axis, 0.0)) # prevent sgn == 0 axis *= tf.expand_dims(sgn_axis, axis=2) # BxKx3 tmp = W * acos_safe_tensorflow(tf.abs(P_minus_apex_normalized_dot_axis)) # BxNxK W_sum = tf.reduce_sum(W, axis=1) # BxK half_angle = tf.reduce_sum(tmp, axis=1) / W_sum # BxK tf.clip_by_value(half_angle, 1e-3, np.pi / 2 - 1e-3) # angle cannot be too weird return apex, axis, half_angle if __name__ == '__main__': batch_size = 100 num_points = 1024 n_max_instances = 12 device = torch.device('cuda:0') np.random.seed(0) P = np.random.randn(batch_size, num_points, 3) W = np.random.rand(batch_size, num_points, n_max_instances) X = np.random.randn(batch_size, num_points, 3) X = X / np.linalg.norm(X, axis=2, keepdims=True) P_torch = torch.from_numpy(P).float().to(device) W_torch = torch.from_numpy(W).float().to(device) X_torch = torch.from_numpy(X).float().to(device) apex_torch, axis_torch, half_angle_torch = compute_parameters(P_torch, W_torch, X_torch) apex_torch = apex_torch.detach().cpu().numpy() axis_torch = axis_torch.detach().cpu().numpy() half_angle_torch = half_angle_torch.detach().cpu().numpy() print('apex_torch', apex_torch) print('axis_torch', axis_torch) print('half_angle_torch', half_angle_torch) # Debugging with Tensorflow P_tensorflow = tf.constant(P, dtype=tf.float32) W_tensorflow = tf.constant(W, dtype=tf.float32) X_tensorflow = tf.constant(X, dtype=tf.float32) apex_tensorflow, axis_tensorflow, half_angle_tensorflow = compute_parameters_tensorflow(P_tensorflow, W_tensorflow, X_tensorflow) sess = tf.Session() apex_tensorflow, axis_tensorflow, half_angle_tensorflow = sess.run([apex_tensorflow, axis_tensorflow, half_angle_tensorflow]) print(np.abs(apex_tensorflow - apex_torch).max()) print(np.minimum(np.abs(axis_tensorflow - axis_torch), np.abs(axis_tensorflow + axis_torch)).max()) print(np.abs(half_angle_tensorflow - half_angle_torch).max()) def compute_residue_single(apex, axis, half_angle, p): # apex: ...x3, axis: ...x3, half_angle: ..., p: ...x30 v = p - apex v_normalized = torch.nn.functional.normalize(v, p=2, dim=-1, eps=1e-12) alpha = acos_safe(torch.sum(v_normalized * axis, dim=-1)) return (torch.sin(torch.clamp(torch.abs(alpha - half_angle), min=None, max=np.pi / 2)))**2 * torch.sum(v * v, dim=-1) def compute_residue_single_tensorflow(apex, axis, half_angle, p): # apex: ...x3, axis: ...x3, half_angle: ..., p: ...x3 v = p - apex v_normalized = tf.nn.l2_normalize(v, axis=-1) alpha = acos_safe_tensorflow(tf.reduce_sum(v_normalized * axis, axis=-1)) return tf.square(tf.sin(tf.minimum(tf.abs(alpha - half_angle), np.pi / 2))) * tf.reduce_sum(v * v, axis=-1) if __name__ == '__main__': batch_size = 100 num_points = 1024 device = torch.device('cuda:0') np.random.seed(0) apex = np.random.randn(batch_size, num_points, 3) axis = np.random.randn(batch_size, num_points, 3) half_angle = np.random.randn(batch_size, num_points) p = np.random.randn(batch_size, num_points, 3) apex_torch = torch.from_numpy(apex).float().to(device) axis_torch = torch.from_numpy(axis).float().to(device) half_angle_torch = torch.from_numpy(half_angle).float().to(device) p_torch = torch.from_numpy(p).float().to(device) loss_torch = compute_residue_single(apex_torch, axis_torch, half_angle_torch, p_torch) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow apex_tensorflow = tf.constant(apex, dtype=tf.float32) axis_tensorflow = tf.constant(axis, dtype=tf.float32) half_angle_tensorflow = tf.constant(half_angle, dtype=tf.float32) p_tensorflow = tf.constant(p, dtype=tf.float32) loss_tensorflow = compute_residue_single_tensorflow(apex_tensorflow, axis_tensorflow, half_angle_tensorflow, p_tensorflow) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) def compute_parameter_loss(predicted_axis, gt_axis, matching_indices, angle_diff): # predicted_axis: BxK1x3 # gt_axis: BXK2x3 # matching indices: BxK2 batch_size, nb_primitives, _ = gt_axis.size() predicted_axis = torch.gather(predicted_axis, 1, matching_indices.unsqueeze(2).expand(batch_size, nb_primitives, 3)) dot_abs = torch.abs(torch.sum(predicted_axis * gt_axis, axis=2)) if angle_diff: return acos_safe(dot_abs) # BxK else: return 1.0 - dot_abs # BxK def batched_gather(data, indices, axis): # data - Bx...xKx..., axis is where dimension K is # indices - BxK # output[b, ..., k, ...] = in[b, ..., indices[b, k], ...] assert axis >= 1 ndims = data.get_shape().ndims # allow dynamic rank if axis > 1: # tranpose data to BxKx... perm = np.arange(ndims) perm[axis] = 1 perm[1] = axis data = tf.transpose(data, perm=perm) batch_size = tf.shape(data)[0] batch_nums = tf.tile(tf.expand_dims(tf.expand_dims(tf.range(batch_size), axis=1), axis=2), multiples=[1, tf.shape(indices)[1], 1]) # BxKx1 indices = tf.concat([batch_nums, tf.expand_dims(indices, axis=2)], axis=2) # BxKx2 gathered_data = tf.gather_nd(data, indices=indices) if axis > 1: gathered_data = tf.transpose(gathered_data, perm=perm) return gathered_data def compute_parameter_loss_tensorflow(predicted_axis, gt_axis, matching_indices, angle_diff): axis = batched_gather(predicted_axis, matching_indices, axis=1) dot_abs = tf.abs(tf.reduce_sum(axis * gt_axis, axis=2)) if angle_diff: return acos_safe_tensorflow(dot_abs) # BxK else: