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
keras
keras-master/keras/applications/mobilenet_v2.py
# Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """MobileNet v2 models for Keras. MobileNetV2 is a general architecture and can be used for multiple use cases. Depending on the use case, it can use different input layer size and different width factors. This allows different width models to reduce the number of multiply-adds and thereby reduce inference cost on mobile devices. MobileNetV2 is very similar to the original MobileNet, except that it uses inverted residual blocks with bottlenecking features. It has a drastically lower parameter count than the original MobileNet. MobileNets support any input size greater than 32 x 32, with larger image sizes offering better performance. The number of parameters and number of multiply-adds can be modified by using the `alpha` parameter, which increases/decreases the number of filters in each layer. By altering the image size and `alpha` parameter, all 22 models from the paper can be built, with ImageNet weights provided. The paper demonstrates the performance of MobileNets using `alpha` values of 1.0 (also called 100 % MobileNet), 0.35, 0.5, 0.75, 1.0, 1.3, and 1.4 For each of these `alpha` values, weights for 5 different input image sizes are provided (224, 192, 160, 128, and 96). The following table describes the performance of MobileNet on various input sizes: ------------------------------------------------------------------------ MACs stands for Multiply Adds Classification Checkpoint|MACs (M)|Parameters (M)|Top 1 Accuracy|Top 5 Accuracy --------------------------|------------|---------------|---------|----|--------- | [mobilenet_v2_1.4_224] | 582 | 6.06 | 75.0 | 92.5 | | [mobilenet_v2_1.3_224] | 509 | 5.34 | 74.4 | 92.1 | | [mobilenet_v2_1.0_224] | 300 | 3.47 | 71.8 | 91.0 | | [mobilenet_v2_1.0_192] | 221 | 3.47 | 70.7 | 90.1 | | [mobilenet_v2_1.0_160] | 154 | 3.47 | 68.8 | 89.0 | | [mobilenet_v2_1.0_128] | 99 | 3.47 | 65.3 | 86.9 | | [mobilenet_v2_1.0_96] | 56 | 3.47 | 60.3 | 83.2 | | [mobilenet_v2_0.75_224] | 209 | 2.61 | 69.8 | 89.6 | | [mobilenet_v2_0.75_192] | 153 | 2.61 | 68.7 | 88.9 | | [mobilenet_v2_0.75_160] | 107 | 2.61 | 66.4 | 87.3 | | [mobilenet_v2_0.75_128] | 69 | 2.61 | 63.2 | 85.3 | | [mobilenet_v2_0.75_96] | 39 | 2.61 | 58.8 | 81.6 | | [mobilenet_v2_0.5_224] | 97 | 1.95 | 65.4 | 86.4 | | [mobilenet_v2_0.5_192] | 71 | 1.95 | 63.9 | 85.4 | | [mobilenet_v2_0.5_160] | 50 | 1.95 | 61.0 | 83.2 | | [mobilenet_v2_0.5_128] | 32 | 1.95 | 57.7 | 80.8 | | [mobilenet_v2_0.5_96] | 18 | 1.95 | 51.2 | 75.8 | | [mobilenet_v2_0.35_224] | 59 | 1.66 | 60.3 | 82.9 | | [mobilenet_v2_0.35_192] | 43 | 1.66 | 58.2 | 81.2 | | [mobilenet_v2_0.35_160] | 30 | 1.66 | 55.7 | 79.1 | | [mobilenet_v2_0.35_128] | 20 | 1.66 | 50.8 | 75.0 | | [mobilenet_v2_0.35_96] | 11 | 1.66 | 45.5 | 70.4 | Reference: - [MobileNetV2: Inverted Residuals and Linear Bottlenecks]( https://arxiv.org/abs/1801.04381) (CVPR 2018) """ from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils import tensorflow.compat.v2 as tf from tensorflow.python.platform import tf_logging as logging from tensorflow.python.util.tf_export import keras_export BASE_WEIGHT_PATH = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/mobilenet_v2/') layers = None @keras_export('keras.applications.mobilenet_v2.MobileNetV2', 'keras.applications.MobileNetV2') def MobileNetV2(input_shape=None, alpha=1.0, include_top=True, weights='imagenet', input_tensor=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): """Instantiates the MobileNetV2 architecture. MobileNetV2 is very similar to the original MobileNet, except that it uses inverted residual blocks with bottlenecking features. It has a drastically lower parameter count than the original MobileNet. MobileNets support any input size greater than 32 x 32, with larger image sizes offering better performance. Reference: - [MobileNetV2: Inverted Residuals and Linear Bottlenecks]( https://arxiv.org/abs/1801.04381) (CVPR 2018) This function returns a Keras image classification model, optionally loaded with weights pre-trained on ImageNet. For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For MobileNetV2, call `tf.keras.applications.mobilenet_v2.preprocess_input` on your inputs before passing them to the model. `mobilenet_v2.preprocess_input` will scale input pixels between -1 and 1. Args: input_shape: Optional shape tuple, to be specified if you would like to use a model with an input image resolution that is not (224, 224, 3). It should have exactly 3 inputs channels (224, 224, 3). You can also omit this option if you would like to infer input_shape from an input_tensor. If you choose to include both input_tensor and input_shape then input_shape will be used if they match, if the shapes do not match then we will throw an error. E.g. `(160, 160, 3)` would be one valid value. alpha: Float, larger than zero, controls the width of the network. This is known as the width multiplier in the MobileNetV2 paper, but the name is kept for consistency with `applications.MobileNetV1` model in Keras. - If `alpha` < 1.0, proportionally decreases the number of filters in each layer. - If `alpha` > 1.0, proportionally increases the number of filters in each layer. - If `alpha` = 1.0, default number of filters from the paper are used at each layer. include_top: Boolean, whether to include the fully-connected layer at the top of the network. Defaults to `True`. weights: String, one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. pooling: String, optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Optional integer number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. **kwargs: For backwards compatibility only. Returns: A `keras.Model` instance. """ global layers if 'layers' in kwargs: layers = kwargs.pop('layers') else: layers = VersionAwareLayers() if kwargs: raise ValueError(f'Unknown argument(s): {kwargs}') if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded. ' f'Received `weights={weights}`') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError( 'If using `weights` as `"imagenet"` with `include_top` ' f'as true, `classes` should be 1000. Received `classes={classes}`') # Determine proper input shape and default size. # If both input_shape and input_tensor are used, they should match if input_shape is not None and input_tensor is not None: try: is_input_t_tensor = backend.is_keras_tensor(input_tensor) except ValueError: try: is_input_t_tensor = backend.is_keras_tensor( layer_utils.get_source_inputs(input_tensor)) except ValueError: raise ValueError( f'input_tensor: {input_tensor}' 'is not type input_tensor. ' f'Received `type(input_tensor)={type(input_tensor)}`' ) if is_input_t_tensor: if backend.image_data_format() == 'channels_first': if backend.int_shape(input_tensor)[1] != input_shape[1]: raise ValueError('input_shape[1] must equal shape(input_tensor)[1] ' 'when `image_data_format` is `channels_first`; ' 'Received `input_tensor.shape=' f'{input_tensor.shape}`' f', `input_shape={input_shape}`') else: if backend.int_shape(input_tensor)[2] != input_shape[1]: raise ValueError( 'input_tensor.shape[2] must equal input_shape[1]; ' 'Received `input_tensor.shape=' f'{input_tensor.shape}`, ' f'`input_shape={input_shape}`') else: raise ValueError('input_tensor is not a Keras tensor; ' f'Received `input_tensor={input_tensor}`') # If input_shape is None, infer shape from input_tensor. if input_shape is None and input_tensor is not None: try: backend.is_keras_tensor(input_tensor) except ValueError: raise ValueError('input_tensor must be a valid Keras tensor type; ' f'Received {input_tensor} of type {type(input_tensor)}') if input_shape is None and not backend.is_keras_tensor(input_tensor): default_size = 224 elif input_shape is None and backend.is_keras_tensor(input_tensor): if backend.image_data_format() == 'channels_first': rows = backend.int_shape(input_tensor)[2] cols = backend.int_shape(input_tensor)[3] else: rows = backend.int_shape(input_tensor)[1] cols = backend.int_shape(input_tensor)[2] if rows == cols and rows in [96, 128, 160, 192, 224]: default_size = rows else: default_size = 224 # If input_shape is None and no input_tensor elif input_shape is None: default_size = 224 # If input_shape is not None, assume default size. else: if backend.image_data_format() == 'channels_first': rows = input_shape[1] cols = input_shape[2] else: rows = input_shape[0] cols = input_shape[1] if rows == cols and rows in [96, 128, 160, 192, 224]: default_size = rows else: default_size = 224 input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=default_size, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if backend.image_data_format() == 'channels_last': row_axis, col_axis = (0, 1) else: row_axis, col_axis = (1, 2) rows = input_shape[row_axis] cols = input_shape[col_axis] if weights == 'imagenet': if alpha not in [0.35, 0.50, 0.75, 1.0, 1.3, 1.4]: raise ValueError('If imagenet weights are being loaded, ' 'alpha must be one of `0.35`, `0.50`, `0.75`, ' '`1.0`, `1.3` or `1.4` only;' f' Received `alpha={alpha}`') if rows != cols or rows not in [96, 128, 160, 192, 224]: rows = 224 logging.warning('`input_shape` is undefined or non-square, ' 'or `rows` is not in [96, 128, 160, 192, 224]. ' 'Weights for input shape (224, 224) will be ' 'loaded as the default.') if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 first_block_filters = _make_divisible(32 * alpha, 8) x = layers.Conv2D( first_block_filters, kernel_size=3, strides=(2, 2), padding='same', use_bias=False, name='Conv1')(img_input) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name='bn_Conv1')( x) x = layers.ReLU(6., name='Conv1_relu')(x) 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=2, 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 = _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 = _inverted_res_block( x, filters=160, alpha=alpha, stride=2, 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 = _inverted_res_block( x, filters=320, alpha=alpha, stride=1, expansion=6, block_id=16) # no alpha applied to last conv as stated in the paper: # if the width multiplier is greater than 1 we increase the number of output # channels. if alpha > 1.0: last_block_filters = _make_divisible(1280 * alpha, 8) else: last_block_filters = 1280 x = layers.Conv2D( last_block_filters, kernel_size=1, use_bias=False, name='Conv_1')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name='Conv_1_bn')( x) x = layers.ReLU(6., name='out_relu')(x) if include_top: x = layers.GlobalAveragePooling2D()(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account any potential predecessors of # `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='mobilenetv2_%0.2f_%s' % (alpha, rows)) # Load weights. if weights == 'imagenet': if include_top: model_name = ('mobilenet_v2_weights_tf_dim_ordering_tf_kernels_' + str(float(alpha)) + '_' + str(rows) + '.h5') weight_path = BASE_WEIGHT_PATH + model_name weights_path = data_utils.get_file( model_name, weight_path, cache_subdir='models') else: model_name = ('mobilenet_v2_weights_tf_dim_ordering_tf_kernels_' + str(float(alpha)) + '_' + str(rows) + '_no_top' + '.h5') weight_path = BASE_WEIGHT_PATH + model_name weights_path = data_utils.get_file( model_name, weight_path, cache_subdir='models') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model def _inverted_res_block(inputs, expansion, stride, alpha, filters, block_id): """Inverted ResNet block.""" channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 in_channels = backend.int_shape(inputs)[channel_axis] pointwise_conv_filters = int(filters * alpha) # Ensure the number of filters on the last 1x1 convolution is divisible by 8. pointwise_filters = _make_divisible(pointwise_conv_filters, 8) x = inputs prefix = 'block_{}_'.format(block_id) if block_id: # Expand with a pointwise 1x1 convolution. x = layers.Conv2D( expansion * in_channels, kernel_size=1, padding='same', use_bias=False, activation=None, name=prefix + 'expand')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name=prefix + 'expand_BN')( x) x = layers.ReLU(6., name=prefix + 'expand_relu')(x) else: prefix = 'expanded_conv_' # Depthwise 3x3 convolution. if stride == 2: x = layers.ZeroPadding2D( padding=imagenet_utils.correct_pad(x, 3), name=prefix + 'pad')(x) x = layers.DepthwiseConv2D( kernel_size=3, strides=stride, activation=None, use_bias=False, padding='same' if stride == 1 else 'valid', name=prefix + 'depthwise')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name=prefix + 'depthwise_BN')( x) x = layers.ReLU(6., name=prefix + 'depthwise_relu')(x) # Project wiht a pointwise 1x1 convolution. x = layers.Conv2D( pointwise_filters, kernel_size=1, padding='same', use_bias=False, activation=None, name=prefix + 'project')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name=prefix + 'project_BN')( x) if in_channels == pointwise_filters and stride == 1: return layers.Add(name=prefix + 'add')([inputs, x]) return x 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) # Make sure that round down does not go down by more than 10%. if new_v < 0.9 * v: new_v += divisor return new_v @keras_export('keras.applications.mobilenet_v2.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input(x, data_format=data_format, mode='tf') @keras_export('keras.applications.mobilenet_v2.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
20,697
38.05283
87
py
keras
keras-master/keras/applications/efficientnet.py
# Copyright 2019 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name # pylint: disable=missing-docstring """EfficientNet models for Keras. Reference: - [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks]( https://arxiv.org/abs/1905.11946) (ICML 2019) """ import tensorflow.compat.v2 as tf import copy import math from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export BASE_WEIGHTS_PATH = 'https://storage.googleapis.com/keras-applications/' WEIGHTS_HASHES = { 'b0': ('902e53a9f72be733fc0bcb005b3ebbac', '50bc09e76180e00e4465e1a485ddc09d'), 'b1': ('1d254153d4ab51201f1646940f018540', '74c4e6b3e1f6a1eea24c589628592432'), 'b2': ('b15cce36ff4dcbd00b6dd88e7857a6ad', '111f8e2ac8aa800a7a99e3239f7bfb39'), 'b3': ('ffd1fdc53d0ce67064dc6a9c7960ede0', 'af6d107764bb5b1abb91932881670226'), 'b4': ('18c95ad55216b8f92d7e70b3a046e2fc', 'ebc24e6d6c33eaebbd558eafbeedf1ba'), 'b5': ('ace28f2a6363774853a83a0b21b9421a', '38879255a25d3c92d5e44e04ae6cec6f'), 'b6': ('165f6e37dce68623721b423839de8be5', '9ecce42647a20130c1f39a5d4cb75743'), 'b7': ('8c03f828fec3ef71311cd463b6759d99', 'cbcfe4450ddf6f3ad90b1b398090fe4a'), } DEFAULT_BLOCKS_ARGS = [{ 'kernel_size': 3, 'repeats': 1, 'filters_in': 32, 'filters_out': 16, 'expand_ratio': 1, 'id_skip': True, 'strides': 1, 'se_ratio': 0.25 }, { 'kernel_size': 3, 'repeats': 2, 'filters_in': 16, 'filters_out': 24, 'expand_ratio': 6, 'id_skip': True, 'strides': 2, 'se_ratio': 0.25 }, { 'kernel_size': 5, 'repeats': 2, 'filters_in': 24, 'filters_out': 40, 'expand_ratio': 6, 'id_skip': True, 'strides': 2, 'se_ratio': 0.25 }, { 'kernel_size': 3, 'repeats': 3, 'filters_in': 40, 'filters_out': 80, 'expand_ratio': 6, 'id_skip': True, 'strides': 2, 'se_ratio': 0.25 }, { 'kernel_size': 5, 'repeats': 3, 'filters_in': 80, 'filters_out': 112, 'expand_ratio': 6, 'id_skip': True, 'strides': 1, 'se_ratio': 0.25 }, { 'kernel_size': 5, 'repeats': 4, 'filters_in': 112, 'filters_out': 192, 'expand_ratio': 6, 'id_skip': True, 'strides': 2, 'se_ratio': 0.25 }, { 'kernel_size': 3, 'repeats': 1, 'filters_in': 192, 'filters_out': 320, 'expand_ratio': 6, 'id_skip': True, 'strides': 1, 'se_ratio': 0.25 }] CONV_KERNEL_INITIALIZER = { 'class_name': 'VarianceScaling', 'config': { 'scale': 2.0, 'mode': 'fan_out', 'distribution': 'truncated_normal' } } DENSE_KERNEL_INITIALIZER = { 'class_name': 'VarianceScaling', 'config': { 'scale': 1. / 3., 'mode': 'fan_out', 'distribution': 'uniform' } } layers = VersionAwareLayers() BASE_DOCSTRING = """Instantiates the {name} architecture. Reference: - [EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks]( https://arxiv.org/abs/1905.11946) (ICML 2019) This function returns a Keras image classification model, optionally loaded with weights pre-trained on ImageNet. For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For EfficientNet, input preprocessing is included as part of the model (as a `Rescaling` layer), and thus `tf.keras.applications.efficientnet.preprocess_input` is actually a pass-through function. EfficientNet models expect their inputs to be float tensors of pixels with values in the [0-255] range. Args: include_top: Whether to include the fully-connected layer at the top of the network. Defaults to True. weights: One of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. Defaults to 'imagenet'. input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: Optional shape tuple, only to be specified if `include_top` is False. It should have exactly 3 inputs channels. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. Defaults to None. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. Defaults to 1000 (number of ImageNet classes). classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. Defaults to 'softmax'. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ def EfficientNet( width_coefficient, depth_coefficient, default_size, dropout_rate=0.2, drop_connect_rate=0.2, depth_divisor=8, activation='swish', blocks_args='default', model_name='efficientnet', include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the EfficientNet architecture using given scaling coefficients. Args: width_coefficient: float, scaling coefficient for network width. depth_coefficient: float, scaling coefficient for network depth. default_size: integer, default input image size. dropout_rate: float, dropout rate before final classifier layer. drop_connect_rate: float, dropout rate at skip connections. depth_divisor: integer, a unit of network width. activation: activation function. blocks_args: list of dicts, parameters to construct block modules. model_name: string, model name. include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False. It should have exactly 3 inputs channels. pooling: optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. Returns: A `keras.Model` instance. Raises: ValueError: in case of invalid argument for `weights`, or invalid input shape. ValueError: if `classifier_activation` is not `softmax` or `None` when using a pretrained top layer. """ if blocks_args == 'default': blocks_args = DEFAULT_BLOCKS_ARGS if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded.') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top`' ' as true, `classes` should be 1000') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=default_size, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 def round_filters(filters, divisor=depth_divisor): """Round number of filters based on depth multiplier.""" filters *= width_coefficient new_filters = max(divisor, int(filters + divisor / 2) // divisor * divisor) # Make sure that round down does not go down by more than 10%. if new_filters < 0.9 * filters: new_filters += divisor return int(new_filters) def round_repeats(repeats): """Round number of repeats based on depth multiplier.""" return int(math.ceil(depth_coefficient * repeats)) # Build stem x = img_input x = layers.Rescaling(1. / 255.)(x) x = layers.Normalization(axis=bn_axis)(x) x = layers.ZeroPadding2D( padding=imagenet_utils.correct_pad(x, 3), name='stem_conv_pad')(x) x = layers.Conv2D( round_filters(32), 3, strides=2, padding='valid', use_bias=False, kernel_initializer=CONV_KERNEL_INITIALIZER, name='stem_conv')(x) x = layers.BatchNormalization(axis=bn_axis, name='stem_bn')(x) x = layers.Activation(activation, name='stem_activation')(x) # Build blocks blocks_args = copy.deepcopy(blocks_args) b = 0 blocks = float(sum(round_repeats(args['repeats']) for args in blocks_args)) for (i, args) in enumerate(blocks_args): assert args['repeats'] > 0 # Update block input and output filters based on depth multiplier. args['filters_in'] = round_filters(args['filters_in']) args['filters_out'] = round_filters(args['filters_out']) for j in range(round_repeats(args.pop('repeats'))): # The first block needs to take care of stride and filter size increase. if j > 0: args['strides'] = 1 args['filters_in'] = args['filters_out'] x = block( x, activation, drop_connect_rate * b / blocks, name='block{}{}_'.format(i + 1, chr(j + 97)), **args) b += 1 # Build top x = layers.Conv2D( round_filters(1280), 1, padding='same', use_bias=False, kernel_initializer=CONV_KERNEL_INITIALIZER, name='top_conv')(x) x = layers.BatchNormalization(axis=bn_axis, name='top_bn')(x) x = layers.Activation(activation, name='top_activation')(x) if include_top: x = layers.GlobalAveragePooling2D(name='avg_pool')(x) if dropout_rate > 0: x = layers.Dropout(dropout_rate, name='top_dropout')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense( classes, activation=classifier_activation, kernel_initializer=DENSE_KERNEL_INITIALIZER, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D(name='avg_pool')(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D(name='max_pool')(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name=model_name) # Load weights. if weights == 'imagenet': if include_top: file_suffix = '.h5' file_hash = WEIGHTS_HASHES[model_name[-2:]][0] else: file_suffix = '_notop.h5' file_hash = WEIGHTS_HASHES[model_name[-2:]][1] file_name = model_name + file_suffix weights_path = data_utils.get_file( file_name, BASE_WEIGHTS_PATH + file_name, cache_subdir='models', file_hash=file_hash) model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model def block(inputs, activation='swish', drop_rate=0., name='', filters_in=32, filters_out=16, kernel_size=3, strides=1, expand_ratio=1, se_ratio=0., id_skip=True): """An inverted residual block. Args: inputs: input tensor. activation: activation function. drop_rate: float between 0 and 1, fraction of the input units to drop. name: string, block label. filters_in: integer, the number of input filters. filters_out: integer, the number of output filters. kernel_size: integer, the dimension of the convolution window. strides: integer, the stride of the convolution. expand_ratio: integer, scaling coefficient for the input filters. se_ratio: float between 0 and 1, fraction to squeeze the input filters. id_skip: boolean. Returns: output tensor for the block. """ bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 # Expansion phase filters = filters_in * expand_ratio if expand_ratio != 1: x = layers.Conv2D( filters, 1, padding='same', use_bias=False, kernel_initializer=CONV_KERNEL_INITIALIZER, name=name + 'expand_conv')( inputs) x = layers.BatchNormalization(axis=bn_axis, name=name + 'expand_bn')(x) x = layers.Activation(activation, name=name + 'expand_activation')(x) else: x = inputs # Depthwise Convolution if strides == 2: x = layers.ZeroPadding2D( padding=imagenet_utils.correct_pad(x, kernel_size), name=name + 'dwconv_pad')(x) conv_pad = 'valid' else: conv_pad = 'same' x = layers.DepthwiseConv2D( kernel_size, strides=strides, padding=conv_pad, use_bias=False, depthwise_initializer=CONV_KERNEL_INITIALIZER, name=name + 'dwconv')(x) x = layers.BatchNormalization(axis=bn_axis, name=name + 'bn')(x) x = layers.Activation(activation, name=name + 'activation')(x) # Squeeze and Excitation phase if 0 < se_ratio <= 1: filters_se = max(1, int(filters_in * se_ratio)) se = layers.GlobalAveragePooling2D(name=name + 'se_squeeze')(x) if bn_axis == 1: se_shape = (filters, 1, 1) else: se_shape = (1, 1, filters) se = layers.Reshape(se_shape, name=name + 'se_reshape')(se) se = layers.Conv2D( filters_se, 1, padding='same', activation=activation, kernel_initializer=CONV_KERNEL_INITIALIZER, name=name + 'se_reduce')( se) se = layers.Conv2D( filters, 1, padding='same', activation='sigmoid', kernel_initializer=CONV_KERNEL_INITIALIZER, name=name + 'se_expand')(se) x = layers.multiply([x, se], name=name + 'se_excite') # Output phase x = layers.Conv2D( filters_out, 1, padding='same', use_bias=False, kernel_initializer=CONV_KERNEL_INITIALIZER, name=name + 'project_conv')(x) x = layers.BatchNormalization(axis=bn_axis, name=name + 'project_bn')(x) if id_skip and strides == 1 and filters_in == filters_out: if drop_rate > 0: x = layers.Dropout( drop_rate, noise_shape=(None, 1, 1, 1), name=name + 'drop')(x) x = layers.add([x, inputs], name=name + 'add') return x @keras_export('keras.applications.efficientnet.EfficientNetB0', 'keras.applications.EfficientNetB0') def EfficientNetB0(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.0, 1.0, 224, 0.2, model_name='efficientnetb0', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB1', 'keras.applications.EfficientNetB1') def EfficientNetB1(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.0, 1.1, 240, 0.2, model_name='efficientnetb1', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB2', 'keras.applications.EfficientNetB2') def EfficientNetB2(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.1, 1.2, 260, 0.3, model_name='efficientnetb2', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB3', 'keras.applications.EfficientNetB3') def EfficientNetB3(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.2, 1.4, 300, 0.3, model_name='efficientnetb3', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB4', 'keras.applications.EfficientNetB4') def EfficientNetB4(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.4, 1.8, 380, 0.4, model_name='efficientnetb4', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB5', 'keras.applications.EfficientNetB5') def EfficientNetB5(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.6, 2.2, 456, 0.4, model_name='efficientnetb5', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB6', 'keras.applications.EfficientNetB6') def EfficientNetB6(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 1.8, 2.6, 528, 0.5, model_name='efficientnetb6', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) @keras_export('keras.applications.efficientnet.EfficientNetB7', 'keras.applications.EfficientNetB7') def EfficientNetB7(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): return EfficientNet( 2.0, 3.1, 600, 0.5, model_name='efficientnetb7', include_top=include_top, weights=weights, input_tensor=input_tensor, input_shape=input_shape, pooling=pooling, classes=classes, classifier_activation=classifier_activation, **kwargs) EfficientNetB0.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB0') EfficientNetB1.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB1') EfficientNetB2.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB2') EfficientNetB3.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB3') EfficientNetB4.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB4') EfficientNetB5.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB5') EfficientNetB6.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB6') EfficientNetB7.__doc__ = BASE_DOCSTRING.format(name='EfficientNetB7') @keras_export('keras.applications.efficientnet.preprocess_input') def preprocess_input(x, data_format=None): # pylint: disable=unused-argument """A placeholder method for backward compatibility. The preprocessing logic has been included in the efficientnet model implementation. Users are no longer required to call this method to normalize the input data. This method does nothing and only kept as a placeholder to align the API surface between old and new version of model. Args: x: A floating point `numpy.array` or a `tf.Tensor`. data_format: Optional data format of the image tensor/array. Defaults to None, in which case the global setting `tf.keras.backend.image_data_format()` is used (unless you changed it, it defaults to "channels_last").{mode} Returns: Unchanged `numpy.array` or `tf.Tensor`. """ return x @keras_export('keras.applications.efficientnet.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
25,089
31.5
87
py
keras
keras-master/keras/applications/resnet.py
# Copyright 2015 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """ResNet models for Keras. Reference: - [Deep Residual Learning for Image Recognition]( https://arxiv.org/abs/1512.03385) (CVPR 2015) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export BASE_WEIGHTS_PATH = ( 'https://storage.googleapis.com/tensorflow/keras-applications/resnet/') WEIGHTS_HASHES = { 'resnet50': ('2cb95161c43110f7111970584f804107', '4d473c1dd8becc155b73f8504c6f6626'), 'resnet101': ('f1aeb4b969a6efcfb50fad2f0c20cfc5', '88cf7a10940856eca736dc7b7e228a21'), 'resnet152': ('100835be76be38e30d865e96f2aaae62', 'ee4c566cf9a93f14d82f913c2dc6dd0c'), 'resnet50v2': ('3ef43a0b657b3be2300d5770ece849e0', 'fac2f116257151a9d068a22e544a4917'), 'resnet101v2': ('6343647c601c52e1368623803854d971', 'c0ed64b8031c3730f411d2eb4eea35b5'), 'resnet152v2': ('a49b44d1979771252814e80f8ec446f9', 'ed17cf2e0169df9d443503ef94b23b33'), 'resnext50': ('67a5b30d522ed92f75a1f16eef299d1a', '62527c363bdd9ec598bed41947b379fc'), 'resnext101': ('34fb605428fcc7aa4d62f44404c11509', '0f678c91647380debd923963594981b3') } layers = None def ResNet(stack_fn, preact, use_bias, model_name='resnet', include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): """Instantiates the ResNet, ResNetV2, and ResNeXt architecture. Args: stack_fn: a function that returns output tensor for the stacked residual blocks. preact: whether to use pre-activation or not (True for ResNetV2, False for ResNet and ResNeXt). use_bias: whether to use biases for convolutional layers or not (True for ResNet and ResNetV2, False for ResNeXt). model_name: string, model name. include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `channels_last` data format) or `(3, 224, 224)` (with `channels_first` data format). It should have exactly 3 inputs channels. pooling: optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. **kwargs: For backwards compatibility only. Returns: A `keras.Model` instance. """ global layers if 'layers' in kwargs: layers = kwargs.pop('layers') else: layers = VersionAwareLayers() if kwargs: raise ValueError('Unknown argument(s): %s' % (kwargs,)) if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded.') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top`' ' as true, `classes` should be 1000') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=224, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 x = layers.ZeroPadding2D( padding=((3, 3), (3, 3)), name='conv1_pad')(img_input) x = layers.Conv2D(64, 7, strides=2, use_bias=use_bias, name='conv1_conv')(x) if not preact: x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name='conv1_bn')(x) x = layers.Activation('relu', name='conv1_relu')(x) x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name='pool1_pad')(x) x = layers.MaxPooling2D(3, strides=2, name='pool1_pool')(x) x = stack_fn(x) if preact: x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name='post_bn')(x) x = layers.Activation('relu', name='post_relu')(x) if include_top: x = layers.GlobalAveragePooling2D(name='avg_pool')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D(name='avg_pool')(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D(name='max_pool')(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name=model_name) # Load weights. if (weights == 'imagenet') and (model_name in WEIGHTS_HASHES): if include_top: file_name = model_name + '_weights_tf_dim_ordering_tf_kernels.h5' file_hash = WEIGHTS_HASHES[model_name][0] else: file_name = model_name + '_weights_tf_dim_ordering_tf_kernels_notop.h5' file_hash = WEIGHTS_HASHES[model_name][1] weights_path = data_utils.get_file( file_name, BASE_WEIGHTS_PATH + file_name, cache_subdir='models', file_hash=file_hash) model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model def block1(x, filters, kernel_size=3, stride=1, conv_shortcut=True, name=None): """A residual block. Args: x: input tensor. filters: integer, filters of the bottleneck layer. kernel_size: default 3, kernel size of the bottleneck layer. stride: default 1, stride of the first layer. conv_shortcut: default True, use convolution shortcut if True, otherwise identity shortcut. name: string, block label. Returns: Output tensor for the residual block. """ bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 if conv_shortcut: shortcut = layers.Conv2D( 4 * filters, 1, strides=stride, name=name + '_0_conv')(x) shortcut = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_0_bn')(shortcut) else: shortcut = x x = layers.Conv2D(filters, 1, strides=stride, name=name + '_1_conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_1_bn')(x) x = layers.Activation('relu', name=name + '_1_relu')(x) x = layers.Conv2D( filters, kernel_size, padding='SAME', name=name + '_2_conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_2_bn')(x) x = layers.Activation('relu', name=name + '_2_relu')(x) x = layers.Conv2D(4 * filters, 1, name=name + '_3_conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_3_bn')(x) x = layers.Add(name=name + '_add')([shortcut, x]) x = layers.Activation('relu', name=name + '_out')(x) return x def stack1(x, filters, blocks, stride1=2, name=None): """A set of stacked residual blocks. Args: x: input tensor. filters: integer, filters of the bottleneck layer in a block. blocks: integer, blocks in the stacked blocks. stride1: default 2, stride of the first layer in the first block. name: string, stack label. Returns: Output tensor for the stacked blocks. """ x = block1(x, filters, stride=stride1, name=name + '_block1') for i in range(2, blocks + 1): x = block1(x, filters, conv_shortcut=False, name=name + '_block' + str(i)) return x def block2(x, filters, kernel_size=3, stride=1, conv_shortcut=False, name=None): """A residual block. Args: x: input tensor. filters: integer, filters of the bottleneck layer. kernel_size: default 3, kernel size of the bottleneck layer. stride: default 1, stride of the first layer. conv_shortcut: default False, use convolution shortcut if True, otherwise identity shortcut. name: string, block label. Returns: Output tensor for the residual block. """ bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 preact = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_preact_bn')(x) preact = layers.Activation('relu', name=name + '_preact_relu')(preact) if conv_shortcut: shortcut = layers.Conv2D( 4 * filters, 1, strides=stride, name=name + '_0_conv')(preact) else: shortcut = layers.MaxPooling2D(1, strides=stride)(x) if stride > 1 else x x = layers.Conv2D( filters, 1, strides=1, use_bias=False, name=name + '_1_conv')(preact) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_1_bn')(x) x = layers.Activation('relu', name=name + '_1_relu')(x) x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name=name + '_2_pad')(x) x = layers.Conv2D( filters, kernel_size, strides=stride, use_bias=False, name=name + '_2_conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_2_bn')(x) x = layers.Activation('relu', name=name + '_2_relu')(x) x = layers.Conv2D(4 * filters, 1, name=name + '_3_conv')(x) x = layers.Add(name=name + '_out')([shortcut, x]) return x def stack2(x, filters, blocks, stride1=2, name=None): """A set of stacked residual blocks. Args: x: input tensor. filters: integer, filters of the bottleneck layer in a block. blocks: integer, blocks in the stacked blocks. stride1: default 2, stride of the first layer in the first block. name: string, stack label. Returns: Output tensor for the stacked blocks. """ x = block2(x, filters, conv_shortcut=True, name=name + '_block1') for i in range(2, blocks): x = block2(x, filters, name=name + '_block' + str(i)) x = block2(x, filters, stride=stride1, name=name + '_block' + str(blocks)) return x def block3(x, filters, kernel_size=3, stride=1, groups=32, conv_shortcut=True, name=None): """A residual block. Args: x: input tensor. filters: integer, filters of the bottleneck layer. kernel_size: default 3, kernel size of the bottleneck layer. stride: default 1, stride of the first layer. groups: default 32, group size for grouped convolution. conv_shortcut: default True, use convolution shortcut if True, otherwise identity shortcut. name: string, block label. Returns: Output tensor for the residual block. """ bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 if conv_shortcut: shortcut = layers.Conv2D( (64 // groups) * filters, 1, strides=stride, use_bias=False, name=name + '_0_conv')(x) shortcut = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_0_bn')(shortcut) else: shortcut = x x = layers.Conv2D(filters, 1, use_bias=False, name=name + '_1_conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_1_bn')(x) x = layers.Activation('relu', name=name + '_1_relu')(x) c = filters // groups x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)), name=name + '_2_pad')(x) x = layers.DepthwiseConv2D( kernel_size, strides=stride, depth_multiplier=c, use_bias=False, name=name + '_2_conv')(x) x_shape = backend.shape(x)[:-1] x = backend.reshape(x, backend.concatenate([x_shape, (groups, c, c)])) x = layers.Lambda( lambda x: sum(x[:, :, :, :, i] for i in range(c)), name=name + '_2_reduce')(x) x = backend.reshape(x, backend.concatenate([x_shape, (filters,)])) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_2_bn')(x) x = layers.Activation('relu', name=name + '_2_relu')(x) x = layers.Conv2D( (64 // groups) * filters, 1, use_bias=False, name=name + '_3_conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_3_bn')(x) x = layers.Add(name=name + '_add')([shortcut, x]) x = layers.Activation('relu', name=name + '_out')(x) return x def stack3(x, filters, blocks, stride1=2, groups=32, name=None): """A set of stacked residual blocks. Args: x: input tensor. filters: integer, filters of the bottleneck layer in a block. blocks: integer, blocks in the stacked blocks. stride1: default 2, stride of the first layer in the first block. groups: default 32, group size for grouped convolution. name: string, stack label. Returns: Output tensor for the stacked blocks. """ x = block3(x, filters, stride=stride1, groups=groups, name=name + '_block1') for i in range(2, blocks + 1): x = block3( x, filters, groups=groups, conv_shortcut=False, name=name + '_block' + str(i)) return x @keras_export('keras.applications.resnet50.ResNet50', 'keras.applications.resnet.ResNet50', 'keras.applications.ResNet50') def ResNet50(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, **kwargs): """Instantiates the ResNet50 architecture.""" def stack_fn(x): x = stack1(x, 64, 3, stride1=1, name='conv2') x = stack1(x, 128, 4, name='conv3') x = stack1(x, 256, 6, name='conv4') return stack1(x, 512, 3, name='conv5') return ResNet(stack_fn, False, True, 'resnet50', include_top, weights, input_tensor, input_shape, pooling, classes, **kwargs) @keras_export('keras.applications.resnet.ResNet101', 'keras.applications.ResNet101') def ResNet101(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, **kwargs): """Instantiates the ResNet101 architecture.""" def stack_fn(x): x = stack1(x, 64, 3, stride1=1, name='conv2') x = stack1(x, 128, 4, name='conv3') x = stack1(x, 256, 23, name='conv4') return stack1(x, 512, 3, name='conv5') return ResNet(stack_fn, False, True, 'resnet101', include_top, weights, input_tensor, input_shape, pooling, classes, **kwargs) @keras_export('keras.applications.resnet.ResNet152', 'keras.applications.ResNet152') def ResNet152(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, **kwargs): """Instantiates the ResNet152 architecture.""" def stack_fn(x): x = stack1(x, 64, 3, stride1=1, name='conv2') x = stack1(x, 128, 8, name='conv3') x = stack1(x, 256, 36, name='conv4') return stack1(x, 512, 3, name='conv5') return ResNet(stack_fn, False, True, 'resnet152', include_top, weights, input_tensor, input_shape, pooling, classes, **kwargs) @keras_export('keras.applications.resnet50.preprocess_input', 'keras.applications.resnet.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input( x, data_format=data_format, mode='caffe') @keras_export('keras.applications.resnet50.decode_predictions', 'keras.applications.resnet.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_CAFFE, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__ DOC = """ Reference: - [Deep Residual Learning for Image Recognition]( https://arxiv.org/abs/1512.03385) (CVPR 2015) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For ResNet, call `tf.keras.applications.resnet.preprocess_input` on your inputs before passing them to the model. `resnet.preprocess_input` will convert the input images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. Args: include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `'channels_last'` data format) or `(3, 224, 224)` (with `'channels_first'` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A Keras model instance. """ setattr(ResNet50, '__doc__', ResNet50.__doc__ + DOC) setattr(ResNet101, '__doc__', ResNet101.__doc__ + DOC) setattr(ResNet152, '__doc__', ResNet152.__doc__ + DOC)
21,207
35.191126
87
py
keras
keras-master/keras/applications/vgg16.py