return 1.0 - dot_abs # BxK if __name__ == '__main__': batch_size = 100 num_primitives1 = 15 num_primitives2 = 5 device = torch.device('cuda:0') np.random.seed(0) predicted_axis = np.random.randn(batch_size, num_primitives1, 3) gt_axis = np.random.randn(batch_size, num_primitives2, 3) matching_indices = np.random.randint(0, 15, (batch_size, num_primitives2)) angle_diff = True predicted_axis_torch = torch.from_numpy(predicted_axis).float().to(device) gt_axis_torch = torch.from_numpy(gt_axis).float().to(device) matching_indices_torch = torch.from_numpy(matching_indices).long().to(device) loss_torch = compute_parameter_loss(predicted_axis_torch, gt_axis_torch, matching_indices_torch, angle_diff) loss_torch = loss_torch.detach().cpu().numpy() print('loss_torch', loss_torch) # Debugging with Tensorflow predicted_axis_tensorflow = tf.constant(predicted_axis, dtype=tf.float32) gt_axis_tensorflow = tf.constant(gt_axis, dtype=tf.float32) matching_indices_tensorflow = tf.constant(matching_indices, dtype=tf.int32) loss_tensorflow = compute_parameter_loss_tensorflow(predicted_axis_tensorflow, gt_axis_tensorflow, matching_indices_tensorflow, angle_diff) sess = tf.Session() loss_tensorflow = sess.run(loss_tensorflow) print(np.abs(loss_torch - loss_tensorflow).max()) def create_primitive_from_dict(d): assert d['type'] == 'cone' apex = np.array([d['apex_x'], d['apex_y'], d['apex_z']], dtype=float) axis = np.array([d['axis_x'], d['axis_y'], d['axis_z']], dtype=float) half_angle = float(d['semi_angle']) return Cone(apex=apex, axis=axis, half_angle=half_angle) def extract_parameter_data_as_dict(primitives, n_max_instances): axis_gt = np.zeros(dtype=float, shape=[n_max_instances, 3]) apex_gt = np.zeros(dtype=float, shape=[n_max_instances, 3]) half_angle_gt = np.zeros(dtype=float, shape=[n_max_instances]) for i, primitive in enumerate(primitives): if isinstance(primitive, Cone): axis_gt[i] = primitive.axis apex_gt[i] = primitive.apex half_angle_gt[i] = primitive.half_angle return { 'cone_axis_gt': axis_gt, } def extract_predicted_parameters_as_json(cone_apex, cone_axis, cone_half_angle, k): cone = Cone(cone_apex, cone_axis, cone_half_angle, z_min=0.0, z_max=5.0) return { 'type': 'cone', 'apex_x': float(cone.apex[0]), 'apex_y': float(cone.apex[1]), 'apex_z': float(cone.apex[2]), 'axis_x': float(cone.axis[0]), 'axis_y': float(cone.axis[1]), 'axis_z': float(cone.axis[2]), 'angle': float(cone.half_angle * 2), 'z_min': float(cone.z_min), 'z_max': float(cone.z_max), 'label': k, }
12,047
49.835443
172
py
CPFN
CPFN-master/Utils/training_utils.py
# Importation of packages import sys import torch import numpy as np from SPFN import losses_implementation # BN Decay def get_batch_norm_decay(global_step, batch_size, bn_decay_step, staircase=True): BN_INIT_DECAY = 0.5 BN_DECAY_RATE = 0.5 BN_DECAY_CLIP = 0.99 p = global_step * batch_size / bn_decay_step if staircase: p = int(np.floor(p)) bn_momentum = max(BN_INIT_DECAY * (BN_DECAY_RATE ** p), 1-BN_DECAY_CLIP) return bn_momentum def update_momentum(module, bn_momentum): for name, module_ in module.named_modules(): if 'bn' in name: module_.momentum = bn_momentum # LR Decay def get_learning_rate(init_learning_rate, global_step, batch_size, decay_step, decay_rate, staircase=True): p = global_step * batch_size / decay_step if staircase: p = int(np.floor(p)) learning_rate = init_learning_rate * (decay_rate ** p) return learning_rate # Train For One Epoch def patch_selection_train_val_epoch(dataloader, patchselec_module, epoch, optimizer, global_step, visualiser, args, conf, device, network_mode='train'): assert(network_mode in ['train', 'val']) # Loading conf information related to current file batch_size = conf.get_batch_size() bn_decay_step = conf.get_bn_decay_step() decay_step = conf.get_decay_step() decay_rate = conf.get_decay_rate() init_learning_rate = conf.get_init_learning_rate() # Iteration over the dataset old_bn_momentum = get_batch_norm_decay(global_step, batch_size, bn_decay_step, staircase=True) old_learning_rate = get_learning_rate(init_learning_rate, global_step, batch_size, decay_step, decay_rate, staircase=True) total_loss = 0 if network_mode == 'train': patchselec_module.train() elif network_mode == 'val': patchselec_module.eval() patchselec_module.train() for batch_id, data in enumerate(dataloader, 0): optimizer.zero_grad() # Updating the BN decay bn_momentum = get_batch_norm_decay(global_step, batch_size, bn_decay_step, staircase=True) if old_bn_momentum != bn_momentum: update_momentum(patchselec_module, bn_momentum) old_bn_momentum = bn_momentum # Updating the LR decay learning_rate = get_learning_rate(init_learning_rate, global_step, batch_size, decay_step, decay_rate, staircase=True) if old_learning_rate != learning_rate: for param_group in optimizer.param_groups: param_group['lr'] = learning_rate old_learning_rate = learning_rate # Proper training points = data[0].type(torch.FloatTensor).to(device) batch_size_current, num_points, _ = points.size() output_labels = data[1].type(torch.LongTensor).to(device) predicted_labels = patchselec_module(points)[0] predicted_labels = predicted_labels.contiguous().view(batch_size_current * num_points, 2) output_labels = output_labels.view(batch_size_current * num_points) loss = torch.nn.functional.cross_entropy(predicted_labels, output_labels) total_loss += batch_size_current * loss.item() # Printing Values if batch_id%100==0: print('[%s][Epoch %d - Iteration %d] Loss: %f' % (network_mode, epoch, batch_id, loss.item())) if network_mode == 'train': # Backward Pass loss.backward() optimizer.step() global_step += 1 # Updating the visualiser visualiser.log_loss(loss.item(), '%s_loss' % network_mode) visualiser.update() return global_step, total_loss def spfn_train_val_epoch(dataloader, spfn_module, epoch, optimizer, global_step, visualiser, args, conf, device, network_mode='train'): assert(network_mode in ['train', 'val']) # Loading conf information related to current file batch_size = conf.get_batch_size() bn_decay_step = conf.get_bn_decay_step() decay_step = conf.get_decay_step() decay_rate = conf.get_decay_rate() init_learning_rate = conf.get_init_learning_rate() # Losses miou_loss_multiplier = conf.get_miou_loss_multiplier() normal_loss_multiplier = conf.get_normal_loss_multiplier() type_loss_multiplier = conf.get_type_loss_multiplier() parameter_loss_multiplier = conf.get_parameter_loss_multiplier() residue_loss_multiplier = conf.get_residue_loss_multiplier() total_loss_multiplier = conf.get_total_loss_multiplier() # Iteration over the dataset old_bn_momentum = get_batch_norm_decay(global_step, batch_size, bn_decay_step, staircase=True) old_learning_rate = get_learning_rate(init_learning_rate, global_step, batch_size, decay_step, decay_rate, staircase=True) total_loss_ = 0 if network_mode == 'train': spfn_module.train() elif network_mode == 'val': spfn_module.eval() for batch_id, data in enumerate(dataloader, 0): if batch_id%100==0: print('[%s][Epoch %d - Iteration %d]' % (network_mode, epoch, batch_id)) optimizer.zero_grad() # Updating the BN decay bn_momentum = get_batch_norm_decay(global_step, batch_size, bn_decay_step, staircase=True) if old_bn_momentum != bn_momentum: update_momentum(spfn_module, bn_momentum) old_bn_momentum = bn_momentum # Updating the LR decay learning_rate = get_learning_rate(init_learning_rate, global_step, batch_size, decay_step, decay_rate, staircase=True) if old_learning_rate != learning_rate: for param_group in optimizer.param_groups: param_group['lr'] = learning_rate old_learning_rate = learning_rate # Loading the inputs P = data[0].type(torch.FloatTensor).to(device) batch_size_current, num_points, _ = P.size() X_gt = data[1].type(torch.FloatTensor).to(device) points_per_instance = data[2].type(torch.FloatTensor).to(device) _, nb_primitives, nb_points_primitives, _ = points_per_instance.size() I_gt = data[3].type(torch.LongTensor).to(device) T_gt = data[4].type(torch.LongTensor).to(device) plane_n_gt = data[5].type(torch.FloatTensor).to(device) cylinder_axis_gt = data[6].type(torch.FloatTensor).to(device) cone_axis_gt = data[7].type(torch.FloatTensor).to(device) gt_parameters = {'plane_normal': plane_n_gt, 'cylinder_axis': cylinder_axis_gt, 'cone_axis': cone_axis_gt} if args.network == 'GlobalSPFN': glob_features = None loc_features = None elif args.network == 'LocalSPFN': glob_features = data[8].type(torch.FloatTensor).to(device) loc_features = data[9].type(torch.FloatTensor).to(device) # Forward Pass X, T, W, _, _ = spfn_module(P, glob_features=glob_features, loc_features=loc_features) X = torch.nn.functional.normalize(X, p=2, dim=2, eps=1e-12) W = torch.softmax(W, dim=2) total_loss, total_normal_loss, total_type_loss, total_miou_loss, total_residue_loss, total_parameter_loss, _, _, _ = losses_implementation.compute_all_losses( P, W, I_gt, X, X_gt, T, T_gt, gt_parameters, points_per_instance, normal_loss_multiplier, type_loss_multiplier, miou_loss_multiplier, residue_loss_multiplier, parameter_loss_multiplier, total_loss_multiplier, False, mode_seg='mIoU', classes=conf.get_list_of_primitives()) total_loss_ += batch_size_current * total_loss.item() if network_mode == 'train': # Backward Pass total_loss.backward() # Unecessary check for the gradient flag = False for param in spfn_module.parameters(): if param.requires_grad and ((torch.any(torch.isinf(param.grad))) or torch.any(torch.isnan(param.grad))): flag = True break if not flag: optimizer.step() global_step += 1 # Printing Values if batch_id%100==0: print('Loss Value: ', total_loss.item()) print('Normal Loss', total_normal_loss.item()) print('Type Loss', total_type_loss.item()) print('mIoU Loss', total_miou_loss.item()) print('Residue Loss', total_residue_loss.item()) print('Parameter Loss', total_parameter_loss.item()) # Updating the visualiser visualiser.log_loss(total_loss.item(), '%s_loss'%network_mode) visualiser.log_loss(total_normal_loss.item(), '%s_normal_loss'%network_mode) visualiser.log_loss(total_type_loss.item(), '%s_type_loss'%network_mode) visualiser.log_loss(total_miou_loss.item(), '%s_miou_loss'%network_mode) visualiser.log_loss(total_residue_loss.item(), '%s_residue_loss'%network_mode) visualiser.log_loss(total_parameter_loss.item(), '%s_parameter_loss'%network_mode) visualiser.update() return global_step, total_loss_
8,974
49.994318
166
py
CPFN