# Copyright 2015 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """VGG16 model for Keras. Reference: - [Very Deep Convolutional Networks for Large-Scale Image Recognition] (https://arxiv.org/abs/1409.1556) (ICLR 2015) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export WEIGHTS_PATH = ('https://storage.googleapis.com/tensorflow/keras-applications/' 'vgg16/vgg16_weights_tf_dim_ordering_tf_kernels.h5') WEIGHTS_PATH_NO_TOP = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/vgg16/' 'vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5') layers = VersionAwareLayers() @keras_export('keras.applications.vgg16.VGG16', 'keras.applications.VGG16') def VGG16( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the VGG16 model. Reference: - [Very Deep Convolutional Networks for Large-Scale Image Recognition]( https://arxiv.org/abs/1409.1556) (ICLR 2015) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). The default input size for this model is 224x224. Note: each Keras Application expects a specific kind of input preprocessing. For VGG16, call `tf.keras.applications.vgg16.preprocess_input` on your inputs before passing them to the model. `vgg16.preprocess_input` will convert the input images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. Args: include_top: whether to include the 3 fully-connected layers at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `channels_last` data format) or `(3, 224, 224)` (with `channels_first` data format). It should have exactly 3 input channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError( 'The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded. Received: ' f'weights={weights}') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top` ' 'as true, `classes` should be 1000. ' f'Received `classes={classes}`') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=224, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor # Block 1 x = layers.Conv2D( 64, (3, 3), activation='relu', padding='same', name='block1_conv1')( img_input) x = layers.Conv2D( 64, (3, 3), activation='relu', padding='same', name='block1_conv2')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x) # Block 2 x = layers.Conv2D( 128, (3, 3), activation='relu', padding='same', name='block2_conv1')(x) x = layers.Conv2D( 128, (3, 3), activation='relu', padding='same', name='block2_conv2')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x) # Block 3 x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv1')(x) x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv2')(x) x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv3')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x) # Block 4 x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv1')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv2')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv3')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x) # Block 5 x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv1')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv2')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv3')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x) if include_top: # Classification block x = layers.Flatten(name='flatten')(x) x = layers.Dense(4096, activation='relu', name='fc1')(x) x = layers.Dense(4096, activation='relu', name='fc2')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='vgg16') # Load weights. if weights == 'imagenet': if include_top: weights_path = data_utils.get_file( 'vgg16_weights_tf_dim_ordering_tf_kernels.h5', WEIGHTS_PATH, cache_subdir='models', file_hash='64373286793e3c8b2b4e3219cbf3544b') else: weights_path = data_utils.get_file( 'vgg16_weights_tf_dim_ordering_tf_kernels_notop.h5', WEIGHTS_PATH_NO_TOP, cache_subdir='models', file_hash='6d6bbae143d832006294945121d1f1fc') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model @keras_export('keras.applications.vgg16.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input( x, data_format=data_format, mode='caffe') @keras_export('keras.applications.vgg16.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_CAFFE, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
9,586
37.971545
87
py
keras
keras-master/keras/applications/densenet.py
# Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """DenseNet models for Keras. Reference: - [Densely Connected Convolutional Networks]( https://arxiv.org/abs/1608.06993) (CVPR 2017) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export BASE_WEIGHTS_PATH = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/densenet/') DENSENET121_WEIGHT_PATH = ( BASE_WEIGHTS_PATH + 'densenet121_weights_tf_dim_ordering_tf_kernels.h5') DENSENET121_WEIGHT_PATH_NO_TOP = ( BASE_WEIGHTS_PATH + 'densenet121_weights_tf_dim_ordering_tf_kernels_notop.h5') DENSENET169_WEIGHT_PATH = ( BASE_WEIGHTS_PATH + 'densenet169_weights_tf_dim_ordering_tf_kernels.h5') DENSENET169_WEIGHT_PATH_NO_TOP = ( BASE_WEIGHTS_PATH + 'densenet169_weights_tf_dim_ordering_tf_kernels_notop.h5') DENSENET201_WEIGHT_PATH = ( BASE_WEIGHTS_PATH + 'densenet201_weights_tf_dim_ordering_tf_kernels.h5') DENSENET201_WEIGHT_PATH_NO_TOP = ( BASE_WEIGHTS_PATH + 'densenet201_weights_tf_dim_ordering_tf_kernels_notop.h5') layers = VersionAwareLayers() def dense_block(x, blocks, name): """A dense block. Args: x: input tensor. blocks: integer, the number of building blocks. name: string, block label. Returns: Output tensor for the block. """ for i in range(blocks): x = conv_block(x, 32, name=name + '_block' + str(i + 1)) return x def transition_block(x, reduction, name): """A transition block. Args: x: input tensor. reduction: float, compression rate at transition layers. name: string, block label. Returns: output tensor for the block. """ bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_bn')( x) x = layers.Activation('relu', name=name + '_relu')(x) x = layers.Conv2D( int(backend.int_shape(x)[bn_axis] * reduction), 1, use_bias=False, name=name + '_conv')( x) x = layers.AveragePooling2D(2, strides=2, name=name + '_pool')(x) return x def conv_block(x, growth_rate, name): """A building block for a dense block. Args: x: input tensor. growth_rate: float, growth rate at dense layers. name: string, block label. Returns: Output tensor for the block. """ bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 x1 = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_0_bn')( x) x1 = layers.Activation('relu', name=name + '_0_relu')(x1) x1 = layers.Conv2D( 4 * growth_rate, 1, use_bias=False, name=name + '_1_conv')( x1) x1 = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name=name + '_1_bn')( x1) x1 = layers.Activation('relu', name=name + '_1_relu')(x1) x1 = layers.Conv2D( growth_rate, 3, padding='same', use_bias=False, name=name + '_2_conv')( x1) x = layers.Concatenate(axis=bn_axis, name=name + '_concat')([x, x1]) return x def DenseNet( blocks, include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the DenseNet architecture. Reference: - [Densely Connected Convolutional Networks]( https://arxiv.org/abs/1608.06993) (CVPR 2017) This function returns a Keras image classification model, optionally loaded with weights pre-trained on ImageNet. For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For DenseNet, call `tf.keras.applications.densenet.preprocess_input` on your inputs before passing them to the model. `densenet.preprocess_input` will scale pixels between 0 and 1 and then will normalize each channel with respect to the ImageNet dataset statistics. Args: blocks: numbers of building blocks for the four dense layers. include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `'channels_last'` data format) or `(3, 224, 224)` (with `'channels_first'` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. pooling: optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded.') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top`' ' as true, `classes` should be 1000') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=224, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor bn_axis = 3 if backend.image_data_format() == 'channels_last' else 1 x = layers.ZeroPadding2D(padding=((3, 3), (3, 3)))(img_input) x = layers.Conv2D(64, 7, strides=2, use_bias=False, name='conv1/conv')(x) x = layers.BatchNormalization( axis=bn_axis, epsilon=1.001e-5, name='conv1/bn')( x) x = layers.Activation('relu', name='conv1/relu')(x) x = layers.ZeroPadding2D(padding=((1, 1), (1, 1)))(x) x = layers.MaxPooling2D(3, strides=2, name='pool1')(x) x = dense_block(x, blocks[0], name='conv2') x = transition_block(x, 0.5, name='pool2') x = dense_block(x, blocks[1], name='conv3') x = transition_block(x, 0.5, name='pool3') x = dense_block(x, blocks[2], name='conv4') x = transition_block(x, 0.5, name='pool4') x = dense_block(x, blocks[3], name='conv5') x = layers.BatchNormalization(axis=bn_axis, epsilon=1.001e-5, name='bn')(x) x = layers.Activation('relu', name='relu')(x) if include_top: x = layers.GlobalAveragePooling2D(name='avg_pool')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D(name='avg_pool')(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D(name='max_pool')(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. if blocks == [6, 12, 24, 16]: model = training.Model(inputs, x, name='densenet121') elif blocks == [6, 12, 32, 32]: model = training.Model(inputs, x, name='densenet169') elif blocks == [6, 12, 48, 32]: model = training.Model(inputs, x, name='densenet201') else: model = training.Model(inputs, x, name='densenet') # Load weights. if weights == 'imagenet': if include_top: if blocks == [6, 12, 24, 16]: weights_path = data_utils.get_file( 'densenet121_weights_tf_dim_ordering_tf_kernels.h5', DENSENET121_WEIGHT_PATH, cache_subdir='models', file_hash='9d60b8095a5708f2dcce2bca79d332c7') elif blocks == [6, 12, 32, 32]: weights_path = data_utils.get_file( 'densenet169_weights_tf_dim_ordering_tf_kernels.h5', DENSENET169_WEIGHT_PATH, cache_subdir='models', file_hash='d699b8f76981ab1b30698df4c175e90b') elif blocks == [6, 12, 48, 32]: weights_path = data_utils.get_file( 'densenet201_weights_tf_dim_ordering_tf_kernels.h5', DENSENET201_WEIGHT_PATH, cache_subdir='models', file_hash='1ceb130c1ea1b78c3bf6114dbdfd8807') else: if blocks == [6, 12, 24, 16]: weights_path = data_utils.get_file( 'densenet121_weights_tf_dim_ordering_tf_kernels_notop.h5', DENSENET121_WEIGHT_PATH_NO_TOP, cache_subdir='models', file_hash='30ee3e1110167f948a6b9946edeeb738') elif blocks == [6, 12, 32, 32]: weights_path = data_utils.get_file( 'densenet169_weights_tf_dim_ordering_tf_kernels_notop.h5', DENSENET169_WEIGHT_PATH_NO_TOP, cache_subdir='models', file_hash='b8c4d4c20dd625c148057b9ff1c1176b') elif blocks == [6, 12, 48, 32]: weights_path = data_utils.get_file( 'densenet201_weights_tf_dim_ordering_tf_kernels_notop.h5', DENSENET201_WEIGHT_PATH_NO_TOP, cache_subdir='models', file_hash='c13680b51ded0fb44dff2d8f86ac8bb1') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model @keras_export('keras.applications.densenet.DenseNet121', 'keras.applications.DenseNet121') def DenseNet121(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000): """Instantiates the Densenet121 architecture.""" return DenseNet([6, 12, 24, 16], include_top, weights, input_tensor, input_shape, pooling, classes) @keras_export('keras.applications.densenet.DenseNet169', 'keras.applications.DenseNet169') def DenseNet169(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000): """Instantiates the Densenet169 architecture.""" return DenseNet([6, 12, 32, 32], include_top, weights, input_tensor, input_shape, pooling, classes) @keras_export('keras.applications.densenet.DenseNet201', 'keras.applications.DenseNet201') def DenseNet201(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000): """Instantiates the Densenet201 architecture.""" return DenseNet([6, 12, 48, 32], include_top, weights, input_tensor, input_shape, pooling, classes) @keras_export('keras.applications.densenet.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input( x, data_format=data_format, mode='torch') @keras_export('keras.applications.densenet.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TORCH, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__ DOC = """ Reference: - [Densely Connected Convolutional Networks]( https://arxiv.org/abs/1608.06993) (CVPR 2017) Optionally loads weights pre-trained on ImageNet. Note that the data format convention used by the model is the one specified in your Keras config at `~/.keras/keras.json`. Note: each Keras Application expects a specific kind of input preprocessing. For DenseNet, call `tf.keras.applications.densenet.preprocess_input` on your inputs before passing them to the model. Args: include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `'channels_last'` data format) or `(3, 224, 224)` (with `'channels_first'` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. Returns: A Keras model instance. """ setattr(DenseNet121, '__doc__', DenseNet121.__doc__ + DOC) setattr(DenseNet169, '__doc__', DenseNet169.__doc__ + DOC) setattr(DenseNet201, '__doc__', DenseNet201.__doc__ + DOC)
16,084
36.320186
87
py
keras
keras-master/keras/applications/imagenet_utils.py
# Copyright 2019 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Utilities for ImageNet data preprocessing & prediction decoding.""" import json import warnings import numpy as np from keras import activations from keras import backend from keras.utils import data_utils from tensorflow.python.util.tf_export import keras_export CLASS_INDEX = None CLASS_INDEX_PATH = ('https://storage.googleapis.com/download.tensorflow.org/' 'data/imagenet_class_index.json') PREPROCESS_INPUT_DOC = """ Preprocesses a tensor or Numpy array encoding a batch of images. Usage example with `applications.MobileNet`: ```python i = tf.keras.layers.Input([None, None, 3], dtype = tf.uint8) x = tf.cast(i, tf.float32) x = tf.keras.applications.mobilenet.preprocess_input(x) core = tf.keras.applications.MobileNet() x = core(x) model = tf.keras.Model(inputs=[i], outputs=[x]) image = tf.image.decode_png(tf.io.read_file('file.png')) result = model(image) ``` Args: x: A floating point `numpy.array` or a `tf.Tensor`, 3D or 4D with 3 color channels, with values in the range [0, 255]. The preprocessed data are written over the input data if the data types are compatible. To avoid this behaviour, `numpy.copy(x)` can be used. data_format: Optional data format of the image tensor/array. Defaults to None, in which case the global setting `tf.keras.backend.image_data_format()` is used (unless you changed it, it defaults to "channels_last").{mode} Returns: Preprocessed `numpy.array` or a `tf.Tensor` with type `float32`. {ret} Raises: {error} """ PREPROCESS_INPUT_MODE_DOC = """ mode: One of "caffe", "tf" or "torch". Defaults to "caffe". - caffe: will convert the images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. - tf: will scale pixels between -1 and 1, sample-wise. - torch: will scale pixels between 0 and 1 and then will normalize each channel with respect to the ImageNet dataset. """ PREPROCESS_INPUT_DEFAULT_ERROR_DOC = """ ValueError: In case of unknown `mode` or `data_format` argument.""" PREPROCESS_INPUT_ERROR_DOC = """ ValueError: In case of unknown `data_format` argument.""" PREPROCESS_INPUT_RET_DOC_TF = """ The inputs pixel values are scaled between -1 and 1, sample-wise.""" PREPROCESS_INPUT_RET_DOC_TORCH = """ The input pixels values are scaled between 0 and 1 and each channel is normalized with respect to the ImageNet dataset.""" PREPROCESS_INPUT_RET_DOC_CAFFE = """ The images are converted from RGB to BGR, then each color channel is zero-centered with respect to the ImageNet dataset, without scaling.""" @keras_export('keras.applications.imagenet_utils.preprocess_input') def preprocess_input(x, data_format=None, mode='caffe'): """Preprocesses a tensor or Numpy array encoding a batch of images.""" if mode not in {'caffe', 'tf', 'torch'}: raise ValueError('Expected mode to be one of `caffe`, `tf` or `torch`. ' f'Received: mode={mode}') if data_format is None: data_format = backend.image_data_format() elif data_format not in {'channels_first', 'channels_last'}: raise ValueError('Expected data_format to be one of `channels_first` or ' f'`channels_last`. Received: data_format={data_format}') if isinstance(x, np.ndarray): return _preprocess_numpy_input( x, data_format=data_format, mode=mode) else: return _preprocess_symbolic_input( x, data_format=data_format, mode=mode) preprocess_input.__doc__ = PREPROCESS_INPUT_DOC.format( mode=PREPROCESS_INPUT_MODE_DOC, ret='', error=PREPROCESS_INPUT_DEFAULT_ERROR_DOC) @keras_export('keras.applications.imagenet_utils.decode_predictions') def decode_predictions(preds, top=5): """Decodes the prediction of an ImageNet model. Args: preds: Numpy array encoding a batch of predictions. top: Integer, how many top-guesses to return. Defaults to 5. Returns: A list of lists of top class prediction tuples `(class_name, class_description, score)`. One list of tuples per sample in batch input. Raises: ValueError: In case of invalid shape of the `pred` array (must be 2D). """ global CLASS_INDEX if len(preds.shape) != 2 or preds.shape[1] != 1000: raise ValueError('`decode_predictions` expects ' 'a batch of predictions ' '(i.e. a 2D array of shape (samples, 1000)). ' 'Found array with shape: ' + str(preds.shape)) if CLASS_INDEX is None: fpath = data_utils.get_file( 'imagenet_class_index.json', CLASS_INDEX_PATH, cache_subdir='models', file_hash='c2c37ea517e94d9795004a39431a14cb') with open(fpath) as f: CLASS_INDEX = json.load(f) results = [] for pred in preds: top_indices = pred.argsort()[-top:][::-1] result = [tuple(CLASS_INDEX[str(i)]) + (pred[i],) for i in top_indices] result.sort(key=lambda x: x[2], reverse=True) results.append(result) return results def _preprocess_numpy_input(x, data_format, mode): """Preprocesses a Numpy array encoding a batch of images. Args: x: Input array, 3D or 4D. data_format: Data format of the image array. mode: One of "caffe", "tf" or "torch". - caffe: will convert the images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. - tf: will scale pixels between -1 and 1, sample-wise. - torch: will scale pixels between 0 and 1 and then will normalize each channel with respect to the ImageNet dataset. Returns: Preprocessed Numpy array. """ if not issubclass(x.dtype.type, np.floating): x = x.astype(backend.floatx(), copy=False) if mode == 'tf': x /= 127.5 x -= 1. return x elif mode == 'torch': x /= 255. mean = [0.485, 0.456, 0.406] std = [0.229, 0.224, 0.225] else: if data_format == 'channels_first': # 'RGB'->'BGR' if x.ndim == 3: x = x[::-1, ...] else: x = x[:, ::-1, ...] else: # 'RGB'->'BGR' x = x[..., ::-1] mean = [103.939, 116.779, 123.68] std = None # Zero-center by mean pixel if data_format == 'channels_first': if x.ndim == 3: x[0, :, :] -= mean[0] x[1, :, :] -= mean[1] x[2, :, :] -= mean[2] if std is not None: x[0, :, :] /= std[0] x[1, :, :] /= std[1] x[2, :, :] /= std[2] else: x[:, 0, :, :] -= mean[0] x[:, 1, :, :] -= mean[1] x[:, 2, :, :] -= mean[2] if std is not None: x[:, 0, :, :] /= std[0] x[:, 1, :, :] /= std[1] x[:, 2, :, :] /= std[2] else: x[..., 0] -= mean[0] x[..., 1] -= mean[1] x[..., 2] -= mean[2] if std is not None: x[..., 0] /= std[0] x[..., 1] /= std[1] x[..., 2] /= std[2] return x def _preprocess_symbolic_input(x, data_format, mode): """Preprocesses a tensor encoding a batch of images. Args: x: Input tensor, 3D or 4D. data_format: Data format of the image tensor. mode: One of "caffe", "tf" or "torch". - caffe: will convert the images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. - tf: will scale pixels between -1 and 1, sample-wise. - torch: will scale pixels between 0 and 1 and then will normalize each channel with respect to the ImageNet dataset. Returns: Preprocessed tensor. """ if mode == 'tf': x /= 127.5 x -= 1. return x elif mode == 'torch': x /= 255. mean = [0.485, 0.456, 0.406] std = [0.229, 0.224, 0.225] else: if data_format == 'channels_first': # 'RGB'->'BGR' if backend.ndim(x) == 3: x = x[::-1, ...] else: x = x[:, ::-1, ...] else: # 'RGB'->'BGR' x = x[..., ::-1] mean = [103.939, 116.779, 123.68] std = None mean_tensor = backend.constant(-np.array(mean)) # Zero-center by mean pixel if backend.dtype(x) != backend.dtype(mean_tensor): x = backend.bias_add( x, backend.cast(mean_tensor, backend.dtype(x)), data_format=data_format) else: x = backend.bias_add(x, mean_tensor, data_format) if std is not None: std_tensor = backend.constant(np.array(std), dtype=backend.dtype(x)) if data_format == 'channels_first': std_tensor = backend.reshape(std_tensor, (-1, 1, 1)) x /= std_tensor return x def obtain_input_shape(input_shape, default_size, min_size, data_format, require_flatten, weights=None): """Internal utility to compute/validate a model's input shape. Args: input_shape: Either None (will return the default network input shape), or a user-provided shape to be validated. default_size: Default input width/height for the model. min_size: Minimum input width/height accepted by the model. data_format: Image data format to use. require_flatten: Whether the model is expected to be linked to a classifier via a Flatten layer. weights: One of `None` (random initialization) or 'imagenet' (pre-training on ImageNet). If weights='imagenet' input channels must be equal to 3. Returns: An integer shape tuple (may include None entries). Raises: ValueError: In case of invalid argument values. """ if weights != 'imagenet' and input_shape and len(input_shape) == 3: if data_format == 'channels_first': if input_shape[0] not in {1, 3}: warnings.warn('This model usually expects 1 or 3 input channels. ' 'However, it was passed an input_shape with ' + str(input_shape[0]) + ' input channels.') default_shape = (input_shape[0], default_size, default_size) else: if input_shape[-1] not in {1, 3}: warnings.warn('This model usually expects 1 or 3 input channels. ' 'However, it was passed an input_shape with ' + str(input_shape[-1]) + ' input channels.') default_shape = (default_size, default_size, input_shape[-1]) else: if data_format == 'channels_first': default_shape = (3, default_size, default_size) else: default_shape = (default_size, default_size, 3) if weights == 'imagenet' and require_flatten: if input_shape is not None: if input_shape != default_shape: raise ValueError('When setting `include_top=True` ' 'and loading `imagenet` weights, ' f'`input_shape` should be {default_shape}. ' f'Received: input_shape={input_shape}') return default_shape if input_shape: if data_format == 'channels_first': if input_shape is not None: if len(input_shape) != 3: raise ValueError('`input_shape` must be a tuple of three integers.') if input_shape[0] != 3 and weights == 'imagenet': raise ValueError('The input must have 3 channels; Received ' f'`input_shape={input_shape}`') if ((input_shape[1] is not None and input_shape[1] < min_size) or (input_shape[2] is not None and input_shape[2] < min_size)): raise ValueError(f'Input size must be at least {min_size}' f'x{min_size}; Received: ' f'input_shape={input_shape}') else: if input_shape is not None: if len(input_shape) != 3: raise ValueError('`input_shape` must be a tuple of three integers.') if input_shape[-1] != 3 and weights == 'imagenet': raise ValueError('The input must have 3 channels; Received ' f'`input_shape={input_shape}`') if ((input_shape[0] is not None and input_shape[0] < min_size) or (input_shape[1] is not None and input_shape[1] < min_size)): raise ValueError('Input size must be at least ' f'{min_size}x{min_size}; Received: ' f'input_shape={input_shape}') else: if require_flatten: input_shape = default_shape else: if data_format == 'channels_first': input_shape = (3, None, None) else: input_shape = (None, None, 3) if require_flatten: if None in input_shape: raise ValueError('If `include_top` is True, ' 'you should specify a static `input_shape`. ' f'Received: input_shape={input_shape}') return input_shape def correct_pad(inputs, kernel_size): """Returns a tuple for zero-padding for 2D convolution with downsampling. Args: inputs: Input tensor. kernel_size: An integer or tuple/list of 2 integers. Returns: A tuple. """ img_dim = 2 if backend.image_data_format() == 'channels_first' else 1 input_size = backend.int_shape(inputs)[img_dim:(img_dim + 2)] if isinstance(kernel_size, int): kernel_size = (kernel_size, kernel_size) if input_size[0] is None: adjust = (1, 1) else: adjust = (1 - input_size[0] % 2, 1 - input_size[1] % 2) correct = (kernel_size[0] // 2, kernel_size[1] // 2) return ((correct[0] - adjust[0], correct[0]), (correct[1] - adjust[1], correct[1])) def validate_activation(classifier_activation, weights): """validates that the classifer_activation is compatible with the weights. Args: classifier_activation: str or callable activation function weights: The pretrained weights to load. Raises: ValueError: if an activation other than `None` or `softmax` are used with pretrained weights. """ if weights is None: return classifier_activation = activations.get(classifier_activation) if classifier_activation not in { activations.get('softmax'), activations.get(None) }: raise ValueError('Only `None` and `softmax` activations are allowed ' 'for the `classifier_activation` argument when using ' 'pretrained weights, with `include_top=True`; Received: ' f'classifier_activation={classifier_activation}')
15,197
33.778032
80
py
keras
keras-master/keras/applications/resnet_v2.py
# Copyright 2019 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """ResNet v2 models for Keras. Reference: - [Identity Mappings in Deep Residual Networks] (https://arxiv.org/abs/1603.05027) (CVPR 2016) """ from keras.applications import imagenet_utils from keras.applications import resnet from tensorflow.python.util.tf_export import keras_export @keras_export('keras.applications.resnet_v2.ResNet50V2', 'keras.applications.ResNet50V2') def ResNet50V2( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the ResNet50V2 architecture.""" def stack_fn(x): x = resnet.stack2(x, 64, 3, name='conv2') x = resnet.stack2(x, 128, 4, name='conv3') x = resnet.stack2(x, 256, 6, name='conv4') return resnet.stack2(x, 512, 3, stride1=1, name='conv5') return resnet.ResNet( stack_fn, True, True, 'resnet50v2', include_top, weights, input_tensor, input_shape, pooling, classes, classifier_activation=classifier_activation) @keras_export('keras.applications.resnet_v2.ResNet101V2', 'keras.applications.ResNet101V2') def ResNet101V2( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the ResNet101V2 architecture.""" def stack_fn(x): x = resnet.stack2(x, 64, 3, name='conv2') x = resnet.stack2(x, 128, 4, name='conv3') x = resnet.stack2(x, 256, 23, name='conv4') return resnet.stack2(x, 512, 3, stride1=1, name='conv5') return resnet.ResNet( stack_fn, True, True, 'resnet101v2', include_top, weights, input_tensor, input_shape, pooling, classes, classifier_activation=classifier_activation) @keras_export('keras.applications.resnet_v2.ResNet152V2', 'keras.applications.ResNet152V2') def ResNet152V2( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the ResNet152V2 architecture.""" def stack_fn(x): x = resnet.stack2(x, 64, 3, name='conv2') x = resnet.stack2(x, 128, 8, name='conv3') x = resnet.stack2(x, 256, 36, name='conv4') return resnet.stack2(x, 512, 3, stride1=1, name='conv5') return resnet.ResNet( stack_fn, True, True, 'resnet152v2', include_top, weights, input_tensor, input_shape, pooling, classes, classifier_activation=classifier_activation) @keras_export('keras.applications.resnet_v2.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input( x, data_format=data_format, mode='tf') @keras_export('keras.applications.resnet_v2.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__ DOC = """ Reference: - [Identity Mappings in Deep Residual Networks] (https://arxiv.org/abs/1603.05027) (CVPR 2016) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For ResNetV2, call `tf.keras.applications.resnet_v2.preprocess_input` on your inputs before passing them to the model. `resnet_v2.preprocess_input` will scale input pixels between -1 and 1. Args: include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `'channels_last'` data format) or `(3, 224, 224)` (with `'channels_first'` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ setattr(ResNet50V2, '__doc__', ResNet50V2.__doc__ + DOC) setattr(ResNet101V2, '__doc__', ResNet101V2.__doc__ + DOC) setattr(ResNet152V2, '__doc__', ResNet152V2.__doc__ + DOC)
6,741
32.879397
87
py
keras
keras-master/keras/applications/vgg19.py
# Copyright 2015 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """VGG19 model for Keras. Reference: - [Very Deep Convolutional Networks for Large-Scale Image Recognition]( https://arxiv.org/abs/1409.1556) (ICLR 2015) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export WEIGHTS_PATH = ('https://storage.googleapis.com/tensorflow/keras-applications/' 'vgg19/vgg19_weights_tf_dim_ordering_tf_kernels.h5') WEIGHTS_PATH_NO_TOP = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/vgg19/' 'vgg19_weights_tf_dim_ordering_tf_kernels_notop.h5') layers = VersionAwareLayers() @keras_export('keras.applications.vgg19.VGG19', 'keras.applications.VGG19') def VGG19( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the VGG19 architecture. Reference: - [Very Deep Convolutional Networks for Large-Scale Image Recognition]( https://arxiv.org/abs/1409.1556) (ICLR 2015) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). The default input size for this model is 224x224. Note: each Keras Application expects a specific kind of input preprocessing. For VGG19, call `tf.keras.applications.vgg19.preprocess_input` on your inputs before passing them to the model. `vgg19.preprocess_input` will convert the input images from RGB to BGR, then will zero-center each color channel with respect to the ImageNet dataset, without scaling. Args: include_top: whether to include the 3 fully-connected layers at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `channels_last` data format) or `(3, 224, 224)` (with `channels_first` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded. ' f'Received: `weights={weights}.`') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top` ' 'as true, `classes` should be 1000. ' f'Received: `classes={classes}.`') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=224, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor # Block 1 x = layers.Conv2D( 64, (3, 3), activation='relu', padding='same', name='block1_conv1')( img_input) x = layers.Conv2D( 64, (3, 3), activation='relu', padding='same', name='block1_conv2')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block1_pool')(x) # Block 2 x = layers.Conv2D( 128, (3, 3), activation='relu', padding='same', name='block2_conv1')(x) x = layers.Conv2D( 128, (3, 3), activation='relu', padding='same', name='block2_conv2')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block2_pool')(x) # Block 3 x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv1')(x) x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv2')(x) x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv3')(x) x = layers.Conv2D( 256, (3, 3), activation='relu', padding='same', name='block3_conv4')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block3_pool')(x) # Block 4 x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv1')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv2')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv3')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block4_conv4')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block4_pool')(x) # Block 5 x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv1')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv2')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv3')(x) x = layers.Conv2D( 512, (3, 3), activation='relu', padding='same', name='block5_conv4')(x) x = layers.MaxPooling2D((2, 2), strides=(2, 2), name='block5_pool')(x) if include_top: # Classification block x = layers.Flatten(name='flatten')(x) x = layers.Dense(4096, activation='relu', name='fc1')(x) x = layers.Dense(4096, activation='relu', name='fc2')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='vgg19') # Load weights. if weights == 'imagenet': if include_top: weights_path = data_utils.get_file( 'vgg19_weights_tf_dim_ordering_tf_kernels.h5', WEIGHTS_PATH, cache_subdir='models', file_hash='cbe5617147190e668d6c5d5026f83318') else: weights_path = data_utils.get_file( 'vgg19_weights_tf_dim_ordering_tf_kernels_notop.h5', WEIGHTS_PATH_NO_TOP, cache_subdir='models', file_hash='253f8cb515780f3b799900260a226db6') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model @keras_export('keras.applications.vgg19.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input( x, data_format=data_format, mode='caffe') @keras_export('keras.applications.vgg19.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_CAFFE, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
9,818
38.276
87
py
keras
keras-master/keras/applications/inception_v3.py
# Copyright 2015 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """Inception V3 model for Keras. Reference: - [Rethinking the Inception Architecture for Computer Vision]( http://arxiv.org/abs/1512.00567) (CVPR 2016) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export WEIGHTS_PATH = ( 'https://storage.googleapis.com/tensorflow/keras-applications/' 'inception_v3/inception_v3_weights_tf_dim_ordering_tf_kernels.h5') WEIGHTS_PATH_NO_TOP = ( 'https://storage.googleapis.com/tensorflow/keras-applications/' 'inception_v3/inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5') layers = VersionAwareLayers() @keras_export('keras.applications.inception_v3.InceptionV3', 'keras.applications.InceptionV3') def InceptionV3( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the Inception v3 architecture. Reference: - [Rethinking the Inception Architecture for Computer Vision]( http://arxiv.org/abs/1512.00567) (CVPR 2016) This function returns a Keras image classification model, optionally loaded with weights pre-trained on ImageNet. For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For `InceptionV3`, call `tf.keras.applications.inception_v3.preprocess_input` on your inputs before passing them to the model. `inception_v3.preprocess_input` will scale input pixels between -1 and 1. Args: include_top: Boolean, whether to include the fully-connected layer at the top, as the last layer of the network. Default to `True`. weights: One of `None` (random initialization), `imagenet` (pre-training on ImageNet), or the path to the weights file to be loaded. Default to `imagenet`. input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. `input_tensor` is useful for sharing inputs between multiple different networks. Default to None. input_shape: Optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(299, 299, 3)` (with `channels_last` data format) or `(3, 299, 299)` (with `channels_first` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 75. E.g. `(150, 150, 3)` would be one valid value. `input_shape` will be ignored if the `input_tensor` is provided. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` (default) means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. Default to 1000. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded; ' f'Received: weights={weights}') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top` ' 'as true, `classes` should be 1000; ' f'Received classes={classes}') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=299, min_size=75, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor if backend.image_data_format() == 'channels_first': channel_axis = 1 else: channel_axis = 3 x = conv2d_bn(img_input, 32, 3, 3, strides=(2, 2), padding='valid') x = conv2d_bn(x, 32, 3, 3, padding='valid') x = conv2d_bn(x, 64, 3, 3) x = layers.MaxPooling2D((3, 3), strides=(2, 2))(x) x = conv2d_bn(x, 80, 1, 1, padding='valid') x = conv2d_bn(x, 192, 3, 3, padding='valid') x = layers.MaxPooling2D((3, 3), strides=(2, 2))(x) # mixed 0: 35 x 35 x 256 branch1x1 = conv2d_bn(x, 64, 1, 1) branch5x5 = conv2d_bn(x, 48, 1, 1) branch5x5 = conv2d_bn(branch5x5, 64, 5, 5) branch3x3dbl = conv2d_bn(x, 64, 1, 1) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch_pool = layers.AveragePooling2D( (3, 3), strides=(1, 1), padding='same')(x) branch_pool = conv2d_bn(branch_pool, 32, 1, 1) x = layers.concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool], axis=channel_axis, name='mixed0') # mixed 1: 35 x 35 x 288 branch1x1 = conv2d_bn(x, 64, 1, 1) branch5x5 = conv2d_bn(x, 48, 1, 1) branch5x5 = conv2d_bn(branch5x5, 64, 5, 5) branch3x3dbl = conv2d_bn(x, 64, 1, 1) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch_pool = layers.AveragePooling2D( (3, 3), strides=(1, 1), padding='same')(x) branch_pool = conv2d_bn(branch_pool, 64, 1, 1) x = layers.concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool], axis=channel_axis, name='mixed1') # mixed 2: 35 x 35 x 288 branch1x1 = conv2d_bn(x, 64, 1, 1) branch5x5 = conv2d_bn(x, 48, 1, 1) branch5x5 = conv2d_bn(branch5x5, 64, 5, 5) branch3x3dbl = conv2d_bn(x, 64, 1, 1) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch_pool = layers.AveragePooling2D( (3, 3), strides=(1, 1), padding='same')(x) branch_pool = conv2d_bn(branch_pool, 64, 1, 1) x = layers.concatenate([branch1x1, branch5x5, branch3x3dbl, branch_pool], axis=channel_axis, name='mixed2') # mixed 3: 17 x 17 x 768 branch3x3 = conv2d_bn(x, 384, 3, 3, strides=(2, 2), padding='valid') branch3x3dbl = conv2d_bn(x, 64, 1, 1) branch3x3dbl = conv2d_bn(branch3x3dbl, 96, 3, 3) branch3x3dbl = conv2d_bn( branch3x3dbl, 96, 3, 3, strides=(2, 2), padding='valid') branch_pool = layers.MaxPooling2D((3, 3), strides=(2, 2))(x) x = layers.concatenate([branch3x3, branch3x3dbl, branch_pool], axis=channel_axis, name='mixed3') # mixed 4: 17 x 17 x 768 branch1x1 = conv2d_bn(x, 192, 1, 1) branch7x7 = conv2d_bn(x, 128, 1, 1) branch7x7 = conv2d_bn(branch7x7, 128, 1, 7) branch7x7 = conv2d_bn(branch7x7, 192, 7, 1) branch7x7dbl = conv2d_bn(x, 128, 1, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 7, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 1, 7) branch7x7dbl = conv2d_bn(branch7x7dbl, 128, 7, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7) branch_pool = layers.AveragePooling2D( (3, 3), strides=(1, 1), padding='same')(x) branch_pool = conv2d_bn(branch_pool, 192, 1, 1) x = layers.concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool], axis=channel_axis, name='mixed4') # mixed 5, 6: 17 x 17 x 768 for i in range(2): branch1x1 = conv2d_bn(x, 192, 1, 1) branch7x7 = conv2d_bn(x, 160, 1, 1) branch7x7 = conv2d_bn(branch7x7, 160, 1, 7) branch7x7 = conv2d_bn(branch7x7, 192, 7, 1) branch7x7dbl = conv2d_bn(x, 160, 1, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 7, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 1, 7) branch7x7dbl = conv2d_bn(branch7x7dbl, 160, 7, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7) branch_pool = layers.AveragePooling2D((3, 3), strides=(1, 1), padding='same')( x) branch_pool = conv2d_bn(branch_pool, 192, 1, 1) x = layers.concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool], axis=channel_axis, name='mixed' + str(5 + i)) # mixed 7: 17 x 17 x 768 branch1x1 = conv2d_bn(x, 192, 1, 1) branch7x7 = conv2d_bn(x, 192, 1, 1) branch7x7 = conv2d_bn(branch7x7, 192, 1, 7) branch7x7 = conv2d_bn(branch7x7, 192, 7, 1) branch7x7dbl = conv2d_bn(x, 192, 1, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 7, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7) branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 7, 1) branch7x7dbl = conv2d_bn(branch7x7dbl, 192, 1, 7) branch_pool = layers.AveragePooling2D( (3, 3), strides=(1, 1), padding='same')(x) branch_pool = conv2d_bn(branch_pool, 192, 1, 1) x = layers.concatenate([branch1x1, branch7x7, branch7x7dbl, branch_pool], axis=channel_axis, name='mixed7') # mixed 8: 8 x 8 x 1280 branch3x3 = conv2d_bn(x, 192, 1, 1) branch3x3 = conv2d_bn(branch3x3, 320, 3, 3, strides=(2, 2), padding='valid') branch7x7x3 = conv2d_bn(x, 192, 1, 1) branch7x7x3 = conv2d_bn(branch7x7x3, 192, 1, 7) branch7x7x3 = conv2d_bn(branch7x7x3, 192, 7, 1) branch7x7x3 = conv2d_bn( branch7x7x3, 192, 3, 3, strides=(2, 2), padding='valid') branch_pool = layers.MaxPooling2D((3, 3), strides=(2, 2))(x) x = layers.concatenate([branch3x3, branch7x7x3, branch_pool], axis=channel_axis, name='mixed8') # mixed 9: 8 x 8 x 2048 for i in range(2): branch1x1 = conv2d_bn(x, 320, 1, 1) branch3x3 = conv2d_bn(x, 384, 1, 1) branch3x3_1 = conv2d_bn(branch3x3, 384, 1, 3) branch3x3_2 = conv2d_bn(branch3x3, 384, 3, 1) branch3x3 = layers.concatenate([branch3x3_1, branch3x3_2], axis=channel_axis, name='mixed9_' + str(i)) branch3x3dbl = conv2d_bn(x, 448, 1, 1) branch3x3dbl = conv2d_bn(branch3x3dbl, 384, 3, 3) branch3x3dbl_1 = conv2d_bn(branch3x3dbl, 384, 1, 3) branch3x3dbl_2 = conv2d_bn(branch3x3dbl, 384, 3, 1) branch3x3dbl = layers.concatenate([branch3x3dbl_1, branch3x3dbl_2], axis=channel_axis) branch_pool = layers.AveragePooling2D((3, 3), strides=(1, 1), padding='same')( x) branch_pool = conv2d_bn(branch_pool, 192, 1, 1) x = layers.concatenate([branch1x1, branch3x3, branch3x3dbl, branch_pool], axis=channel_axis, name='mixed' + str(9 + i)) if include_top: # Classification block x = layers.GlobalAveragePooling2D(name='avg_pool')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='inception_v3') # Load weights. if weights == 'imagenet': if include_top: weights_path = data_utils.get_file( 'inception_v3_weights_tf_dim_ordering_tf_kernels.h5', WEIGHTS_PATH, cache_subdir='models', file_hash='9a0d58056eeedaa3f26cb7ebd46da564') else: weights_path = data_utils.get_file( 'inception_v3_weights_tf_dim_ordering_tf_kernels_notop.h5', WEIGHTS_PATH_NO_TOP, cache_subdir='models', file_hash='bcbd6486424b2319ff4ef7d526e38f63') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model def conv2d_bn(x, filters, num_row, num_col, padding='same', strides=(1, 1), name=None): """Utility function to apply conv + BN. Args: x: input tensor. filters: filters in `Conv2D`. num_row: height of the convolution kernel. num_col: width of the convolution kernel. padding: padding mode in `Conv2D`. strides: strides in `Conv2D`. name: name of the ops; will become `name + '_conv'` for the convolution and `name + '_bn'` for the batch norm layer. Returns: Output tensor after applying `Conv2D` and `BatchNormalization`. """ if name is not None: bn_name = name + '_bn' conv_name = name + '_conv' else: bn_name = None conv_name = None if backend.image_data_format() == 'channels_first': bn_axis = 1 else: bn_axis = 3 x = layers.Conv2D( filters, (num_row, num_col), strides=strides, padding=padding, use_bias=False, name=conv_name)( x) x = layers.BatchNormalization(axis=bn_axis, scale=False, name=bn_name)(x) x = layers.Activation('relu', name=name)(x) return x @keras_export('keras.applications.inception_v3.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input(x, data_format=data_format, mode='tf') @keras_export('keras.applications.inception_v3.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
16,038
36.562061
87
py
keras
keras-master/keras/applications/mobilenet_v3.py
# Copyright 2020 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name # pylint: disable=missing-function-docstring """MobileNet v3 models for Keras.""" import tensorflow.compat.v2 as tf from keras import backend from keras import models from keras.applications import imagenet_utils from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.platform import tf_logging as logging from tensorflow.python.util.tf_export import keras_export # TODO(scottzhu): Change this to the GCS path. BASE_WEIGHT_PATH = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/mobilenet_v3/') WEIGHTS_HASHES = { 'large_224_0.75_float': ('765b44a33ad4005b3ac83185abf1d0eb', 'e7b4d1071996dd51a2c2ca2424570e20'), 'large_224_1.0_float': ('59e551e166be033d707958cf9e29a6a7', '037116398e07f018c0005ffcb0406831'), 'large_minimalistic_224_1.0_float': ('675e7b876c45c57e9e63e6d90a36599c', 'a2c33aed672524d1d0b4431808177695'), 'small_224_0.75_float': ('cb65d4e5be93758266aa0a7f2c6708b7', '4d2fe46f1c1f38057392514b0df1d673'), 'small_224_1.0_float': ('8768d4c2e7dee89b9d02b2d03d65d862', 'be7100780f875c06bcab93d76641aa26'), 'small_minimalistic_224_1.0_float': ('99cd97fb2fcdad2bf028eb838de69e37', '20d4e357df3f7a6361f3a288857b1051'), } layers = VersionAwareLayers() BASE_DOCSTRING = """Instantiates the {name} architecture. Reference: - [Searching for MobileNetV3]( https://arxiv.org/pdf/1905.02244.pdf) (ICCV 2019) The following table describes the performance of MobileNets v3: ------------------------------------------------------------------------ MACs stands for Multiply Adds |Classification Checkpoint|MACs(M)|Parameters(M)|Top1 Accuracy|Pixel1 CPU(ms)| |---|---|---|---|---| | mobilenet_v3_large_1.0_224 | 217 | 5.4 | 75.6 | 51.2 | | mobilenet_v3_large_0.75_224 | 155 | 4.0 | 73.3 | 39.8 | | mobilenet_v3_large_minimalistic_1.0_224 | 209 | 3.9 | 72.3 | 44.1 | | mobilenet_v3_small_1.0_224 | 66 | 2.9 | 68.1 | 15.8 | | mobilenet_v3_small_0.75_224 | 44 | 2.4 | 65.4 | 12.8 | | mobilenet_v3_small_minimalistic_1.0_224 | 65 | 2.0 | 61.9 | 12.2 | For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For ModelNetV3, by default input preprocessing is included as a part of the model (as a `Rescaling` layer), and thus `tf.keras.applications.mobilenet_v3.preprocess_input` is actually a pass-through function. In this use case, ModelNetV3 models expect their inputs to be float tensors of pixels with values in the [0-255] range. At the same time, preprocessing as a part of the model (i.e. `Rescaling` layer) can be disabled by setting `include_preprocessing` argument to False. With preprocessing disabled ModelNetV3 models expect their inputs to be float tensors of pixels with values in the [-1, 1] range. Args: input_shape: Optional shape tuple, to be specified if you would like to use a model with an input image resolution that is not (224, 224, 3). It should have exactly 3 inputs channels (224, 224, 3). You can also omit this option if you would like to infer input_shape from an input_tensor. If you choose to include both input_tensor and input_shape then input_shape will be used if they match, if the shapes do not match then we will throw an error. E.g. `(160, 160, 3)` would be one valid value. alpha: controls the width of the network. This is known as the depth multiplier in the MobileNetV3 paper, but the name is kept for consistency with MobileNetV1 in Keras. - If `alpha` < 1.0, proportionally decreases the number of filters in each layer. - If `alpha` > 1.0, proportionally increases the number of filters in each layer. - If `alpha` = 1, default number of filters from the paper are used at each layer. minimalistic: In addition to large and small models this module also contains so-called minimalistic models, these models have the same per-layer dimensions characteristic as MobilenetV3 however, they don't utilize any of the advanced blocks (squeeze-and-excite units, hard-swish, and 5x5 convolutions). While these models are less efficient on CPU, they are much more performant on GPU/DSP. include_top: Boolean, whether to include the fully-connected layer at the top of the network. Defaults to `True`. weights: String, one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. pooling: String, optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Integer, optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. dropout_rate: fraction of the input units to drop on the last layer. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. include_preprocessing: Boolean, whether to include the preprocessing layer (`Rescaling`) at the bottom of the network. Defaults to `True`. Call arguments: inputs: A floating point `numpy.array` or a `tf.Tensor`, 4D with 3 color channels, with values in the range [0, 255] if `include_preprocessing` is True and in the range [-1, 1] otherwise. Returns: A `keras.Model` instance. """ def MobileNetV3(stack_fn, last_point_ch, input_shape=None, alpha=1.0, model_type='large', minimalistic=False, include_top=True, weights='imagenet', input_tensor=None, classes=1000, pooling=None, dropout_rate=0.2, classifier_activation='softmax', include_preprocessing=True): if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded. ' f'Received weights={weights}') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top` ' 'as true, `classes` should be 1000. ' f'Received classes={classes}') # Determine proper input shape and default size. # If both input_shape and input_tensor are used, they should match if input_shape is not None and input_tensor is not None: try: is_input_t_tensor = backend.is_keras_tensor(input_tensor) except ValueError: try: is_input_t_tensor = backend.is_keras_tensor( layer_utils.get_source_inputs(input_tensor)) except ValueError: raise ValueError('input_tensor: ', input_tensor, 'is not type input_tensor. ' f'Received type(input_tensor)={type(input_tensor)}') if is_input_t_tensor: if backend.image_data_format() == 'channels_first': if backend.int_shape(input_tensor)[1] != input_shape[1]: raise ValueError('When backend.image_data_format()=channels_first, ' 'input_shape[1] must equal ' 'backend.int_shape(input_tensor)[1]. Received ' f'input_shape={input_shape}, ' 'backend.int_shape(input_tensor)=' f'{backend.int_shape(input_tensor)}') else: if backend.int_shape(input_tensor)[2] != input_shape[1]: raise ValueError('input_shape[1] must equal ' 'backend.int_shape(input_tensor)[2]. Received ' f'input_shape={input_shape}, ' 'backend.int_shape(input_tensor)=' f'{backend.int_shape(input_tensor)}') else: raise ValueError('input_tensor specified: ', input_tensor, 'is not a keras tensor') # If input_shape is None, infer shape from input_tensor if input_shape is None and input_tensor is not None: try: backend.is_keras_tensor(input_tensor) except ValueError: raise ValueError('input_tensor: ', input_tensor, 'is type: ', type(input_tensor), 'which is not a valid type') if backend.is_keras_tensor(input_tensor): if backend.image_data_format() == 'channels_first': rows = backend.int_shape(input_tensor)[2] cols = backend.int_shape(input_tensor)[3] input_shape = (3, cols, rows) else: rows = backend.int_shape(input_tensor)[1] cols = backend.int_shape(input_tensor)[2] input_shape = (cols, rows, 3) # If input_shape is None and input_tensor is None using standard shape if input_shape is None and input_tensor is None: input_shape = (None, None, 3) if backend.image_data_format() == 'channels_last': row_axis, col_axis = (0, 1) else: row_axis, col_axis = (1, 2) rows = input_shape[row_axis] cols = input_shape[col_axis] if rows and cols and (rows < 32 or cols < 32): raise ValueError('Input size must be at least 32x32; Received `input_shape=' f'{input_shape}`') if weights == 'imagenet': if (not minimalistic and alpha not in [0.75, 1.0] or minimalistic and alpha != 1.0): raise ValueError('If imagenet weights are being loaded, ' 'alpha can be one of `0.75`, `1.0` for non minimalistic ' 'or `1.0` for minimalistic only.') if rows != cols or rows != 224: logging.warning('`input_shape` is undefined or non-square, ' 'or `rows` is not 224. ' 'Weights for input shape (224, 224) will be ' 'loaded as the default.') if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 if minimalistic: kernel = 3 activation = relu se_ratio = None else: kernel = 5 activation = hard_swish se_ratio = 0.25 x = img_input if include_preprocessing: x = layers.Rescaling(scale=1. / 127.5, offset=-1.)(x) x = layers.Conv2D( 16, kernel_size=3, strides=(2, 2), padding='same', use_bias=False, name='Conv')(x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name='Conv/BatchNorm')(x) x = activation(x) x = stack_fn(x, kernel, activation, se_ratio) last_conv_ch = _depth(backend.int_shape(x)[channel_axis] * 6) # if the width multiplier is greater than 1 we # increase the number of output channels if alpha > 1.0: last_point_ch = _depth(last_point_ch * alpha) x = layers.Conv2D( last_conv_ch, kernel_size=1, padding='same', use_bias=False, name='Conv_1')(x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name='Conv_1/BatchNorm')(x) x = activation(x) x = layers.GlobalAveragePooling2D(keepdims=True)(x) x = layers.Conv2D( last_point_ch, kernel_size=1, padding='same', use_bias=True, name='Conv_2')(x) x = activation(x) if include_top: if dropout_rate > 0: x = layers.Dropout(dropout_rate)(x) x = layers.Conv2D(classes, kernel_size=1, padding='same', name='Logits')(x) x = layers.Flatten()(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Activation(activation=classifier_activation, name='Predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D(name='avg_pool')(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D(name='max_pool')(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = models.Model(inputs, x, name='MobilenetV3' + model_type) # Load weights. if weights == 'imagenet': model_name = '{}{}_224_{}_float'.format( model_type, '_minimalistic' if minimalistic else '', str(alpha)) if include_top: file_name = 'weights_mobilenet_v3_' + model_name + '.h5' file_hash = WEIGHTS_HASHES[model_name][0] else: file_name = 'weights_mobilenet_v3_' + model_name + '_no_top.h5' file_hash = WEIGHTS_HASHES[model_name][1] weights_path = data_utils.get_file( file_name, BASE_WEIGHT_PATH + file_name, cache_subdir='models', file_hash=file_hash) model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model @keras_export('keras.applications.MobileNetV3Small') def MobileNetV3Small(input_shape=None, alpha=1.0, minimalistic=False, include_top=True, weights='imagenet', input_tensor=None, classes=1000, pooling=None, dropout_rate=0.2, classifier_activation='softmax', include_preprocessing=True): def stack_fn(x, kernel, activation, se_ratio): def depth(d): return _depth(d * alpha) x = _inverted_res_block(x, 1, depth(16), 3, 2, se_ratio, relu, 0) x = _inverted_res_block(x, 72. / 16, depth(24), 3, 2, None, relu, 1) x = _inverted_res_block(x, 88. / 24, depth(24), 3, 1, None, relu, 2) x = _inverted_res_block(x, 4, depth(40), kernel, 2, se_ratio, activation, 3) x = _inverted_res_block(x, 6, depth(40), kernel, 1, se_ratio, activation, 4) x = _inverted_res_block(x, 6, depth(40), kernel, 1, se_ratio, activation, 5) x = _inverted_res_block(x, 3, depth(48), kernel, 1, se_ratio, activation, 6) x = _inverted_res_block(x, 3, depth(48), kernel, 1, se_ratio, activation, 7) x = _inverted_res_block(x, 6, depth(96), kernel, 2, se_ratio, activation, 8) x = _inverted_res_block(x, 6, depth(96), kernel, 1, se_ratio, activation, 9) x = _inverted_res_block(x, 6, depth(96), kernel, 1, se_ratio, activation, 10) return x return MobileNetV3(stack_fn, 1024, input_shape, alpha, 'small', minimalistic, include_top, weights, input_tensor, classes, pooling, dropout_rate, classifier_activation, include_preprocessing) @keras_export('keras.applications.MobileNetV3Large') def MobileNetV3Large(input_shape=None, alpha=1.0, minimalistic=False, include_top=True, weights='imagenet', input_tensor=None, classes=1000, pooling=None, dropout_rate=0.2, classifier_activation='softmax', include_preprocessing=True): def stack_fn(x, kernel, activation, se_ratio): def depth(d): return _depth(d * alpha) x = _inverted_res_block(x, 1, depth(16), 3, 1, None, relu, 0) x = _inverted_res_block(x, 4, depth(24), 3, 2, None, relu, 1) x = _inverted_res_block(x, 3, depth(24), 3, 1, None, relu, 2) x = _inverted_res_block(x, 3, depth(40), kernel, 2, se_ratio, relu, 3) x = _inverted_res_block(x, 3, depth(40), kernel, 1, se_ratio, relu, 4) x = _inverted_res_block(x, 3, depth(40), kernel, 1, se_ratio, relu, 5) x = _inverted_res_block(x, 6, depth(80), 3, 2, None, activation, 6) x = _inverted_res_block(x, 2.5, depth(80), 3, 1, None, activation, 7) x = _inverted_res_block(x, 2.3, depth(80), 3, 1, None, activation, 8) x = _inverted_res_block(x, 2.3, depth(80), 3, 1, None, activation, 9) x = _inverted_res_block(x, 6, depth(112), 3, 1, se_ratio, activation, 10) x = _inverted_res_block(x, 6, depth(112), 3, 1, se_ratio, activation, 11) x = _inverted_res_block(x, 6, depth(160), kernel, 2, se_ratio, activation, 12) x = _inverted_res_block(x, 6, depth(160), kernel, 1, se_ratio, activation, 13) x = _inverted_res_block(x, 6, depth(160), kernel, 1, se_ratio, activation, 14) return x return MobileNetV3(stack_fn, 1280, input_shape, alpha, 'large', minimalistic, include_top, weights, input_tensor, classes, pooling, dropout_rate, classifier_activation, include_preprocessing) MobileNetV3Small.__doc__ = BASE_DOCSTRING.format(name='MobileNetV3Small') MobileNetV3Large.__doc__ = BASE_DOCSTRING.format(name='MobileNetV3Large') def relu(x): return layers.ReLU()(x) def hard_sigmoid(x): return layers.ReLU(6.)(x + 3.) * (1. / 6.) def hard_swish(x): return layers.Multiply()([x, hard_sigmoid(x)]) # This function is taken from the original tf repo. # It ensures that all layers have a channel number that is divisible by 8 # It can be seen here: # https://github.com/tensorflow/models/blob/master/research/ # slim/nets/mobilenet/mobilenet.py def _depth(v, divisor=8, min_value=None): if min_value is None: min_value = divisor new_v = max(min_value, int(v + divisor / 2) // divisor * divisor) # Make sure that round down does not go down by more than 10%. if new_v < 0.9 * v: new_v += divisor return new_v def _se_block(inputs, filters, se_ratio, prefix): x = layers.GlobalAveragePooling2D( keepdims=True, name=prefix + 'squeeze_excite/AvgPool')( inputs) x = layers.Conv2D( _depth(filters * se_ratio), kernel_size=1, padding='same', name=prefix + 'squeeze_excite/Conv')( x) x = layers.ReLU(name=prefix + 'squeeze_excite/Relu')(x) x = layers.Conv2D( filters, kernel_size=1, padding='same', name=prefix + 'squeeze_excite/Conv_1')( x) x = hard_sigmoid(x) x = layers.Multiply(name=prefix + 'squeeze_excite/Mul')([inputs, x]) return x def _inverted_res_block(x, expansion, filters, kernel_size, stride, se_ratio, activation, block_id): channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 shortcut = x prefix = 'expanded_conv/' infilters = backend.int_shape(x)[channel_axis] if block_id: # Expand prefix = 'expanded_conv_{}/'.format(block_id) x = layers.Conv2D( _depth(infilters * expansion), kernel_size=1, padding='same', use_bias=False, name=prefix + 'expand')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name=prefix + 'expand/BatchNorm')( x) x = activation(x) if stride == 2: x = layers.ZeroPadding2D( padding=imagenet_utils.correct_pad(x, kernel_size), name=prefix + 'depthwise/pad')( x) x = layers.DepthwiseConv2D( kernel_size, strides=stride, padding='same' if stride == 1 else 'valid', use_bias=False, name=prefix + 'depthwise')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name=prefix + 'depthwise/BatchNorm')( x) x = activation(x) if se_ratio: x = _se_block(x, _depth(infilters * expansion), se_ratio, prefix) x = layers.Conv2D( filters, kernel_size=1, padding='same', use_bias=False, name=prefix + 'project')( x) x = layers.BatchNormalization( axis=channel_axis, epsilon=1e-3, momentum=0.999, name=prefix + 'project/BatchNorm')( x) if stride == 1 and infilters == filters: x = layers.Add(name=prefix + 'Add')([shortcut, x]) return x @keras_export('keras.applications.mobilenet_v3.preprocess_input') def preprocess_input(x, data_format=None): # pylint: disable=unused-argument """A placeholder method for backward compatibility. The preprocessing logic has been included in the mobilenet_v3 model implementation. Users are no longer required to call this method to normalize the input data. This method does nothing and only kept as a placeholder to align the API surface between old and new version of model. Args: x: A floating point `numpy.array` or a `tf.Tensor`. data_format: Optional data format of the image tensor/array. Defaults to None, in which case the global setting `tf.keras.backend.image_data_format()` is used (unless you changed it, it defaults to "channels_last").{mode} Returns: Unchanged `numpy.array` or `tf.Tensor`. """ return x @keras_export('keras.applications.mobilenet_v3.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
23,489
38.412752
87
py
keras
keras-master/keras/applications/nasnet.py
# Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """NASNet-A models for Keras. NASNet refers to Neural Architecture Search Network, a family of models that were designed automatically by learning the model architectures directly on the dataset of interest. Here we consider NASNet-A, the highest performance model that was found for the CIFAR-10 dataset, and then extended to ImageNet 2012 dataset, obtaining state of the art performance on CIFAR-10 and ImageNet 2012. Only the NASNet-A models, and their respective weights, which are suited for ImageNet 2012 are provided. The below table describes the performance on ImageNet 2012: -------------------------------------------------------------------------------- Architecture | Top-1 Acc | Top-5 Acc | Multiply-Adds | Params (M) -------------------------------------------------------------------------------- | NASNet-A (4 @ 1056) | 74.0 % | 91.6 % | 564 M | 5.3 | | NASNet-A (6 @ 4032) | 82.7 % | 96.2 % | 23.8 B | 88.9 | -------------------------------------------------------------------------------- Reference: - [Learning Transferable Architectures for Scalable Image Recognition]( https://arxiv.org/abs/1707.07012) (CVPR 2018) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.platform import tf_logging as logging from tensorflow.python.util.tf_export import keras_export BASE_WEIGHTS_PATH = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/nasnet/') NASNET_MOBILE_WEIGHT_PATH = BASE_WEIGHTS_PATH + 'NASNet-mobile.h5' NASNET_MOBILE_WEIGHT_PATH_NO_TOP = BASE_WEIGHTS_PATH + 'NASNet-mobile-no-top.h5' NASNET_LARGE_WEIGHT_PATH = BASE_WEIGHTS_PATH + 'NASNet-large.h5' NASNET_LARGE_WEIGHT_PATH_NO_TOP = BASE_WEIGHTS_PATH + 'NASNet-large-no-top.h5' layers = VersionAwareLayers() def NASNet(input_shape=None, penultimate_filters=4032, num_blocks=6, stem_block_filters=96, skip_reduction=True, filter_multiplier=2, include_top=True, weights='imagenet', input_tensor=None, pooling=None, classes=1000, default_size=None, classifier_activation='softmax'): """Instantiates a NASNet model. Reference: - [Learning Transferable Architectures for Scalable Image Recognition]( https://arxiv.org/abs/1707.07012) (CVPR 2018) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For NasNet, call `tf.keras.applications.nasnet.preprocess_input` on your inputs before passing them to the model. `nasnet.preprocess_input` will scale input pixels between -1 and 1. Args: input_shape: Optional shape tuple, the input shape is by default `(331, 331, 3)` for NASNetLarge and `(224, 224, 3)` for NASNetMobile. It should have exactly 3 input channels, and width and height should be no smaller than 32. E.g. `(224, 224, 3)` would be one valid value. penultimate_filters: Number of filters in the penultimate layer. NASNet models use the notation `NASNet (N @ P)`, where: - N is the number of blocks - P is the number of penultimate filters num_blocks: Number of repeated blocks of the NASNet model. NASNet models use the notation `NASNet (N @ P)`, where: - N is the number of blocks - P is the number of penultimate filters stem_block_filters: Number of filters in the initial stem block skip_reduction: Whether to skip the reduction step at the tail end of the network. filter_multiplier: Controls the width of the network. - If `filter_multiplier` < 1.0, proportionally decreases the number of filters in each layer. - If `filter_multiplier` > 1.0, proportionally increases the number of filters in each layer. - If `filter_multiplier` = 1, default number of filters from the paper are used at each layer. include_top: Whether to include the fully-connected layer at the top of the network. weights: `None` (random initialization) or `imagenet` (ImageNet weights) input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. default_size: Specifies the default image size of the model classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded.') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top` ' 'as true, `classes` should be 1000') if (isinstance(input_shape, tuple) and None in input_shape and weights == 'imagenet'): raise ValueError('When specifying the input shape of a NASNet' ' and loading `ImageNet` weights, ' 'the input_shape argument must be static ' '(no None entries). Got: `input_shape=' + str(input_shape) + '`.') if default_size is None: default_size = 331 # Determine proper input shape and default size. input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=default_size, min_size=32, data_format=backend.image_data_format(), require_flatten=True, weights=weights) if backend.image_data_format() != 'channels_last': logging.warning('The NASNet family of models is only available ' 'for the input data format "channels_last" ' '(width, height, channels). ' 'However your settings specify the default ' 'data format "channels_first" (channels, width, height).' ' You should set `image_data_format="channels_last"` ' 'in your Keras config located at ~/.keras/keras.json. ' 'The model being returned right now will expect inputs ' 'to follow the "channels_last" data format.') backend.set_image_data_format('channels_last') old_data_format = 'channels_first' else: old_data_format = None if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor if penultimate_filters % (24 * (filter_multiplier**2)) != 0: raise ValueError( 'For NASNet-A models, the `penultimate_filters` must be a multiple ' 'of 24 * (`filter_multiplier` ** 2). Current value: %d' % penultimate_filters) channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1 filters = penultimate_filters // 24 x = layers.Conv2D( stem_block_filters, (3, 3), strides=(2, 2), padding='valid', use_bias=False, name='stem_conv1', kernel_initializer='he_normal')( img_input) x = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='stem_bn1')( x) p = None x, p = _reduction_a_cell( x, p, filters // (filter_multiplier**2), block_id='stem_1') x, p = _reduction_a_cell( x, p, filters // filter_multiplier, block_id='stem_2') for i in range(num_blocks): x, p = _normal_a_cell(x, p, filters, block_id='%d' % (i)) x, p0 = _reduction_a_cell( x, p, filters * filter_multiplier, block_id='reduce_%d' % (num_blocks)) p = p0 if not skip_reduction else p for i in range(num_blocks): x, p = _normal_a_cell( x, p, filters * filter_multiplier, block_id='%d' % (num_blocks + i + 1)) x, p0 = _reduction_a_cell( x, p, filters * filter_multiplier**2, block_id='reduce_%d' % (2 * num_blocks)) p = p0 if not skip_reduction else p for i in range(num_blocks): x, p = _normal_a_cell( x, p, filters * filter_multiplier**2, block_id='%d' % (2 * num_blocks + i + 1)) x = layers.Activation('relu')(x) if include_top: x = layers.GlobalAveragePooling2D()(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input model = training.Model(inputs, x, name='NASNet') # Load weights. if weights == 'imagenet': if default_size == 224: # mobile version if include_top: weights_path = data_utils.get_file( 'nasnet_mobile.h5', NASNET_MOBILE_WEIGHT_PATH, cache_subdir='models', file_hash='020fb642bf7360b370c678b08e0adf61') else: weights_path = data_utils.get_file( 'nasnet_mobile_no_top.h5', NASNET_MOBILE_WEIGHT_PATH_NO_TOP, cache_subdir='models', file_hash='1ed92395b5b598bdda52abe5c0dbfd63') model.load_weights(weights_path) elif default_size == 331: # large version if include_top: weights_path = data_utils.get_file( 'nasnet_large.h5', NASNET_LARGE_WEIGHT_PATH, cache_subdir='models', file_hash='11577c9a518f0070763c2b964a382f17') else: weights_path = data_utils.get_file( 'nasnet_large_no_top.h5', NASNET_LARGE_WEIGHT_PATH_NO_TOP, cache_subdir='models', file_hash='d81d89dc07e6e56530c4e77faddd61b5') model.load_weights(weights_path) else: raise ValueError('ImageNet weights can only be loaded with NASNetLarge' ' or NASNetMobile') elif weights is not None: model.load_weights(weights) if old_data_format: backend.set_image_data_format(old_data_format) return model @keras_export('keras.applications.nasnet.NASNetMobile', 'keras.applications.NASNetMobile') def NASNetMobile(input_shape=None, include_top=True, weights='imagenet', input_tensor=None, pooling=None, classes=1000): """Instantiates a Mobile NASNet model in ImageNet mode. Reference: - [Learning Transferable Architectures for Scalable Image Recognition]( https://arxiv.org/abs/1707.07012) (CVPR 2018) Optionally loads weights pre-trained on ImageNet. Note that the data format convention used by the model is the one specified in your Keras config at `~/.keras/keras.json`. Note: each Keras Application expects a specific kind of input preprocessing. For NASNet, call `tf.keras.applications.nasnet.preprocess_input` on your inputs before passing them to the model. Args: input_shape: Optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` for NASNetMobile It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(224, 224, 3)` would be one valid value. include_top: Whether to include the fully-connected layer at the top of the network. weights: `None` (random initialization) or `imagenet` (ImageNet weights) For loading `imagenet` weights, `input_shape` should be (224, 224, 3) input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. Returns: A Keras model instance. Raises: ValueError: In case of invalid argument for `weights`, or invalid input shape. RuntimeError: If attempting to run this model with a backend that does not support separable convolutions. """ return NASNet( input_shape, penultimate_filters=1056, num_blocks=4, stem_block_filters=32, skip_reduction=False, filter_multiplier=2, include_top=include_top, weights=weights, input_tensor=input_tensor, pooling=pooling, classes=classes, default_size=224) @keras_export('keras.applications.nasnet.NASNetLarge', 'keras.applications.NASNetLarge') def NASNetLarge(input_shape=None, include_top=True, weights='imagenet', input_tensor=None, pooling=None, classes=1000): """Instantiates a NASNet model in ImageNet mode. Reference: - [Learning Transferable Architectures for Scalable Image Recognition]( https://arxiv.org/abs/1707.07012) (CVPR 2018) Optionally loads weights pre-trained on ImageNet. Note that the data format convention used by the model is the one specified in your Keras config at `~/.keras/keras.json`. Note: each Keras Application expects a specific kind of input preprocessing. For NASNet, call `tf.keras.applications.nasnet.preprocess_input` on your inputs before passing them to the model. Args: input_shape: Optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(331, 331, 3)` for NASNetLarge. It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(224, 224, 3)` would be one valid value. include_top: Whether to include the fully-connected layer at the top of the network. weights: `None` (random initialization) or `imagenet` (ImageNet weights) For loading `imagenet` weights, `input_shape` should be (331, 331, 3) input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional layer. - `avg` means that global average pooling will be applied to the output of the last convolutional layer, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. Returns: A Keras model instance. Raises: ValueError: in case of invalid argument for `weights`, or invalid input shape. RuntimeError: If attempting to run this model with a backend that does not support separable convolutions. """ return NASNet( input_shape, penultimate_filters=4032, num_blocks=6, stem_block_filters=96, skip_reduction=True, filter_multiplier=2, include_top=include_top, weights=weights, input_tensor=input_tensor, pooling=pooling, classes=classes, default_size=331) def _separable_conv_block(ip, filters, kernel_size=(3, 3), strides=(1, 1), block_id=None): """Adds 2 blocks of [relu-separable conv-batchnorm]. Args: ip: Input tensor filters: Number of output filters per layer kernel_size: Kernel size of separable convolutions strides: Strided convolution for downsampling block_id: String block_id Returns: A Keras tensor """ channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1 with backend.name_scope('separable_conv_block_%s' % block_id): x = layers.Activation('relu')(ip) if strides == (2, 2): x = layers.ZeroPadding2D( padding=imagenet_utils.correct_pad(x, kernel_size), name='separable_conv_1_pad_%s' % block_id)(x) conv_pad = 'valid' else: conv_pad = 'same' x = layers.SeparableConv2D( filters, kernel_size, strides=strides, name='separable_conv_1_%s' % block_id, padding=conv_pad, use_bias=False, kernel_initializer='he_normal')( x) x = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='separable_conv_1_bn_%s' % (block_id))( x) x = layers.Activation('relu')(x) x = layers.SeparableConv2D( filters, kernel_size, name='separable_conv_2_%s' % block_id, padding='same', use_bias=False, kernel_initializer='he_normal')( x) x = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='separable_conv_2_bn_%s' % (block_id))( x) return x def _adjust_block(p, ip, filters, block_id=None): """Adjusts the input `previous path` to match the shape of the `input`. Used in situations where the output number of filters needs to be changed. Args: p: Input tensor which needs to be modified ip: Input tensor whose shape needs to be matched filters: Number of output filters to be matched block_id: String block_id Returns: Adjusted Keras tensor """ channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1 img_dim = 2 if backend.image_data_format() == 'channels_first' else -2 ip_shape = backend.int_shape(ip) if p is not None: p_shape = backend.int_shape(p) with backend.name_scope('adjust_block'): if p is None: p = ip elif p_shape[img_dim] != ip_shape[img_dim]: with backend.name_scope('adjust_reduction_block_%s' % block_id): p = layers.Activation('relu', name='adjust_relu_1_%s' % block_id)(p) p1 = layers.AveragePooling2D((1, 1), strides=(2, 2), padding='valid', name='adjust_avg_pool_1_%s' % block_id)( p) p1 = layers.Conv2D( filters // 2, (1, 1), padding='same', use_bias=False, name='adjust_conv_1_%s' % block_id, kernel_initializer='he_normal')( p1) p2 = layers.ZeroPadding2D(padding=((0, 1), (0, 1)))(p) p2 = layers.Cropping2D(cropping=((1, 0), (1, 0)))(p2) p2 = layers.AveragePooling2D((1, 1), strides=(2, 2), padding='valid', name='adjust_avg_pool_2_%s' % block_id)( p2) p2 = layers.Conv2D( filters // 2, (1, 1), padding='same', use_bias=False, name='adjust_conv_2_%s' % block_id, kernel_initializer='he_normal')( p2) p = layers.concatenate([p1, p2], axis=channel_dim) p = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='adjust_bn_%s' % block_id)( p) elif p_shape[channel_dim] != filters: with backend.name_scope('adjust_projection_block_%s' % block_id): p = layers.Activation('relu')(p) p = layers.Conv2D( filters, (1, 1), strides=(1, 1), padding='same', name='adjust_conv_projection_%s' % block_id, use_bias=False, kernel_initializer='he_normal')( p) p = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='adjust_bn_%s' % block_id)( p) return p def _normal_a_cell(ip, p, filters, block_id=None): """Adds a Normal cell for NASNet-A (Fig. 4 in the paper). Args: ip: Input tensor `x` p: Input tensor `p` filters: Number of output filters block_id: String block_id Returns: A Keras tensor """ channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1 with backend.name_scope('normal_A_block_%s' % block_id): p = _adjust_block(p, ip, filters, block_id) h = layers.Activation('relu')(ip) h = layers.Conv2D( filters, (1, 1), strides=(1, 1), padding='same', name='normal_conv_1_%s' % block_id, use_bias=False, kernel_initializer='he_normal')( h) h = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='normal_bn_1_%s' % block_id)( h) with backend.name_scope('block_1'): x1_1 = _separable_conv_block( h, filters, kernel_size=(5, 5), block_id='normal_left1_%s' % block_id) x1_2 = _separable_conv_block( p, filters, block_id='normal_right1_%s' % block_id) x1 = layers.add([x1_1, x1_2], name='normal_add_1_%s' % block_id) with backend.name_scope('block_2'): x2_1 = _separable_conv_block( p, filters, (5, 5), block_id='normal_left2_%s' % block_id) x2_2 = _separable_conv_block( p, filters, (3, 3), block_id='normal_right2_%s' % block_id) x2 = layers.add([x2_1, x2_2], name='normal_add_2_%s' % block_id) with backend.name_scope('block_3'): x3 = layers.AveragePooling2D((3, 3), strides=(1, 1), padding='same', name='normal_left3_%s' % (block_id))( h) x3 = layers.add([x3, p], name='normal_add_3_%s' % block_id) with backend.name_scope('block_4'): x4_1 = layers.AveragePooling2D((3, 3), strides=(1, 1), padding='same', name='normal_left4_%s' % (block_id))( p) x4_2 = layers.AveragePooling2D((3, 3), strides=(1, 1), padding='same', name='normal_right4_%s' % (block_id))( p) x4 = layers.add([x4_1, x4_2], name='normal_add_4_%s' % block_id) with backend.name_scope('block_5'): x5 = _separable_conv_block( h, filters, block_id='normal_left5_%s' % block_id) x5 = layers.add([x5, h], name='normal_add_5_%s' % block_id) x = layers.concatenate([p, x1, x2, x3, x4, x5], axis=channel_dim, name='normal_concat_%s' % block_id) return x, ip def _reduction_a_cell(ip, p, filters, block_id=None): """Adds a Reduction cell for NASNet-A (Fig. 4 in the paper). Args: ip: Input tensor `x` p: Input tensor `p` filters: Number of output filters block_id: String block_id Returns: A Keras tensor """ channel_dim = 1 if backend.image_data_format() == 'channels_first' else -1 with backend.name_scope('reduction_A_block_%s' % block_id): p = _adjust_block(p, ip, filters, block_id) h = layers.Activation('relu')(ip) h = layers.Conv2D( filters, (1, 1), strides=(1, 1), padding='same', name='reduction_conv_1_%s' % block_id, use_bias=False, kernel_initializer='he_normal')( h) h = layers.BatchNormalization( axis=channel_dim, momentum=0.9997, epsilon=1e-3, name='reduction_bn_1_%s' % block_id)( h) h3 = layers.ZeroPadding2D( padding=imagenet_utils.correct_pad(h, 3), name='reduction_pad_1_%s' % block_id)( h) with backend.name_scope('block_1'): x1_1 = _separable_conv_block( h, filters, (5, 5), strides=(2, 2), block_id='reduction_left1_%s' % block_id) x1_2 = _separable_conv_block( p, filters, (7, 7), strides=(2, 2), block_id='reduction_right1_%s' % block_id) x1 = layers.add([x1_1, x1_2], name='reduction_add_1_%s' % block_id) with backend.name_scope('block_2'): x2_1 = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='valid', name='reduction_left2_%s' % block_id)( h3) x2_2 = _separable_conv_block( p, filters, (7, 7), strides=(2, 2), block_id='reduction_right2_%s' % block_id) x2 = layers.add([x2_1, x2_2], name='reduction_add_2_%s' % block_id) with backend.name_scope('block_3'): x3_1 = layers.AveragePooling2D((3, 3), strides=(2, 2), padding='valid', name='reduction_left3_%s' % block_id)( h3) x3_2 = _separable_conv_block( p, filters, (5, 5), strides=(2, 2), block_id='reduction_right3_%s' % block_id) x3 = layers.add([x3_1, x3_2], name='reduction_add3_%s' % block_id) with backend.name_scope('block_4'): x4 = layers.AveragePooling2D((3, 3), strides=(1, 1), padding='same', name='reduction_left4_%s' % block_id)( x1) x4 = layers.add([x2, x4]) with backend.name_scope('block_5'): x5_1 = _separable_conv_block( x1, filters, (3, 3), block_id='reduction_left4_%s' % block_id) x5_2 = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='valid', name='reduction_right5_%s' % block_id)( h3) x5 = layers.add([x5_1, x5_2], name='reduction_add4_%s' % block_id) x = layers.concatenate([x2, x3, x4, x5], axis=channel_dim, name='reduction_concat_%s' % block_id) return x, ip @keras_export('keras.applications.nasnet.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input(x, data_format=data_format, mode='tf') @keras_export('keras.applications.nasnet.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
30,441
36.215159
87
py
keras
keras-master/keras/applications/imagenet_utils_test.py