CPFN-master/Utils/training_visualisation.py
import torch import numpy as np from torch import nn from visdom import Visdom ORANGE = np.array([[255, 105, 0]]) BLUE = np.array([[40, 40, 255]]) RED = np.array([[255, 40, 40]]) class Visualiser(object): def __init__(self, plotting_interval, port=8097): self.vis = Visdom(port=port) self.line_plotter = VisdomLinePlotter(self.vis) self.plotting_interval = plotting_interval self.plotting_step = 0 self.loss_history_dict = {} self.image_dict = {} self.window_elements = [] def log_image(self, image, name): image = torch.clamp(image, 0, 1) image = image.cpu().detach().numpy() self.image_dict[name] = image if not name in self.window_elements: self.window_elements.append(name) def log_loss(self, loss, name): current_history = self.loss_history_dict.get(name, [np.nan] * self.plotting_interval) updated_history = current_history[1:] + [loss] self.loss_history_dict[name] = updated_history if not name in self.window_elements: self.window_elements.append(name) def update(self): if self.plotting_step % self.plotting_interval == 0: loss_avg_dict = {k: torch.tensor(self.loss_history_dict[k]).mean().item() for k in self.loss_history_dict} for name in loss_avg_dict: loss_avg = loss_avg_dict[name] self.line_plotter.plot(name, name, name, self.plotting_step, loss_avg, color=ORANGE) for name in self.image_dict: self.vis.image(self.image_dict[name], opts=dict(title=name), win=self.window_elements.index(name)) self.plotting_step += 1 class VisdomLinePlotter(object): def __init__(self, vis, env_name='main'): self.vis = vis self.env = env_name self.plots = {} def plot(self, var_name, split_name, title_name, x, y, color): if var_name not in self.plots: self.plots[var_name] = self.vis.line(X=np.array([x, x]), Y=np.array([y, y]), env=self.env, opts=dict( legend=[split_name], title=title_name, linecolor=color, xlabel='Training steps', ylabel=var_name )) else: self.vis.line(X=np.array([x]), Y=np.array([y]), env=self.env, opts=dict( legend=[split_name], title=title_name, linecolor=color, xlabel='Training steps', ylabel=var_name ), win=self.plots[var_name], name=split_name, update = 'append')
2,658
38.102941
118
py
CPFN
CPFN-master/Utils/merging_utils.py
# Importation of packages import torch import numba import numpy as np def similarity_soft(spfn_labels, predicted_labels, point_indices): num_points_per_object, max_label_per_object = spfn_labels.size() nb_patches, num_points_per_patch, max_label_per_patch = predicted_labels.size() point2primitive_prediction = torch.zeros([num_points_per_object, nb_patches*max_label_per_patch+max_label_per_object]).to(predicted_labels.device) for b in range(nb_patches): predicted_labels_b = predicted_labels[b] point2primitive_prediction[point_indices[b],b*max_label_per_patch:(b+1)*max_label_per_patch] += predicted_labels_b point2primitive_prediction[:,(b+1)*max_label_per_patch:] = spfn_labels intersection_primitives = torch.mm(point2primitive_prediction.transpose(0,1),point2primitive_prediction) return intersection_primitives @numba.jit(numba.int64[:](numba.int64[:,:], numba.int64[:], numba.float64[:]), nopython=True) def heuristic_merging(pairs_id, patch_id, penalty_value): pairs_id1 = pairs_id[:,0] pairs_id2 = pairs_id[:,1] segment_id = np.arange(len(patch_id), dtype=numba.int64) patch_1hot = np.eye(patch_id.max()+1)[patch_id] while len(pairs_id1) > 0: pair_id1 = pairs_id1[np.argmax(penalty_value)] pair_id2 = pairs_id2[np.argmax(penalty_value)] segment_id[segment_id==segment_id[pair_id2]] = segment_id[pair_id1] selection_row = segment_id==segment_id[pair_id1] patch_1hot[selection_row] = np.sum(patch_1hot[selection_row], axis=0) intersection = np.sum(patch_1hot[pairs_id1] * patch_1hot[pairs_id2], axis=1) pairs_id1 = pairs_id1[intersection==0] pairs_id2 = pairs_id2[intersection==0] penalty_value = penalty_value[intersection==0] return segment_id def run_heuristic_solver(similarity_matrix, nb_patches, max_label_per_object, max_label_per_patch, threshold=0): # Building the Gurobi optimisation problem indices = np.where(similarity_matrix>threshold) penalty_array = np.stack((indices[0], indices[1], similarity_matrix[indices[0], indices[1]]), axis=1) penalty_array = penalty_array[penalty_array[:,0]<penalty_array[:,1]] # Heuristic patch_id = np.concatenate((np.repeat(np.arange(nb_patches), repeats=max_label_per_patch, axis=0), nb_patches*np.ones([max_label_per_object], dtype=int)), axis=0) glob_output_labels_heuristic = heuristic_merging(penalty_array[:,:2].astype(int), patch_id, penalty_array[:,2]) flag = np.diag(similarity_matrix) replacement_values = np.concatenate((np.tile(np.arange(-max_label_per_patch, 0), nb_patches), np.arange(-max_label_per_object, 0)), axis=0) glob_output_labels_heuristic[flag<threshold] = replacement_values[flag<threshold] _, glob_output_labels_heuristic = np.unique(glob_output_labels_heuristic, return_inverse=True) return glob_output_labels_heuristic def get_point_final(point2primitive_prediction, output_labels_heuristic): output_labels_heuristic = torch.eye(output_labels_heuristic.max()+1).to(output_labels_heuristic.device)[output_labels_heuristic.long()] output_labels_heuristic = output_labels_heuristic / (torch.sum(output_labels_heuristic, dim=0, keepdim=True) + 1e-10) final_output_labels_heuristic = torch.mm(point2primitive_prediction, output_labels_heuristic) return final_output_labels_heuristic
3,371