# Copyright 2019 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Tests for imagenet_utils.""" import tensorflow.compat.v2 as tf from absl.testing import parameterized import numpy as np import keras from keras import keras_parameterized from keras.applications import imagenet_utils as utils from keras.mixed_precision.policy import set_policy class TestImageNetUtils(keras_parameterized.TestCase): def test_preprocess_input(self): # Test invalid mode check x = np.random.uniform(0, 255, (10, 10, 3)) with self.assertRaises(ValueError): utils.preprocess_input(x, mode='some_unknown_mode') # Test image batch with float and int image input x = np.random.uniform(0, 255, (2, 10, 10, 3)) xint = x.astype('int32') self.assertEqual(utils.preprocess_input(x).shape, x.shape) self.assertEqual(utils.preprocess_input(xint).shape, xint.shape) out1 = utils.preprocess_input(x, 'channels_last') out1int = utils.preprocess_input(xint, 'channels_last') out2 = utils.preprocess_input( np.transpose(x, (0, 3, 1, 2)), 'channels_first') out2int = utils.preprocess_input( np.transpose(xint, (0, 3, 1, 2)), 'channels_first') self.assertAllClose(out1, out2.transpose(0, 2, 3, 1)) self.assertAllClose(out1int, out2int.transpose(0, 2, 3, 1)) # Test single image x = np.random.uniform(0, 255, (10, 10, 3)) xint = x.astype('int32') self.assertEqual(utils.preprocess_input(x).shape, x.shape) self.assertEqual(utils.preprocess_input(xint).shape, xint.shape) out1 = utils.preprocess_input(x, 'channels_last') out1int = utils.preprocess_input(xint, 'channels_last') out2 = utils.preprocess_input(np.transpose(x, (2, 0, 1)), 'channels_first') out2int = utils.preprocess_input( np.transpose(xint, (2, 0, 1)), 'channels_first') self.assertAllClose(out1, out2.transpose(1, 2, 0)) self.assertAllClose(out1int, out2int.transpose(1, 2, 0)) # Test that writing over the input data works predictably for mode in ['torch', 'tf']: x = np.random.uniform(0, 255, (2, 10, 10, 3)) xint = x.astype('int') x2 = utils.preprocess_input(x, mode=mode) xint2 = utils.preprocess_input(xint) self.assertAllClose(x, x2) self.assertNotEqual(xint.astype('float').max(), xint2.max()) # Caffe mode works differently from the others x = np.random.uniform(0, 255, (2, 10, 10, 3)) xint = x.astype('int') x2 = utils.preprocess_input(x, data_format='channels_last', mode='caffe') xint2 = utils.preprocess_input(xint) self.assertAllClose(x, x2[..., ::-1]) self.assertNotEqual(xint.astype('float').max(), xint2.max()) @parameterized.named_parameters([ { 'testcase_name': 'mode_torch', 'mode': 'torch' }, { 'testcase_name': 'mode_tf', 'mode': 'tf' }, { 'testcase_name': 'mode_caffe', 'mode': 'caffe' }, ]) def test_preprocess_input_symbolic(self, mode): # Test image batch x = np.random.uniform(0, 255, (2, 10, 10, 3)) inputs = keras.layers.Input(shape=x.shape[1:]) outputs = keras.layers.Lambda( lambda x: utils.preprocess_input(x, mode=mode), output_shape=x.shape[1:])( inputs) model = keras.Model(inputs, outputs) self.assertEqual(model.predict(x).shape, x.shape) outputs1 = keras.layers.Lambda( lambda x: utils.preprocess_input(x, 'channels_last', mode=mode), output_shape=x.shape[1:])( inputs) model1 = keras.Model(inputs, outputs1) out1 = model1.predict(x) x2 = np.transpose(x, (0, 3, 1, 2)) inputs2 = keras.layers.Input(shape=x2.shape[1:]) outputs2 = keras.layers.Lambda( lambda x: utils.preprocess_input(x, 'channels_first', mode=mode), output_shape=x2.shape[1:])( inputs2) model2 = keras.Model(inputs2, outputs2) out2 = model2.predict(x2) self.assertAllClose(out1, out2.transpose(0, 2, 3, 1)) # Test single image x = np.random.uniform(0, 255, (10, 10, 3)) inputs = keras.layers.Input(shape=x.shape) outputs = keras.layers.Lambda( lambda x: utils.preprocess_input(x, mode=mode), output_shape=x.shape)( inputs) model = keras.Model(inputs, outputs) self.assertEqual(model.predict(x[np.newaxis])[0].shape, x.shape) outputs1 = keras.layers.Lambda( lambda x: utils.preprocess_input(x, 'channels_last', mode=mode), output_shape=x.shape)( inputs) model1 = keras.Model(inputs, outputs1) out1 = model1.predict(x[np.newaxis])[0] x2 = np.transpose(x, (2, 0, 1)) inputs2 = keras.layers.Input(shape=x2.shape) outputs2 = keras.layers.Lambda( lambda x: utils.preprocess_input(x, 'channels_first', mode=mode), output_shape=x2.shape)( inputs2) model2 = keras.Model(inputs2, outputs2) out2 = model2.predict(x2[np.newaxis])[0] self.assertAllClose(out1, out2.transpose(1, 2, 0)) @parameterized.named_parameters([ { 'testcase_name': 'mode_torch', 'mode': 'torch' }, { 'testcase_name': 'mode_tf', 'mode': 'tf' }, { 'testcase_name': 'mode_caffe', 'mode': 'caffe' }, ]) def test_preprocess_input_symbolic_mixed_precision(self, mode): set_policy('mixed_float16') shape = (20, 20, 3) inputs = keras.layers.Input(shape=shape) try: keras.layers.Lambda( lambda x: utils.preprocess_input(x, mode=mode), output_shape=shape)( inputs) finally: set_policy('float32') @parameterized.named_parameters([ {'testcase_name': 'channels_last_format', 'data_format': 'channels_last'}, {'testcase_name': 'channels_first_format', 'data_format': 'channels_first'}, ]) def test_obtain_input_shape(self, data_format): # input_shape and default_size are not identical. with self.assertRaises(ValueError): utils.obtain_input_shape( input_shape=(224, 224, 3), default_size=299, min_size=139, data_format='channels_last', require_flatten=True, weights='imagenet') # Test invalid use cases shape = (139, 139) if data_format == 'channels_last': input_shape = shape + (99,) else: input_shape = (99,) + shape # input_shape is smaller than min_size. shape = (100, 100) if data_format == 'channels_last': input_shape = shape + (3,) else: input_shape = (3,) + shape with self.assertRaises(ValueError): utils.obtain_input_shape( input_shape=input_shape, default_size=None, min_size=139, data_format=data_format, require_flatten=False) # shape is 1D. shape = (100,) if data_format == 'channels_last': input_shape = shape + (3,) else: input_shape = (3,) + shape with self.assertRaises(ValueError): utils.obtain_input_shape( input_shape=input_shape, default_size=None, min_size=139, data_format=data_format, require_flatten=False) # the number of channels is 5 not 3. shape = (100, 100) if data_format == 'channels_last': input_shape = shape + (5,) else: input_shape = (5,) + shape with self.assertRaises(ValueError): utils.obtain_input_shape( input_shape=input_shape, default_size=None, min_size=139, data_format=data_format, require_flatten=False) # require_flatten=True with dynamic input shape. with self.assertRaises(ValueError): utils.obtain_input_shape( input_shape=None, default_size=None, min_size=139, data_format='channels_first', require_flatten=True) # test include top self.assertEqual(utils.obtain_input_shape( input_shape=(3, 200, 200), default_size=None, min_size=139, data_format='channels_first', require_flatten=True), (3, 200, 200)) self.assertEqual(utils.obtain_input_shape( input_shape=None, default_size=None, min_size=139, data_format='channels_last', require_flatten=False), (None, None, 3)) self.assertEqual(utils.obtain_input_shape( input_shape=None, default_size=None, min_size=139, data_format='channels_first', require_flatten=False), (3, None, None)) self.assertEqual(utils.obtain_input_shape( input_shape=None, default_size=None, min_size=139, data_format='channels_last', require_flatten=False), (None, None, 3)) self.assertEqual(utils.obtain_input_shape( input_shape=(150, 150, 3), default_size=None, min_size=139, data_format='channels_last', require_flatten=False), (150, 150, 3)) self.assertEqual(utils.obtain_input_shape( input_shape=(3, None, None), default_size=None, min_size=139, data_format='channels_first', require_flatten=False), (3, None, None)) if __name__ == '__main__': tf.test.main()
9,851
32.39661
80
py
keras
keras-master/keras/applications/applications_load_weight_test.py
# Copyright 2020 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Integration tests for Keras applications.""" import tensorflow.compat.v2 as tf from absl import flags from absl.testing import parameterized import numpy as np from keras.applications import densenet from keras.applications import efficientnet from keras.applications import inception_resnet_v2 from keras.applications import inception_v3 from keras.applications import mobilenet from keras.applications import mobilenet_v2 from keras.applications import mobilenet_v3 from keras.applications import nasnet from keras.applications import resnet from keras.applications import resnet_v2 from keras.applications import vgg16 from keras.applications import vgg19 from keras.applications import xception from keras.preprocessing import image from keras.utils import data_utils ARG_TO_MODEL = { 'resnet': (resnet, [resnet.ResNet50, resnet.ResNet101, resnet.ResNet152]), 'resnet_v2': (resnet_v2, [resnet_v2.ResNet50V2, resnet_v2.ResNet101V2, resnet_v2.ResNet152V2]), 'vgg16': (vgg16, [vgg16.VGG16]), 'vgg19': (vgg19, [vgg19.VGG19]), 'xception': (xception, [xception.Xception]), 'inception_v3': (inception_v3, [inception_v3.InceptionV3]), 'inception_resnet_v2': (inception_resnet_v2, [inception_resnet_v2.InceptionResNetV2]), 'mobilenet': (mobilenet, [mobilenet.MobileNet]), 'mobilenet_v2': (mobilenet_v2, [mobilenet_v2.MobileNetV2]), 'mobilenet_v3_small': (mobilenet_v3, [mobilenet_v3.MobileNetV3Small]), 'mobilenet_v3_large': (mobilenet_v3, [mobilenet_v3.MobileNetV3Large]), 'densenet': (densenet, [densenet.DenseNet121, densenet.DenseNet169, densenet.DenseNet201]), 'nasnet_mobile': (nasnet, [nasnet.NASNetMobile]), 'nasnet_large': (nasnet, [nasnet.NASNetLarge]), 'efficientnet': (efficientnet, [efficientnet.EfficientNetB0, efficientnet.EfficientNetB1, efficientnet.EfficientNetB2, efficientnet.EfficientNetB3, efficientnet.EfficientNetB4, efficientnet.EfficientNetB5, efficientnet.EfficientNetB6, efficientnet.EfficientNetB7]) } TEST_IMAGE_PATH = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/tests/elephant.jpg') _IMAGENET_CLASSES = 1000 # Add a flag to define which application module file is tested. # This is set as an 'arg' in the build target to guarantee that # it only triggers the tests of the application models in the module # if that module file has been modified. FLAGS = flags.FLAGS flags.DEFINE_string('module', None, 'Application module used in this test.') def _get_elephant(target_size): # For models that don't include a Flatten step, # the default is to accept variable-size inputs # even when loading ImageNet weights (since it is possible). # In this case, default to 299x299. if target_size[0] is None: target_size = (299, 299) test_image = data_utils.get_file('elephant.jpg', TEST_IMAGE_PATH) img = image.load_img(test_image, target_size=tuple(target_size)) x = image.img_to_array(img) return np.expand_dims(x, axis=0) class ApplicationsLoadWeightTest(tf.test.TestCase, parameterized.TestCase): def assertShapeEqual(self, shape1, shape2): if len(shape1) != len(shape2): raise AssertionError( 'Shapes are different rank: %s vs %s' % (shape1, shape2)) if shape1 != shape2: raise AssertionError('Shapes differ: %s vs %s' % (shape1, shape2)) def test_application_pretrained_weights_loading(self): app_module = ARG_TO_MODEL[FLAGS.module][0] apps = ARG_TO_MODEL[FLAGS.module][1] for app in apps: model = app(weights='imagenet') self.assertShapeEqual(model.output_shape, (None, _IMAGENET_CLASSES)) x = _get_elephant(model.input_shape[1:3]) x = app_module.preprocess_input(x) preds = model.predict(x) names = [p[1] for p in app_module.decode_predictions(preds)[0]] # Test correct label is in top 3 (weak correctness test). self.assertIn('African_elephant', names[:3]) if __name__ == '__main__': tf.test.main()
4,840
40.732759
80
py
keras
keras-master/keras/applications/xception.py
# Copyright 2016 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """Xception V1 model for Keras. On ImageNet, this model gets to a top-1 validation accuracy of 0.790 and a top-5 validation accuracy of 0.945. Reference: - [Xception: Deep Learning with Depthwise Separable Convolutions]( https://arxiv.org/abs/1610.02357) (CVPR 2017) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export TF_WEIGHTS_PATH = ( 'https://storage.googleapis.com/tensorflow/keras-applications/' 'xception/xception_weights_tf_dim_ordering_tf_kernels.h5') TF_WEIGHTS_PATH_NO_TOP = ( 'https://storage.googleapis.com/tensorflow/keras-applications/' 'xception/xception_weights_tf_dim_ordering_tf_kernels_notop.h5') layers = VersionAwareLayers() @keras_export('keras.applications.xception.Xception', 'keras.applications.Xception') def Xception( include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax'): """Instantiates the Xception architecture. Reference: - [Xception: Deep Learning with Depthwise Separable Convolutions]( https://arxiv.org/abs/1610.02357) (CVPR 2017) For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). The default input image size for this model is 299x299. Note: each Keras Application expects a specific kind of input preprocessing. For Xception, call `tf.keras.applications.xception.preprocess_input` on your inputs before passing them to the model. `xception.preprocess_input` will scale input pixels between -1 and 1. Args: include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(299, 299, 3)`. It should have exactly 3 inputs channels, and width and height should be no smaller than 71. E.g. `(150, 150, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. Returns: A `keras.Model` instance. """ if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded.') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top`' ' as true, `classes` should be 1000') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=299, min_size=71, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 x = layers.Conv2D( 32, (3, 3), strides=(2, 2), use_bias=False, name='block1_conv1')(img_input) x = layers.BatchNormalization(axis=channel_axis, name='block1_conv1_bn')(x) x = layers.Activation('relu', name='block1_conv1_act')(x) x = layers.Conv2D(64, (3, 3), use_bias=False, name='block1_conv2')(x) x = layers.BatchNormalization(axis=channel_axis, name='block1_conv2_bn')(x) x = layers.Activation('relu', name='block1_conv2_act')(x) residual = layers.Conv2D( 128, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) residual = layers.BatchNormalization(axis=channel_axis)(residual) x = layers.SeparableConv2D( 128, (3, 3), padding='same', use_bias=False, name='block2_sepconv1')(x) x = layers.BatchNormalization(axis=channel_axis, name='block2_sepconv1_bn')(x) x = layers.Activation('relu', name='block2_sepconv2_act')(x) x = layers.SeparableConv2D( 128, (3, 3), padding='same', use_bias=False, name='block2_sepconv2')(x) x = layers.BatchNormalization(axis=channel_axis, name='block2_sepconv2_bn')(x) x = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='same', name='block2_pool')(x) x = layers.add([x, residual]) residual = layers.Conv2D( 256, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) residual = layers.BatchNormalization(axis=channel_axis)(residual) x = layers.Activation('relu', name='block3_sepconv1_act')(x) x = layers.SeparableConv2D( 256, (3, 3), padding='same', use_bias=False, name='block3_sepconv1')(x) x = layers.BatchNormalization(axis=channel_axis, name='block3_sepconv1_bn')(x) x = layers.Activation('relu', name='block3_sepconv2_act')(x) x = layers.SeparableConv2D( 256, (3, 3), padding='same', use_bias=False, name='block3_sepconv2')(x) x = layers.BatchNormalization(axis=channel_axis, name='block3_sepconv2_bn')(x) x = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='same', name='block3_pool')(x) x = layers.add([x, residual]) residual = layers.Conv2D( 728, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) residual = layers.BatchNormalization(axis=channel_axis)(residual) x = layers.Activation('relu', name='block4_sepconv1_act')(x) x = layers.SeparableConv2D( 728, (3, 3), padding='same', use_bias=False, name='block4_sepconv1')(x) x = layers.BatchNormalization(axis=channel_axis, name='block4_sepconv1_bn')(x) x = layers.Activation('relu', name='block4_sepconv2_act')(x) x = layers.SeparableConv2D( 728, (3, 3), padding='same', use_bias=False, name='block4_sepconv2')(x) x = layers.BatchNormalization(axis=channel_axis, name='block4_sepconv2_bn')(x) x = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='same', name='block4_pool')(x) x = layers.add([x, residual]) for i in range(8): residual = x prefix = 'block' + str(i + 5) x = layers.Activation('relu', name=prefix + '_sepconv1_act')(x) x = layers.SeparableConv2D( 728, (3, 3), padding='same', use_bias=False, name=prefix + '_sepconv1')(x) x = layers.BatchNormalization( axis=channel_axis, name=prefix + '_sepconv1_bn')(x) x = layers.Activation('relu', name=prefix + '_sepconv2_act')(x) x = layers.SeparableConv2D( 728, (3, 3), padding='same', use_bias=False, name=prefix + '_sepconv2')(x) x = layers.BatchNormalization( axis=channel_axis, name=prefix + '_sepconv2_bn')(x) x = layers.Activation('relu', name=prefix + '_sepconv3_act')(x) x = layers.SeparableConv2D( 728, (3, 3), padding='same', use_bias=False, name=prefix + '_sepconv3')(x) x = layers.BatchNormalization( axis=channel_axis, name=prefix + '_sepconv3_bn')(x) x = layers.add([x, residual]) residual = layers.Conv2D( 1024, (1, 1), strides=(2, 2), padding='same', use_bias=False)(x) residual = layers.BatchNormalization(axis=channel_axis)(residual) x = layers.Activation('relu', name='block13_sepconv1_act')(x) x = layers.SeparableConv2D( 728, (3, 3), padding='same', use_bias=False, name='block13_sepconv1')(x) x = layers.BatchNormalization( axis=channel_axis, name='block13_sepconv1_bn')(x) x = layers.Activation('relu', name='block13_sepconv2_act')(x) x = layers.SeparableConv2D( 1024, (3, 3), padding='same', use_bias=False, name='block13_sepconv2')(x) x = layers.BatchNormalization( axis=channel_axis, name='block13_sepconv2_bn')(x) x = layers.MaxPooling2D((3, 3), strides=(2, 2), padding='same', name='block13_pool')(x) x = layers.add([x, residual]) x = layers.SeparableConv2D( 1536, (3, 3), padding='same', use_bias=False, name='block14_sepconv1')(x) x = layers.BatchNormalization( axis=channel_axis, name='block14_sepconv1_bn')(x) x = layers.Activation('relu', name='block14_sepconv1_act')(x) x = layers.SeparableConv2D( 2048, (3, 3), padding='same', use_bias=False, name='block14_sepconv2')(x) x = layers.BatchNormalization( axis=channel_axis, name='block14_sepconv2_bn')(x) x = layers.Activation('relu', name='block14_sepconv2_act')(x) if include_top: x = layers.GlobalAveragePooling2D(name='avg_pool')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='xception') # Load weights. if weights == 'imagenet': if include_top: weights_path = data_utils.get_file( 'xception_weights_tf_dim_ordering_tf_kernels.h5', TF_WEIGHTS_PATH, cache_subdir='models', file_hash='0a58e3b7378bc2990ea3b43d5981f1f6') else: weights_path = data_utils.get_file( 'xception_weights_tf_dim_ordering_tf_kernels_notop.h5', TF_WEIGHTS_PATH_NO_TOP, cache_subdir='models', file_hash='b0042744bf5b25fce3cb969f33bebb97') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model @keras_export('keras.applications.xception.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input(x, data_format=data_format, mode='tf') @keras_export('keras.applications.xception.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
13,000
38.159639
87
py
keras
keras-master/keras/applications/inception_resnet_v2.py
# Copyright 2017 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """Inception-ResNet V2 model for Keras. Reference: - [Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning](https://arxiv.org/abs/1602.07261) (AAAI 2017) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.util.tf_export import keras_export BASE_WEIGHT_URL = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/inception_resnet_v2/') layers = None @keras_export('keras.applications.inception_resnet_v2.InceptionResNetV2', 'keras.applications.InceptionResNetV2') def InceptionResNetV2(include_top=True, weights='imagenet', input_tensor=None, input_shape=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): """Instantiates the Inception-ResNet v2 architecture. Reference: - [Inception-v4, Inception-ResNet and the Impact of Residual Connections on Learning](https://arxiv.org/abs/1602.07261) (AAAI 2017) This function returns a Keras image classification model, optionally loaded with weights pre-trained on ImageNet. For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For InceptionResNetV2, call `tf.keras.applications.inception_resnet_v2.preprocess_input` on your inputs before passing them to the model. `inception_resnet_v2.preprocess_input` will scale input pixels between -1 and 1. Args: include_top: whether to include the fully-connected layer at the top of the network. weights: one of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. input_tensor: optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. input_shape: optional shape tuple, only to be specified if `include_top` is `False` (otherwise the input shape has to be `(299, 299, 3)` (with `'channels_last'` data format) or `(3, 299, 299)` (with `'channels_first'` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 75. E.g. `(150, 150, 3)` would be one valid value. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` means that the output of the model will be the 4D tensor output of the last convolutional block. - `'avg'` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `'max'` means that global max pooling will be applied. classes: optional number of classes to classify images into, only to be specified if `include_top` is `True`, and if no `weights` argument is specified. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. **kwargs: For backwards compatibility only. Returns: A `keras.Model` instance. """ global layers if 'layers' in kwargs: layers = kwargs.pop('layers') else: layers = VersionAwareLayers() if kwargs: raise ValueError('Unknown argument(s): %s' % (kwargs,)) if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded.') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top`' ' as true, `classes` should be 1000') # Determine proper input shape input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=299, min_size=75, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor # Stem block: 35 x 35 x 192 x = conv2d_bn(img_input, 32, 3, strides=2, padding='valid') x = conv2d_bn(x, 32, 3, padding='valid') x = conv2d_bn(x, 64, 3) x = layers.MaxPooling2D(3, strides=2)(x) x = conv2d_bn(x, 80, 1, padding='valid') x = conv2d_bn(x, 192, 3, padding='valid') x = layers.MaxPooling2D(3, strides=2)(x) # Mixed 5b (Inception-A block): 35 x 35 x 320 branch_0 = conv2d_bn(x, 96, 1) branch_1 = conv2d_bn(x, 48, 1) branch_1 = conv2d_bn(branch_1, 64, 5) branch_2 = conv2d_bn(x, 64, 1) branch_2 = conv2d_bn(branch_2, 96, 3) branch_2 = conv2d_bn(branch_2, 96, 3) branch_pool = layers.AveragePooling2D(3, strides=1, padding='same')(x) branch_pool = conv2d_bn(branch_pool, 64, 1) branches = [branch_0, branch_1, branch_2, branch_pool] channel_axis = 1 if backend.image_data_format() == 'channels_first' else 3 x = layers.Concatenate(axis=channel_axis, name='mixed_5b')(branches) # 10x block35 (Inception-ResNet-A block): 35 x 35 x 320 for block_idx in range(1, 11): x = inception_resnet_block( x, scale=0.17, block_type='block35', block_idx=block_idx) # Mixed 6a (Reduction-A block): 17 x 17 x 1088 branch_0 = conv2d_bn(x, 384, 3, strides=2, padding='valid') branch_1 = conv2d_bn(x, 256, 1) branch_1 = conv2d_bn(branch_1, 256, 3) branch_1 = conv2d_bn(branch_1, 384, 3, strides=2, padding='valid') branch_pool = layers.MaxPooling2D(3, strides=2, padding='valid')(x) branches = [branch_0, branch_1, branch_pool] x = layers.Concatenate(axis=channel_axis, name='mixed_6a')(branches) # 20x block17 (Inception-ResNet-B block): 17 x 17 x 1088 for block_idx in range(1, 21): x = inception_resnet_block( x, scale=0.1, block_type='block17', block_idx=block_idx) # Mixed 7a (Reduction-B block): 8 x 8 x 2080 branch_0 = conv2d_bn(x, 256, 1) branch_0 = conv2d_bn(branch_0, 384, 3, strides=2, padding='valid') branch_1 = conv2d_bn(x, 256, 1) branch_1 = conv2d_bn(branch_1, 288, 3, strides=2, padding='valid') branch_2 = conv2d_bn(x, 256, 1) branch_2 = conv2d_bn(branch_2, 288, 3) branch_2 = conv2d_bn(branch_2, 320, 3, strides=2, padding='valid') branch_pool = layers.MaxPooling2D(3, strides=2, padding='valid')(x) branches = [branch_0, branch_1, branch_2, branch_pool] x = layers.Concatenate(axis=channel_axis, name='mixed_7a')(branches) # 10x block8 (Inception-ResNet-C block): 8 x 8 x 2080 for block_idx in range(1, 10): x = inception_resnet_block( x, scale=0.2, block_type='block8', block_idx=block_idx) x = inception_resnet_block( x, scale=1., activation=None, block_type='block8', block_idx=10) # Final convolution block: 8 x 8 x 1536 x = conv2d_bn(x, 1536, 1, name='conv_7b') if include_top: # Classification block x = layers.GlobalAveragePooling2D(name='avg_pool')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Dense(classes, activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='inception_resnet_v2') # Load weights. if weights == 'imagenet': if include_top: fname = 'inception_resnet_v2_weights_tf_dim_ordering_tf_kernels.h5' weights_path = data_utils.get_file( fname, BASE_WEIGHT_URL + fname, cache_subdir='models', file_hash='e693bd0210a403b3192acc6073ad2e96') else: fname = ('inception_resnet_v2_weights_' 'tf_dim_ordering_tf_kernels_notop.h5') weights_path = data_utils.get_file( fname, BASE_WEIGHT_URL + fname, cache_subdir='models', file_hash='d19885ff4a710c122648d3b5c3b684e4') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model def conv2d_bn(x, filters, kernel_size, strides=1, padding='same', activation='relu', use_bias=False, name=None): """Utility function to apply conv + BN. Args: x: input tensor. filters: filters in `Conv2D`. kernel_size: kernel size as in `Conv2D`. strides: strides in `Conv2D`. padding: padding mode in `Conv2D`. activation: activation in `Conv2D`. use_bias: whether to use a bias in `Conv2D`. name: name of the ops; will become `name + '_ac'` for the activation and `name + '_bn'` for the batch norm layer. Returns: Output tensor after applying `Conv2D` and `BatchNormalization`. """ x = layers.Conv2D( filters, kernel_size, strides=strides, padding=padding, use_bias=use_bias, name=name)( x) if not use_bias: bn_axis = 1 if backend.image_data_format() == 'channels_first' else 3 bn_name = None if name is None else name + '_bn' x = layers.BatchNormalization(axis=bn_axis, scale=False, name=bn_name)(x) if activation is not None: ac_name = None if name is None else name + '_ac' x = layers.Activation(activation, name=ac_name)(x) return x def inception_resnet_block(x, scale, block_type, block_idx, activation='relu'): """Adds an Inception-ResNet block. This function builds 3 types of Inception-ResNet blocks mentioned in the paper, controlled by the `block_type` argument (which is the block name used in the official TF-slim implementation): - Inception-ResNet-A: `block_type='block35'` - Inception-ResNet-B: `block_type='block17'` - Inception-ResNet-C: `block_type='block8'` Args: x: input tensor. scale: scaling factor to scale the residuals (i.e., the output of passing `x` through an inception module) before adding them to the shortcut branch. Let `r` be the output from the residual branch, the output of this block will be `x + scale * r`. block_type: `'block35'`, `'block17'` or `'block8'`, determines the network structure in the residual branch. block_idx: an `int` used for generating layer names. The Inception-ResNet blocks are repeated many times in this network. We use `block_idx` to identify each of the repetitions. For example, the first Inception-ResNet-A block will have `block_type='block35', block_idx=0`, and the layer names will have a common prefix `'block35_0'`. activation: activation function to use at the end of the block (see [activations](../activations.md)). When `activation=None`, no activation is applied (i.e., "linear" activation: `a(x) = x`). Returns: Output tensor for the block. Raises: ValueError: if `block_type` is not one of `'block35'`, `'block17'` or `'block8'`. """ if block_type == 'block35': branch_0 = conv2d_bn(x, 32, 1) branch_1 = conv2d_bn(x, 32, 1) branch_1 = conv2d_bn(branch_1, 32, 3) branch_2 = conv2d_bn(x, 32, 1) branch_2 = conv2d_bn(branch_2, 48, 3) branch_2 = conv2d_bn(branch_2, 64, 3) branches = [branch_0, branch_1, branch_2] elif block_type == 'block17': branch_0 = conv2d_bn(x, 192, 1) branch_1 = conv2d_bn(x, 128, 1) branch_1 = conv2d_bn(branch_1, 160, [1, 7]) branch_1 = conv2d_bn(branch_1, 192, [7, 1]) branches = [branch_0, branch_1] elif block_type == 'block8': branch_0 = conv2d_bn(x, 192, 1) branch_1 = conv2d_bn(x, 192, 1) branch_1 = conv2d_bn(branch_1, 224, [1, 3]) branch_1 = conv2d_bn(branch_1, 256, [3, 1]) branches = [branch_0, branch_1] else: raise ValueError('Unknown Inception-ResNet block type. ' 'Expects "block35", "block17" or "block8", ' 'but got: ' + str(block_type)) block_name = block_type + '_' + str(block_idx) channel_axis = 1 if backend.image_data_format() == 'channels_first' else 3 mixed = layers.Concatenate( axis=channel_axis, name=block_name + '_mixed')( branches) up = conv2d_bn( mixed, backend.int_shape(x)[channel_axis], 1, activation=None, use_bias=True, name=block_name + '_conv') x = layers.Lambda( lambda inputs, scale: inputs[0] + inputs[1] * scale, output_shape=backend.int_shape(x)[1:], arguments={'scale': scale}, name=block_name)([x, up]) if activation is not None: x = layers.Activation(activation, name=block_name + '_ac')(x) return x @keras_export('keras.applications.inception_resnet_v2.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input(x, data_format=data_format, mode='tf') @keras_export('keras.applications.inception_resnet_v2.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
15,195
37.470886
87
py
keras
keras-master/keras/applications/mobilenet.py
# Copyright 2015 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== # pylint: disable=invalid-name """MobileNet v1 models for Keras. MobileNet is a general architecture and can be used for multiple use cases. Depending on the use case, it can use different input layer size and different width factors. This allows different width models to reduce the number of multiply-adds and thereby reduce inference cost on mobile devices. MobileNets support any input size greater than 32 x 32, with larger image sizes offering better performance. The number of parameters and number of multiply-adds can be modified by using the `alpha` parameter, which increases/decreases the number of filters in each layer. By altering the image size and `alpha` parameter, all 16 models from the paper can be built, with ImageNet weights provided. The paper demonstrates the performance of MobileNets using `alpha` values of 1.0 (also called 100 % MobileNet), 0.75, 0.5 and 0.25. For each of these `alpha` values, weights for 4 different input image sizes are provided (224, 192, 160, 128). The following table describes the size and accuracy of the 100% MobileNet on size 224 x 224: ---------------------------------------------------------------------------- Width Multiplier (alpha) | ImageNet Acc | Multiply-Adds (M) | Params (M) ---------------------------------------------------------------------------- | 1.0 MobileNet-224 | 70.6 % | 529 | 4.2 | | 0.75 MobileNet-224 | 68.4 % | 325 | 2.6 | | 0.50 MobileNet-224 | 63.7 % | 149 | 1.3 | | 0.25 MobileNet-224 | 50.6 % | 41 | 0.5 | ---------------------------------------------------------------------------- The following table describes the performance of the 100 % MobileNet on various input sizes: ------------------------------------------------------------------------ Resolution | ImageNet Acc | Multiply-Adds (M) | Params (M) ------------------------------------------------------------------------ | 1.0 MobileNet-224 | 70.6 % | 569 | 4.2 | | 1.0 MobileNet-192 | 69.1 % | 418 | 4.2 | | 1.0 MobileNet-160 | 67.2 % | 290 | 4.2 | | 1.0 MobileNet-128 | 64.4 % | 186 | 4.2 | ------------------------------------------------------------------------ Reference: - [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications]( https://arxiv.org/abs/1704.04861) """ import tensorflow.compat.v2 as tf from keras import backend from keras.applications import imagenet_utils from keras.engine import training from keras.layers import VersionAwareLayers from keras.utils import data_utils from keras.utils import layer_utils from tensorflow.python.platform import tf_logging as logging from tensorflow.python.util.tf_export import keras_export BASE_WEIGHT_PATH = ('https://storage.googleapis.com/tensorflow/' 'keras-applications/mobilenet/') layers = None @keras_export('keras.applications.mobilenet.MobileNet', 'keras.applications.MobileNet') def MobileNet(input_shape=None, alpha=1.0, depth_multiplier=1, dropout=1e-3, include_top=True, weights='imagenet', input_tensor=None, pooling=None, classes=1000, classifier_activation='softmax', **kwargs): """Instantiates the MobileNet architecture. Reference: - [MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications]( https://arxiv.org/abs/1704.04861) This function returns a Keras image classification model, optionally loaded with weights pre-trained on ImageNet. For image classification use cases, see [this page for detailed examples]( https://keras.io/api/applications/#usage-examples-for-image-classification-models). For transfer learning use cases, make sure to read the [guide to transfer learning & fine-tuning]( https://keras.io/guides/transfer_learning/). Note: each Keras Application expects a specific kind of input preprocessing. For MobileNet, call `tf.keras.applications.mobilenet.preprocess_input` on your inputs before passing them to the model. `mobilenet.preprocess_input` will scale input pixels between -1 and 1. Args: input_shape: Optional shape tuple, only to be specified if `include_top` is False (otherwise the input shape has to be `(224, 224, 3)` (with `channels_last` data format) or (3, 224, 224) (with `channels_first` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(200, 200, 3)` would be one valid value. Default to `None`. `input_shape` will be ignored if the `input_tensor` is provided. alpha: Controls the width of the network. This is known as the width multiplier in the MobileNet paper. - If `alpha` < 1.0, proportionally decreases the number of filters in each layer. - If `alpha` > 1.0, proportionally increases the number of filters in each layer. - If `alpha` = 1, default number of filters from the paper are used at each layer. Default to 1.0. depth_multiplier: Depth multiplier for depthwise convolution. This is called the resolution multiplier in the MobileNet paper. Default to 1.0. dropout: Dropout rate. Default to 0.001. include_top: Boolean, whether to include the fully-connected layer at the top of the network. Default to `True`. weights: One of `None` (random initialization), 'imagenet' (pre-training on ImageNet), or the path to the weights file to be loaded. Default to `imagenet`. input_tensor: Optional Keras tensor (i.e. output of `layers.Input()`) to use as image input for the model. `input_tensor` is useful for sharing inputs between multiple different networks. Default to None. pooling: Optional pooling mode for feature extraction when `include_top` is `False`. - `None` (default) means that the output of the model will be the 4D tensor output of the last convolutional block. - `avg` means that global average pooling will be applied to the output of the last convolutional block, and thus the output of the model will be a 2D tensor. - `max` means that global max pooling will be applied. classes: Optional number of classes to classify images into, only to be specified if `include_top` is True, and if no `weights` argument is specified. Defaults to 1000. classifier_activation: A `str` or callable. The activation function to use on the "top" layer. Ignored unless `include_top=True`. Set `classifier_activation=None` to return the logits of the "top" layer. When loading pretrained weights, `classifier_activation` can only be `None` or `"softmax"`. **kwargs: For backwards compatibility only. Returns: A `keras.Model` instance. """ global layers if 'layers' in kwargs: layers = kwargs.pop('layers') else: layers = VersionAwareLayers() if kwargs: raise ValueError(f'Unknown argument(s): {(kwargs,)}') if not (weights in {'imagenet', None} or tf.io.gfile.exists(weights)): raise ValueError('The `weights` argument should be either ' '`None` (random initialization), `imagenet` ' '(pre-training on ImageNet), ' 'or the path to the weights file to be loaded. ' f'Received weights={weights}') if weights == 'imagenet' and include_top and classes != 1000: raise ValueError('If using `weights` as `"imagenet"` with `include_top` ' 'as true, `classes` should be 1000. ' f'Received classes={classes}') # Determine proper input shape and default size. if input_shape is None: default_size = 224 else: if backend.image_data_format() == 'channels_first': rows = input_shape[1] cols = input_shape[2] else: rows = input_shape[0] cols = input_shape[1] if rows == cols and rows in [128, 160, 192, 224]: default_size = rows else: default_size = 224 input_shape = imagenet_utils.obtain_input_shape( input_shape, default_size=default_size, min_size=32, data_format=backend.image_data_format(), require_flatten=include_top, weights=weights) if backend.image_data_format() == 'channels_last': row_axis, col_axis = (0, 1) else: row_axis, col_axis = (1, 2) rows = input_shape[row_axis] cols = input_shape[col_axis] if weights == 'imagenet': if depth_multiplier != 1: raise ValueError('If imagenet weights are being loaded, ' 'depth multiplier must be 1. ' 'Received depth_multiplier={depth_multiplier}') if alpha not in [0.25, 0.50, 0.75, 1.0]: raise ValueError('If imagenet weights are being loaded, ' 'alpha can be one of' '`0.25`, `0.50`, `0.75` or `1.0` only. ' f'Received alpha={alpha}') if rows != cols or rows not in [128, 160, 192, 224]: rows = 224 logging.warning('`input_shape` is undefined or non-square, ' 'or `rows` is not in [128, 160, 192, 224]. ' 'Weights for input shape (224, 224) will be ' 'loaded as the default.') if input_tensor is None: img_input = layers.Input(shape=input_shape) else: if not backend.is_keras_tensor(input_tensor): img_input = layers.Input(tensor=input_tensor, shape=input_shape) else: img_input = input_tensor x = _conv_block(img_input, 32, alpha, strides=(2, 2)) x = _depthwise_conv_block(x, 64, alpha, depth_multiplier, block_id=1) x = _depthwise_conv_block( x, 128, alpha, depth_multiplier, strides=(2, 2), block_id=2) x = _depthwise_conv_block(x, 128, alpha, depth_multiplier, block_id=3) x = _depthwise_conv_block( x, 256, alpha, depth_multiplier, strides=(2, 2), block_id=4) x = _depthwise_conv_block(x, 256, alpha, depth_multiplier, block_id=5) x = _depthwise_conv_block( x, 512, alpha, depth_multiplier, strides=(2, 2), block_id=6) x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=7) x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=8) x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=9) x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=10) x = _depthwise_conv_block(x, 512, alpha, depth_multiplier, block_id=11) x = _depthwise_conv_block( x, 1024, alpha, depth_multiplier, strides=(2, 2), block_id=12) x = _depthwise_conv_block(x, 1024, alpha, depth_multiplier, block_id=13) if include_top: x = layers.GlobalAveragePooling2D(keepdims=True)(x) x = layers.Dropout(dropout, name='dropout')(x) x = layers.Conv2D(classes, (1, 1), padding='same', name='conv_preds')(x) x = layers.Reshape((classes,), name='reshape_2')(x) imagenet_utils.validate_activation(classifier_activation, weights) x = layers.Activation(activation=classifier_activation, name='predictions')(x) else: if pooling == 'avg': x = layers.GlobalAveragePooling2D()(x) elif pooling == 'max': x = layers.GlobalMaxPooling2D()(x) # Ensure that the model takes into account # any potential predecessors of `input_tensor`. if input_tensor is not None: inputs = layer_utils.get_source_inputs(input_tensor) else: inputs = img_input # Create model. model = training.Model(inputs, x, name='mobilenet_%0.2f_%s' % (alpha, rows)) # Load weights. if weights == 'imagenet': if alpha == 1.0: alpha_text = '1_0' elif alpha == 0.75: alpha_text = '7_5' elif alpha == 0.50: alpha_text = '5_0' else: alpha_text = '2_5' if include_top: model_name = 'mobilenet_%s_%d_tf.h5' % (alpha_text, rows) weight_path = BASE_WEIGHT_PATH + model_name weights_path = data_utils.get_file( model_name, weight_path, cache_subdir='models') else: model_name = 'mobilenet_%s_%d_tf_no_top.h5' % (alpha_text, rows) weight_path = BASE_WEIGHT_PATH + model_name weights_path = data_utils.get_file( model_name, weight_path, cache_subdir='models') model.load_weights(weights_path) elif weights is not None: model.load_weights(weights) return model def _conv_block(inputs, filters, alpha, kernel=(3, 3), strides=(1, 1)): """Adds an initial convolution layer (with batch normalization and relu6). Args: inputs: Input tensor of shape `(rows, cols, 3)` (with `channels_last` data format) or (3, rows, cols) (with `channels_first` data format). It should have exactly 3 inputs channels, and width and height should be no smaller than 32. E.g. `(224, 224, 3)` would be one valid value. filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the convolution). alpha: controls the width of the network. - If `alpha` < 1.0, proportionally decreases the number of filters in each layer. - If `alpha` > 1.0, proportionally increases the number of filters in each layer. - If `alpha` = 1, default number of filters from the paper are used at each layer. kernel: An integer or tuple/list of 2 integers, specifying the width and height of the 2D convolution window. Can be a single integer to specify the same value for all spatial dimensions. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the width and height. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. # Input shape 4D tensor with shape: `(samples, channels, rows, cols)` if data_format='channels_first' or 4D tensor with shape: `(samples, rows, cols, channels)` if data_format='channels_last'. # Output shape 4D tensor with shape: `(samples, filters, new_rows, new_cols)` if data_format='channels_first' or 4D tensor with shape: `(samples, new_rows, new_cols, filters)` if data_format='channels_last'. `rows` and `cols` values might have changed due to stride. Returns: Output tensor of block. """ channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 filters = int(filters * alpha) x = layers.Conv2D( filters, kernel, padding='same', use_bias=False, strides=strides, name='conv1')(inputs) x = layers.BatchNormalization(axis=channel_axis, name='conv1_bn')(x) return layers.ReLU(6., name='conv1_relu')(x) def _depthwise_conv_block(inputs, pointwise_conv_filters, alpha, depth_multiplier=1, strides=(1, 1), block_id=1): """Adds a depthwise convolution block. A depthwise convolution block consists of a depthwise conv, batch normalization, relu6, pointwise convolution, batch normalization and relu6 activation. Args: inputs: Input tensor of shape `(rows, cols, channels)` (with `channels_last` data format) or (channels, rows, cols) (with `channels_first` data format). pointwise_conv_filters: Integer, the dimensionality of the output space (i.e. the number of output filters in the pointwise convolution). alpha: controls the width of the network. - If `alpha` < 1.0, proportionally decreases the number of filters in each layer. - If `alpha` > 1.0, proportionally increases the number of filters in each layer. - If `alpha` = 1, default number of filters from the paper are used at each layer. depth_multiplier: The number of depthwise convolution output channels for each input channel. The total number of depthwise convolution output channels will be equal to `filters_in * depth_multiplier`. strides: An integer or tuple/list of 2 integers, specifying the strides of the convolution along the width and height. Can be a single integer to specify the same value for all spatial dimensions. Specifying any stride value != 1 is incompatible with specifying any `dilation_rate` value != 1. block_id: Integer, a unique identification designating the block number. # Input shape 4D tensor with shape: `(batch, channels, rows, cols)` if data_format='channels_first' or 4D tensor with shape: `(batch, rows, cols, channels)` if data_format='channels_last'. # Output shape 4D tensor with shape: `(batch, filters, new_rows, new_cols)` if data_format='channels_first' or 4D tensor with shape: `(batch, new_rows, new_cols, filters)` if data_format='channels_last'. `rows` and `cols` values might have changed due to stride. Returns: Output tensor of block. """ channel_axis = 1 if backend.image_data_format() == 'channels_first' else -1 pointwise_conv_filters = int(pointwise_conv_filters * alpha) if strides == (1, 1): x = inputs else: x = layers.ZeroPadding2D(((0, 1), (0, 1)), name='conv_pad_%d' % block_id)( inputs) x = layers.DepthwiseConv2D((3, 3), padding='same' if strides == (1, 1) else 'valid', depth_multiplier=depth_multiplier, strides=strides, use_bias=False, name='conv_dw_%d' % block_id)( x) x = layers.BatchNormalization( axis=channel_axis, name='conv_dw_%d_bn' % block_id)( x) x = layers.ReLU(6., name='conv_dw_%d_relu' % block_id)(x) x = layers.Conv2D( pointwise_conv_filters, (1, 1), padding='same', use_bias=False, strides=(1, 1), name='conv_pw_%d' % block_id)( x) x = layers.BatchNormalization( axis=channel_axis, name='conv_pw_%d_bn' % block_id)( x) return layers.ReLU(6., name='conv_pw_%d_relu' % block_id)(x) @keras_export('keras.applications.mobilenet.preprocess_input') def preprocess_input(x, data_format=None): return imagenet_utils.preprocess_input(x, data_format=data_format, mode='tf') @keras_export('keras.applications.mobilenet.decode_predictions') def decode_predictions(preds, top=5): return imagenet_utils.decode_predictions(preds, top=top) preprocess_input.__doc__ = imagenet_utils.PREPROCESS_INPUT_DOC.format( mode='', ret=imagenet_utils.PREPROCESS_INPUT_RET_DOC_TF, error=imagenet_utils.PREPROCESS_INPUT_ERROR_DOC) decode_predictions.__doc__ = imagenet_utils.decode_predictions.__doc__
19,722