62.622642
165
py
CPFN
CPFN-master/Dataset/dataloaders.py
# Importation of packages import os import re import h5py import torch import pickle import random import numpy as np import pandas as pd from tqdm import tqdm import torch.utils.data as data # Importing Utils files from Utils import dataset_utils class Dataset_PatchSelection(data.Dataset): def __init__(self, csv_path, lowres_folder, highres_folder, scale, n_points=None, normalisation=True): self.lowres_folder = lowres_folder self.highres_folder = highres_folder self.scale = scale self.n_points = n_points self.normalisation = normalisation csv_raw = pd.read_csv(csv_path, delimiter=',', header=None)[0] self.hdf5_file_list = np.sort([file_ for file_ in csv_raw]) self.hdf5_file_list_lowres = [os.path.join(self.lowres_folder, file_.split('.')[0] + '.h5') for file_ in self.hdf5_file_list] self.hdf5_file_list_highres = [os.path.join(self.highres_folder, file_.split('.')[0] + '.h5') for file_ in self.hdf5_file_list] self.n_data = len(self.hdf5_file_list) self.preload_dataset() def preload_dataset(self): self.list_points = [] self.list_output_labels = [] self.list_shuffled_indices = [] print('Preloading Dataset:') for i in tqdm(range(self.n_data)): points, output_labels, shuffled_indices = dataset_utils.create_unit_data_from_hdf5_patch_selection(self.hdf5_file_list_lowres[i], self.hdf5_file_list_highres[i], normalisation=self.normalisation, scale=self.scale, n_points=self.n_points) self.list_points.append(points) self.list_output_labels.append(output_labels) self.list_shuffled_indices.append(shuffled_indices) def __getitem__(self, index): # find shape that contains the point with given global index points = self.list_points[index] points = torch.from_numpy(points).float() output_labels = self.list_output_labels[index] output_labels = torch.from_numpy(output_labels).long() shuffled_indices = self.list_shuffled_indices[index] shuffled_indices = torch.from_numpy(shuffled_indices).long() return points, output_labels, shuffled_indices def __len__(self): return self.n_data class Dataset_GlobalSPFN(data.Dataset): def __init__(self, n_max_global_instances, csv_path, lowres_folder, highres_folder, path_patches, noisy, n_points=8192, test=False, first_n=-1, fixed_order=False): self.n_max_global_instances = n_max_global_instances self.lowres_folder = lowres_folder self.highres_folder = highres_folder if not test: self.dir_files = self.lowres_folder self.path_patches = None else: self.dir_files = self.highres_folder self.path_patches = path_patches self.noisy = noisy self.n_points = n_points self.test = test self.first_n = first_n self.fixed_order = fixed_order csv_raw = pd.read_csv(csv_path, delimiter=',', header=None)[0] self.hdf5_file_list = np.sort(csv_raw) if not fixed_order: random.shuffle(self.hdf5_file_list) if first_n != -1: self.hdf5_file_list = self.hdf5_file_list[:first_n] self.n_data = len(self.hdf5_file_list) if not self.test: self.preload_dataset() def preload_dataset(self): print(f'Preloading Dataset:') for index in tqdm(range(self.__len__())): data_elt = self.fetch_data_at_index(index) if not hasattr(self, 'data_matrix'): self.data_matrix = {} for key in data_elt.keys(): trailing_ones = np.full([len(data_elt[key].shape)], 1, dtype=int) self.data_matrix[key] = np.tile(np.expand_dims(np.zeros_like(data_elt[key]), axis=0), [self.n_data, *trailing_ones]) for key in data_elt.keys(): self.data_matrix[key][index, ...] = data_elt[key] def fetch_data_at_index(self, i): file_ = self.hdf5_file_list[i] with h5py.File(os.path.join(self.dir_files, file_), 'r') as f: data = dataset_utils.create_unit_data_from_hdf5_spfn(f, self.n_max_global_instances, self.noisy, n_points=self.n_points, use_glob_features=False, use_loc_features=False, fixed_order=self.fixed_order, shuffle=not self.fixed_order) assert data is not None # assume data are all clean if self.test: if os.path.isfile(os.path.join(self.path_patches, file_.replace('.h5','_indices.npy'))): data['patch_centers'] = np.load(os.path.join(self.path_patches, file_.replace('.h5','_indices.npy')))[:,0] else: data['patch_centers'] = np.array([]) return data def __getitem__(self, index): # find shape that contains the point with given global index if not self.test: data = {} for key in self.data_matrix.keys(): data[key] = self.data_matrix[key][index,...] else: data = self.fetch_data_at_index(index) P = torch.from_numpy(data['P'].astype(np.float32)) normal_gt = torch.from_numpy(data['normal_gt'].astype(np.float32)) P_gt = torch.from_numpy(data['P_gt'].astype(np.float32)) I_gt = torch.from_numpy(data['I_gt'].astype(np.int64)) T_gt = torch.from_numpy(data['T_gt'].astype(np.int64)) plane_n_gt = torch.from_numpy(data['plane_n_gt'].astype(np.float32)) cylinder_axis_gt = torch.from_numpy(data['cylinder_axis_gt'].astype(np.float32)) cone_axis_gt = torch.from_numpy(data['cone_axis_gt'].astype(np.float32)) if self.test: patch_centers = torch.from_numpy(data['patch_centers'].astype(np.int64)) return