42.157549
87
py
keras
keras-master/keras/applications/applications_test.py
# Copyright 2018 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== """Integration tests for Keras applications.""" import tensorflow.compat.v2 as tf from absl.testing import parameterized from keras import backend from keras.applications import densenet from keras.applications import efficientnet from keras.applications import inception_resnet_v2 from keras.applications import inception_v3 from keras.applications import mobilenet from keras.applications import mobilenet_v2 from keras.applications import mobilenet_v3 from keras.applications import nasnet from keras.applications import resnet from keras.applications import resnet_v2 from keras.applications import vgg16 from keras.applications import vgg19 from keras.applications import xception MODEL_LIST_NO_NASNET = [ (resnet.ResNet50, 2048), (resnet.ResNet101, 2048), (resnet.ResNet152, 2048), (resnet_v2.ResNet50V2, 2048), (resnet_v2.ResNet101V2, 2048), (resnet_v2.ResNet152V2, 2048), (vgg16.VGG16, 512), (vgg19.VGG19, 512), (xception.Xception, 2048), (inception_v3.InceptionV3, 2048), (inception_resnet_v2.InceptionResNetV2, 1536), (mobilenet.MobileNet, 1024), (mobilenet_v2.MobileNetV2, 1280), (mobilenet_v3.MobileNetV3Small, 1024), (mobilenet_v3.MobileNetV3Large, 1280), (densenet.DenseNet121, 1024), (densenet.DenseNet169, 1664), (densenet.DenseNet201, 1920), (efficientnet.EfficientNetB0, 1280), (efficientnet.EfficientNetB1, 1280), (efficientnet.EfficientNetB2, 1408), (efficientnet.EfficientNetB3, 1536), (efficientnet.EfficientNetB4, 1792), (efficientnet.EfficientNetB5, 2048), (efficientnet.EfficientNetB6, 2304), (efficientnet.EfficientNetB7, 2560), ] NASNET_LIST = [ (nasnet.NASNetMobile, 1056), (nasnet.NASNetLarge, 4032), ] MODEL_LIST = MODEL_LIST_NO_NASNET + NASNET_LIST class ApplicationsTest(tf.test.TestCase, parameterized.TestCase): def assertShapeEqual(self, shape1, shape2): if len(shape1) != len(shape2): raise AssertionError( 'Shapes are different rank: %s vs %s' % (shape1, shape2)) for v1, v2 in zip(shape1, shape2): if v1 != v2: raise AssertionError('Shapes differ: %s vs %s' % (shape1, shape2)) @parameterized.parameters(*MODEL_LIST) def test_application_base(self, app, _): # Can be instantiated with default arguments model = app(weights=None) # Can be serialized and deserialized config = model.get_config() reconstructed_model = model.__class__.from_config(config) self.assertEqual(len(model.weights), len(reconstructed_model.weights)) backend.clear_session() @parameterized.parameters(*MODEL_LIST) def test_application_notop(self, app, last_dim): if 'NASNet' or 'MobileNetV3' in app.__name__: only_check_last_dim = True else: only_check_last_dim = False output_shape = _get_output_shape( lambda: app(weights=None, include_top=False)) if only_check_last_dim: self.assertEqual(output_shape[-1], last_dim) else: self.assertShapeEqual(output_shape, (None, None, None, last_dim)) backend.clear_session() @parameterized.parameters(MODEL_LIST) def test_application_pooling(self, app, last_dim): output_shape = _get_output_shape( lambda: app(weights=None, include_top=False, pooling='avg')) self.assertShapeEqual(output_shape, (None, last_dim)) @parameterized.parameters(*MODEL_LIST_NO_NASNET) def test_application_variable_input_channels(self, app, last_dim): if backend.image_data_format() == 'channels_first': input_shape = (1, None, None) else: input_shape = (None, None, 1) output_shape = _get_output_shape( lambda: app(weights=None, include_top=False, input_shape=input_shape)) if 'MobileNetV3' in app.__name__: self.assertShapeEqual(output_shape, (None, 1, 1, last_dim)) else: self.assertShapeEqual(output_shape, (None, None, None, last_dim)) backend.clear_session() if backend.image_data_format() == 'channels_first': input_shape = (4, None, None) else: input_shape = (None, None, 4) output_shape = _get_output_shape( lambda: app(weights=None, include_top=False, input_shape=input_shape)) if 'MobileNetV3' in app.__name__: self.assertShapeEqual(output_shape, (None, 1, 1, last_dim)) else: self.assertShapeEqual(output_shape, (None, None, None, last_dim)) backend.clear_session() def _get_output_shape(model_fn): model = model_fn() return model.output_shape if __name__ == '__main__': tf.test.main()
5,219
34.27027
80
py
keras
keras-master/keras/applications/efficientnet_weight_update_util.py
# Copyright 2020 The TensorFlow Authors. All Rights Reserved. # # Licensed under the Apache License, Version 2.0 (the "License"); # you may not use this file except in compliance with the License. # You may obtain a copy of the License at # # http://www.apache.org/licenses/LICENSE-2.0 # # Unless required by applicable law or agreed to in writing, software # distributed under the License is distributed on an "AS IS" BASIS, # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. # See the License for the specific language governing permissions and # limitations under the License. # ============================================================================== r"""Utils for EfficientNet models for Keras. Write weights from ckpt file as in original repo (https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet) to h5 file for keras implementation of the models. Usage: # use checkpoint efficientnet-b0/model.ckpt (can be downloaded from # https://storage.googleapis.com/cloud-tpu-checkpoints/ # efficientnet/ckptsaug/efficientnet-b0.tar.gz) # to update weight without top layers, saving to efficientnetb0_notop.h5 python efficientnet_weight_update_util.py --model b0 --notop \ --ckpt efficientnet-b0/model.ckpt --o efficientnetb0_notop.h5 # use checkpoint noisy_student_efficientnet-b3/model.ckpt (providing # improved result for b3, can be downloaded from # https://storage.googleapis.com/cloud-tpu-checkpoints/ # efficientnet/noisystudent/noisy_student_efficientnet-b3.tar.gz) # to update weight with top layers, saving to efficientnetb3_new.h5 python efficientnet_weight_update_util.py --model b3 --notop \ --ckpt noisy_student_efficientnet-b3/model.ckpt --o efficientnetb3_new.h5 """ import tensorflow.compat.v2 as tf import argparse import warnings from tensorflow.keras.applications import efficientnet def write_ckpt_to_h5(path_h5, path_ckpt, keras_model, use_ema=True): """Map the weights in checkpoint file (tf) to h5 file (keras). Args: path_h5: str, path to output hdf5 file to write weights loaded from ckpt files. path_ckpt: str, path to the ckpt files (e.g. 'efficientnet-b0/model.ckpt') that records efficientnet weights from original repo https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet keras_model: keras model, built from keras.applications efficientnet functions (e.g. EfficientNetB0) use_ema: Bool, whether to use ExponentialMovingAverage result or not """ model_name_keras = keras_model.name model_name_tf = model_name_keras.replace('efficientnet', 'efficientnet-') keras_weight_names = [w.name for w in keras_model.weights] tf_weight_names = get_variable_names_from_ckpt(path_ckpt) keras_blocks = get_keras_blocks(keras_weight_names) tf_blocks = get_tf_blocks(tf_weight_names) print('check variables match in each block') for keras_block, tf_block in zip(keras_blocks, tf_blocks): check_match(keras_block, tf_block, keras_weight_names, tf_weight_names, model_name_tf) print('{} and {} match.'.format(tf_block, keras_block)) block_mapping = {x[0]: x[1] for x in zip(keras_blocks, tf_blocks)} changed_weights = 0 for w in keras_model.weights: if 'block' in w.name: # example: 'block1a_dwconv/depthwise_kernel:0' -> 'block1a' keras_block = w.name.split('/')[0].split('_')[0] tf_block = block_mapping[keras_block] tf_name = keras_name_to_tf_name_block( w.name, keras_block=keras_block, tf_block=tf_block, use_ema=use_ema, model_name_tf=model_name_tf) elif any([x in w.name for x in ['stem', 'top', 'predictions', 'probs']]): tf_name = keras_name_to_tf_name_stem_top( w.name, use_ema=use_ema, model_name_tf=model_name_tf) elif 'normalization' in w.name: print('skipping variable {}: normalization is a layer' 'in keras implementation, but preprocessing in ' 'TF implementation.'.format(w.name)) continue else: raise ValueError('{} failed to parse.'.format(w.name)) try: w_tf = tf.train.load_variable(path_ckpt, tf_name) if (w.value().numpy() != w_tf).any(): w.assign(w_tf) changed_weights += 1 except ValueError as e: if any([x in w.name for x in ['top', 'predictions', 'probs']]): warnings.warn('Fail to load top layer variable {}' 'from {} because of {}.'.format(w.name, tf_name, e)) else: raise ValueError('Fail to load {} from {}'.format(w.name, tf_name)) total_weights = len(keras_model.weights) print('{}/{} weights updated'.format(changed_weights, total_weights)) keras_model.save_weights(path_h5) def get_variable_names_from_ckpt(path_ckpt, use_ema=True): """Get list of tensor names from checkpoint. Args: path_ckpt: str, path to the ckpt files use_ema: Bool, whether to use ExponentialMovingAverage result or not. Returns: List of variable names from checkpoint. """ v_all = tf.train.list_variables(path_ckpt) # keep name only v_name_all = [x[0] for x in v_all] if use_ema: v_name_all = [x for x in v_name_all if 'ExponentialMovingAverage' in x] else: v_name_all = [x for x in v_name_all if 'ExponentialMovingAverage' not in x] # remove util variables used for RMSprop v_name_all = [x for x in v_name_all if 'RMS' not in x] return v_name_all def get_tf_blocks(tf_weight_names): """Extract the block names from list of full weight names.""" # Example: 'efficientnet-b0/blocks_0/conv2d/kernel' -> 'blocks_0' tf_blocks = {x.split('/')[1] for x in tf_weight_names if 'block' in x} # sort by number tf_blocks = sorted(tf_blocks, key=lambda x: int(x.split('_')[1])) return tf_blocks def get_keras_blocks(keras_weight_names): """Extract the block names from list of full weight names.""" # example: 'block1a_dwconv/depthwise_kernel:0' -> 'block1a' keras_blocks = {x.split('_')[0] for x in keras_weight_names if 'block' in x} return sorted(keras_blocks) def keras_name_to_tf_name_stem_top(keras_name, use_ema=True, model_name_tf='efficientnet-b0'): """Mapping name in h5 to ckpt that is in stem or top (head). we map name keras_name that points to a weight in h5 file to a name of weight in ckpt file. Args: keras_name: str, the name of weight in the h5 file of keras implementation use_ema: Bool, use the ExponentialMovingAverage resuolt in ckpt or not model_name_tf: str, the name of model in ckpt. Returns: String for the name of weight as in ckpt file. Raises: KeyError: if we cannot parse the keras_name. """ if use_ema: ema = '/ExponentialMovingAverage' else: ema = '' stem_top_dict = { 'probs/bias:0': '{}/head/dense/bias{}', 'probs/kernel:0': '{}/head/dense/kernel{}', 'predictions/bias:0': '{}/head/dense/bias{}', 'predictions/kernel:0': '{}/head/dense/kernel{}', 'stem_conv/kernel:0': '{}/stem/conv2d/kernel{}', 'top_conv/kernel:0': '{}/head/conv2d/kernel{}', } for x in stem_top_dict: stem_top_dict[x] = stem_top_dict[x].format(model_name_tf, ema) # stem batch normalization for bn_weights in ['beta', 'gamma', 'moving_mean', 'moving_variance']: tf_name = '{}/stem/tpu_batch_normalization/{}{}'.format( model_name_tf, bn_weights, ema) stem_top_dict['stem_bn/{}:0'.format(bn_weights)] = tf_name # top / head batch normalization for bn_weights in ['beta', 'gamma', 'moving_mean', 'moving_variance']: tf_name = '{}/head/tpu_batch_normalization/{}{}'.format( model_name_tf, bn_weights, ema) stem_top_dict['top_bn/{}:0'.format(bn_weights)] = tf_name if keras_name in stem_top_dict: return stem_top_dict[keras_name] raise KeyError('{} from h5 file cannot be parsed'.format(keras_name)) def keras_name_to_tf_name_block(keras_name, keras_block='block1a', tf_block='blocks_0', use_ema=True, model_name_tf='efficientnet-b0'): """Mapping name in h5 to ckpt that belongs to a block. we map name keras_name that points to a weight in h5 file to a name of weight in ckpt file. Args: keras_name: str, the name of weight in the h5 file of keras implementation keras_block: str, the block name for keras implementation (e.g. 'block1a') tf_block: str, the block name for tf implementation (e.g. 'blocks_0') use_ema: Bool, use the ExponentialMovingAverage resuolt in ckpt or not model_name_tf: str, the name of model in ckpt. Returns: String for the name of weight as in ckpt file. Raises: ValueError if keras_block does not show up in keras_name """ if keras_block not in keras_name: raise ValueError('block name {} not found in {}'.format( keras_block, keras_name)) # all blocks in the first group will not have expand conv and bn is_first_blocks = (keras_block[5] == '1') tf_name = [model_name_tf, tf_block] # depthwide conv if 'dwconv' in keras_name: tf_name.append('depthwise_conv2d') tf_name.append('depthwise_kernel') # conv layers if is_first_blocks: # first blocks only have one conv2d if 'project_conv' in keras_name: tf_name.append('conv2d') tf_name.append('kernel') else: if 'project_conv' in keras_name: tf_name.append('conv2d_1') tf_name.append('kernel') elif 'expand_conv' in keras_name: tf_name.append('conv2d') tf_name.append('kernel') # squeeze expansion layers if '_se_' in keras_name: if 'reduce' in keras_name: tf_name.append('se/conv2d') elif 'expand' in keras_name: tf_name.append('se/conv2d_1') if 'kernel' in keras_name: tf_name.append('kernel') elif 'bias' in keras_name: tf_name.append('bias') # batch normalization layers if 'bn' in keras_name: if is_first_blocks: if 'project' in keras_name: tf_name.append('tpu_batch_normalization_1') else: tf_name.append('tpu_batch_normalization') else: if 'project' in keras_name: tf_name.append('tpu_batch_normalization_2') elif 'expand' in keras_name: tf_name.append('tpu_batch_normalization') else: tf_name.append('tpu_batch_normalization_1') for x in ['moving_mean', 'moving_variance', 'beta', 'gamma']: if x in keras_name: tf_name.append(x) if use_ema: tf_name.append('ExponentialMovingAverage') return '/'.join(tf_name) def check_match(keras_block, tf_block, keras_weight_names, tf_weight_names, model_name_tf): """Check if the weights in h5 and ckpt match. we match each name from keras_weight_names that is in keras_block and check if there is 1-1 correspondence to names from tf_weight_names that is in tf_block Args: keras_block: str, the block name for keras implementation (e.g. 'block1a') tf_block: str, the block name for tf implementation (e.g. 'blocks_0') keras_weight_names: list of str, weight names in keras implementation tf_weight_names: list of str, weight names in tf implementation model_name_tf: str, the name of model in ckpt. """ names_from_keras = set() for x in keras_weight_names: if keras_block in x: y = keras_name_to_tf_name_block( x, keras_block=keras_block, tf_block=tf_block, model_name_tf=model_name_tf) names_from_keras.add(y) names_from_tf = set() for x in tf_weight_names: if tf_block in x and x.split('/')[1].endswith(tf_block): names_from_tf.add(x) names_missing = names_from_keras - names_from_tf if names_missing: raise ValueError('{} variables not found in checkpoint file: {}'.format( len(names_missing), names_missing)) names_unused = names_from_tf - names_from_keras if names_unused: warnings.warn('{} variables from checkpoint file are not used: {}'.format( len(names_unused), names_unused)) if __name__ == '__main__': arg_to_model = { 'b0': efficientnet.EfficientNetB0, 'b1': efficientnet.EfficientNetB1, 'b2': efficientnet.EfficientNetB2, 'b3': efficientnet.EfficientNetB3, 'b4': efficientnet.EfficientNetB4, 'b5': efficientnet.EfficientNetB5, 'b6': efficientnet.EfficientNetB6, 'b7': efficientnet.EfficientNetB7 } p = argparse.ArgumentParser(description='write weights from checkpoint to h5') p.add_argument( '--model', required=True, type=str, help='name of efficient model', choices=arg_to_model.keys()) p.add_argument( '--notop', action='store_true', help='do not include top layers', default=False) p.add_argument('--ckpt', required=True, type=str, help='checkpoint path') p.add_argument( '--output', '-o', required=True, type=str, help='output (h5) file path') args = p.parse_args() include_top = not args.notop model = arg_to_model[args.model](include_top=include_top) write_ckpt_to_h5(args.output, args.ckpt, keras_model=model)
13,222
34.834688
80
py
ReCO
ReCO-master/test.py
# -*- coding: utf-8 -*- """ @Time : 2020/6/23 下午1:43 @FileName: test.py @author: 王炳宁 @contact: wangbingning@sogou-inc.com """ import argparse import torch from model import Bert4ReCO from utils import * parser = argparse.ArgumentParser() parser.add_argument("--model_type", type=str, default='bert-base-chinese') parser.add_argument("--batch_size", type=int, default=16) parser.add_argument( "--fp16", action="store_true", default=True, ) args = parser.parse_args() model_type = args.model_type batch_size = args.batch_size test_data = load_file('data/test.{}.obj'.format(model_type.replace('/', '.'))) test_data = sorted(test_data, key=lambda x: len(x[0])) model = Bert4ReCO(model_type) model.load_state_dict(torch.load('checkpoint.{}.th'.format(model_type.replace('/', '.')), map_location='cpu')) model.cuda() if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") [model] = amp.initialize([model], opt_level='O1', verbosity=0) model.eval() total = len(test_data) right = 0 with torch.no_grad(): for i in tqdm(range(0, total, batch_size)): seq = [x[0] for x in test_data[i:i + batch_size]] labels = [x[1] for x in test_data[i:i + batch_size]] seq = padding(seq, pads=0, max_len=512) seq = torch.LongTensor(seq).cuda() predictions = model([seq, None]) predictions = predictions.cpu() right += predictions.eq(torch.LongTensor(labels)).sum().item() acc = 100 * right / total print('test acc is {}'.format(acc))
1,629
30.346154
110
py
ReCO
ReCO-master/model.py
# -*- coding: utf-8 -*- """ @Time : 2020/6/23 上午10:13 @FileName: model.py @author: 王炳宁 @contact: wangbingning@sogou-inc.com """ import torch import torch.nn as nn import torch.nn.functional as F from transformers import AutoModel class Bert4ReCO(nn.Module): def __init__(self, model_type): super().__init__() self.encoder = AutoModel.from_pretrained(model_type) self.n_hidden = self.encoder.config.hidden_size self.prediction = nn.Linear(self.n_hidden, 1, bias=False) def forward(self, inputs): [seq, label] = inputs hidden = self.encoder(seq)[0] mask_idx = torch.eq(seq, 1) # 1 is the index in the seq we separate each candidates. hidden = hidden.masked_select(mask_idx.unsqueeze(2).expand_as(hidden)).view( -1, 3, self.n_hidden) hidden = self.prediction(hidden).squeeze(-1) if label is None: return hidden.argmax(1) return F.cross_entropy(hidden, label) if __name__ == '__main__': model = Bert4ReCO('voidful/albert_chinese_xxlarge')
1,073
28.833333
93
py
ReCO
ReCO-master/train.py
# -*- coding: utf-8 -*- """ @Time : 2019/11/21 下午7:14 @FileName: train.py @author: 王炳宁 @contact: wangbingning@sogou-inc.com """ import argparse import torch from model import Bert4ReCO from prepare_data import prepare_bert_data from utils import * import torch.distributed as dist torch.manual_seed(100) np.random.seed(100) parser = argparse.ArgumentParser() parser.add_argument("--batch_size", type=int, default=16) parser.add_argument("--epoch", type=int, default=10) parser.add_argument("--lr", type=float, default=2.0e-5) parser.add_argument("--max_grad_norm", type=float, default=0.2) parser.add_argument("--model_type", type=str, default="voidful/albert_chinese_base") parser.add_argument( "--fp16", action="store_true", default=True, ) parser.add_argument("--local_rank", type=int, default=-1) args = parser.parse_args() model_type = args.model_type local_rank = args.local_rank if local_rank >= 0: torch.distributed.init_process_group(backend='nccl', init_method='env://') torch.cuda.set_device(args.local_rank) if local_rank in [-1, 0]: prepare_bert_data(model_type) if local_rank >= 0: dist.barrier() # wait for the first gpu to load data data = load_file('data/train.{}.obj'.format(model_type.replace('/', '.'))) valid_data = load_file('data/valid.{}.obj'.format(model_type.replace('/', '.'))) valid_data = sorted(valid_data, key=lambda x: len(x[0])) batch_size = args.batch_size model = Bert4ReCO(model_type).cuda() optimizer = torch.optim.AdamW(model.parameters(), weight_decay=0.01, lr=args.lr) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level='O1', verbosity=0) if local_rank >= 0: try: import apex except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use parallel training.") model = apex.parallel.DistributedDataParallel(model) def get_shuffle_data(): pool = {} for one in data: length = len(one[0]) // 5 if length not in pool: pool[length] = [] pool[length].append(one) for one in pool: np.random.shuffle(pool[one]) length_lst = list(pool.keys()) np.random.shuffle(length_lst) whole_data = [x for y in length_lst for x in pool[y]] if local_rank >= 0: remove_data_size = len(whole_data) % dist.get_world_size() thread_data = [whole_data[x + args.local_rank] for x in range(0, len(whole_data) - remove_data_size, dist.get_world_size())] return thread_data return whole_data def iter_printer(total, epoch): if local_rank >= 0: if local_rank == 0: return tqdm(range(0, total, batch_size), desc='epoch {}'.format(epoch)) else: return range(0, total, batch_size) else: return tqdm(range(0, total, batch_size), desc='epoch {}'.format(epoch)) def train(epoch): model.train() train_data = get_shuffle_data() total = len(train_data) for i in iter_printer(total, epoch): seq = [x[0] for x in train_data[i:i + batch_size]] label = [x[1] for x in train_data[i:i + batch_size]] seq = padding(seq, pads=0, max_len=512) seq = torch.LongTensor(seq).cuda() label = torch.LongTensor(label).cuda() loss = model([seq, label]) if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() optimizer.zero_grad() def evaluation(epoch): model.eval() total = len(valid_data) right = 0 with torch.no_grad(): for i in iter_printer(total, epoch): seq = [x[0] for x in valid_data[i:i + batch_size]] labels = [x[1] for x in valid_data[i:i + batch_size]] seq = padding(seq, pads=0, max_len=512) seq = torch.LongTensor(seq).cuda() predictions = model([seq, None]) predictions = predictions.cpu() right += predictions.eq(torch.LongTensor(labels)).sum().item() acc = 100 * right / total print('epoch {} eval acc is {}'.format(epoch, acc)) return acc best_acc = 0.0 for epo in range(args.epoch): train(epo) if local_rank == -1 or local_rank == 0: accuracy = evaluation(epo) if accuracy > best_acc: best_acc = accuracy with open('checkpoint.{}.th'.format(model_type.replace('/', '.')), 'wb') as f: state_dict = model.module.state_dict() if args.fp16 else model.state_dict() torch.save(state_dict, f)
5,074
33.290541
114
py
ReCO
ReCO-master/BiDAF/inference.py
# -*- coding: utf-8 -*- import argparse import cPickle import codecs import torch from utils import * from preprocess import seg_data, transform_data_to_id parser = argparse.ArgumentParser(description='inference procedure, note you should train the data at first') parser.add_argument('--data', type=str, default='data/ai_challenger_oqmrc_testa_20180816/ai_challenger_oqmrc_testa.json', help='location of the test data') parser.add_argument('--word_path', type=str, default='data/word2id.obj', help='location of the word2id.obj') parser.add_argument('--output', type=str, default='data/prediction.a.txt', help='prediction path') parser.add_argument('--model', type=str, default='model.pt', help='model path') parser.add_argument('--batch_size', type=int, default=32, metavar='N', help='batch size') parser.add_argument('--cuda', action='store_true',default=True, help='use CUDA') args = parser.parse_args() with open(args.model, 'rb') as f: model = torch.load(f) if args.cuda: model.cuda() with open(args.word_path, 'rb') as f: word2id = cPickle.load(f) raw_data = seg_data(args.data) transformed_data = transform_data_to_id(raw_data, word2id) data = [x + [y[2]] for x, y in zip(transformed_data, raw_data)] data = sorted(data, key=lambda x: len(x[1])) print 'test data size {:d}'.format(len(data)) def inference(): model.eval() predictions = [] with torch.no_grad(): for i in range(0, len(data), args.batch_size): one = data[i:i + args.batch_size] query, _ = padding([x[0] for x in one], max_len=50) passage, _ = padding([x[1] for x in one], max_len=300) answer = pad_answer([x[2] for x in one]) str_words = [x[-1] for x in one] ids = [x[3] for x in one] query, passage, answer = torch.LongTensor(query), torch.LongTensor(passage), torch.LongTensor(answer) if args.cuda: query = query.cuda() passage = passage.cuda() answer = answer.cuda() output = model([query, passage, answer, False]) for q_id, prediction, candidates in zip(ids, output, str_words): prediction_answer = u''.join(candidates[prediction]) predictions.append(str(q_id) + '\t' + prediction_answer) outputs = u'\n'.join(predictions) with codecs.open(args.output, 'w',encoding='utf-8') as f: f.write(outputs) print 'done!' if __name__ == '__main__': inference()
2,636
34.635135
113
py
ReCO
ReCO-master/BiDAF/MwAN.py
# -*- coding: utf-8 -*- import torch from torch import nn from torch.nn import functional as F class MwAN(nn.Module): def __init__(self, vocab_size, embedding_size, encoder_size, drop_out=0.2): super(MwAN, self).__init__() self.drop_out=drop_out self.embedding = nn.Embedding(vocab_size + 1, embedding_dim=embedding_size) self.q_encoder = nn.GRU(input_size=embedding_size, hidden_size=encoder_size, batch_first=True, bidirectional=True) self.p_encoder = nn.GRU(input_size=embedding_size, hidden_size=encoder_size, batch_first=True, bidirectional=True) self.a_encoder = nn.GRU(input_size=embedding_size, hidden_size=embedding_size / 2, batch_first=True, bidirectional=True) self.a_attention = nn.Linear(embedding_size, 1, bias=False) # Concat Attention self.Wc1 = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.Wc2 = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vc = nn.Linear(encoder_size, 1, bias=False) # Bilinear Attention self.Wb = nn.Linear(2 * encoder_size, 2 * encoder_size, bias=False) # Dot Attention : self.Wd = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vd = nn.Linear(encoder_size, 1, bias=False) # Minus Attention : self.Wm = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vm = nn.Linear(encoder_size, 1, bias=False) self.Ws = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vs = nn.Linear(encoder_size, 1, bias=False) self.gru_agg = nn.GRU(12 * encoder_size, encoder_size, batch_first=True, bidirectional=True) """ prediction layer """ self.Wq = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vq = nn.Linear(encoder_size, 1, bias=False) self.Wp1 = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.Wp2 = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vp = nn.Linear(encoder_size, 1, bias=False) self.prediction = nn.Linear(2 * encoder_size, embedding_size, bias=False) self.initiation() def initiation(self): initrange = 0.1 nn.init.uniform_(self.embedding.weight, -initrange, initrange) for module in self.modules(): if isinstance(module, nn.Linear): nn.init.xavier_uniform_(module.weight, 0.1) def forward(self, inputs): [query, passage, answer, is_train] = inputs q_embedding = self.embedding(query) p_embedding = self.embedding(passage) a_embeddings = self.embedding(answer) a_embedding, _ = self.a_encoder(a_embeddings.view(-1, a_embeddings.size(2), a_embeddings.size(3))) a_score = F.softmax(self.a_attention(a_embedding), 1) a_output = a_score.transpose(2, 1).bmm(a_embedding).squeeze() a_embedding = a_output.view(a_embeddings.size(0), 3, -1) hq, _ = self.q_encoder(p_embedding) hq=F.dropout(hq,self.drop_out) hp, _ = self.p_encoder(q_embedding) hp=F.dropout(hp,self.drop_out) _s1 = self.Wc1(hq).unsqueeze(1) _s2 = self.Wc2(hp).unsqueeze(2) sjt = self.vc(torch.tanh(_s1 + _s2)).squeeze() ait = F.softmax(sjt, 2) qtc = ait.bmm(hq) _s1 = self.Wb(hq).transpose(2, 1) sjt = hp.bmm(_s1) ait = F.softmax(sjt, 2) qtb = ait.bmm(hq) _s1 = hq.unsqueeze(1) _s2 = hp.unsqueeze(2) sjt = self.vd(torch.tanh(self.Wd(_s1 * _s2))).squeeze() ait = F.softmax(sjt, 2) qtd = ait.bmm(hq) sjt = self.vm(torch.tanh(self.Wm(_s1 - _s2))).squeeze() ait = F.softmax(sjt, 2) qtm = ait.bmm(hq) _s1 = hp.unsqueeze(1) _s2 = hp.unsqueeze(2) sjt = self.vs(torch.tanh(self.Ws(_s1 * _s2))).squeeze() ait = F.softmax(sjt, 2) qts = ait.bmm(hp) aggregation = torch.cat([hp, qts, qtc, qtd, qtb, qtm], 2) aggregation_representation, _ = self.gru_agg(aggregation) sj = self.vq(torch.tanh(self.Wq(hq))).transpose(2, 1) rq = F.softmax(sj, 2).bmm(hq) sj = F.softmax(self.vp(self.Wp1(aggregation_representation) + self.Wp2(rq)).transpose(2, 1), 2) rp = sj.bmm(aggregation_representation) encoder_output = F.dropout(F.leaky_relu(self.prediction(rp)),self.drop_out) score = F.softmax(a_embedding.bmm(encoder_output.transpose(2, 1)).squeeze(), 1) if not is_train: return score.argmax(1) loss = -torch.log(score[:, 0]).mean() return loss
4,704
44.679612
108
py
ReCO
ReCO-master/BiDAF/train.py
# -*- coding: utf-8 -*- import argparse import cPickle import torch from MwAN import MwAN from preprocess import process_data from utils import * parser = argparse.ArgumentParser(description='PyTorch implementation for Multiway Attention Networks for Modeling ' 'Sentence Pairs of the AI-Challenges') parser.add_argument('--data', type=str, default='data/', help='location directory of the data corpus') parser.add_argument('--threshold', type=int, default=5, help='threshold count of the word') parser.add_argument('--epoch', type=int, default=50, help='training epochs') parser.add_argument('--emsize', type=int, default=128, help='size of word embeddings') parser.add_argument('--nhid', type=int, default=128, help='hidden size of the model') parser.add_argument('--batch_size', type=int, default=32, metavar='N', help='batch size') parser.add_argument('--log_interval', type=int, default=300, help='# of batches to see the training error') parser.add_argument('--dropout', type=float, default=0.2, help='dropout applied to layers (0 = no dropout)') parser.add_argument('--cuda', action='store_true', help='use CUDA') parser.add_argument('--save', type=str, default='model.pt', help='path to save the final model') args = parser.parse_args() # vocab_size = process_data(args.data, args.threshold) vocab_size = 98745 model = MwAN(vocab_size=vocab_size, embedding_size=args.emsize, encoder_size=args.nhid, drop_out=args.dropout) print('Model total parameters:', get_model_parameters(model)) if args.cuda: model.cuda() optimizer = torch.optim.Adamax(model.parameters()) with open(args.data + 'train.pickle', 'rb') as f: train_data = cPickle.load(f) with open(args.data + 'dev.pickle', 'rb') as f: dev_data = cPickle.load(f) dev_data = sorted(dev_data, key=lambda x: len(x[1])) print('train data size {:d}, dev data size {:d}'.format(len(train_data), len(dev_data))) def train(epoch): model.train() data = shuffle_data(train_data, 1) total_loss = 0.0 for num, i in enumerate(range(0, len(data), args.batch_size)): one = data[i:i + args.batch_size] query, _ = padding([x[0] for x in one], max_len=50) passage, _ = padding([x[1] for x in one], max_len=350) answer = pad_answer([x[2] for x in one]) query, passage, answer = torch.LongTensor(query), torch.LongTensor(passage), torch.LongTensor(answer) if args.cuda: query = query.cuda() passage = passage.cuda() answer = answer.cuda() optimizer.zero_grad() loss = model([query, passage, answer, True]) loss.backward() total_loss += loss.item() optimizer.step() if (num + 1) % args.log_interval == 0: print '|------epoch {:d} train error is {:f} eclipse {:.2f}%------|'.format(epoch, total_loss / args.log_interval, i * 100.0 / len(data)) total_loss = 0 def test(): model.eval() r, a = 0.0, 0.0 with torch.no_grad(): for i in range(0, len(dev_data), args.batch_size): one = dev_data[i:i + args.batch_size] query, _ = padding([x[0] for x in one], max_len=50) passage, _ = padding([x[1] for x in one], max_len=500) answer = pad_answer([x[2] for x in one]) query, passage, answer = torch.LongTensor(query), torch.LongTensor(passage), torch.LongTensor(answer) if args.cuda: query = query.cuda() passage = passage.cuda() answer = answer.cuda() output = model([query, passage, answer, False]) r += torch.eq(output, 0).sum().item() a += len(one) return r * 100.0 / a def main(): best = 0.0 for epoch in range(args.epoch): train(epoch) acc = test() if acc > best: best = acc with open(args.save, 'wb') as f: torch.save(model, f) print 'epcoh {:d} dev acc is {:f}, best dev acc {:f}'.format(epoch, acc, best) if __name__ == '__main__': main()
4,466
38.184211
120
py
ReCO
ReCO-master/BiDAF/BiDAF.py
# -*- coding: utf-8 -*- """ @Time : 2019/11/21 下午4:42 @FileName: BiDAF.py @author: 王炳宁 @contact: wangbingning@sogou-inc.com """ import torch import torch.nn as nn import torch.nn.functional as F class BiDAF(nn.Module): def __init__(self, vocab_size, embedding_size, encoder_size, drop_out=0.2): super(BiDAF, self).__init__() # 2. Word Embedding Layer # initialize word embedding with GloVe self.word_emb = nn.Embedding(vocab_size, embedding_size) self.a_encoder = nn.GRU(input_size=embedding_size, hidden_size=embedding_size // 2, batch_first=True, bidirectional=True) self.a_attention = nn.Linear(embedding_size, 1, bias=False) # 3. Contextual Embedding Layer self.context_LSTM = nn.LSTM(input_size=embedding_size, hidden_size=encoder_size, bidirectional=True, batch_first=True, ) # 4. Attention Flow Layer self.att_weight_c = nn.Linear(encoder_size * 2, 1) self.att_weight_q = nn.Linear(encoder_size * 2, 1) self.att_weight_cq = nn.Linear(encoder_size * 2, 1) # 5. Modeling Layer self.modeling_LSTM = nn.LSTM(input_size=encoder_size * 8, hidden_size=encoder_size, bidirectional=True, batch_first=True, num_layers=2 ) self.Wq = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vq = nn.Linear(encoder_size, 1, bias=False) self.Wp1 = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.Wp2 = nn.Linear(2 * encoder_size, encoder_size, bias=False) self.vp = nn.Linear(encoder_size, 1, bias=False) self.prediction = nn.Linear(2 * encoder_size, embedding_size, bias=False) self.drop_out = drop_out def forward(self, inputs): def att_flow_layer(c, q): """ :param c: (batch, c_len, hidden_size * 2) :param q: (batch, q_len, hidden_size * 2) :return: (batch, c_len, q_len) """ c_len = c.size(1) q_len = q.size(1) cq = [] for i in range(q_len): # (batch, 1, hidden_size * 2) qi = q.select(1, i).unsqueeze(1) # (batch, c_len, 1) ci = self.att_weight_cq(c * qi).squeeze() cq.append(ci) # (batch, c_len, q_len) cq = torch.stack(cq, dim=-1) # (batch, c_len, q_len) s = self.att_weight_c(c).expand(-1, -1, q_len) + \ self.att_weight_q(q).permute(0, 2, 1).expand(-1, c_len, -1) + \ cq # (batch, c_len, q_len) a = F.softmax(s, dim=2) # (batch, c_len, q_len) * (batch, q_len, hidden_size * 2) -> (batch, c_len, hidden_size * 2) c2q_att = torch.bmm(a, q) # (batch, 1, c_len) b = F.softmax(torch.max(s, dim=2)[0], dim=1).unsqueeze(1) # (batch, 1, c_len) * (batch, c_len, hidden_size * 2) -> (batch, hidden_size * 2) q2c_att = torch.bmm(b, c).squeeze() # (batch, c_len, hidden_size * 2) (tiled) q2c_att = q2c_att.unsqueeze(1).expand(-1, c_len, -1) # q2c_att = torch.stack([q2c_att] * c_len, dim=1) # (batch, c_len, hidden_size * 8) x = torch.cat([c, c2q_att, c * c2q_att, c * q2c_att], dim=-1) return x # 2. Word Embedding Layer [query, passage, answer, is_train] = inputs c_word = self.word_emb(passage) q_word = self.word_emb(query) a_embeddings = self.word_emb(answer) a_embedding, _ = self.a_encoder(a_embeddings.view(-1, a_embeddings.size(2), a_embeddings.size(3))) a_score = F.softmax(self.a_attention(a_embedding), 1) a_output = a_score.transpose(2, 1).bmm(a_embedding).squeeze() a_embedding = a_output.view(a_embeddings.size(0), 3, -1) # Highway network # 3. Contextual Embedding Layer c, _ = self.context_LSTM(c_word) q, _ = self.context_LSTM(q_word) # 4. Attention Flow Layer g = att_flow_layer(c, q) # 5. Modeling Layer m, _ = self.modeling_LSTM(g) # 6. Output Layer sj = F.softmax(self.vp(self.Wp1(m)).transpose(2, 1), 2) rp = sj.bmm(m) encoder_output = F.dropout(F.leaky_relu(self.prediction(rp)), self.drop_out) score = F.softmax(a_embedding.bmm(encoder_output.transpose(2, 1)).squeeze(), 1) if not is_train: return score.argmax(1) loss = -torch.log(score[:, 0]).mean() return loss
4,904
40.923077
109
py
ReCO
ReCO-master/InHouseBert/model.py
# -*- coding: utf-8 -*- """ @Time : 2020/6/24 下午6:18 @FileName: model.py @author: 王炳宁 @contact: wangbingning@sogou-inc.com """ import warnings import apex import torch import torch.nn as nn from apex.contrib.multihead_attn import SelfMultiheadAttn from apex.mlp import MLP from torch.nn import functional as F warnings.filterwarnings("ignore") layer_norm = apex.normalization.FusedLayerNorm class TransformerEncoderLayer(nn.Module): def __init__(self, d_model, nhead, dim_feedforward=2048, dropout=0.1): super(TransformerEncoderLayer, self).__init__() self.self_attn = SelfMultiheadAttn(d_model, nhead, dropout=dropout, impl='fast') self.feed_forward = MLP([d_model, dim_feedforward, d_model]) self.d_model = d_model self.norm1 = layer_norm(d_model) self.norm2 = layer_norm(d_model) self.activation = F.gelu def __setstate__(self, state): if 'activation' not in state: state['activation'] = F.relu super(TransformerEncoderLayer, self).__setstate__(state) def forward(self, src, src_mask=None, src_key_padding_mask=None): # type: (Tensor, Optional[Tensor], Optional[Tensor]) -> Tensor src = self.norm2(src) src2 = self.self_attn(src, src, src, attn_mask=src_mask, key_padding_mask=src_key_padding_mask, is_training=self.training)[0] src = src + src2 src = self.norm1(src) src2 = self.feed_forward(src.view(-1, self.d_model)).view(src.size()) src = src + src2 return src class SelfAttention(nn.Module): def __init__(self, n_hidden, n_layer, n_head=6): super().__init__() self.att = nn.ModuleList() for l in range(n_layer): en = TransformerEncoderLayer(n_hidden, n_head, n_hidden * 4) self.att.append(en) self.output_ln = layer_norm(n_hidden) def forward(self, representations): representations = representations.transpose(0, 1).contiguous() for one in self.att: representations = one(representations) return self.output_ln(representations.transpose(0, 1)) class BERTLSTM(nn.Module): def __init__(self, vocab_size, n_embedding, n_hidden, n_layer, n_head): super().__init__() vocabulary_size = (2 + vocab_size // 8) * 8 self.word_embedding = nn.Embedding(vocabulary_size, embedding_dim=n_embedding) self.encoder = nn.LSTM(input_size=n_embedding, hidden_size=n_hidden // 2, bidirectional=True, batch_first=True) self.n_embedding = n_embedding self.n_hidden = n_hidden self.attention = SelfAttention(n_hidden, n_layer, n_head=n_head) self.output = nn.Sequential(nn.Linear(n_hidden, n_embedding), nn.LeakyReLU(inplace=True), apex.normalization.FusedLayerNorm(n_embedding)) self.trans = nn.Linear(n_embedding, vocabulary_size, bias=False) self.word_embedding.weight = self.trans.weight def inference(self, seq): word_embedding = self.word_embedding(seq) encoder_representations, _ = self.encoder(word_embedding) encoder_representations = self.attention(encoder_representations) return encoder_representations class BERT(BERTLSTM): def __init__(self, vocab_size, n_embedding, n_hidden, n_layer, n_head): super().__init__(vocab_size, n_embedding, n_hidden, n_layer, n_head) del self.trans del self.output self.prediction = nn.Sequential( nn.Linear(self.n_hidden, self.n_hidden // 2), nn.GELU(), nn.Linear(self.n_hidden // 2, 1, bias=False), ) def forward(self, inputs): [seq, label] = inputs hidden = self.inference(seq) mask_idx = torch.eq(seq, 1) # 1 is the index in the seq we separate each candidates. hidden = hidden.masked_select(mask_idx.unsqueeze(2).expand_as(hidden)).view( -1, 3, self.n_hidden) hidden = self.prediction(hidden).squeeze(-1) if label is None: return hidden.argmax(1) return F.cross_entropy(hidden, label)
4,192
37.118182
119
py
ReCO
ReCO-master/InHouseBert/train.py
# -*- coding: utf-8 -*- """ @Time : 2020/6/24 下午6:16 @FileName: train.py @author: 王炳宁 @contact: wangbingning@sogou-inc.com """ import argparse import sys sys.path.append("../..") sys.path.append("..") from tasks.ReCO.model import BERT from utils import * import torch.distributed as dist torch.manual_seed(100) np.random.seed(200) parser = argparse.ArgumentParser() parser.add_argument("--batch_size", type=int, default=16) parser.add_argument("--epoch", type=int, default=10) parser.add_argument("--lr", type=float, default=4.0e-5) parser.add_argument("--max_grad_norm", type=float, default=0.2) parser.add_argument("--model_type", type=str, default="bert-base-chinese-new") parser.add_argument( "--fp16", action="store_true", default=True, ) parser.add_argument("--local_rank", type=int, default=-1) args = parser.parse_args() model_type = args.model_type local_rank = args.local_rank if local_rank >= 0: torch.distributed.init_process_group(backend='nccl', init_method='env://') torch.cuda.set_device(args.local_rank) data = load_file('data/train.{}.obj'.format(model_type.replace('/', '.'))) valid_data = load_file('data/valid.{}.obj'.format(model_type.replace('/', '.'))) valid_data = sorted(valid_data, key=lambda x: len(x[0])) batch_size = args.batch_size n_embedding = 128 n_hidden = 768 n_layer = 12 n_head = 12 vocab_size = 50000 model = BERT(vocab_size, n_embedding, n_hidden, n_layer, n_head) state_dict = load_file('model.bert.base.th') for name, para in model.named_parameters(): if name not in state_dict: print('{} not load'.format(name)) continue para.data = torch.FloatTensor(state_dict[name]) model.cuda() optimizer = torch.optim.AdamW(model.parameters(), weight_decay=0.01, lr=args.lr) if args.fp16: try: from apex import amp except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use fp16 training.") model, optimizer = amp.initialize(model, optimizer, opt_level='O2', verbosity=0) if local_rank >= 0: try: import apex except ImportError: raise ImportError("Please install apex from https://www.github.com/nvidia/apex to use parallel training.") model = apex.parallel.DistributedDataParallel(model) def get_shuffle_data(): pool = {} for one in data: length = len(one[0]) // 5 if length not in pool: pool[length] = [] pool[length].append(one) for one in pool: np.random.shuffle(pool[one]) length_lst = list(pool.keys()) np.random.shuffle(length_lst) whole_data = [x for y in length_lst for x in pool[y]] if local_rank >= 0: remove_data_size = len(whole_data) % dist.get_world_size() thread_data = [whole_data[x + args.local_rank] for x in range(0, len(whole_data) - remove_data_size, dist.get_world_size())] return thread_data return whole_data def iter_printer(total, epoch): if local_rank >= 0: if local_rank == 0: return tqdm(range(0, total, batch_size), desc='epoch {}'.format(epoch)) else: return range(0, total, batch_size) else: return tqdm(range(0, total, batch_size), desc='epoch {}'.format(epoch)) def train(epoch): model.train() train_data = get_shuffle_data() total = len(train_data) for i in iter_printer(total, epoch): seq = [x[0] for x in train_data[i:i + batch_size]] label = [x[1] for x in train_data[i:i + batch_size]] seq, _ = padding(seq, pads=0, max_len=512) seq = torch.LongTensor(seq).cuda() label = torch.LongTensor(label).cuda() loss = model([seq, label]) if args.fp16: with amp.scale_loss(loss, optimizer) as scaled_loss: scaled_loss.backward() torch.nn.utils.clip_grad_norm_(amp.master_params(optimizer), args.max_grad_norm) else: loss.backward() torch.nn.utils.clip_grad_norm_(model.parameters(), args.max_grad_norm) optimizer.step() optimizer.zero_grad() def evaluation(epoch): model.eval() total = len(valid_data) right = 0 with torch.no_grad(): for i in iter_printer(total, epoch): seq = [x[0] for x in valid_data[i:i + batch_size]] labels = [x[1] for x in valid_data[i:i + batch_size]] seq, _ = padding(seq, pads=0, max_len=512) seq = torch.LongTensor(seq).cuda() predictions = model([seq, None]) predictions = predictions.cpu() right += predictions.eq(torch.LongTensor(labels)).sum().item() acc = 100 * right / total print('epoch {} eval acc is {}'.format(epoch, acc)) return acc best_acc = 0.0 for epo in range(args.epoch): train(epo) if local_rank == -1 or local_rank == 0: accuracy = evaluation(epo) if accuracy > best_acc: best_acc = accuracy with open('checkpoint.{}.th'.format(model_type.replace('/', '.')), 'wb') as f: state_dict = model.module.state_dict() if args.fp16 else model.state_dict() torch.save(state_dict, f)
5,292
32.713376
114
py
mining-legal-arguments
mining-legal-arguments-main/evaluate.py
#!/usr/bin/env python # coding: utf-8 from collections import Counter from prettytable import PrettyTable import os from transformers import AutoTokenizer import torch from torch.utils.data import Dataset import pandas as pd from datasets import load_dataset, load_metric import csv from ast import literal_eval import numpy as np import torch import torch.nn as nn import transformers from datasets import load_dataset, load_metric import logging import dataclasses from torch.utils.data.dataloader import DataLoader from transformers.training_args import is_torch_tpu_available from transformers.trainer_pt_utils import get_tpu_sampler from transformers.data.data_collator import DataCollator, InputDataClass from torch.utils.data.distributed import DistributedSampler from torch.utils.data.sampler import RandomSampler, SequentialSampler from typing import List, Union, Dict from transformers import DataCollatorForTokenClassification from transformers.tokenization_utils_base import PreTrainedTokenizerBase from transformers.file_utils import PaddingStrategy from typing import Optional, Any from sklearn.metrics import confusion_matrix from multiTaskModel import MultitaskModel, StrIgnoreDevice, DataLoaderWithTaskname, MultitaskDataloader, MultitaskTrainer, MyDataCollatorForTokenClassification, compute_f1, compute_macro_f1, eval_f1 import argparse logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) id2label_argType = ['B-Distinguishing', 'B-Einschätzungsspielraum', 'B-Entscheidung des EGMR', 'B-Konsens der prozessualen Parteien', 'B-Overruling', 'B-Rechtsvergleichung', 'B-Sinn & Zweck Auslegung', 'B-Subsumtion', 'B-Systematische Auslegung', 'B-Verhältnismäßigkeitsprüfung – Angemessenheit', 'B-Verhältnismäßigkeitsprüfung – Geeignetheit', 'B-Verhältnismäßigkeitsprüfung – Legitimer Zweck', 'B-Verhältnismäßigkeitsprüfung – Rechtsgrundlage', 'B-Vorherige Rechtsprechung des EGMR', 'B-Wortlaut Auslegung', 'I-Distinguishing', 'I-Einschätzungsspielraum', 'I-Entscheidung des EGMR', 'I-Konsens der prozessualen Parteien', 'I-Overruling', 'I-Rechtsvergleichung', 'I-Sinn & Zweck Auslegung', 'I-Subsumtion', 'I-Systematische Auslegung', 'I-Verhältnismäßigkeitsprüfung – Angemessenheit', 'I-Verhältnismäßigkeitsprüfung – Geeignetheit', 'I-Verhältnismäßigkeitsprüfung – Legitimer Zweck', 'I-Verhältnismäßigkeitsprüfung – Rechtsgrundlage', 'I-Vorherige Rechtsprechung des EGMR', 'I-Wortlaut Auslegung', 'O'] label2id_argType = {} for i, label in enumerate(id2label_argType): label2id_argType[label] = i id2label_agent = ['B-Beschwerdeführer', 'B-Dritte', 'B-EGMR', 'B-Kommission/Kammer', 'B-Staat', 'I-Beschwerdeführer', 'I-Dritte', 'I-EGMR', 'I-Kommission/Kammer', 'I-Staat', 'O'] label2id_agent = {} for i, label in enumerate(id2label_agent): label2id_agent[label] = i def tokenize_and_align_labels_argType(examples, label_all_tokens=False): """ Tokenizes the input using the tokenizer and aligns the argument type labels to the subwords. :param examples: input dataset :param label_all_tokens: Whether to label all subwords of a token or only the first subword :return: Tokenized input""" tokenized_inputs = tokenizer(examples['tokens'], truncation=True, is_split_into_words=True) labels = [] for i, label in enumerate(examples['labels']): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label2id_argType[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label2id_argType[label[word_idx]] if label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs def tokenize_and_align_labels_agent(examples, label_all_tokens=False): """ Tokenizes the input using the tokenizer and aligns the agent labels to the subwords. :param examples: input dataset :param label_all_tokens: Whether to label all subwords of a token or only the first subword :return: Tokenized input""" tokenized_inputs = tokenizer(examples['tokens'], truncation=True, is_split_into_words=True) labels = [] for i, label in enumerate(examples['labels']): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label2id_agent[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label2id_agent[label[word_idx]] if label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs def get_subset_df(tokens, predictions, labels, predlabel=None, truelabel=None): """ Can filter model predictions by predicted and true label. If none provided, just postprocesses and returns output. :param tokens: text tokens :param predictions: predictions of the model :param labels: true annotator labels :param predlabel: optional, filters the model predictions if a label is provided, e.g. label2id_agent['I-EGMR'] :param truelabels: optional, filters the true annotator labels if a label is provided, e.g. label2id_agent['I-EGMR'] :return: DataFrame with the (filtered) predictions, labels and tokens.""" pred = [] label = [] for p,l in zip(predictions, labels): p = np.array(p) l = np.array(l) ind = np.logical_and(p > -1, l > -1) pred.append(p[ind].tolist()) label.append(l[ind].tolist()) if predlabel is None and truelabel is None: return pd.DataFrame({'Predictions': pred, 'Labels': label, 'Tokens': tokens}) elif predlabel is None: preds = [] labels = [] toks = [] for i,l in enumerate(label): if truelabel in l[2:] and not truelabel in pred[i][2:]: preds.append(pred[i]) labels.append(l) toks.append(tokens[i]) return pd.DataFrame({'Predictions': preds, 'Labels': labels, 'Tokens': toks}) elif truelabel is None: preds = [] labels = [] toks = [] for i,p in enumerate(pred): if predlabel in p[2:] and not predlabel in label[i][2:]: preds.append(p) labels.append(llabel[i]) toks.append(tokens[i]) return pd.DataFrame({'Predictions': preds, 'Labels': labels, 'Tokens': toks}) else: preds = [] labels = [] toks = [] for i,p in enumerate(pred): if predlabel in p[2:] and truelabel in label[i][2:]: preds.append(p) labels.append(label[i]) toks.append(tokens[i]) return pd.DataFrame({'Predictions': preds, 'Labels': labels, 'Tokens': toks}) def save_predictions(tokens, predictions, labels, file): """ Saves the model predictions as csv after postprocessing them. :param tokens: text tokens :param predictions: predictions of the model :param labels: true annotator labels :param file: path of the output file """ df = get_subset_df(tokens, predictions, labels) df.to_csv(file , sep='\t', encoding='utf-8', index=False) if __name__ == '__main__': # parse optional args parser = argparse.ArgumentParser(description='Evaluate a MultiTask model and save its predictions') parser.add_argument('--pathprefix', help='path to the project directory') parser.add_argument('--models', nargs='*' ,help='paths to the models to evaluate') parser.add_argument('--test_dir', help='path to the directory with the test files') parser.add_argument('--val_dir', help='path to the directory with the dev files') parser.add_argument('--output_dir', help='path to the output directory for saving the predictions') parser.add_argument('--do_val', default=False, type=lambda x: (str(x).lower() == 'true'), help='whether to evaluate the validation/dev dataset') parser.add_argument('--do_test', default=True, type=lambda x: (str(x).lower() == 'true'), help='whether to evaluate the test dataset') args = parser.parse_args() # project directory pathprefix = '/ukp-storage-1/dfaber/' pathprefix = '../Uni/masterthesis/' pathprefix = '' if args.pathprefix: pathprefix = args.pathprefix #test_dir = 'data/article_3/' test_dir = 'data/test/' if args.test_dir: test_dir = args.test_dir val_dir = 'data/val/' if args.val_dir: val_dir = args.val_dir output_dir = 'predictions/' if args.output_dir: output_dir = args.output_dir # load datasets testfiles = [f for f in os.listdir(os.path.join(pathprefix, test_dir, 'argType/')) if f.endswith('.csv')] valfiles = [f for f in os.listdir(os.path.join(pathprefix, val_dir, 'argType/')) if f.endswith('.csv')] testfiles = testfiles[:2] valfiles = valfiles[:2] dataset_argType = load_dataset('csv', data_files={'test': [os.path.join(pathprefix, test_dir, 'argType/', file) for file in testfiles], 'validation': [os.path.join(pathprefix, val_dir, 'argType/', file) for file in valfiles]}, delimiter='\t') dataset_actor = load_dataset('csv', data_files={'test': [os.path.join(pathprefix, test_dir, 'agent/', file) for file in testfiles], 'validation': [os.path.join(pathprefix, val_dir, 'agent/', file) for file in valfiles]}, delimiter='\t') dataset_argType = dataset_argType.map(lambda x: {'tokens': literal_eval(x['tokens']), 'labels': literal_eval(x['labels'])}) dataset_actor = dataset_actor.map(lambda x: {'tokens': literal_eval(x['tokens']), 'labels': literal_eval(x['labels'])}) # models to evaluate ''' models = ['/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-39820/bert', '/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-47784/bert', '/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-55748/bert', '/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-71676/bert', '/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-79640/bert', '/ukp-storage-1/dfaber/models/multitask/roberta-large-final/checkpoint-111482/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-final/checkpoint-143334/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-final/checkpoint-159260/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-13000/checkpoint-95556/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-13000/checkpoint-127408/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-13000/checkpoint-143334/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-13000/checkpoint-159260/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-111482/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-127408/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-143334/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-159260/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-final/checkpoint-143334/roberta'] ''' models = ['/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-39820/bert', '/ukp-storage-1/dfaber/models/multitask/legal-bert-final/checkpoint-47784/bert', '/ukp-storage-1/dfaber/models/multitask/roberta-large-final/checkpoint-111482/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-final/checkpoint-143334/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-13000/checkpoint-95556/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-13000/checkpoint-143334/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-143334/roberta', '/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-159260/roberta'] models = ['/ukp-storage-1/dfaber/models/multitask/roberta-large-fp-15000/checkpoint-143334/roberta'] if args.models: models = args.models # Evaluate each model for model in models: print('\n\n\n\n********************Evaluating ', model, '********************\n\n\n\n') # load model and tokenizer multitask_model = torch.load(model) tokenizer = AutoTokenizer.from_pretrained(multitask_model.encoder.name_or_path) if model.split('/')[-1] == 'roberta': tokenizer.add_prefix_space = True if tokenizer.model_max_length > 1024: tokenizer.model_max_length = 512 # preprocess data and create datasets tokenized_dataset_argType = dataset_argType.map(tokenize_and_align_labels_argType, batched=True) tokenized_dataset_actor = dataset_actor.map(tokenize_and_align_labels_agent, batched=True) dataset_dict = { "ArgType": tokenized_dataset_argType, "Actor": tokenized_dataset_actor, } data_collator= MyDataCollatorForTokenClassification(tokenizer) test_dataset = { task_name: dataset["test"] for task_name, dataset in dataset_dict.items() } val_dataset = { task_name: dataset["validation"] for task_name, dataset in dataset_dict.items() } # initialize Trainer batch_size = 8 train_args = transformers.TrainingArguments( 'test_bert/legal_bert/', per_device_train_batch_size=batch_size, per_device_eval_batch_size=batch_size, ) trainer = MultitaskTrainer( model=multitask_model, args=train_args, data_collator=data_collator, eval_dataset=val_dataset, tokenizer=tokenizer, compute_metrics=eval_f1, ) # evaluate validation data if specified if args.do_val: print('\n\n*****VALIDATION DATASET*****\n\n') eval_dataloader_argType = DataLoaderWithTaskname( 'ArgType', data_loader=DataLoader( val_dataset['ArgType'], batch_size=trainer.args.eval_batch_size, collate_fn=trainer.data_collator.collate_batch, ), ) preds_arg = trainer.prediction_loop(eval_dataloader_argType, description='Validation ArgType') eval_dataloader_agent = DataLoaderWithTaskname( 'Actor', data_loader=DataLoader( val_dataset['Actor'], batch_size=trainer.args.eval_batch_size, collate_fn=trainer.data_collator.collate_batch, ), ) preds_agent = trainer.prediction_loop(eval_dataloader_agent, description='Validation Agent') # postprocess (remove -100 indices) labels_argType_wordlevel = [] preds_argType_wordlevel = [] for l,p in zip(preds_arg.label_ids, np.argmax(preds_arg.predictions, axis=2)): ind = np.logical_and(p > -1, l > -1) labels_argType_wordlevel.append(l[ind]) preds_argType_wordlevel.append(p[ind]) print('ArgType:') print('Macro F1: ', compute_macro_f1(gold=labels_argType_wordlevel, pred=preds_argType_wordlevel, id2label=id2label_argType)) # postprocess (remove -100 indices) labels_agent_wordlevel = [] preds_agent_wordlevel = [] for l,p in zip(preds_agent.label_ids, np.argmax(preds_agent.predictions, axis=2)): ind = np.logical_and(p > -1, l > -1) labels_agent_wordlevel.append(l[ind]) preds_agent_wordlevel.append(p[ind]) print('Agent:') print('Macro F1: ', compute_macro_f1(gold=labels_agent_wordlevel, pred=preds_agent_wordlevel, id2label=id2label_agent)) # save predictions save_predictions(val_dataset['ArgType']['tokens'], np.argmax(preds_arg.predictions, axis=2), preds_arg.label_ids, os.path.join(pathprefix, output_dir, 'val_preds/', '_'.join(model.split('/')[-3:]) + '-argType.csv')) save_predictions(val_dataset['Actor']['tokens'], np.argmax(preds_agent.predictions, axis=2), preds_agent.label_ids, os.path.join(pathprefix, output_dir, 'val_preds/', '_'.join(model.split('/')[-3:]) + '-agent.csv')) # evaluate test data if specified if args.do_test: print('\n\n*****TEST DATASET*****\n\n') eval_dataloader_argType = DataLoaderWithTaskname( 'ArgType', data_loader=DataLoader( test_dataset['ArgType'], batch_size=trainer.args.eval_batch_size, collate_fn=trainer.data_collator.collate_batch, ), ) preds_arg = trainer.prediction_loop(eval_dataloader_argType, description='Validation ArgType') eval_dataloader_agent = DataLoaderWithTaskname( 'Actor', data_loader=DataLoader( test_dataset['Actor'], batch_size=trainer.args.eval_batch_size, collate_fn=trainer.data_collator.collate_batch, ), ) preds_agent = trainer.prediction_loop(eval_dataloader_agent, description='Validation Agent') # postprocess (remove -100 indices) labels_argType_wordlevel = [] preds_argType_wordlevel = [] for l,p in zip(preds_arg.label_ids, np.argmax(preds_arg.predictions, axis=2)): ind = np.logical_and(p > -1, l > -1) labels_argType_wordlevel.append(l[ind]) preds_argType_wordlevel.append(p[ind]) print('ArgType:') print('Macro F1: ', compute_macro_f1(gold=labels_argType_wordlevel, pred=preds_argType_wordlevel, id2label=id2label_argType)) # postprocess (remove -100 indices) labels_agent_wordlevel = [] preds_agent_wordlevel = [] for l,p in zip(preds_agent.label_ids, np.argmax(preds_agent.predictions, axis=2)): ind = np.logical_and(p > -1, l > -1) labels_agent_wordlevel.append(l[ind]) preds_agent_wordlevel.append(p[ind]) print('Agent:') print('Macro F1: ', compute_macro_f1(gold=labels_agent_wordlevel, pred=preds_agent_wordlevel, id2label=id2label_agent)) # save predictions save_predictions(test_dataset['ArgType']['tokens'], np.argmax(preds_arg.predictions, axis=2), preds_arg.label_ids, os.path.join(pathprefix, output_dir, '_'.join(model.split('/')[-3:]) + '-argType.csv')) save_predictions(test_dataset['Actor']['tokens'], np.argmax(preds_agent.predictions, axis=2), preds_agent.label_ids, os.path.join(pathprefix, output_dir, '_'.join(model.split('/')[-3:]) + '-agent.csv'))
20,634
44.855556
227
py
mining-legal-arguments
mining-legal-arguments-main/multiTaskModel.py
#!/usr/bin/env python # coding: utf-8 from collections import Counter from prettytable import PrettyTable import os from transformers import AutoTokenizer import torch from torch.utils.data import Dataset import pandas as pd from datasets import load_dataset, load_metric import csv from ast import literal_eval import numpy as np import torch import torch.nn as nn import transformers from datasets import load_dataset, load_metric import logging import dataclasses from torch.utils.data.dataloader import DataLoader from transformers.training_args import is_torch_tpu_available from transformers.trainer_pt_utils import get_tpu_sampler from transformers.data.data_collator import DataCollator, InputDataClass from torch.utils.data.distributed import DistributedSampler from torch.utils.data.sampler import RandomSampler from typing import List, Union, Dict from transformers import DataCollatorForTokenClassification from transformers.tokenization_utils_base import PreTrainedTokenizerBase from transformers.file_utils import PaddingStrategy from typing import Optional, Any import argparse from tabulate import tabulate logging.basicConfig(level=logging.INFO) logger = logging.getLogger(__name__) id2label_argType = ['B-Distinguishing', 'B-Einschätzungsspielraum', 'B-Entscheidung des EGMR', 'B-Konsens der prozessualen Parteien', 'B-Overruling', 'B-Rechtsvergleichung', 'B-Sinn & Zweck Auslegung', 'B-Subsumtion', 'B-Systematische Auslegung', 'B-Verhältnismäßigkeitsprüfung – Angemessenheit', 'B-Verhältnismäßigkeitsprüfung – Geeignetheit', 'B-Verhältnismäßigkeitsprüfung – Legitimer Zweck', 'B-Verhältnismäßigkeitsprüfung – Rechtsgrundlage', 'B-Vorherige Rechtsprechung des EGMR', 'B-Wortlaut Auslegung', 'I-Distinguishing', 'I-Einschätzungsspielraum', 'I-Entscheidung des EGMR', 'I-Konsens der prozessualen Parteien', 'I-Overruling', 'I-Rechtsvergleichung', 'I-Sinn & Zweck Auslegung', 'I-Subsumtion', 'I-Systematische Auslegung', 'I-Verhältnismäßigkeitsprüfung – Angemessenheit', 'I-Verhältnismäßigkeitsprüfung – Geeignetheit', 'I-Verhältnismäßigkeitsprüfung – Legitimer Zweck', 'I-Verhältnismäßigkeitsprüfung – Rechtsgrundlage', 'I-Vorherige Rechtsprechung des EGMR', 'I-Wortlaut Auslegung', 'O'] label2id_argType = {} for i, label in enumerate(id2label_argType): label2id_argType[label] = i id2label_agent = ['B-Beschwerdeführer', 'B-Dritte', 'B-EGMR', 'B-Kommission/Kammer', 'B-Staat', 'I-Beschwerdeführer', 'I-Dritte', 'I-EGMR', 'I-Kommission/Kammer', 'I-Staat', 'O'] label2id_agent = {} for i, label in enumerate(id2label_agent): label2id_agent[label] = i def tokenize_and_align_labels_argType(examples, label_all_tokens=False): """ Tokenizes the input using the tokenizer and aligns the argument type labels to the subwords. :param examples: input dataset :param label_all_tokens: Whether to label all subwords of a token or only the first subword :return: Tokenized input""" tokenized_inputs = tokenizer(examples['tokens'], truncation=True, is_split_into_words=True) labels = [] for i, label in enumerate(examples['labels']): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label2id_argType[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label2id_argType[label[word_idx]] if label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs def tokenize_and_align_labels_agent(examples, label_all_tokens=False): """ Tokenizes the input using the tokenizer and aligns the agent labels to the subwords. :param examples: input dataset :param label_all_tokens: Whether to label all subwords of a token or only the first subword :return: Tokenized input""" tokenized_inputs = tokenizer(examples['tokens'], truncation=True, is_split_into_words=True) labels = [] for i, label in enumerate(examples['labels']): word_ids = tokenized_inputs.word_ids(batch_index=i) previous_word_idx = None label_ids = [] for word_idx in word_ids: # Special tokens have a word id that is None. We set the label to -100 so they are automatically # ignored in the loss function. if word_idx is None: label_ids.append(-100) # We set the label for the first token of each word. elif word_idx != previous_word_idx: label_ids.append(label2id_agent[label[word_idx]]) # For the other tokens in a word, we set the label to either the current label or -100, depending on # the label_all_tokens flag. else: label_ids.append(label2id_agent[label[word_idx]] if label_all_tokens else -100) previous_word_idx = word_idx labels.append(label_ids) tokenized_inputs["labels"] = labels return tokenized_inputs class MultitaskModel(transformers.PreTrainedModel): def __init__(self, encoder, taskmodels_dict): """ Setting MultitaskModel up as a PretrainedModel allows us to take better advantage of Trainer features """ super().__init__(transformers.PretrainedConfig()) self.encoder = encoder self.taskmodels_dict = nn.ModuleDict(taskmodels_dict) @classmethod def create(cls, model_name, model_type_dict, model_config_dict): """ This creates a MultitaskModel using the model class and config objects from single-task models. We do this by creating each single-task model, and having them share the same encoder transformer. """ shared_encoder = None taskmodels_dict = {} for task_name, model_type in model_type_dict.items(): model = model_type.from_pretrained( model_name, config=model_config_dict[task_name], ) if shared_encoder is None: shared_encoder = getattr(model, cls.get_encoder_attr_name(model)) else: setattr(model, cls.get_encoder_attr_name(model), shared_encoder) taskmodels_dict[task_name] = model return cls(encoder=shared_encoder, taskmodels_dict=taskmodels_dict) @classmethod def get_encoder_attr_name(cls, model): """ The encoder transformer is named differently in each model "architecture". This method lets us get the name of the encoder attribute """ model_class_name = model.__class__.__name__ if model_class_name.startswith("Bert"): return "bert" elif model_class_name.startswith("Roberta"): return "roberta" elif model_class_name.startswith("Albert"): return "albert" elif model_class_name.startswith("DistilBert"): return "distilbert" else: raise KeyError(f"Add support for new model {model_class_name}") def forward(self, task_name, **kwargs): return self.taskmodels_dict[task_name](**kwargs) class StrIgnoreDevice(str): """ This is a hack. The Trainer is going call .to(device) on every input value, but we need to pass in an additional `task_name` string. This prevents it from throwing an error """ def to(self, device): return self class DataLoaderWithTaskname: """ Wrapper around a DataLoader to also yield a task name """ def __init__(self, task_name, data_loader): self.task_name = task_name self.data_loader = data_loader self.batch_size = data_loader.batch_size self.dataset = data_loader.dataset def __len__(self): return len(self.data_loader) def __iter__(self): for batch in self.data_loader: batch["task_name"] = StrIgnoreDevice(self.task_name) yield batch class MultitaskDataloader: """ Data loader that combines and samples from multiple single-task data loaders. """ def __init__(self, dataloader_dict): self.dataloader_dict = dataloader_dict self.num_batches_dict = { task_name: len(dataloader) for task_name, dataloader in self.dataloader_dict.items() } self.task_name_list = list(self.dataloader_dict) self.dataset = [None] * sum( len(dataloader.dataset) for dataloader in self.dataloader_dict.values() ) def __len__(self): return sum(self.num_batches_dict.values()) def __iter__(self): """ For each batch, sample a task, and yield a batch from the respective task Dataloader. We use size-proportional sampling, but you could easily modify this to sample from some-other distribution. """ task_choice_list = [] for i, task_name in enumerate(self.task_name_list): task_choice_list += [i] * self.num_batches_dict[task_name] task_choice_list = np.array(task_choice_list) np.random.shuffle(task_choice_list) dataloader_iter_dict = { task_name: iter(dataloader) for task_name, dataloader in self.dataloader_dict.items() } for task_choice in task_choice_list: task_name = self.task_name_list[task_choice] yield next(dataloader_iter_dict[task_name]) class MultitaskTrainer(transformers.Trainer): def get_single_train_dataloader(self, task_name, train_dataset): """ Create a single-task data loader that also yields task names """ if self.train_dataset is None: raise ValueError("Trainer: training requires a train_dataset.") if is_torch_tpu_available(): train_sampler = get_tpu_sampler(train_dataset) else: train_sampler = ( RandomSampler(train_dataset) if self.args.local_rank == -1 else DistributedSampler(train_dataset) ) data_loader = DataLoaderWithTaskname( task_name=task_name, data_loader=DataLoader( train_dataset, batch_size=self.args.train_batch_size, sampler=train_sampler, collate_fn=self.data_collator.collate_batch, ), ) if is_torch_tpu_available(): data_loader = pl.ParallelLoader( data_loader, [self.args.device] ).per_device_loader(self.args.device) return data_loader def get_train_dataloader(self): """ Returns a MultitaskDataloader, which is not actually a Dataloader but an iterable that returns a generator that samples from each task Dataloader. """ return MultitaskDataloader({ task_name: self.get_single_train_dataloader(task_name, task_dataset) for task_name, task_dataset in self.train_dataset.items() }) def get_eval_dataloader(self, q): """ Returns a DataLoaderWithTaskname for the argument type task for evaluation of it during the training. """ eval_dataloader_argType = DataLoaderWithTaskname( 'ArgType', data_loader=DataLoader( eval_dataset['ArgType'], batch_size=trainer.args.eval_batch_size, collate_fn=trainer.data_collator.collate_batch, ), ) return eval_dataloader_argType def save_model(self, output_dir: Optional[str] = None): """ Saving best-practices: if you use default names for the model, you can reload it using from_pretrained(). Will only save from the world_master process (unless in TPUs). """ if is_torch_tpu_available(): self._save_tpu(output_dir) elif self.is_world_process_zero(): self._save(output_dir) def _save_tpu(self, output_dir: Optional[str] = None): output_dir = output_dir if output_dir is not None else self.args.output_dir logger.info("Saving model checkpoint to %s", output_dir) if xm.is_master_ordinal(): os.makedirs(output_dir, exist_ok=True) torch.save(self.args, os.path.join(output_dir, "training_args.bin")) xm.rendezvous("saving_checkpoint") torch.save(self.model, os.path.join(output_dir, self.model.encoder.base_model_prefix)) def _save(self, output_dir: Optional[str] = None): output_dir = output_dir if output_dir is not None else self.args.output_dir os.makedirs(output_dir, exist_ok=True) logger.info("Saving model checkpoint to %s", output_dir) # Low-Level workaround for MultiTaskModel torch.save(self.model, os.path.join(output_dir, self.model.encoder.base_model_prefix)) # Good practice: save your training arguments together with the trained model torch.save(self.args, os.path.join(output_dir, "training_args.bin")) @dataclasses.dataclass class MyDataCollatorForTokenClassification: """ Data collator that will dynamically pad the inputs received, as well as the labels. Args: tokenizer (:class:`~transformers.PreTrainedTokenizer` or :class:`~transformers.PreTrainedTokenizerFast`): The tokenizer used for encoding the data. padding (:obj:`bool`, :obj:`str` or :class:`~transformers.file_utils.PaddingStrategy`, `optional`, defaults to :obj:`True`): Select a strategy to pad the returned sequences (according to the model's padding side and padding index) among: * :obj:`True` or :obj:`'longest'`: Pad to the longest sequence in the batch (or no padding if only a single sequence if provided). * :obj:`'max_length'`: Pad to a maximum length specified with the argument :obj:`max_length` or to the maximum acceptable input length for the model if that argument is not provided. * :obj:`False` or :obj:`'do_not_pad'` (default): No padding (i.e., can output a batch with sequences of different lengths). max_length (:obj:`int`, `optional`): Maximum length of the returned list and optionally padding length (see above). pad_to_multiple_of (:obj:`int`, `optional`): If set will pad the sequence to a multiple of the provided value. This is especially useful to enable the use of Tensor Cores on NVIDIA hardware with compute capability >= 7.5 (Volta). label_pad_token_id (:obj:`int`, `optional`, defaults to -100): The id to use when padding the labels (-100 will be automatically ignore by PyTorch loss functions). """ tokenizer: PreTrainedTokenizerBase padding: Union[bool, str, PaddingStrategy] = True max_length: Optional[int] = None pad_to_multiple_of: Optional[int] = None label_pad_token_id: int = -100 # call not used? def __call__(self, features): label_name = "label" if "label" in features[0].keys() else "labels" labels = [feature[label_name] for feature in features] if label_name in features[0].keys() else None batch = self.tokenizer.pad( features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, # Conversion to tensors will fail if we have labels as they are not of the same length yet. return_tensors="pt" if labels is None else None, ) if labels is None: return batch sequence_length = torch.tensor(batch["input_ids"]).shape[1] padding_side = self.tokenizer.padding_side if padding_side == "right": batch["labels"] = [label + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels] else: batch["labels"] = [[self.label_pad_token_id] * (sequence_length - len(label)) + label for label in labels] batch = {k: torch.tensor(v, dtype=torch.int64) for k, v in batch.items()} return batch def collate_batch(self, features, pad_to_multiple_of: Optional[int] = None): label_name = "label" if "label" in features[0].keys() else "labels" labels = [feature[label_name] for feature in features] if label_name in features[0].keys() else None batch = self.tokenizer.pad( features, padding=self.padding, max_length=self.max_length, pad_to_multiple_of=self.pad_to_multiple_of, # Conversion to tensors will fail if we have labels as they are not of the same length yet. return_tensors="pt" if labels is None else None, ) if labels is None: return batch del batch['tokens'] sequence_length = torch.tensor(batch["input_ids"]).shape[1] padding_side = self.tokenizer.padding_side if padding_side == "right": batch["labels"] = [label + [self.label_pad_token_id] * (sequence_length - len(label)) for label in labels] else: batch["labels"] = [[self.label_pad_token_id] * (sequence_length - len(label)) + label for label in labels] batch = {k: torch.tensor(v, dtype=torch.int64) for k, v in batch.items()} return batch def compute_f1(label, gold, pred): """ Computes the F1 Score for a single class. :param labal: the class to compute the score for :param gold: the gold standard :param pred: the model predictions :return: the F1 score for the label""" tp = 0 fp = 0 fn = 0 for i, sent in enumerate(pred): for j, tag in enumerate(sent): # check for relevant label to compute F1 if tag == label: # if relevant and equals gold -> true positive if tag == gold[i][j]: tp += 1 # if it differs from gold -> false positive else: fp += 1 # we have a negative, so check if it's a false negative else: if gold[i][j] == label: fn += 1 # use epsilon to avoid division by zero precision = tp / (tp + fp + 1e-10) recall = tp / (tp + fn + 1e-10) f1 = 2 * precision * recall / (precision + recall + 1e-10) return f1 def compute_macro_f1(gold, pred, id2label): """ Computes the Macro F1 Score over all classes. :param gold: the gold standard :param pred: the model predictions :param id2label: the mapping list for the current labels :return: the Macro F1 score""" f1s = [(tag, compute_f1(tag, gold, pred)) for tag in range(len(id2label))] all_f1s = [(id2label[idx], score) for idx, score in f1s] df = pd.DataFrame(all_f1s, columns=['Label', 'F1']) df['F1'] = np.around(df['F1'], decimals=4) print(tabulate(df, headers='keys', tablefmt='pretty', showindex=False)) f1_scores = [f1[1] for f1 in f1s] macro_f1 = np.sum(f1_scores) / len(f1_scores) #print('Macro F1: ', macro_f1) return macro_f1 def eval_f1(evalpred): """ Computes the Macro F1 Score over all argument type classes during train evaluation. :param evalpred: evalpred from the trainer :return: the Macro F1 score""" pred = [] gold = [] for p,l in zip(np.argmax(evalpred.predictions, axis=2), evalpred.label_ids): ind = np.logical_and(p > -1, l > -1) pred.append(p[ind]) gold.append(l[ind]) f1s = [(tag, compute_f1(tag, gold, pred)) for tag in range(len(id2label_argType))] all_f1s = [(id2label_argType[idx], score) for idx, score in f1s] #print('F1 for each Class: ', all_f1s) f1_scores = [f1[1] for f1 in f1s] macro_f1 = np.sum(f1_scores) / len(f1_scores) return {"F1 ArgType": macro_f1} if __name__ == '__main__': # parse optional args parser = argparse.ArgumentParser(description='Train a MultiTask model') parser.add_argument('--pathprefix', help='path to the project directory') parser.add_argument('--model', help='name of the model or path to the model') parser.add_argument('--tokenizer', help='name of the model or path to the tokenizer') parser.add_argument('--batch_size', type=int, help='batch size of the model') parser.add_argument('--output_dir', help='path to the output directory') args = parser.parse_args() # path to working directory pathprefix = '/ukp-storage-1/dfaber/' #pathprefix = '' if args.pathprefix: pathprefix = args.pathprefix # load datasets trainfiles = [f for f in os.listdir(pathprefix + 'data/train/argType/') if f.endswith('.csv')] valfiles = [f for f in os.listdir(pathprefix + 'data/val/argType/') if f.endswith('.csv')] dataset_argType = load_dataset('csv', data_files={'train': [pathprefix + 'data/train/argType/' + file for file in trainfiles], 'validation': [pathprefix + 'data/val/argType/' + file for file in valfiles]}, delimiter='\t') dataset_actor = load_dataset('csv', data_files={'train': [pathprefix + 'data/train/agent/' + file for file in trainfiles], 'validation': [pathprefix + 'data/val/agent/' + file for file in valfiles]}, delimiter='\t') dataset_argType = dataset_argType.map(lambda x: {'tokens': literal_eval(x['tokens']), 'labels': literal_eval(x['labels'])}) dataset_actor = dataset_actor.map(lambda x: {'tokens': literal_eval(x['tokens']), 'labels': literal_eval(x['labels'])}) # select the model with the correspronding tokenizer #model_name = "/ukp-storage-1/dfaber/models/court_bert/checkpoint-20000" #tokenizer = AutoTokenizer.from_pretrained('/ukp-storage-1/dfaber/legal_tokenizer_bert', do_lower_case=False) model_name = "/ukp-storage-1/dfaber/models/roberta-large-finetuned/checkpoint-15000" #model_name = 'roberta-large' tokenizer = AutoTokenizer.from_pretrained('roberta-large') #model_name = 'nlpaueb/legal-bert-base-uncased' #tokenizer = AutoTokenizer.from_pretrained('nlpaueb/legal-bert-base-uncased') # use parsed args if provided if args.model: model_name = args.model if args.tokenizer: tokenizer = AutoTokenizer.from_pretrained(args.tokenizer) # need prefix space for already tokenized data if 'roberta' in model_name: tokenizer.add_prefix_space = True if tokenizer.model_max_length > 1024: tokenizer.model_max_length = 512 # tokenize and align labels tokenized_dataset_argType = dataset_argType.map(tokenize_and_align_labels_argType, batched=True) tokenized_dataset_actor = dataset_actor.map(tokenize_and_align_labels_agent, batched=True) # create multitask dataset dataset_dict = { "ArgType": tokenized_dataset_argType, "Actor": tokenized_dataset_actor, } # create multitask model multitask_model = MultitaskModel.create( model_name=model_name, model_type_dict={ "ArgType": transformers.AutoModelForTokenClassification, "Actor": transformers.AutoModelForTokenClassification, }, model_config_dict={ "ArgType": transformers.AutoConfig.from_pretrained(model_name, num_labels=len(id2label_argType)), "Actor": transformers.AutoConfig.from_pretrained(model_name, num_labels=len(id2label_agent)), }, ) # create data collator data_collator= MyDataCollatorForTokenClassification(tokenizer) # split dataset into training and evaluation (dev) dataset train_dataset = { task_name: dataset["train"] for task_name, dataset in dataset_dict.items() } eval_dataset = { task_name: dataset["validation"] for task_name, dataset in dataset_dict.items() } # set training parameter and train the model output_dir = pathprefix + 'models/multitask/roberta-large-fp-15000' batch_size = 4 # use parsed if provided if args.output_dir: output_dir = args.output_dir if args.batch_size: batch_size = args.batch_size train_args = transformers.TrainingArguments( output_dir, evaluation_strategy = "epoch", logging_steps=1592, learning_rate=1e-5, per_device_train_batch_size=batch_size, per_device_eval_batch_size=batch_size, num_train_epochs=10, weight_decay=0.01, warmup_steps=1000, save_steps=15926, save_total_limit = 10, logging_dir=pathprefix + 'logs', ) trainer = MultitaskTrainer( model=multitask_model, args=train_args, data_collator=data_collator, train_dataset=train_dataset, eval_dataset=eval_dataset, tokenizer=tokenizer, compute_metrics=eval_f1, ) trainer.train()
25,615
37.232836
144
py
imbalanced-learn
imbalanced-learn-master/conftest.py
# This file is here so that when running from the root folder # ./imblearn is added to sys.path by pytest. # See https://docs.pytest.org/en/latest/pythonpath.html for more details. # For example, this allows to build extensions in place and run pytest # doc/modules/clustering.rst and use imblearn from the local folder # rather than the one from site-packages. import os import pytest def pytest_runtest_setup(item): fname = item.fspath.strpath if ( fname.endswith(os.path.join("keras", "_generator.py")) or fname.endswith(os.path.join("tensorflow", "_generator.py")) or fname.endswith("miscellaneous.rst") ): try: import tensorflow # noqa except ImportError: pytest.skip("The tensorflow package is not installed.")