P, normal_gt, P_gt, I_gt, T_gt, plane_n_gt, cylinder_axis_gt, cone_axis_gt, patch_centers else: return P, normal_gt, P_gt, I_gt, T_gt, plane_n_gt, cylinder_axis_gt, cone_axis_gt def __len__(self): return self.n_data class Dataset_TrainLocalSPFN(data.Dataset): def __init__(self, n_max_local_instances, csv_path, patch_folder, noisy, first_n=-1, fixed_order=False, lean=False): self.n_max_local_instances = n_max_local_instances self.noisy = noisy self.first_n = first_n self.fixed_order = fixed_order self.lean = lean self.patch_folder = patch_folder csv_raw = pd.read_csv(csv_path, delimiter=',', header=None)[0] self.hdf5_file_list = np.sort(csv_raw) self.n_data = 0 self.hdf5_file_list = np.sort([elt for elt in self.hdf5_file_list if self.check_dataset(elt)]) if not fixed_order: random.shuffle(self.hdf5_file_list) if self.lean: nb_patch_file = np.zeros([len(self.hdf5_file_list)]) for i, file_ in enumerate(self.hdf5_file_list): patch_files = [os.path.join(self.patch_folder, file_.split('.')[0], file_) for file_ in os.listdir(os.path.join(self.patch_folder, file_.split('.')[0])) if file_.split('.')[1] == 'h5'] nb_patch_file[i] = len(patch_files) self.nb_patch_file = nb_patch_file if first_n != -1: self.hdf5_file_list = self.hdf5_file_list[:first_n] if not self.lean: self.preload_data() def check_dataset(self, file_): cond = os.path.isdir(os.path.join(self.patch_folder, file_.split('.')[0])) if not cond: return False patch_files = [os.path.join(self.patch_folder, file_.split('.')[0], file_) for file_ in os.listdir(os.path.join(self.patch_folder, file_.split('.')[0])) if file_.split('.')[1] == 'h5'] self.n_data += len(patch_files) return True def preload_data(self): cpt = 0 print('Preloading Dataset:') for i, file_ in tqdm(enumerate(self.hdf5_file_list)): if i%100==0: print('%d / %d'%(i, len(self.hdf5_file_list))) patch_files = [os.path.join(self.patch_folder, file_.split('.')[0], file_) for file_ in os.listdir(os.path.join(self.patch_folder, file_.split('.')[0])) if file_.split('.')[1] == 'h5'] patch_files = np.sort(patch_files) for j in range(len(patch_files)): patch_file = os.path.join(self.patch_folder, file_.split('.')[0], file_.replace('.h5','_patch%d.h5'%j)) data_elt = self.fetch_data_at_index(patch_file) if not hasattr(self, 'data_matrix'): self.data_matrix = {} for key in data_elt.keys(): trailing_ones = np.full([len(data_elt[key].shape)], 1, dtype=int) self.data_matrix[key] = np.tile(np.expand_dims(np.zeros_like(data_elt[key]), axis=0), [self.n_data, *trailing_ones]) for key in data_elt.keys(): self.data_matrix[key][cpt, ...] = data_elt[key] cpt += 1 def fetch_data_at_index(self, patch_file): with h5py.File(patch_file, 'r') as f: data = dataset_utils.create_unit_data_from_hdf5_spfn(f, self.n_max_local_instances, noisy=self.noisy, n_points=None, use_glob_features=True, use_loc_features=True, fixed_order=self.fixed_order, shuffle=not self.fixed_order) assert data is not None # assume data are all clean return data def __getitem__(self, index): # find shape that contains the point with given global index if not self.lean: data = {} for key in self.data_matrix.keys(): data[key] = self.data_matrix[key][index, ...] else: cumsum = np.cumsum(self.nb_patch_file) index_ = np.where(index<cumsum)[0][0] file_ = self.hdf5_file_list[index_] if index_ == 0: j = index else: j = int(index - cumsum[index_-1]) patch_file = os.path.join(self.patch_folder, file_.split('.')[0], file_.replace('.h5', '_patch%d.h5' % j)) data = self.fetch_data_at_index(patch_file) P = torch.from_numpy(data['P'].astype(np.float32)) normal_gt = torch.from_numpy(data['normal_gt'].astype(np.float32)) P_gt = torch.from_numpy(data['P_gt'].astype(np.float32)) I_gt = torch.from_numpy(data['I_gt'].astype(np.int64)) T_gt = torch.from_numpy(data['T_gt'].astype(np.int64)) plane_n_gt = torch.from_numpy(data['plane_n_gt'].astype(np.float32)) cylinder_axis_gt = torch.from_numpy(data['cylinder_axis_gt'].astype(np.float32)) cone_axis_gt = torch.from_numpy(data['cone_axis_gt'].astype(np.float32)) glob_features = torch.from_numpy(data['glob_features'].astype(np.float32)) loc_features = torch.from_numpy(data['loc_features'].astype(np.float32)) output_tuple = (P, normal_gt, P_gt, I_gt, T_gt, plane_n_gt, cylinder_axis_gt, cone_axis_gt, glob_features, loc_features) return output_tuple def __len__(self): return self.n_data class Dataset_TestLocalSPFN(data.Dataset): def __init__(self, n_max_global_instances, n_max_local_instances, csv_path, dir_spfn, dir_lowres, dir_highres, dir_indices, noisy, first_n=-1, fixed_order=False): self.n_max_global_instances = n_max_global_instances self.n_max_local_instances = n_max_local_instances self.dir_spfn = dir_spfn self.dir_lowres = dir_lowres self.dir_highres = dir_highres self.dir_indices = dir_indices self.noisy = noisy self.first_n = first_n self.fixed_order = fixed_order csv_raw = pd.read_csv(csv_path, delimiter=',', header=None)[0] self.hdf5_file_list = np.sort(csv_raw) self.n_data = len(self.hdf5_file_list) self.hdf5_file_list_improvement = [elt for elt in self.hdf5_file_list if self.check_dataset(elt)] def check_dataset(self, file_): cond = os.path.isfile(os.path.join(self.dir_indices, file_.split('.')