798
32.291667
73
py
imbalanced-learn
imbalanced-learn-master/examples/applications/porto_seguro_keras_under_sampling.py
""" ========================================================== Porto Seguro: balancing samples in mini-batches with Keras ========================================================== This example compares two strategies to train a neural-network on the Porto Seguro Kaggle data set [1]_. The data set is imbalanced and we show that balancing each mini-batch allows to improve performance and reduce the training time. References ---------- .. [1] https://www.kaggle.com/c/porto-seguro-safe-driver-prediction/data """ # Authors: Guillaume Lemaitre <g.lemaitre58@gmail.com> # License: MIT print(__doc__) ############################################################################### # Data loading ############################################################################### from collections import Counter import numpy as np import pandas as pd ############################################################################### # First, you should download the Porto Seguro data set from Kaggle. See the # link in the introduction. training_data = pd.read_csv("./input/train.csv") testing_data = pd.read_csv("./input/test.csv") y_train = training_data[["id", "target"]].set_index("id") X_train = training_data.drop(["target"], axis=1).set_index("id") X_test = testing_data.set_index("id") ############################################################################### # The data set is imbalanced and it will have an effect on the fitting. print(f"The data set is imbalanced: {Counter(y_train['target'])}") ############################################################################### # Define the pre-processing pipeline ############################################################################### from sklearn.compose import ColumnTransformer from sklearn.impute import SimpleImputer from sklearn.pipeline import make_pipeline from sklearn.preprocessing import FunctionTransformer, OneHotEncoder, StandardScaler def convert_float64(X): return X.astype(np.float64) ############################################################################### # We want to standard scale the numerical features while we want to one-hot # encode the categorical features. In this regard, we make use of the # :class:`~sklearn.compose.ColumnTransformer`. numerical_columns = [ name for name in X_train.columns if "_calc_" in name and "_bin" not in name ] numerical_pipeline = make_pipeline( FunctionTransformer(func=convert_float64, validate=False), StandardScaler() ) categorical_columns = [name for name in X_train.columns if "_cat" in name] categorical_pipeline = make_pipeline( SimpleImputer(missing_values=-1, strategy="most_frequent"), OneHotEncoder(categories="auto"), ) preprocessor = ColumnTransformer( [ ("numerical_preprocessing", numerical_pipeline, numerical_columns), ( "categorical_preprocessing", categorical_pipeline, categorical_columns, ), ], remainder="drop", ) # Create an environment variable to avoid using the GPU. This can be changed. import os os.environ["CUDA_VISIBLE_DEVICES"] = "-1" from tensorflow.keras.layers import Activation, BatchNormalization, Dense, Dropout ############################################################################### # Create a neural-network ############################################################################### from tensorflow.keras.models import Sequential def make_model(n_features): model = Sequential() model.add(Dense(200, input_shape=(n_features,), kernel_initializer="glorot_normal")) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(Dropout(0.5)) model.add(Dense(100, kernel_initializer="glorot_normal", use_bias=False)) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(Dropout(0.25)) model.add(Dense(50, kernel_initializer="glorot_normal", use_bias=False)) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(Dropout(0.15)) model.add(Dense(25, kernel_initializer="glorot_normal", use_bias=False)) model.add(BatchNormalization()) model.add(Activation("relu")) model.add(Dropout(0.1)) model.add(Dense(1, activation="sigmoid")) model.compile(loss="binary_crossentropy", optimizer="adam", metrics=["accuracy"]) return model ############################################################################### # We create a decorator to report the computation time import time from functools import wraps def timeit(f): @wraps(f) def wrapper(*args, **kwds): start_time = time.time() result = f(*args, **kwds) elapsed_time = time.time() - start_time print(f"Elapsed computation time: {elapsed_time:.3f} secs") return (elapsed_time, result) return wrapper ############################################################################### # The first model will be trained using the ``fit`` method and with imbalanced # mini-batches. import tensorflow from sklearn.metrics import roc_auc_score from sklearn.utils import parse_version tf_version = parse_version(tensorflow.__version__) @timeit def fit_predict_imbalanced_model(X_train, y_train, X_test, y_test): model = make_model(X_train.shape[1]) model.fit(X_train, y_train, epochs=2, verbose=1, batch_size=1000) if tf_version < parse_version("2.6"): # predict_proba was removed in tensorflow 2.6 predict_method = "predict_proba" else: predict_method = "predict" y_pred = getattr(model, predict_method)(X_test, batch_size=1000) return roc_auc_score(y_test, y_pred) ############################################################################### # In the contrary, we will use imbalanced-learn to create a generator of # mini-batches which will yield balanced mini-batches. from imblearn.keras import BalancedBatchGenerator @timeit def fit_predict_balanced_model(X_train, y_train, X_test, y_test): model = make_model(X_train.shape[1]) training_generator = BalancedBatchGenerator( X_train, y_train, batch_size=1000, random_state=42 ) model.fit(training_generator, epochs=5, verbose=1) y_pred = model.predict(X_test, batch_size=1000) return roc_auc_score(y_test, y_pred) ############################################################################### # Classification loop ############################################################################### ############################################################################### # We will perform a 10-fold cross-validation and train the neural-network with # the two different strategies previously presented. from sklearn.model_selection import StratifiedKFold skf = StratifiedKFold(n_splits=10) cv_results_imbalanced = [] cv_time_imbalanced = [] cv_results_balanced = [] cv_time_balanced = [] for train_idx, valid_idx in skf.split(X_train, y_train): X_local_train = preprocessor.fit_transform(X_train.iloc[train_idx]) y_local_train = y_train.iloc[train_idx].values.ravel() X_local_test = preprocessor.transform(X_train.iloc[valid_idx]) y_local_test = y_train.iloc[valid_idx].values.ravel() elapsed_time, roc_auc = fit_predict_imbalanced_model( X_local_train, y_local_train, X_local_test, y_local_test ) cv_time_imbalanced.append(elapsed_time) cv_results_imbalanced.append(roc_auc) elapsed_time, roc_auc = fit_predict_balanced_model( X_local_train, y_local_train, X_local_test, y_local_test ) cv_time_balanced.append(elapsed_time) cv_results_balanced.append(roc_auc) ############################################################################### # Plot of the results and computation time ############################################################################### df_results = pd.DataFrame( { "Balanced model": cv_results_balanced, "Imbalanced model": cv_results_imbalanced, } ) df_results = df_results.unstack().reset_index() df_time = pd.DataFrame( {"Balanced model": cv_time_balanced, "Imbalanced model": cv_time_imbalanced} ) df_time = df_time.unstack().reset_index() import matplotlib.pyplot as plt import seaborn as sns plt.figure() sns.boxplot(y="level_0", x=0, data=df_time) sns.despine(top=True, right=True, left=True) plt.xlabel("time [s]") plt.ylabel("") plt.title("Computation time difference using a random under-sampling") plt.figure() sns.boxplot(y="level_0", x=0, data=df_results, whis=10.0) sns.despine(top=True, right=True, left=True) ax = plt.gca() ax.xaxis.set_major_formatter(plt.FuncFormatter(lambda x, pos: "%i%%" % (100 * x))) plt.xlabel("ROC-AUC") plt.ylabel("") plt.title("Difference in terms of ROC-AUC using a random under-sampling")
8,747
32.776062
88
py
imbalanced-learn
imbalanced-learn-master/imblearn/_min_dependencies.py
"""All minimum dependencies for imbalanced-learn.""" import argparse NUMPY_MIN_VERSION = "1.17.3" SCIPY_MIN_VERSION = "1.5.0" PANDAS_MIN_VERSION = "1.0.5" SKLEARN_MIN_VERSION = "1.0.2" TENSORFLOW_MIN_VERSION = "2.4.3" KERAS_MIN_VERSION = "2.4.3" JOBLIB_MIN_VERSION = "1.1.1" THREADPOOLCTL_MIN_VERSION = "2.0.0" PYTEST_MIN_VERSION = "5.0.1" # 'build' and 'install' is included to have structured metadata for CI. # It will NOT be included in setup's extras_require # The values are (version_spec, comma separated tags) dependent_packages = { "numpy": (NUMPY_MIN_VERSION, "install"), "scipy": (SCIPY_MIN_VERSION, "install"), "scikit-learn": (SKLEARN_MIN_VERSION, "install"), "joblib": (JOBLIB_MIN_VERSION, "install"), "threadpoolctl": (THREADPOOLCTL_MIN_VERSION, "install"), "pandas": (PANDAS_MIN_VERSION, "optional, docs, examples, tests"), "tensorflow": (TENSORFLOW_MIN_VERSION, "optional, docs, examples, tests"), "keras": (KERAS_MIN_VERSION, "optional, docs, examples, tests"), "matplotlib": ("3.1.2", "docs, examples"), "seaborn": ("0.9.0", "docs, examples"), "memory_profiler": ("0.57.0", "docs"), "pytest": (PYTEST_MIN_VERSION, "tests"), "pytest-cov": ("2.9.0", "tests"), "flake8": ("3.8.2", "tests"), "black": ("23.3.0", "tests"), "mypy": ("1.3.0", "tests"), "sphinx": ("6.0.0", "docs"), "sphinx-gallery": ("0.13.0", "docs"), "sphinx-copybutton": ("0.5.2", "docs"), "numpydoc": ("1.5.0", "docs"), "sphinxcontrib-bibtex": ("2.4.1", "docs"), "pydata-sphinx-theme": ("0.13.3", "docs"), } # create inverse mapping for setuptools tag_to_packages: dict = { extra: [] for extra in ["install", "optional", "docs", "examples", "tests"] } for package, (min_version, extras) in dependent_packages.items(): for extra in extras.split(", "): tag_to_packages[extra].append("{}>={}".format(package, min_version)) # Used by CI to get the min dependencies if __name__ == "__main__": parser = argparse.ArgumentParser(description="Get min dependencies for a package") parser.add_argument("package", choices=dependent_packages) args = parser.parse_args() min_version = dependent_packages[args.package][0] print(min_version)
2,240
36.35
86
py
imbalanced-learn
imbalanced-learn-master/imblearn/__init__.py
"""Toolbox for imbalanced dataset in machine learning. ``imbalanced-learn`` is a set of python methods to deal with imbalanced datset in machine learning and pattern recognition. Subpackages ----------- combine Module which provides methods based on over-sampling and under-sampling. ensemble Module which provides methods generating an ensemble of under-sampled subsets. exceptions Module including custom warnings and error clases used across imbalanced-learn. keras Module which provides custom generator, layers for deep learning using keras. metrics Module which provides metrics to quantified the classification performance with imbalanced dataset. over_sampling Module which provides methods to over-sample a dataset. tensorflow Module which provides custom generator, layers for deep learning using tensorflow. under-sampling Module which provides methods to under-sample a dataset. utils Module including various utilities. pipeline Module which allowing to create pipeline with scikit-learn estimators. """ import importlib import sys import types try: # This variable is injected in the __builtins__ by the build # process. It is used to enable importing subpackages of sklearn when # the binaries are not built # mypy error: Cannot determine type of '__SKLEARN_SETUP__' __IMBLEARN_SETUP__ # type: ignore except NameError: __IMBLEARN_SETUP__ = False if __IMBLEARN_SETUP__: sys.stderr.write("Partial import of imblearn during the build process.\n") # We are not importing the rest of scikit-learn during the build # process, as it may not be compiled yet else: from . import ( combine, ensemble, exceptions, metrics, over_sampling, pipeline, tensorflow, under_sampling, utils, ) from ._version import __version__ from .base import FunctionSampler from .utils._show_versions import show_versions # noqa: F401 # FIXME: When we get Python 3.7 as minimal version, we will need to switch to # the following solution: # https://snarky.ca/lazy-importing-in-python-3-7/ class LazyLoader(types.ModuleType): """Lazily import a module, mainly to avoid pulling in large dependencies. Adapted from TensorFlow: https://github.com/tensorflow/tensorflow/blob/master/tensorflow/ python/util/lazy_loader.py """ def __init__(self, local_name, parent_module_globals, name, warning=None): self._local_name = local_name self._parent_module_globals = parent_module_globals self._warning = warning super(LazyLoader, self).__init__(name) def _load(self): """Load the module and insert it into the parent's globals.""" # Import the target module and insert it into the parent's namespace module = importlib.import_module(self.__name__) self._parent_module_globals[self._local_name] = module # Update this object's dict so that if someone keeps a reference to the # LazyLoader, lookups are efficient (__getattr__ is only called on # lookups that fail). self.__dict__.update(module.__dict__) return module def __getattr__(self, item): module = self._load() return getattr(module, item) def __dir__(self): module = self._load() return dir(module) # delay the import of keras since we are going to import either tensorflow # or keras keras = LazyLoader("keras", globals(), "imblearn.keras") __all__ = [ "combine", "ensemble", "exceptions", "keras", "metrics", "over_sampling", "tensorflow", "under_sampling", "utils", "pipeline", "FunctionSampler", "__version__", ]
3,963
30.967742
83
py
imbalanced-learn
imbalanced-learn-master/imblearn/keras/_generator.py
"""Implement generators for ``keras`` which will balance the data.""" # This is a trick to avoid an error during tests collection with pytest. We # avoid the error when importing the package raise the error at the moment of # creating the instance. # This is a trick to avoid an error during tests collection with pytest. We # avoid the error when importing the package raise the error at the moment of # creating the instance. def import_keras(): """Try to import keras from keras and tensorflow. This is possible to import the sequence from keras or tensorflow. """ def import_from_keras(): try: import keras # noqa if hasattr(keras.utils, "Sequence"): return (keras.utils.Sequence,), True else: return (keras.utils.data_utils.Sequence,), True except ImportError: return tuple(), False def import_from_tensforflow(): try: from tensorflow import keras if hasattr(keras.utils, "Sequence"): return (keras.utils.Sequence,), True else: return (keras.utils.data_utils.Sequence,), True except ImportError: return tuple(), False ParentClassKeras, has_keras_k = import_from_keras() ParentClassTensorflow, has_keras_tf = import_from_tensforflow() has_keras = has_keras_k or has_keras_tf if has_keras: if has_keras_k: ParentClass = ParentClassKeras else: ParentClass = ParentClassTensorflow else: ParentClass = (object,) return ParentClass, has_keras ParentClass, HAS_KERAS = import_keras() from scipy.sparse import issparse # noqa from sklearn.base import clone # noqa from sklearn.utils import _safe_indexing # noqa from sklearn.utils import check_random_state # noqa from ..tensorflow import balanced_batch_generator as tf_bbg # noqa from ..under_sampling import RandomUnderSampler # noqa from ..utils import Substitution # noqa from ..utils._docstring import _random_state_docstring # noqa class BalancedBatchGenerator(*ParentClass): # type: ignore """Create balanced batches when training a keras model. Create a keras ``Sequence`` which is given to ``fit``. The sampler defines the sampling strategy used to balance the dataset ahead of creating the batch. The sampler should have an attribute ``sample_indices_``. .. versionadded:: 0.4 Parameters ---------- X : ndarray of shape (n_samples, n_features) Original imbalanced dataset. y : ndarray of shape (n_samples,) or (n_samples, n_classes) Associated targets. sample_weight : ndarray of shape (n_samples,) Sample weight. sampler : sampler object, default=None A sampler instance which has an attribute ``sample_indices_``. By default, the sampler used is a :class:`~imblearn.under_sampling.RandomUnderSampler`. batch_size : int, default=32 Number of samples per gradient update. keep_sparse : bool, default=False Either or not to conserve or not the sparsity of the input (i.e. ``X``, ``y``, ``sample_weight``). By default, the returned batches will be dense. random_state : int, RandomState instance or None, default=None Control the randomization of the algorithm: - If int, ``random_state`` is the seed used by the random number generator; - If ``RandomState`` instance, random_state is the random number generator; - If ``None``, the random number generator is the ``RandomState`` instance used by ``np.random``. Attributes ---------- sampler_ : sampler object The sampler used to balance the dataset. indices_ : ndarray of shape (n_samples, n_features) The indices of the samples selected during sampling. Examples -------- >>> from sklearn.datasets import load_iris >>> iris = load_iris() >>> from imblearn.datasets import make_imbalance >>> class_dict = dict() >>> class_dict[0] = 30; class_dict[1] = 50; class_dict[2] = 40 >>> X, y = make_imbalance(iris.data, iris.target, sampling_strategy=class_dict) >>> import tensorflow >>> y = tensorflow.keras.utils.to_categorical(y, 3) >>> model = tensorflow.keras.models.Sequential() >>> model.add( ... tensorflow.keras.layers.Dense( ... y.shape[1], input_dim=X.shape[1], activation='softmax' ... ) ... ) >>> model.compile(optimizer='sgd', loss='categorical_crossentropy', ... metrics=['accuracy']) >>> from imblearn.keras import BalancedBatchGenerator >>> from imblearn.under_sampling import NearMiss >>> training_generator = BalancedBatchGenerator( ... X, y, sampler=NearMiss(), batch_size=10, random_state=42) >>> callback_history = model.fit(training_generator, epochs=10, verbose=0) """ # flag for keras sequence duck-typing use_sequence_api = True def __init__( self, X, y, *, sample_weight=None, sampler=None, batch_size=32, keep_sparse=False, random_state=None, ): if not HAS_KERAS: raise ImportError("'No module named 'keras'") self.X = X self.y = y self.sample_weight = sample_weight self.sampler = sampler self.batch_size = batch_size self.keep_sparse = keep_sparse self.random_state = random_state self._sample() def _sample(self): random_state = check_random_state(self.random_state) if self.sampler is None: self.sampler_ = RandomUnderSampler(random_state=random_state) else: self.sampler_ = clone(self.sampler) self.sampler_.fit_resample(self.X, self.y) if not hasattr(self.sampler_, "sample_indices_"): raise ValueError("'sampler' needs to have an attribute 'sample_indices_'.") self.indices_ = self.sampler_.sample_indices_ # shuffle the indices since the sampler are packing them by class random_state.shuffle(self.indices_) def __len__(self): return int(self.indices_.size // self.batch_size) def __getitem__(self, index): X_resampled = _safe_indexing( self.X, self.indices_[index * self.batch_size : (index + 1) * self.batch_size], ) y_resampled = _safe_indexing( self.y, self.indices_[index * self.batch_size : (index + 1) * self.batch_size], ) if issparse(X_resampled) and not self.keep_sparse: X_resampled = X_resampled.toarray() if self.sample_weight is not None: sample_weight_resampled = _safe_indexing( self.sample_weight, self.indices_[index * self.batch_size : (index + 1) * self.batch_size], ) if self.sample_weight is None: return X_resampled, y_resampled else: return X_resampled, y_resampled, sample_weight_resampled @Substitution(random_state=_random_state_docstring) def balanced_batch_generator( X, y, *, sample_weight=None, sampler=None, batch_size=32, keep_sparse=False, random_state=None, ): """Create a balanced batch generator to train keras model. Returns a generator --- as well as the number of step per epoch --- which is given to ``fit``. The sampler defines the sampling strategy used to balance the dataset ahead of creating the batch. The sampler should have an attribute ``sample_indices_``. Parameters ---------- X : ndarray of shape (n_samples, n_features) Original imbalanced dataset. y : ndarray of shape (n_samples,) or (n_samples, n_classes) Associated targets. sample_weight : ndarray of shape (n_samples,), default=None Sample weight. sampler : sampler object, default=None A sampler instance which has an attribute ``sample_indices_``. By default, the sampler used is a :class:`~imblearn.under_sampling.RandomUnderSampler`. batch_size : int, default=32 Number of samples per gradient update. keep_sparse : bool, default=False Either or not to conserve or not the sparsity of the input (i.e. ``X``, ``y``, ``sample_weight``). By default, the returned batches will be dense. {random_state} Returns ------- generator : generator of tuple Generate batch of data. The tuple generated are either (X_batch, y_batch) or (X_batch, y_batch, sampler_weight_batch). steps_per_epoch : int The number of samples per epoch. Required by ``fit_generator`` in keras. Examples -------- >>> from sklearn.datasets import load_iris >>> X, y = load_iris(return_X_y=True) >>> from imblearn.datasets import make_imbalance >>> class_dict = dict() >>> class_dict[0] = 30; class_dict[1] = 50; class_dict[2] = 40 >>> from imblearn.datasets import make_imbalance >>> X, y = make_imbalance(X, y, sampling_strategy=class_dict) >>> import tensorflow >>> y = tensorflow.keras.utils.to_categorical(y, 3) >>> model = tensorflow.keras.models.Sequential() >>> model.add( ... tensorflow.keras.layers.Dense( ... y.shape[1], input_dim=X.shape[1], activation='softmax' ... ) ... ) >>> model.compile(optimizer='sgd', loss='categorical_crossentropy', ... metrics=['accuracy']) >>> from imblearn.keras import balanced_batch_generator >>> from imblearn.under_sampling import NearMiss >>> training_generator, steps_per_epoch = balanced_batch_generator( ... X, y, sampler=NearMiss(), batch_size=10, random_state=42) >>> callback_history = model.fit(training_generator, ... steps_per_epoch=steps_per_epoch, ... epochs=10, verbose=0) """ return tf_bbg( X=X, y=y, sample_weight=sample_weight, sampler=sampler, batch_size=batch_size, keep_sparse=keep_sparse, random_state=random_state, )
10,276
33.719595
87
py
imbalanced-learn
imbalanced-learn-master/imblearn/keras/__init__.py
"""The :mod:`imblearn.keras` provides utilities to deal with imbalanced dataset in keras.""" from ._generator import BalancedBatchGenerator, balanced_batch_generator __all__ = ["BalancedBatchGenerator", "balanced_batch_generator"]
233
32.428571
79
py
imbalanced-learn
imbalanced-learn-master/imblearn/keras/tests/test_generator.py
import numpy as np import pytest from scipy import sparse from sklearn.cluster import KMeans from sklearn.datasets import load_iris from sklearn.preprocessing import LabelBinarizer keras = pytest.importorskip("keras") from keras.layers import Dense # noqa: E402 from keras.models import Sequential # noqa: E402 from imblearn.datasets import make_imbalance # noqa: E402 from imblearn.keras import ( BalancedBatchGenerator, # noqa: E402 balanced_batch_generator, # noqa: E402 ) from imblearn.over_sampling import RandomOverSampler # noqa: E402 from imblearn.under_sampling import ( ClusterCentroids, # noqa: E402 NearMiss, # noqa: E402 ) 3 @pytest.fixture def data(): iris = load_iris() X, y = make_imbalance( iris.data, iris.target, sampling_strategy={0: 30, 1: 50, 2: 40} ) y = LabelBinarizer().fit_transform(y) return X, y def _build_keras_model(n_classes, n_features): model = Sequential() model.add(Dense(n_classes, input_dim=n_features, activation="softmax")) model.compile( optimizer="sgd", loss="categorical_crossentropy", metrics=["accuracy"] ) return model def test_balanced_batch_generator_class_no_return_indices(data): with pytest.raises(ValueError, match="needs to have an attribute"): BalancedBatchGenerator( *data, sampler=ClusterCentroids(estimator=KMeans(n_init=1)), batch_size=10 ) @pytest.mark.filterwarnings("ignore:`wait_time` is not used") # keras 2.2.4 @pytest.mark.parametrize( "sampler, sample_weight", [ (None, None), (RandomOverSampler(), None), (NearMiss(), None), (None, np.random.uniform(size=120)), ], ) def test_balanced_batch_generator_class(data, sampler, sample_weight): X, y = data model = _build_keras_model(y.shape[1], X.shape[1]) training_generator = BalancedBatchGenerator( X, y, sample_weight=sample_weight, sampler=sampler, batch_size=10, random_state=42, ) model.fit_generator(generator=training_generator, epochs=10) @pytest.mark.parametrize("keep_sparse", [True, False]) def test_balanced_batch_generator_class_sparse(data, keep_sparse): X, y = data training_generator = BalancedBatchGenerator( sparse.csr_matrix(X), y, batch_size=10, keep_sparse=keep_sparse, random_state=42, ) for idx in range(len(training_generator)): X_batch, _ = training_generator.__getitem__(idx) if keep_sparse: assert sparse.issparse(X_batch) else: assert not sparse.issparse(X_batch) def test_balanced_batch_generator_function_no_return_indices(data): with pytest.raises(ValueError, match="needs to have an attribute"): balanced_batch_generator( *data, sampler=ClusterCentroids(estimator=KMeans(n_init=10)), batch_size=10, random_state=42, ) @pytest.mark.filterwarnings("ignore:`wait_time` is not used") # keras 2.2.4 @pytest.mark.parametrize( "sampler, sample_weight", [ (None, None), (RandomOverSampler(), None), (NearMiss(), None), (None, np.random.uniform(size=120)), ], ) def test_balanced_batch_generator_function(data, sampler, sample_weight): X, y = data model = _build_keras_model(y.shape[1], X.shape[1]) training_generator, steps_per_epoch = balanced_batch_generator( X, y, sample_weight=sample_weight, sampler=sampler, batch_size=10, random_state=42, ) model.fit_generator( generator=training_generator, steps_per_epoch=steps_per_epoch, epochs=10, ) @pytest.mark.parametrize("keep_sparse", [True, False]) def test_balanced_batch_generator_function_sparse(data, keep_sparse): X, y = data training_generator, steps_per_epoch = balanced_batch_generator( sparse.csr_matrix(X), y, keep_sparse=keep_sparse, batch_size=10, random_state=42, ) for _ in range(steps_per_epoch): X_batch, _ = next(training_generator) if keep_sparse: assert sparse.issparse(X_batch) else: assert not sparse.issparse(X_batch)
4,289
27.986486
86
py
imbalanced-learn
imbalanced-learn-master/imblearn/utils/_show_versions.py
""" Utility method which prints system info to help with debugging, and filing issues on GitHub. Adapted from :func:`sklearn.show_versions`, which was adapted from :func:`pandas.show_versions` """ # Author: Alexander L. Hayes <hayesall@iu.edu> # License: MIT from .. import __version__ def _get_deps_info(): """Overview of the installed version of main dependencies Returns ------- deps_info: dict version information on relevant Python libraries """ deps = [ "imbalanced-learn", "pip", "setuptools", "numpy", "scipy", "scikit-learn", "Cython", "pandas", "keras", "tensorflow", "joblib", ] deps_info = { "imbalanced-learn": __version__, } from importlib.metadata import PackageNotFoundError, version for modname in deps: try: deps_info[modname] = version(modname) except PackageNotFoundError: deps_info[modname] = None return deps_info def show_versions(github=False): """Print debugging information. .. versionadded:: 0.5 Parameters ---------- github : bool, If true, wrap system info with GitHub markup. """ from sklearn.utils._show_versions import _get_sys_info _sys_info = _get_sys_info() _deps_info = _get_deps_info() _github_markup = ( "<details>" "<summary>System, Dependency Information</summary>\n\n" "**System Information**\n\n" "{0}\n" "**Python Dependencies**\n\n" "{1}\n" "</details>" ) if github: _sys_markup = "" _deps_markup = "" for k, stat in _sys_info.items(): _sys_markup += f"* {k:<10}: `{stat}`\n" for k, stat in _deps_info.items(): _deps_markup += f"* {k:<10}: `{stat}`\n" print(_github_markup.format(_sys_markup, _deps_markup)) else: print("\nSystem:") for k, stat in _sys_info.items(): print(f"{k:>11}: {stat}") print("\nPython dependencies:") for k, stat in _deps_info.items(): print(f"{k:>11}: {stat}")
2,176
22.408602
64
py
imbalanced-learn
imbalanced-learn-master/imblearn/utils/tests/test_show_versions.py
"""Test for the show_versions helper. Based on the sklearn tests.""" # Author: Alexander L. Hayes <hayesall@iu.edu> # License: MIT from imblearn.utils._show_versions import _get_deps_info, show_versions def test_get_deps_info(): _deps_info = _get_deps_info() assert "pip" in _deps_info assert "setuptools" in _deps_info assert "imbalanced-learn" in _deps_info assert "scikit-learn" in _deps_info assert "numpy" in _deps_info assert "scipy" in _deps_info assert "Cython" in _deps_info assert "pandas" in _deps_info assert "joblib" in _deps_info def test_show_versions_default(capsys): show_versions() out, err = capsys.readouterr() assert "python" in out assert "executable" in out assert "machine" in out assert "pip" in out assert "setuptools" in out assert "imbalanced-learn" in out assert "scikit-learn" in out assert "numpy" in out assert "scipy" in out assert "Cython" in out assert "pandas" in out assert "keras" in out assert "tensorflow" in out assert "joblib" in out def test_show_versions_github(capsys): show_versions(github=True) out, err = capsys.readouterr() assert "<details><summary>System, Dependency Information</summary>" in out assert "**System Information**" in out assert "* python" in out assert "* executable" in out assert "* machine" in out assert "**Python Dependencies**" in out assert "* pip" in out assert "* setuptools" in out assert "* imbalanced-learn" in out assert "* scikit-learn" in out assert "* numpy" in out assert "* scipy" in out assert "* Cython" in out assert "* pandas" in out assert "* keras" in out assert "* tensorflow" in out assert "* joblib" in out assert "</details>" in out
1,818
28.819672
78
py
ML-Doctor
ML-Doctor-main/demo.py
import os import sys import torch import argparse import torch.nn as nn import torchvision.models as models from doctor.meminf import * from doctor.modinv import * from doctor.attrinf import * from doctor.modsteal import * from demoloader.train import * from demoloader.DCGAN import * from utils.define_models import * from demoloader.dataloader import * def train_model(PATH, device, train_set, test_set, model, use_DP, noise, norm, delta): train_loader = torch.utils.data.DataLoader( train_set, batch_size=64, shuffle=True, num_workers=2) test_loader = torch.utils.data.DataLoader( test_set, batch_size=64, shuffle=True, num_workers=2) model = model_training(train_loader, test_loader, model, device, use_DP, noise, norm, delta) acc_train = 0 acc_test = 0 for i in range(100): print("<======================= Epoch " + str(i+1) + " =======================>") print("target training") acc_train = model.train() print("target testing") acc_test = model.test() overfitting = round(acc_train - acc_test, 6) print('The overfitting rate is %s' % overfitting) FILE_PATH = PATH + "_target.pth" model.saveModel(FILE_PATH) print("Saved target model!!!") print("Finished training!!!") return acc_train, acc_test, overfitting def train_DCGAN(PATH, device, train_set, name): train_loader = torch.utils.data.DataLoader( train_set, batch_size=128, shuffle=True, num_workers=2) if name.lower() == 'fmnist': D = FashionDiscriminator().eval() G = FashionGenerator().eval() else: D = Discriminator(ngpu=1).eval() G = Generator(ngpu=1).eval() print("Starting Training DCGAN...") GAN = GAN_training(train_loader, D, G, device) for i in range(200): print("<======================= Epoch " + str(i+1) + " =======================>") GAN.train() GAN.saveModel(PATH + "_discriminator.pth", PATH + "_generator.pth") def test_meminf(PATH, device, num_classes, target_train, target_test, shadow_train, shadow_test, target_model, shadow_model, train_shadow, use_DP, noise, norm, delta, mode): batch_size = 64 if train_shadow: shadow_trainloader = torch.utils.data.DataLoader( shadow_train, batch_size=batch_size, shuffle=True, num_workers=2) shadow_testloader = torch.utils.data.DataLoader( shadow_test, batch_size=batch_size, shuffle=True, num_workers=2) loss = nn.CrossEntropyLoss() optimizer = optim.SGD(shadow_model.parameters(), lr=1e-2, momentum=0.9, weight_decay=5e-4) train_shadow_model(PATH, device, shadow_model, shadow_trainloader, shadow_testloader, use_DP, noise, norm, loss, optimizer, delta) if mode == 0 or mode == 3: attack_trainloader, attack_testloader = get_attack_dataset_with_shadow( target_train, target_test, shadow_train, shadow_test, batch_size) else: attack_trainloader, attack_testloader = get_attack_dataset_without_shadow(target_train, target_test, batch_size) #for white box if mode == 2 or mode == 3: gradient_size = get_gradient_size(target_model) total = gradient_size[0][0] // 2 * gradient_size[0][1] // 2 if mode == 0: attack_model = ShadowAttackModel(num_classes) attack_mode0(PATH + "_target.pth", PATH + "_shadow.pth", PATH, device, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, 1, num_classes) elif mode == 1: attack_model = PartialAttackModel(num_classes) attack_mode1(PATH + "_target.pth", PATH, device, attack_trainloader, attack_testloader, target_model, attack_model, 1, num_classes) elif mode == 2: attack_model = WhiteBoxAttackModel(num_classes, total) attack_mode2(PATH + "_target.pth", PATH, device, attack_trainloader, attack_testloader, target_model, attack_model, 1, num_classes) elif mode == 3: attack_model = WhiteBoxAttackModel(num_classes, total) attack_mode3(PATH + "_target.pth", PATH + "_shadow.pth", PATH, device, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, 1, num_classes) else: raise Exception("Wrong mode") # attack_mode0(PATH + "_target.pth", PATH + "_shadow.pth", PATH, device, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, 1, num_classes) # attack_mode1(PATH + "_target.pth", PATH, device, attack_trainloader, attack_testloader, target_model, attack_model, 1, num_classes) # attack_mode2(PATH + "_target.pth", PATH, device, attack_trainloader, attack_testloader, target_model, attack_model, 1, num_classes) def test_modinv(PATH, device, num_classes, target_train, target_model, name): size = (1,) + tuple(target_train[0][0].shape) target_model, evaluation_model = load_data(PATH + "_target.pth", PATH + "_eval.pth", target_model, models.resnet18(num_classes=num_classes)) # CCS 15 modinv_ccs = ccs_inversion(target_model, size, num_classes, 1, 3000, 100, 0.001, 0.003, device) train_loader = torch.utils.data.DataLoader(target_train, batch_size=1, shuffle=False) ccs_result = modinv_ccs.reverse_mse(train_loader) # Secret Revealer if name.lower() == 'fmnist': D = FashionDiscriminator(ngpu=1).eval() G = FashionGenerator(ngpu=1).eval() else: D = Discriminator(ngpu=1).eval() G = Generator(ngpu=1).eval() PATH_D = PATH + "_discriminator.pth" PATH_G = PATH + "_generator.pth" D, G, iden = prepare_GAN(name, D, G, PATH_D, PATH_G) modinv_revealer = revealer_inversion(G, D, target_model, evaluation_model, iden, device) def test_attrinf(PATH, device, num_classes, target_train, target_test, target_model): attack_length = int(0.5 * len(target_train)) rest = len(target_train) - attack_length attack_train, _ = torch.utils.data.random_split(target_train, [attack_length, rest]) attack_test = target_test attack_trainloader = torch.utils.data.DataLoader( attack_train, batch_size=64, shuffle=True, num_workers=2) attack_testloader = torch.utils.data.DataLoader( attack_test, batch_size=64, shuffle=True, num_workers=2) image_size = [1] + list(target_train[0][0].shape) train_attack_model( PATH + "_target.pth", PATH, num_classes[1], device, target_model, attack_trainloader, attack_testloader, image_size) def test_modsteal(PATH, device, train_set, test_set, target_model, attack_model): train_loader = torch.utils.data.DataLoader( train_set, batch_size=64, shuffle=True, num_workers=2) test_loader = torch.utils.data.DataLoader( test_set, batch_size=64, shuffle=True, num_workers=2) loss = nn.MSELoss() optimizer = optim.SGD(attack_model.parameters(), lr=0.01, momentum=0.9) attacking = train_steal_model( train_loader, test_loader, target_model, attack_model, PATH + "_target.pth", PATH + "_modsteal.pth", device, 64, loss, optimizer) for i in range(100): print("[Epoch %d/%d] attack training"%((i+1), 100)) attacking.train_with_same_distribution() print("Finished training!!!") attacking.saveModel() acc_test, agreement_test = attacking.test() print("Saved Target Model!!!\nstolen test acc = %.3f, stolen test agreement = %.3f\n"%(acc_test, agreement_test)) def str_to_bool(string): if isinstance(string, bool): return string if string.lower() in ('yes', 'true', 't', 'y', '1'): return True elif string.lower() in ('no', 'false', 'f', 'n', '0'): return False else: raise argparse.ArgumentTypeError('Boolean value expected.') def main(): parser = argparse.ArgumentParser() parser.add_argument('-g', '--gpu', type=str, default="0") parser.add_argument('-a', '--attributes', type=str, default="race", help="For attrinf, two attributes should be in format x_y e.g. race_gender") parser.add_argument('-dn', '--dataset_name', type=str, default="UTKFace") parser.add_argument('-at', '--attack_type', type=int, default=0) parser.add_argument('-tm', '--train_model', action='store_true') parser.add_argument('-ts', '--train_shadow', action='store_true') parser.add_argument('-ud', '--use_DP', action='store_true',) parser.add_argument('-ne', '--noise', type=float, default=1.3) parser.add_argument('-nm', '--norm', type=float, default=1.5) parser.add_argument('-d', '--delta', type=float, default=1e-5) parser.add_argument('-m', '--mode', type=int, default=0) args = parser.parse_args() os.environ["CUDA_VISIBLE_DEVICES"] = args.gpu device = torch.device("cuda:0") dataset_name = args.dataset_name attr = args.attributes if "_" in attr: attr = attr.split("_") root = "../data" use_DP = args.use_DP noise = args.noise norm = args.norm delta = args.delta mode = args.mode train_shadow = args.train_shadow TARGET_ROOT = "./demoloader/trained_model/" if not os.path.exists(TARGET_ROOT): print(f"Create directory named {TARGET_ROOT}") os.makedirs(TARGET_ROOT) TARGET_PATH = TARGET_ROOT + dataset_name num_classes, target_train, target_test, shadow_train, shadow_test, target_model, shadow_model = prepare_dataset(dataset_name, attr, root) if args.train_model: train_model(TARGET_PATH, device, target_train, target_test, target_model, use_DP, noise, norm, delta) # membership inference if args.attack_type == 0: test_meminf(TARGET_PATH, device, num_classes, target_train, target_test, shadow_train, shadow_test, target_model, shadow_model, train_shadow, use_DP, noise, norm, delta, mode) # model inversion elif args.attack_type == 1: train_DCGAN(TARGET_PATH, device, shadow_test + shadow_train, dataset_name) test_modinv(TARGET_PATH, device, num_classes, target_train, target_model, dataset_name) # attribut inference elif args.attack_type == 2: test_attrinf(TARGET_PATH, device, num_classes, target_train, target_test, target_model) # model stealing elif args.attack_type == 3: test_modsteal(TARGET_PATH, device, shadow_train+shadow_test, target_test, target_model, shadow_model) else: sys.exit("we have not supported this mode yet! 0c0") # target_model = models.resnet18(num_classes=num_classes) # train_model(TARGET_PATH, device, target_train + shadow_train, target_test + shadow_test, target_model) if __name__ == "__main__": main()
10,614
41.290837
183
py
ML-Doctor
ML-Doctor-main/demoloader/dataloader.py
import os import torch import pandas import torchvision torch.manual_seed(0) import torch.nn as nn import PIL.Image as Image import torchvision.transforms as transforms from functools import partial from typing import Any, Callable, List, Optional, Union, Tuple class CNN(nn.Module): def __init__(self, input_channel=3, num_classes=10): super(CNN, self).__init__() self.features = nn.Sequential( nn.Conv2d(input_channel, 32, kernel_size=3), nn.ReLU(), nn.MaxPool2d(kernel_size=2), nn.Conv2d(32, 64, kernel_size=3), nn.ReLU(), nn.MaxPool2d(kernel_size=2), nn.Conv2d(64, 128, kernel_size=3), nn.ReLU(), nn.MaxPool2d(kernel_size=2), ) self.classifier = nn.Sequential( nn.Linear(128*6*6, 512), nn.ReLU(), nn.Linear(512, num_classes), ) def forward(self, x): x = self.features(x) x = torch.flatten(x, 1) x = self.classifier(x) return x class UTKFaceDataset(torch.utils.data.Dataset): def __init__(self, root, attr: Union[List[str], str] = "gender", transform=None, target_transform=None)-> None: self.root = root self.transform = transform self.target_transform = target_transform self.processed_path = os.path.join(self.root, 'UTKFace/processed/') self.files = os.listdir(self.processed_path) if isinstance(attr, list): self.attr = attr else: self.attr = [attr] self.lines = [] for txt_file in self.files: txt_file_path = os.path.join(self.processed_path, txt_file) with open(txt_file_path, 'r') as f: assert f is not None for i in f: image_name = i.split('jpg ')[0] attrs = image_name.split('_') if len(attrs) < 4 or int(attrs[2]) >= 4 or '' in attrs: continue self.lines.append(image_name+'jpg') def __len__(self): return len(self.lines) def __getitem__(self, index:int)-> Tuple[Any, Any]: attrs = self.lines[index].split('_') age = int(attrs[0]) gender = int(attrs[1]) race = int(attrs[2]) image_path = os.path.join(self.root, 'UTKFace/raw/', self.lines[index]+'.chip.jpg').rstrip() image = Image.open(image_path).convert('RGB') target: Any = [] for t in self.attr: if t == "age": target.append(age) elif t == "gender": target.append(gender) elif t == "race": target.append(race) else: raise ValueError("Target type \"{}\" is not recognized.".format(t)) if self.transform: image = self.transform(image) if target: target = tuple(target) if len(target) > 1 else target[0] if self.target_transform is not None: target = self.target_transform(target) else: target = None return image, target class CelebA(torch.utils.data.Dataset): base_folder = "celeba" def __init__( self, root: str, attr_list: str, target_type: Union[List[str], str] = "attr", transform: Optional[Callable] = None, target_transform: Optional[Callable] = None, ) -> None: if isinstance(target_type, list): self.target_type = target_type else: self.target_type = [target_type] self.root = root self.transform = transform self.target_transform =target_transform self.attr_list = attr_list fn = partial(os.path.join, self.root, self.base_folder) splits = pandas.read_csv(fn("list_eval_partition.txt"), delim_whitespace=True, header=None, index_col=0) attr = pandas.read_csv(fn("list_attr_celeba.txt"), delim_whitespace=True, header=1) mask = slice(None) self.filename = splits[mask].index.values self.attr = torch.as_tensor(attr[mask].values) self.attr = (self.attr + 1) // 2 # map from {-1, 1} to {0, 1} self.attr_names = list(attr.columns) def __getitem__(self, index: int) -> Tuple[Any, Any]: X = Image.open(os.path.join(self.root, self.base_folder, "img_celeba", self.filename[index])) target: Any = [] for t, nums in zip(self.target_type, self.attr_list): if t == "attr": final_attr = 0 for i in range(len(nums)): final_attr += 2 ** i * self.attr[index][nums[i]] target.append(final_attr) else: # TODO: refactor with utils.verify_str_arg raise ValueError("Target type \"{}\" is not recognized.".format(t)) if self.transform is not None: X = self.transform(X) if target: target = tuple(target) if len(target) > 1 else target[0] if self.target_transform is not None: target = self.target_transform(target) else: target = None return X, target def __len__(self) -> int: return len(self.attr) def extra_repr(self) -> str: lines = ["Target type: {target_type}", "Split: {split}"] return '\n'.join(lines).format(**self.__dict__) def prepare_dataset(dataset, attr, root): num_classes, dataset, target_model, shadow_model = get_model_dataset(dataset, attr=attr, root=root) length = len(dataset) each_length = length//4 target_train, target_test, shadow_train, shadow_test, _ = torch.utils.data.random_split(dataset, [each_length, each_length, each_length, each_length, len(dataset)-(each_length*4)]) return num_classes, target_train, target_test, shadow_train, shadow_test, target_model, shadow_model def get_model_dataset(dataset_name, attr, root): if dataset_name.lower() == "utkface": if isinstance(attr, list): num_classes = [] for a in attr: if a == "age": num_classes.append(117) elif a == "gender": num_classes.append(2) elif a == "race": num_classes.append(4) else: raise ValueError("Target type \"{}\" is not recognized.".format(a)) else: if attr == "age": num_classes = 117 elif attr == "gender": num_classes = 2 elif attr == "race": num_classes = 4 else: raise ValueError("Target type \"{}\" is not recognized.".format(attr)) transform = transforms.Compose([ transforms.RandomHorizontalFlip(), transforms.Resize((64, 64)), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), ]) dataset = UTKFaceDataset(root=root, attr=attr, transform=transform) input_channel = 3 elif dataset_name.lower() == "celeba": if isinstance(attr, list): for a in attr: if a != "attr": raise ValueError("Target type \"{}\" is not recognized.".format(a)) num_classes = [8, 4] # heavyMakeup MouthSlightlyOpen Smiling, Male Young attr_list = [[18, 21, 31], [20, 39]] else: if attr == "attr": num_classes = 8 attr_list = [[18, 21, 31]] else: raise ValueError("Target type \"{}\" is not recognized.".format(attr)) transform = transforms.Compose([ transforms.RandomHorizontalFlip(), transforms.Resize((64, 64)), transforms.ToTensor(), transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)), ]) dataset = CelebA(root=root, attr_list=attr_list, target_type=attr, transform=transform) input_channel = 3 elif dataset_name.lower() == "stl10": num_classes = 10 transform = transforms.Compose([ transforms.Resize((64, 64)), transforms.ToTensor(), transforms.Normalize((0.4914, 0.4822, 0.4465), (0.2023, 0.1994, 0.2010)), ]) train_set = torchvision.datasets.STL10( root=root, split='train', transform=transform, download=True) test_set = torchvision.datasets.STL10( root=root, split='test', transform=transform, download=True) dataset = train_set + test_set input_channel = 3 elif dataset_name.lower() == "fmnist": num_classes = 10 transform = transforms.Compose([ transforms.Resize((64, 64)), transforms.ToTensor(), transforms.Normalize((0.1307,), (0.3081,)) ]) train_set = torchvision.datasets.FashionMNIST( root=root, train=True, download=True, transform=transform) test_set = torchvision.datasets.FashionMNIST( root=root, train=False, download=True, transform=transform) dataset = train_set + test_set input_channel = 1 if isinstance(num_classes, int): target_model = CNN(input_channel=input_channel, num_classes=num_classes) shadow_model = CNN(input_channel=input_channel, num_classes=num_classes) else: target_model = CNN(input_channel=input_channel, num_classes=num_classes[0]) shadow_model = CNN(input_channel=input_channel, num_classes=num_classes[0]) return num_classes, dataset, target_model, shadow_model