[0] + '_indices.npy')) if not cond: return False return True def fetch_data_at_index(self, patch_file): with h5py.File(patch_file, 'r') as f: data = dataset_utils.create_unit_data_from_hdf5_spfn(f, self.n_max_global_instances, self.noisy, n_points=None, fixed_order=True, shuffle=False) assert data is not None # assume data are all clean return data def __getitem__(self, index): # find shape that contains the point with given global index folder = self.hdf5_file_list[index] # Loading the highres file data_elt = self.fetch_data_at_index(os.path.join(self.dir_highres, folder)) P_global = data_elt['P'] normal_gt_global = data_elt['normal_gt'] P_gt_global = data_elt['P_gt'] I_gt_global = data_elt['I_gt'] T_gt_global = data_elt['T_gt'] plane_n_gt_global = data_elt['plane_n_gt'] cylinder_axis_gt_global = data_elt['cylinder_axis_gt'] cone_axis_gt_global = data_elt['cone_axis_gt'] if (folder in self.hdf5_file_list_improvement): # Loading the patch indices patch_indices = np.load(os.path.join(self.dir_indices, folder.replace('.h5', '_indices.npy'))) nb_patches, _ = patch_indices.shape P_unormalised = P_global[patch_indices] mean = np.mean(P_unormalised, axis=1, keepdims=True) P = P_unormalised - mean norm = np.linalg.norm(P, axis=2, keepdims=True).max(axis=1, keepdims=True) P = P / norm _, num_local_points, _ = P.shape normal_gt = normal_gt_global[patch_indices] I_gt = I_gt_global[patch_indices] P_gt = np.zeros((nb_patches,) + P_gt_global[:self.n_max_local_instances].shape) T_gt = np.zeros((nb_patches,) + T_gt_global[:self.n_max_local_instances].shape) plane_n_gt = np.zeros((nb_patches,) + plane_n_gt_global[:self.n_max_local_instances].shape) cylinder_axis_gt = np.zeros((nb_patches,) + cylinder_axis_gt_global[:self.n_max_local_instances].shape) cone_axis_gt = np.zeros((nb_patches,) + cone_axis_gt_global[:self.n_max_local_instances].shape) for i in range(nb_patches): flag = -1 in I_gt[i] unique_values, inverse_values = np.unique(I_gt[i], return_inverse=True) if flag: inverse_values = inverse_values - 1 I_gt[i] = inverse_values P_gt[i,np.arange(len(unique_values))] = P_gt_global[unique_values] T_gt[i, np.arange(len(unique_values))] = T_gt_global[unique_values] plane_n_gt[i, np.arange(len(unique_values))] = plane_n_gt_global[unique_values] cylinder_axis_gt[i, np.arange(len(unique_values))] = cylinder_axis_gt_global[unique_values] cone_axis_gt[i, np.arange(len(unique_values))] = cone_axis_gt_global[unique_values] # Loading the features glob_features = np.load(os.path.join(self.dir_spfn, folder.replace('.h5',''), 'global_feat.npy')) loc_features = np.load(os.path.join(self.dir_spfn, folder.replace('.h5',''), 'local_feat_full.npy')) list_glob_features = [] list_loc_features = [] for patch_id in range(nb_patches): list_glob_features.append(glob_features) list_loc_features.append(loc_features[:,patch_id]) glob_features = np.stack(list_glob_features, axis=0) loc_features = np.stack(list_loc_features, axis=0) else: nb_patches = 0 P = np.zeros([0, 8192, 3]).astype(np.float32) normal_gt = np.zeros([0, 8192, 3]).astype(np.float32) I_gt = np.zeros([0, 8192]).astype(np.int64) glob_features = np.zeros([0, 1024]).astype(np.float32) loc_features = np.zeros([0, 128]).astype(np.float32) patch_indices = np.zeros([0, 8192]).astype(np.int64) P_unormalised = np.zeros([0, 8192, 3]).astype(np.float32) P_gt = np.zeros([0, 21, 512, 3]).astype(np.float32) T_gt = np.zeros([0, 21]).astype(np.int64) plane_n_gt = np.zeros([0, 21, 3]).astype(np.float32) cylinder_axis_gt = np.zeros([0, 21, 3]).astype(np.float32) cone_axis_gt = np.zeros([0, 21, 3]).astype(np.float32) # Loading the SPFN output spfn_labels = np.load(os.path.join(self.dir_spfn, folder.replace('.h5', ''), 'object_seg.npy')) spfn_normals = np.load(os.path.join(self.dir_spfn, folder.replace('.h5', ''), 'object_normals.npy')) spfn_type = np.load(os.path.join(self.dir_spfn, folder.replace('.h5', ''), 'object_type.npy')) # Shuffling the output for i in range(nb_patches): perm = np.random.permutation(num_local_points) P[i] = P[i, perm] P_unormalised[i] = P_unormalised[i, perm] normal_gt[i] = normal_gt[i, perm] I_gt[i] = I_gt[i, perm] patch_indices[i] = patch_indices[i, perm] # Exporting all the data P = torch.from_numpy(P.astype(np.float32)) normal_gt = torch.from_numpy(normal_gt.astype(np.float32)) P_gt = torch.from_numpy(P_gt.astype(np.float32)) I_gt = torch.from_numpy(I_gt.astype(np.int64)) T_gt = torch.from_numpy(T_gt.astype(np.int64)) plane_n_gt = torch.from_numpy(plane_n_gt.astype(np.float32)) cylinder_axis_gt = torch.from_numpy(cylinder_axis_gt.astype(np.float32)) cone_axis_gt = torch.from_numpy(cone_axis_gt.astype(np.float32)) patch_indices = torch.from_numpy(patch_indices.astype(np.float32)) spfn_labels = torch.from_numpy(spfn_labels.astype(np.int64)) spfn_normals = torch.from_numpy(spfn_normals.astype(np.float32)) spfn_type = torch.from_numpy(spfn_type.astype(np.float32)) glob_features = torch.from_numpy(glob_features.astype(np.float32)) loc_features = torch.from_numpy(loc_features.astype(np.float32)) I_gt_global = torch.from_numpy(I_gt_global.astype(np.int64)) return P, normal_gt, P_gt_global, I_gt, T_gt_global, patch_indices, spfn_labels, spfn_normals, spfn_type, glob_features, loc_features, P_global, normal_gt_global, I_gt_global, plane_n_gt_global, cylinder_axis_gt_global, cone_axis_gt_global, P_unormalised, P_gt, T_gt, plane_n_gt, cylinder_axis_gt, cone_axis_gt def __len__(self): return self.n_data class RandomSampler(data.sampler.Sampler): def __init__(self, data_source, seed=None, identical_epochs=False): self.data_source = data_source self.seed = seed if self.seed is None: self.seed = np.random.randint(0, 2 ** 32 - 1, 1)[0] self.identical_epochs = identical_epochs self.rng = np.random.RandomState(self.seed) self.total_samples_count = len(self.data_source) def __iter__(self): if self.identical_epochs: self.rng.seed(self.seed) return iter(self.rng.choice(self.total_samples_count, size=self.total_samples_count, replace=False)) def __len__(self): return self.total_samples_count class Sampler(data.sampler.Sampler): def __init__(self, data_source): self.data_source = data_source self.total_samples_count = len(self.data_source) def __iter__(self): return iter(np.arange(0, self.total_samples_count)) def __len__(self): return self.total_samples_count
19,560
54.729345
318
py
CPFN
CPFN-master/PointNet2/pn2_network.py
# Importation of packages import os import sys import torch import numpy as np from SPFN.losses_implementation import compute_all_losses from PointNet2.pointnet2_ops.modules.pointset_abstraction import PointsetAbstraction from PointNet2.pointnet2_ops.modules.pointset_feature_propagation import PointsetFeaturePropagation class PointNet2(torch.nn.Module): def __init__(self, dim_input=3, dim_pos=3, output_sizes=[16], use_glob_features=False, use_loc_features=False, features_extractor=False): super(PointNet2, self).__init__() self.dim_pos = dim_pos self.use_glob_features = use_glob_features self.use_loc_features = use_loc_features self.features_extractor = features_extractor # Encoding stage self.sa1 = PointsetAbstraction(num_points=512, dim_pos=dim_pos, dim_feats=dim_input-dim_pos, radius_list=[0.2], num_samples_list=[64], mlp_list=[[64,64,128]], group_all=False) self.sa2 = PointsetAbstraction(num_points=128, dim_pos=dim_pos, dim_feats=128, radius_list=[0.4], num_samples_list=[64], mlp_list=[[128,128,256]], group_all=False) self.sa3 = PointsetAbstraction(num_points=None, dim_pos=dim_pos, dim_feats=256, radius_list=None, num_samples_list=None, mlp_list=[256, 512, 1024], group_all=True) # Decoding stage offset = 0 if self.use_glob_features: offset += 1024 if self.use_loc_features: offset += 128 self.sfp1 = PointsetFeaturePropagation(dim_feats=1024+offset+256, mlp=[256,256]) self.sfp2 = PointsetFeaturePropagation(dim_feats=256+128, mlp=[256,128]) self.sfp3 = PointsetFeaturePropagation(dim_feats=128+dim_input-dim_pos, mlp=[128,128,128]) # FC stage self.fc1 = torch.nn.Conv1d(128, 128, 1) if not self.features_extractor: self.bn1 = torch.nn.BatchNorm1d(128) self.fc2 = torch.nn.ModuleList() for output_size in output_sizes: self.fc2.append(torch.nn.Conv1d(128, output_size, 1)) def forward(self, x, glob_features=None, loc_features=None, fast=True): x = x.transpose(2,1) batch_size = x.shape[0] input_pos = x[:,:self.dim_pos,:] if x.shape[1] > self.dim_pos: input_feats = x[:,self.dim_pos:,:] else: input_feats = None # Encoding stage l1_pos, l1_feats = self.sa1(input_pos, input_feats, fast=fast) l2_pos, l2_feats = self.sa2(l1_pos, l1_feats, fast=fast) l3_pos, l3_feats = self.sa3(l2_pos, l2_feats, fast=fast) # Adding additional features if self.use_glob_features: l3_feats = torch.cat((l3_feats, glob_features.unsqueeze(2)), dim=1) if self.use_loc_features: l3_feats = torch.cat((l3_feats, loc_features.unsqueeze(2)), dim=1) # Decoding stage l4_feats = self.sfp1(l2_pos, l3_pos, l2_feats, l3_feats, fast=fast) l5_feats = self.sfp2(l1_pos, l2_pos, l1_feats, l4_feats, fast=fast) l6_feats = self.sfp3(input_pos, l1_pos, input_feats, l5_feats, fast=fast) # FC stage output_feat = self.fc1(l6_feats) if not self.features_extractor: output_feat = torch.nn.functional.relu(self.bn1(output_feat)) output_feat = torch.nn.functional.dropout(output_feat, p=0.5) results = [] for fc2_layer in self.fc2: result = fc2_layer(output_feat) result = result.transpose(1,2) results.append(result) results.append(l3_feats) results.append(output_feat) return results else: return l3_feats, output_feat
3,692
49.589041
183
py
CPFN
CPFN-master/PointNet2/pointnet2_ops/setup.py
import os import glob import setuptools from torch.utils.cpp_extension import BuildExtension, CUDAExtension sources = glob.glob("cuda_ops/src/*.cpp") + glob.glob("cuda_ops/src/*.cu") headers = os.path.join(os.path.dirname(os.path.abspath(__file__)), 'cuda_ops/include') setuptools.setup( name="pointnet2_ops", version="1.0", description="PointNet++ modules", packages=setuptools.find_packages(), include_package_data=True, classifiers=[ "Programming Language :: Python :: 3", "Operating System :: OS Independent", ], python_requires='>=3.7', ext_modules=[ CUDAExtension( name='cuda_ops', sources=sources, extra_compile_args={ "cxx": ["-O2", "-I{}".format(headers)], "nvcc": ["-O2", "-I{}".format(headers)], }, ) ], cmdclass={"build_ext": BuildExtension}, )
915
28.548387
86
py