9,806
33.900356
184
py
ML-Doctor
ML-Doctor-main/demoloader/train.py
import torch import numpy as np import torch.nn as nn import torch.optim as optim import torch.nn.functional as F import torch.backends.cudnn as cudnn np.set_printoptions(threshold=np.inf) from opacus import PrivacyEngine from torch.optim import lr_scheduler def GAN_init(m): classname = m.__class__.__name__ if classname.find('Conv') != -1: nn.init.normal_(m.weight.data, 0.0, 0.02) elif classname.find('BatchNorm') != -1: nn.init.normal_(m.weight.data, 1.0, 0.02) nn.init.constant_(m.bias.data, 0) class model_training(): def __init__(self, trainloader, testloader, model, device, use_DP, noise, norm, delta): self.use_DP = use_DP self.device = device self.delta = delta self.net = model.to(self.device) self.trainloader = trainloader self.testloader = testloader if self.device == 'cuda': self.net = torch.nn.DataParallel(self.net) cudnn.benchmark = True self.criterion = nn.CrossEntropyLoss() self.optimizer = optim.SGD(self.net.parameters(), lr=1e-2, momentum=0.9, weight_decay=5e-4) self.noise_multiplier, self.max_grad_norm = noise, norm if self.use_DP: self.privacy_engine = PrivacyEngine() self.model, self.optimizer, self.trainloader = self.privacy_engine.make_private( module=model, optimizer=self.optimizer, data_loader=self.trainloader, noise_multiplier=self.noise_multiplier, max_grad_norm=self.max_grad_norm, ) # self.net = module_modification.convert_batchnorm_modules(self.net) # inspector = DPModelInspector() # inspector.validate(self.net) # privacy_engine = PrivacyEngine( # self.net, # batch_size=64, # sample_size=len(self.trainloader.dataset), # alphas=[1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64)), # noise_multiplier=self.noise_multiplier, # max_grad_norm=self.max_grad_norm, # secure_rng=False, # ) print( 'noise_multiplier: %.3f | max_grad_norm: %.3f' % (self.noise_multiplier, self.max_grad_norm)) # privacy_engine.attach(self.optimizer) self.scheduler = lr_scheduler.MultiStepLR(self.optimizer, [50, 75], 0.1) # Training def train(self): self.net.train() train_loss = 0 correct = 0 total = 0 for batch_idx, (inputs, targets) in enumerate(self.trainloader): if isinstance(targets, list): targets = targets[0] if str(self.criterion) != "CrossEntropyLoss()": targets = torch.from_numpy(np.eye(self.num_classes)[targets]).float() inputs, targets = inputs.to(self.device), targets.to(self.device) self.optimizer.zero_grad() outputs = self.net(inputs) loss = self.criterion(outputs, targets) loss.backward() self.optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) if str(self.criterion) != "CrossEntropyLoss()": _, targets= targets.max(1) correct += predicted.eq(targets).sum().item() if self.use_DP: epsilon = self.privacy_engine.accountant.get_epsilon(delta=self.delta) # epsilon, best_alpha = self.optimizer.privacy_engine.get_privacy_spent(1e-5) print("\u03B5: %.3f \u03B4: 1e-5" % (epsilon)) self.scheduler.step() print( 'Train Acc: %.3f%% (%d/%d) | Loss: %.3f' % (100.*correct/total, correct, total, 1.*train_loss/batch_idx)) return 1.*correct/total def saveModel(self, path): torch.save(self.net.state_dict(), path) def get_noise_norm(self): return self.noise_multiplier, self.max_grad_norm def test(self): self.net.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for inputs, targets in self.testloader: if isinstance(targets, list): targets = targets[0] if str(self.criterion) != "CrossEntropyLoss()": targets = torch.from_numpy(np.eye(self.num_classes)[targets]).float() inputs, targets = inputs.to(self.device), targets.to(self.device) outputs = self.net(inputs) loss = self.criterion(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) if str(self.criterion) != "CrossEntropyLoss()": _, targets= targets.max(1) correct += predicted.eq(targets).sum().item() print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/total, correct, total)) return 1.*correct/total class distillation_training(): def __init__(self, PATH, trainloader, testloader, model, teacher, device): self.device = device self.model = model.to(self.device) self.trainloader = trainloader self.testloader = testloader self.PATH = PATH self.teacher = teacher.to(self.device) self.teacher.load_state_dict(torch.load(self.PATH)) if self.device == 'cuda': self.model = torch.nn.DataParallel(self.model) cudnn.benchmark = True self.criterion = nn.KLDivLoss(reduction='batchmean') self.optimizer = optim.SGD(self.model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4) self.scheduler = lr_scheduler.MultiStepLR(self.optimizer, [50, 100], 0.1) def distillation_loss(self, y, labels, teacher_scores, T, alpha): loss = self.criterion(F.log_softmax(y/T, dim=1), F.softmax(teacher_scores/T, dim=1)) loss = loss * (T*T * 2.0 * alpha) + F.cross_entropy(y, labels) * (1. - alpha) return loss def train(self): self.model.train() self.teacher.eval() train_loss = 0 correct = 0 total = 0 for batch_idx, (inputs, targets) in enumerate(self.trainloader): inputs, targets = inputs.to(self.device), targets.to(self.device) self.optimizer.zero_grad() outputs = self.model(inputs) teacher_output = self.teacher(inputs) teacher_output = teacher_output.detach() loss = self.distillation_loss(outputs, targets, teacher_output, T=20.0, alpha=0.7) loss.backward() self.optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() self.scheduler.step() print( 'Train Acc: %.3f%% (%d/%d) | Loss: %.3f' % (100.*correct/total, correct, total, 1.*train_loss/batch_idx)) return 1.*correct/total def saveModel(self, path): torch.save(self.model.state_dict(), path) def test(self): self.model.eval() correct = 0 total = 0 with torch.no_grad(): for inputs, targets in self.testloader: inputs, targets = inputs.to(self.device), targets.to(self.device) outputs = self.model(inputs) _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/total, correct, total)) return 1.*correct/total class GAN_training(): def __init__(self, trainloader, model_discriminator, model_generator, device): self.device = device self.trainloader = trainloader self.model_discriminator = model_discriminator.to(self.device) self.model_generator = model_generator.to(self.device) self.model_discriminator.apply(GAN_init) self.model_generator.apply(GAN_init) self.criterion = nn.BCELoss() self.optimizer_discriminator = optim.Adam(model_discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999)) self.optimizer_generator = optim.Adam(model_generator.parameters(), lr=0.0002, betas=(0.5, 0.999)) self.real_label = 1. self.fake_label = 0. def train(self): # For each batch in the dataloader for i, data in enumerate(self.trainloader, 0): self.model_discriminator.zero_grad() # Format batch real_cpu = data[0].to(self.device) b_size = real_cpu.size(0) label = torch.full((b_size,), self.real_label, dtype=torch.float, device=self.device) # Forward pass real batch through D output = self.model_discriminator(real_cpu).view(-1) # Calculate loss on all-real batch errD_real = self.criterion(output, label) # Calculate gradients for D in backward pass errD_real.backward() D_x = output.mean().item() ## Train with all-fake batch # Generate batch of latent vectors noise = torch.randn(b_size, 100, 1, 1, device=self.device) # Generate fake image batch with G fake = self.model_generator(noise) label.fill_(self.fake_label) # Classify all fake batch with D output = self.model_discriminator(fake.detach()).view(-1) # Calculate D's loss on the all-fake batch errD_fake = self.criterion(output, label) # Calculate the gradients for this batch, accumulated (summed) with previous gradients errD_fake.backward() D_G_z1 = output.mean().item() # Compute error of D as sum over the fake and the real batches errD = errD_real + errD_fake # Update D self.optimizer_discriminator.step() ############################ # (2) Update G network: maximize log(D(G(z))) ########################### self.model_generator.zero_grad() label.fill_(self.real_label) # fake labels are real for generator cost # Since we just updated D, perform another forward pass of all-fake batch through D output = self.model_discriminator(fake).view(-1) # Calculate G's loss based on this output errG = self.criterion(output, label) # Calculate gradients for G errG.backward() D_G_z2 = output.mean().item() # Update G self.optimizer_generator.step() # Output training stats if i % 50 == 0: print('[%d/%d]\tLoss_D: %.4f\tLoss_G: %.4f\tD(x): %.4f\tD(G(z)): %.4f / %.4f' % (i, len(self.trainloader), errD.item(), errG.item(), D_x, D_G_z1, D_G_z2)) # Save Losses for plotting later # G_losses.append(errG.item()) # D_losses.append(errD.item()) # Check how the generator is doing by saving G's output on fixed_noise # if (iters % 500 == 0) or ((epoch == 4) and (i == len(dataloader)-1)): # with torch.no_grad(): # fake = netG(fixed_noise).detach().cpu() # img_list.append(vutils.make_grid(fake, padding=2, normalize=True)) def saveModel(self, path_d, path_g): torch.save(self.model_discriminator.state_dict(), path_d) torch.save(self.model_generator.state_dict(), path_g)
11,748
37.270358
120
py
ML-Doctor
ML-Doctor-main/demoloader/DCGAN.py
import torch.nn as nn class Generator(nn.Module): def __init__(self, ngpu=1, nc=3, nz=100, ngf=64): super(Generator, self).__init__() self.ngpu = ngpu self.main = nn.Sequential( # input is Z, going into a convolution nn.ConvTranspose2d(nz, ngf * 8, 4, 1, 0, bias=False), nn.BatchNorm2d(ngf * 8), nn.ReLU(True), # state size. (ngf*8) x 4 x 4 nn.ConvTranspose2d(ngf * 8, ngf * 4, 4, 2, 1, bias=False), nn.BatchNorm2d(ngf * 4), nn.ReLU(True), # state size. (ngf*4) x 8 x 8 nn.ConvTranspose2d(ngf * 4, ngf * 2, 4, 2, 1, bias=False), nn.BatchNorm2d(ngf * 2), nn.ReLU(True), # state size. (ngf*2) x 16 x 16 nn.ConvTranspose2d(ngf * 2, ngf, 4, 2, 1, bias=False), nn.BatchNorm2d(ngf), nn.ReLU(True), # state size. (ngf) x 32 x 32 nn.ConvTranspose2d(ngf, nc, 4, 2, 1, bias=False), nn.Tanh() # state size. (nc) x 64 x 64 ) def forward(self, input): if input.is_cuda and self.ngpu > 1: output = nn.parallel.data_parallel(self.main, input, range(self.ngpu)) else: output = self.main(input) return output class Discriminator(nn.Module): def __init__(self, ngpu=1, nc=3, ndf=64): super(Discriminator, self).__init__() self.ngpu = ngpu self.main = nn.Sequential( # input is (nc) x 64 x 64 nn.Conv2d(nc, ndf, 4, 2, 1, bias=False), nn.LeakyReLU(0.2, inplace=True), # state size. (ndf) x 32 x 32 nn.Conv2d(ndf, ndf * 2, 4, 2, 1, bias=False), nn.BatchNorm2d(ndf * 2), nn.LeakyReLU(0.2, inplace=True), # state size. (ndf*2) x 16 x 16 nn.Conv2d(ndf * 2, ndf * 4, 4, 2, 1, bias=False), nn.BatchNorm2d(ndf * 4), nn.LeakyReLU(0.2, inplace=True), # state size. (ndf*4) x 8 x 8 nn.Conv2d(ndf * 4, ndf * 8, 4, 2, 1, bias=False), nn.BatchNorm2d(ndf * 8), nn.LeakyReLU(0.2, inplace=True), # state size. (ndf*8) x 4 x 4 nn.Conv2d(ndf * 8, 1, 4, 1, 0, bias=False), nn.Sigmoid() ) def forward(self, input): if input.is_cuda and self.ngpu > 1: output = nn.parallel.data_parallel(self.main, input, range(self.ngpu)) else: output = self.main(input) return output.view(-1, 1).squeeze(1) class FashionGenerator(nn.Module): def __init__(self): super(FashionGenerator, self).__init__() d_input = 100 d_output = 64 * 64 self.input = nn.Sequential( nn.Linear(d_input, 256), nn.ReLU() ) self.hidden1 = nn.Sequential( nn.Linear(256, 512), nn.ReLU() ) self.hidden2 = nn.Sequential( nn.Linear(512, 1024), nn.ReLU() ) self.output = nn.Sequential( nn.Linear(1024, d_output), nn.Tanh() ) def forward(self, x): x = x.view(-1, 100) x = self.input(x) x = self.hidden1(x) x = self.hidden2(x) x = self.output(x) return x.reshape(-1, 1, 64, 64) class FashionDiscriminator(nn.Module): def __init__(self): super(FashionDiscriminator, self).__init__() d_input = 64 * 64 d_output = 1 self.input = nn.Sequential( nn.Linear(d_input, 1024), nn.ReLU(), nn.Dropout(0.2) ) self.hidden1 = nn.Sequential( nn.Linear(1024, 512), nn.ReLU(), nn.Dropout(0.2) ) self.hidden2 = nn.Sequential( nn.Linear(512, 256), nn.ReLU(), nn.Dropout(0.2) ) self.output = nn.Sequential( nn.Linear(256, d_output), nn.Sigmoid() ) def forward(self, x): x = x.reshape(-1, 64 * 64) x = self.input(x) x = self.hidden1(x) x = self.hidden2(x) x = self.output(x) return x
4,237
29.934307
82
py
ML-Doctor
ML-Doctor-main/utils/define_models.py
import torch import torch.nn as nn import torch.nn.functional as F class attrinf_attack_model(nn.Module): def __init__(self, inputs, outputs): super(attrinf_attack_model, self).__init__() self.classifier = nn.Linear(inputs, outputs) def forward(self, x): x = torch.flatten(x, 1) x = self.classifier(x) return x class ShadowAttackModel(nn.Module): def __init__(self, class_num): super(ShadowAttackModel, self).__init__() self.Output_Component = nn.Sequential( # nn.Dropout(p=0.2), nn.Linear(class_num, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Prediction_Component = nn.Sequential( # nn.Dropout(p=0.2), nn.Linear(1, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Encoder_Component = nn.Sequential( # nn.Dropout(p=0.2), nn.Linear(128, 256), nn.ReLU(), # nn.Dropout(p=0.2), nn.Linear(256, 128), nn.ReLU(), # nn.Dropout(p=0.2), nn.Linear(128, 64), nn.ReLU(), nn.Linear(64, 2), ) def forward(self, output, prediction): Output_Component_result = self.Output_Component(output) Prediction_Component_result = self.Prediction_Component(prediction) final_inputs = torch.cat((Output_Component_result, Prediction_Component_result), 1) final_result = self.Encoder_Component(final_inputs) return final_result class PartialAttackModel(nn.Module): def __init__(self, class_num): super(PartialAttackModel, self).__init__() self.Output_Component = nn.Sequential( # nn.Dropout(p=0.2), nn.Linear(class_num, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Prediction_Component = nn.Sequential( # nn.Dropout(p=0.2), nn.Linear(1, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Encoder_Component = nn.Sequential( # nn.Dropout(p=0.2), nn.Linear(128, 256), nn.ReLU(), # nn.Dropout(p=0.2), nn.Linear(256, 128), nn.ReLU(), # nn.Dropout(p=0.2), nn.Linear(128, 64), nn.ReLU(), nn.Linear(64, 2), ) def forward(self, output, prediction): Output_Component_result = self.Output_Component(output) Prediction_Component_result = self.Prediction_Component(prediction) final_inputs = torch.cat((Output_Component_result, Prediction_Component_result), 1) final_result = self.Encoder_Component(final_inputs) return final_result class WhiteBoxAttackModel(nn.Module): def __init__(self, class_num, total): super(WhiteBoxAttackModel, self).__init__() self.Output_Component = nn.Sequential( nn.Dropout(p=0.2), nn.Linear(class_num, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Loss_Component = nn.Sequential( nn.Dropout(p=0.2), nn.Linear(1, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Gradient_Component = nn.Sequential( nn.Dropout(p=0.2), nn.Conv2d(1, 1, kernel_size=5, padding=2), nn.BatchNorm2d(1), nn.ReLU(), nn.MaxPool2d(kernel_size=2), nn.Flatten(), nn.Dropout(p=0.2), nn.Linear(total, 256), nn.ReLU(), nn.Dropout(p=0.2), nn.Linear(256, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Label_Component = nn.Sequential( nn.Dropout(p=0.2), nn.Linear(class_num, 128), nn.ReLU(), nn.Linear(128, 64), ) self.Encoder_Component = nn.Sequential( nn.Dropout(p=0.2), nn.Linear(256, 256), nn.ReLU(), nn.Dropout(p=0.2), nn.Linear(256, 128), nn.ReLU(), nn.Dropout(p=0.2), nn.Linear(128, 64), nn.ReLU(), nn.Linear(64, 2), ) def forward(self, output, loss, gradient, label): Output_Component_result = self.Output_Component(output) Loss_Component_result = self.Loss_Component(loss) Gradient_Component_result = self.Gradient_Component(gradient) Label_Component_result = self.Label_Component(label) # Loss_Component_result = F.softmax(Loss_Component_result, dim=1) # Gradient_Component_result = F.softmax(Gradient_Component_result, dim=1) # final_inputs = Output_Component_result # final_inputs = Loss_Component_result # final_inputs = Gradient_Component_result # final_inputs = Label_Component_result final_inputs = torch.cat((Output_Component_result, Loss_Component_result, Gradient_Component_result, Label_Component_result), 1) final_result = self.Encoder_Component(final_inputs) return final_result
4,227
24.017751
130
py
ML-Doctor
ML-Doctor-main/doctor/modsteal.py
import torch import torch.nn.functional as F from math import * from tqdm import tqdm class train_steal_model(): def __init__(self, train_loader, test_loader, target_model, attack_model, TARGET_PATH, ATTACK_PATH, device, batch_size, loss, optimizer): self.device = device self.batch_size = batch_size self.train_loader = train_loader self.test_loader = test_loader self.TARGET_PATH = TARGET_PATH self.target_model = target_model.to(self.device) self.target_model.load_state_dict(torch.load(self.TARGET_PATH, map_location=self.device)) self.target_model.eval() self.ATTACK_PATH = ATTACK_PATH self.attack_model = attack_model.to(self.device) self.criterion = loss self.optimizer = optimizer self.index = 0 self.count = [0 for i in range(10)] self.dataset = [] def train(self, train_set, train_out): self.attack_model.train() for inputs, targets in tqdm(zip(train_set, train_out)): inputs, targets = inputs.to(self.device), targets.to(self.device) self.optimizer.zero_grad() outputs = self.attack_model(inputs) outputs = F.softmax(outputs, dim=1) loss = self.criterion(outputs, targets) loss.backward() self.optimizer.step() def train_with_same_distribution(self): self.attack_model.train() train_loss = 0 correct = 0 total = 0 correct_target = 0 total_target = 0 for inputs, targets in tqdm(self.train_loader): inputs, targets = inputs.to(self.device), targets.to(self.device) target_model_logit = self.target_model(inputs) _,target_model_output = target_model_logit.max(1) target_model_posterior = F.softmax(target_model_logit, dim=1) # print(inputs, targets) self.optimizer.zero_grad() outputs = self.attack_model(inputs) # outputs = F.softmax(outputs, dim=1) # loss = self.criterion(outputs, targets) loss = self.criterion(outputs, target_model_posterior) loss.backward() self.optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() total_target += targets.size(0) correct_target += predicted.eq(target_model_output).sum().item() print( 'Train Acc: %.3f%% (%d/%d)' % (100.*correct/total, correct, total)) print( 'Train Agreement: %.3f%% (%d/%d)' % (100.*correct_target/total_target, correct_target, total_target)) def test(self): self.attack_model.eval() correct = 0 target_correct = 0 total = 0 agreement_correct = 0 agreement_total = 0 with torch.no_grad(): for inputs, targets in self.test_loader: inputs, targets = inputs.to(self.device), targets.to(self.device) outputs = self.attack_model(inputs) _, predicted = outputs.max(1) target_model_logit = self.target_model(inputs) _,target_predicted = target_model_logit.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() target_correct += target_predicted.eq(targets).sum().item() output_target = self.target_model(inputs) _, predicted_target = output_target.max(1) agreement_total += targets.size(0) agreement_correct += predicted.eq(predicted_target).sum().item() print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/total, correct, total)) print( 'Target Test Acc: %.3f%% (%d/%d)' % (100.*target_correct/total, target_correct, total)) print( 'Test Agreement: %.3f%% (%d/%d)' % (100.*agreement_correct/agreement_total, agreement_correct, agreement_total)) acc_test = correct/total agreemenet_test = agreement_correct / agreement_total return acc_test, agreemenet_test def saveModel(self): torch.save(self.attack_model.state_dict(), self.ATTACK_PATH)
4,341
33.188976
141
py
ML-Doctor
ML-Doctor-main/doctor/attrinf.py
import torch import pickle import torch.nn as nn import torch.optim as optim from utils.define_models import * from sklearn.metrics import f1_score class attack_training(): def __init__(self, device, attack_trainloader, attack_testloader, target_model, TARGET_PATH, ATTACK_PATH): self.device = device self.TARGET_PATH = TARGET_PATH self.ATTACK_PATH = ATTACK_PATH self.target_model = target_model.to(self.device) self.target_model.load_state_dict(torch.load(self.TARGET_PATH)) self.target_model.eval() self.attack_model = None self.attack_trainloader = attack_trainloader self.attack_testloader = attack_testloader self.criterion = nn.CrossEntropyLoss() self.optimizer = None # self.scheduler = lr_scheduler.MultiStepLR(self.optimizer, [50, 100], 0.1) self.dataset_type = None def _get_activation(self, name, activation): def hook(model, input, output): activation[name] = output.detach() return hook def init_attack_model(self, size, output_classes): x = torch.rand(size).to(self.device) input_classes = self.get_middle_output(x).flatten().shape[0] self.attack_model = attrinf_attack_model(inputs=input_classes, outputs=output_classes) self.attack_model.to(self.device) self.optimizer = optim.Adam(self.attack_model.parameters(), lr=1e-3) if output_classes == 2: self.dataset_type = "binary" else: self.dataset_type = "macro" def get_middle_output(self, x): temp = [] for name, _ in self.target_model.named_parameters(): if "weight" in name: temp.append(name) if 1 > len(temp): raise IndexError('layer is out of range') name = temp[-2].split('.') var = eval('self.target_model.' + name[0]) out = {} var[int(name[1])].register_forward_hook(self._get_activation(name[1], out)) _ = self.target_model(x) return out[name[1]] # Training def train(self, epoch): self.attack_model.train() train_loss = 0 correct = 0 total = 0 final_result = [] final_gndtrth = [] final_predict = [] final_probabe = [] for batch_idx, (inputs, [_, targets]) in enumerate(self.attack_trainloader): inputs, targets = inputs.to(self.device), targets.to(self.device) self.optimizer.zero_grad() oracles = self.get_middle_output(inputs) outputs = self.attack_model(oracles) outputs = F.softmax(outputs, dim=1) loss = self.criterion(outputs, targets) loss.backward() self.optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() if epoch: final_gndtrth.append(targets) final_predict.append(predicted) final_probabe.append(outputs[:, 1]) if epoch: final_gndtrth = torch.cat(final_gndtrth, dim=0).cpu().detach().numpy() final_predict = torch.cat(final_predict, dim=0).cpu().detach().numpy() final_probabe = torch.cat(final_probabe, dim=0).cpu().detach().numpy() test_f1_score = f1_score(final_gndtrth, final_predict, average=self.dataset_type) final_result.append(test_f1_score) with open(self.ATTACK_PATH + "_attrinf_train.p", "wb") as f: pickle.dump((final_gndtrth, final_predict, final_probabe), f) print("Saved Attack Test Ground Truth and Predict Sets") print("Test F1: %f" % (test_f1_score)) final_result.append(1.*correct/total) print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/(1.0*total), correct, total)) return final_result def test(self, epoch): self.attack_model.eval() correct = 0 total = 0 final_result = [] final_gndtrth = [] final_predict = [] final_probabe = [] with torch.no_grad(): for inputs, [_, targets] in self.attack_testloader: inputs, targets = inputs.to(self.device), targets.to(self.device) oracles = self.get_middle_output(inputs) outputs = self.attack_model(oracles) outputs = F.softmax(outputs, dim=1) _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() if epoch: final_gndtrth.append(targets) final_predict.append(predicted) final_probabe.append(outputs[:, 1]) if epoch: final_gndtrth = torch.cat(final_gndtrth, dim=0).cpu().numpy() final_predict = torch.cat(final_predict, dim=0).cpu().numpy() final_probabe = torch.cat(final_probabe, dim=0).cpu().numpy() test_f1_score = f1_score(final_gndtrth, final_predict, average=self.dataset_type) final_result.append(test_f1_score) with open(self.ATTACK_PATH + "_attrinf_test.p", "wb") as f: pickle.dump((final_gndtrth, final_predict, final_probabe), f) print("Saved Attack Test Ground Truth and Predict Sets") print("Test F1: %f" % (test_f1_score)) final_result.append(1.*correct/total) print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/(1.0*total), correct, total)) return final_result def saveModel(self): torch.save(self.attack_model.state_dict(), self.ATTACK_PATH + "_attrinf_attack_model.pth") def train_attack_model(TARGET_PATH, ATTACK_PATH, output_classes, device, target_model, train_loader, test_loader, size): attack = attack_training(device, train_loader, test_loader, target_model, TARGET_PATH, ATTACK_PATH) attack.init_attack_model(size, output_classes) for epoch in range(100): flag = 1 if epoch==99 else 0 print("<======================= Epoch " + str(epoch+1) + " =======================>") print("attack training") acc_train = attack.train(flag) print("attack testing") acc_test = attack.test(flag) attack.saveModel() print("Saved Attack Model") print("Finished!!!") return acc_train, acc_test
6,547
34.978022
120
py
ML-Doctor
ML-Doctor-main/doctor/modinv.py
import time import torch import random import numpy as np import torch.nn as nn import torch.utils.data from torch.autograd import Variable class ccs_inversion(object): ''' Model inversion is a kind of data reconstruct attack. This class we implement the attack on neural network, the attack goal is to generate data that is close to original data distribution. This attack was first described in Fredrikson's paper (Algorithm 1): "Model Inversion Attacks that Exploit Confidence Information and Basic Countermeasures" (CCS2015) -----------------------------NOTICE--------------------------- If the model's output layer doesn't contain Softmax layer, please add it manually. And parameters will influence the quality of the reconstructed data significantly. -------------------------------------------------------------- Args: ------------------------ :param target_model: the target model which we are trying to reconstruct its training dataset :param input_size: the size of the model's input :param output_size: the size of the model's output :param target_label: the reconstructed output is belong to this class :param param_alpha: the number of iteration round :param param_beta, gamma, lambda: the hyperparameters in paper ''' def __init__(self, target_model, input_size, output_size, target_label, param_alpha, param_beta, param_gamma, param_lambda, device): self.target_model = target_model self.input_size = input_size self.output_size = output_size self.target_label = target_label self.param_alpha = param_alpha self.param_beta = param_beta self.param_gamma = param_gamma self.param_lambda = param_lambda self.device = device self.target_model.to(self.device).eval() def model_invert(self): current_x = [] cost_x = [] current_x.append(Variable(torch.from_numpy(np.zeros(self.input_size, dtype=np.uint8))).float().to(self.device)) for i in range(self.param_alpha): cost_x.append(self.invert_cost(current_x[i]).to(self.device)) cost_x[i].backward() current_x.append((current_x[i] - self.param_lambda * current_x[i].grad).data) if self.invert_cost(current_x[i + 1]) <= self.param_gamma: print('Target cost value achieved') break elif i >= self.param_beta and self.invert_cost(current_x[i + 1]) >= max(cost_x[self.param_beta:i + 1]): print('Exceed beta') break i = cost_x.index(min(cost_x)) return current_x[i] def invert_cost(self, input_x): return 1 - self.target_model(input_x.requires_grad_(True))[0][self.target_label] def reverse_mse(self, ori_dataset): ''' output the average MSE value of different classes :param ori_dataset: the data used to train the target model, please make sure setting the batch size as 1. :return: MSE value ''' reverse_data = [] for i in range(self.output_size): self.target_label = i a = self.model_invert() reverse_data.append(a) class_avg = [Variable(torch.from_numpy(np.zeros(self.input_size, dtype=np.uint8))).float().to(self.device) for _ in range(self.output_size)] class_mse = [0 for _ in range(self.output_size)] class_count = [0 for _ in range(self.output_size)] for x, y in ori_dataset: x, y = x.to(self.device), y.to(self.device) class_avg[y] = class_avg[y] + x class_count[y] = class_count[y] + 1 for i in range(self.output_size): class_mse[i] = self.figure_mse(class_avg[i] / class_count[i], (reverse_data[i])) all_class_avg_mse = 0 for i in range(self.output_size): all_class_avg_mse = all_class_avg_mse + class_mse[i] return all_class_avg_mse / self.output_size def figure_mse(self, recover_fig, ori_fig): ''' :param recover_fig: figure recovered by model inversion attack, type: :param ori_fig: figure in the training dataset :return: MSE value of these two figures ''' diff = nn.MSELoss() return diff(recover_fig, ori_fig) def revealer_inversion(G, D, T, E, iden, device, noise = 100,lr=1e-3, momentum=0.9, lamda=100, iter_times=1500, clip_range=1): ''' This model inversion attack was proposed by Zhang et al. in CVPR20 "The Secret Revealer: Generative Model-Inversion Attacks Against Deep Neural Networks" ''' iden = iden.view(-1).long().to(device) G, D, T, E = G.to(device), D.to(device), T.to(device), E.to(device) criterion = nn.CrossEntropyLoss().to(device) bs = iden.shape[0] G.eval() D.eval() T.eval() max_score = torch.zeros(bs) max_iden = torch.zeros(bs) z_hat = torch.zeros(bs, noise,1,1) cnt = 0 for random_seed_sudo in range(10): tf = time.time() random_seed = random.randint(0,200) torch.manual_seed(random_seed) torch.cuda.manual_seed(random_seed) np.random.seed(random_seed) random.seed(random_seed) z = torch.randn(bs, noise, 1, 1).to(device).float() z.requires_grad = True v = torch.zeros(bs, noise, 1, 1).to(device).float() for i in range(iter_times): fake = G(z) label = D(fake) out = T(fake) if z.grad is not None: z.grad.data.zero_() Prior_Loss = - label.mean() Iden_Loss = criterion(out, iden) Total_Loss = Prior_Loss + lamda * Iden_Loss Total_Loss.backward() v_prev = v.clone() gradient = z.grad.data v = momentum * v - lr * gradient z = z + ( - momentum * v_prev + (1 + momentum) * v) z = torch.clamp(z.detach(), -clip_range, clip_range).float() z.requires_grad = True Prior_Loss_val = Prior_Loss.item() Iden_Loss_val = Iden_Loss.item() if (i + 1) % 300 == 0: fake_img = G(z.detach()) eval_prob = E(fake_img) eval_iden = torch.argmax(eval_prob, dim=1).view(-1) acc = iden.eq(eval_iden.long()).sum().item() * 1.0 / bs fake = G(z) score = T(fake) eval_prob = E(fake) _, eval_iden = torch.max(eval_prob, dim=1) for i in range(bs): _, gtl = torch.max(score, 1) gt = gtl[i].item() if score[i, gt].item() > max_score[i].item(): max_score[i] = score[i, gt] max_iden[i] = eval_iden[i] z_hat[i, :] = z[i, :] if eval_iden[i].item() == gt: cnt += 1 print("Acc:{:.2f}\t".format(cnt * 1.0 / (bs*10))) return cnt * 1.0 / (bs*10) def load_data(PATH_target, PATH_evaluation, target_model, evaluate_model): ''' Evaluate model is used to predict the identity based on the input reconstructed image. If the evaluation classifier achieves high accuracy, the reconstructed image is considered to expose private information about the target label. The evaluate model should be different from the target network because the reconstructed images may incorporate features that overfit the target network while being semantically meaningless. Moreover, the evaluation classifier should be highly performant. ''' target_model.load_state_dict(torch.load(PATH_target)) evaluate_model.load_state_dict(torch.load(PATH_evaluation)) print("Finished Loading") return target_model, evaluate_model def prepare_GAN(data_type, discriminator, generator, PATH_1, PATH_2): discriminator.load_state_dict(torch.load(PATH_1)) generator.load_state_dict(torch.load(PATH_2)) iden = torch.zeros(10) if data_type.lower() == 'stl10' or data_type.lower() == 'fmnist': for i in range(10): iden[i] = i elif data_type.lower() == 'utkface': for i in range(10): iden[i] = i % 4 elif data_type.lower() == 'celeba': for i in range(10): iden[i] = i % 8 return discriminator, generator, iden
8,409
36.713004
148
py
ML-Doctor
ML-Doctor-main/doctor/meminf.py
import os import glob import torch import pickle import numpy as np import torch.nn as nn import torch.optim as optim import torch.nn.functional as F import torch.backends.cudnn as cudnn np.set_printoptions(threshold=np.inf) from opacus import PrivacyEngine from torch.optim import lr_scheduler from sklearn.metrics import f1_score, roc_auc_score def weights_init(m): if isinstance(m, nn.Conv2d): nn.init.normal_(m.weight.data) m.bias.data.fill_(0) elif isinstance(m,nn.Linear): nn.init.xavier_normal_(m.weight) nn.init.constant_(m.bias, 0) class shadow(): def __init__(self, trainloader, testloader, model, device, use_DP, noise, norm, loss, optimizer, delta): self.delta = delta self.use_DP = use_DP self.device = device self.model = model.to(self.device) self.trainloader = trainloader self.testloader = testloader self.criterion = loss self.optimizer = optimizer self.noise_multiplier, self.max_grad_norm = noise, norm if self.use_DP: self.privacy_engine = PrivacyEngine() self.model, self.optimizer, self.trainloader = self.privacy_engine.make_private( module=self.model, optimizer=self.optimizer, data_loader=self.trainloader, noise_multiplier=self.noise_multiplier, max_grad_norm=self.max_grad_norm, ) # self.model = module_modification.convert_batchnorm_modules(self.model) # inspector = DPModelInspector() # inspector.validate(self.model) # privacy_engine = PrivacyEngine( # self.model, # batch_size=batch_size, # sample_size=len(self.trainloader.dataset), # alphas=[1 + x / 10.0 for x in range(1, 100)] + list(range(12, 64)), # noise_multiplier=self.noise_multiplier, # max_grad_norm=self.max_grad_norm, # secure_rng=False, # ) print( 'noise_multiplier: %.3f | max_grad_norm: %.3f' % (self.noise_multiplier, self.max_grad_norm)) # privacy_engine.attach(self.optimizer) self.scheduler = lr_scheduler.MultiStepLR(self.optimizer, [50, 100], 0.1) # Training def train(self): self.model.train() train_loss = 0 correct = 0 total = 0 for batch_idx, (inputs, targets) in enumerate(self.trainloader): inputs, targets = inputs.to(self.device), targets.to(self.device) self.optimizer.zero_grad() outputs = self.model(inputs) loss = self.criterion(outputs, targets) loss.backward() self.optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() if self.use_DP: epsilon = self.privacy_engine.accountant.get_epsilon(delta=self.delta) # epsilon, best_alpha = self.optimizer.privacy_engine.get_privacy_spent(1e-5) print("\u03B5: %.3f \u03B4: 1e-5" % (epsilon)) self.scheduler.step() print( 'Train Acc: %.3f%% (%d/%d) | Loss: %.3f' % (100.*correct/total, correct, total, 1.*train_loss/batch_idx)) return 1.*correct/total def saveModel(self, path): torch.save(self.model.state_dict(), path) def get_noise_norm(self): return self.noise_multiplier, self.max_grad_norm def test(self): self.model.eval() test_loss = 0 correct = 0 total = 0 with torch.no_grad(): for inputs, targets in self.testloader: inputs, targets = inputs.to(self.device), targets.to(self.device) outputs = self.model(inputs) loss = self.criterion(outputs, targets) test_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/total, correct, total)) return 1.*correct/total class distillation_training(): def __init__(self, PATH, trainloader, testloader, model, teacher, device, optimizer, T, alpha): self.device = device self.model = model.to(self.device) self.trainloader = trainloader self.testloader = testloader self.PATH = PATH self.teacher = teacher.to(self.device) self.teacher.load_state_dict(torch.load(self.PATH)) self.teacher.eval() self.criterion = nn.KLDivLoss(reduction='batchmean') self.optimizer = optimizer self.scheduler = lr_scheduler.MultiStepLR(self.optimizer, [50, 100], 0.1) self.T = T self.alpha = alpha def distillation_loss(self, y, labels, teacher_scores, T, alpha): loss = self.criterion(F.log_softmax(y/T, dim=1), F.softmax(teacher_scores/T, dim=1)) loss = loss * (T*T * alpha) + F.cross_entropy(y, labels) * (1. - alpha) return loss def train(self): self.model.train() train_loss = 0 correct = 0 total = 0 for batch_idx, (inputs, [targets, _]) in enumerate(self.trainloader): inputs, targets = inputs.to(self.device), targets.to(self.device) self.optimizer.zero_grad() outputs = self.model(inputs) teacher_output = self.teacher(inputs) teacher_output = teacher_output.detach() loss = self.distillation_loss(outputs, targets, teacher_output, T=self.T, alpha=self.alpha) loss.backward() self.optimizer.step() train_loss += loss.item() _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() self.scheduler.step() print( 'Train Acc: %.3f%% (%d/%d) | Loss: %.3f' % (100.*correct/total, correct, total, 1.*train_loss/batch_idx)) return 1.*correct/total def saveModel(self, path): torch.save(self.model.state_dict(), path) def test(self): self.model.eval() correct = 0 total = 0 with torch.no_grad(): for inputs, [targets, _] in self.testloader: inputs, targets = inputs.to(self.device), targets.to(self.device) outputs = self.model(inputs) _, predicted = outputs.max(1) total += targets.size(0) correct += predicted.eq(targets).sum().item() print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/total, correct, total)) return 1.*correct/total class attack_for_blackbox(): def __init__(self, SHADOW_PATH, TARGET_PATH, ATTACK_SETS, attack_train_loader, attack_test_loader, target_model, shadow_model, attack_model, device): self.device = device self.TARGET_PATH = TARGET_PATH self.SHADOW_PATH = SHADOW_PATH self.ATTACK_SETS = ATTACK_SETS self.target_model = target_model.to(self.device) self.shadow_model = shadow_model.to(self.device) self.target_model.load_state_dict(torch.load(self.TARGET_PATH)) self.shadow_model.load_state_dict(torch.load(self.SHADOW_PATH)) self.target_model.eval() self.shadow_model.eval() self.attack_train_loader = attack_train_loader self.attack_test_loader = attack_test_loader self.attack_model = attack_model.to(self.device) torch.manual_seed(0) self.attack_model.apply(weights_init) self.criterion = nn.CrossEntropyLoss() self.optimizer = optim.Adam(self.attack_model.parameters(), lr=1e-5) def _get_data(self, model, inputs, targets): result = model(inputs) output, _ = torch.sort(result, descending=True) # results = F.softmax(results[:,:5], dim=1) _, predicts = result.max(1) prediction = predicts.eq(targets).float() # prediction = [] # for predict in predicts: # prediction.append([1,] if predict else [0,]) # prediction = torch.Tensor(prediction) # final_inputs = torch.cat((results, prediction), 1) # print(final_inputs.shape) return output, prediction.unsqueeze(-1) def prepare_dataset(self): with open(self.ATTACK_SETS + "train.p", "wb") as f: for inputs, targets, members in self.attack_train_loader: inputs, targets = inputs.to(self.device), targets.to(self.device) output, prediction = self._get_data(self.shadow_model, inputs, targets) # output = output.cpu().detach().numpy() pickle.dump((output, prediction, members), f) print("Finished Saving Train Dataset") with open(self.ATTACK_SETS + "test.p", "wb") as f: for inputs, targets, members in self.attack_test_loader: inputs, targets = inputs.to(self.device), targets.to(self.device) output, prediction = self._get_data(self.target_model, inputs, targets) # output = output.cpu().detach().numpy() pickle.dump((output, prediction, members), f) print("Finished Saving Test Dataset") def train(self, epoch, result_path): self.attack_model.train() batch_idx = 1 train_loss = 0 correct = 0 total = 0 final_train_gndtrth = [] final_train_predict = [] final_train_probabe = [] final_result = [] with open(self.ATTACK_SETS + "train.p", "rb") as f: while(True): try: output, prediction, members = pickle.load(f) output, prediction, members = output.to(self.device), prediction.to(self.device), members.to(self.device) results = self.attack_model(output, prediction) results = F.softmax(results, dim=1) losses = self.criterion(results, members) losses.backward() self.optimizer.step() train_loss += losses.item() _, predicted = results.max(1) total += members.size(0) correct += predicted.eq(members).sum().item() if epoch: final_train_gndtrth.append(members) final_train_predict.append(predicted) final_train_probabe.append(results[:, 1]) batch_idx += 1 except EOFError: break if epoch: final_train_gndtrth = torch.cat(final_train_gndtrth, dim=0).cpu().detach().numpy() final_train_predict = torch.cat(final_train_predict, dim=0).cpu().detach().numpy() final_train_probabe = torch.cat(final_train_probabe, dim=0).cpu().detach().numpy() train_f1_score = f1_score(final_train_gndtrth, final_train_predict) train_roc_auc_score = roc_auc_score(final_train_gndtrth, final_train_probabe) final_result.append(train_f1_score) final_result.append(train_roc_auc_score) with open(result_path, "wb") as f: pickle.dump((final_train_gndtrth, final_train_predict, final_train_probabe), f) print("Saved Attack Train Ground Truth and Predict Sets") print("Train F1: %f\nAUC: %f" % (train_f1_score, train_roc_auc_score)) final_result.append(1.*correct/total) print( 'Train Acc: %.3f%% (%d/%d) | Loss: %.3f' % (100.*correct/total, correct, total, 1.*train_loss/batch_idx)) return final_result def test(self, epoch, result_path): self.attack_model.eval() batch_idx = 1 correct = 0 total = 0 final_test_gndtrth = [] final_test_predict = [] final_test_probabe = [] final_result = [] with torch.no_grad(): with open(self.ATTACK_SETS + "test.p", "rb") as f: while(True): try: output, prediction, members = pickle.load(f) output, prediction, members = output.to(self.device), prediction.to(self.device), members.to(self.device) results = self.attack_model(output, prediction) _, predicted = results.max(1) total += members.size(0) correct += predicted.eq(members).sum().item() results = F.softmax(results, dim=1) if epoch: final_test_gndtrth.append(members) final_test_predict.append(predicted) final_test_probabe.append(results[:, 1]) batch_idx += 1 except EOFError: break if epoch: final_test_gndtrth = torch.cat(final_test_gndtrth, dim=0).cpu().numpy() final_test_predict = torch.cat(final_test_predict, dim=0).cpu().numpy() final_test_probabe = torch.cat(final_test_probabe, dim=0).cpu().numpy() test_f1_score = f1_score(final_test_gndtrth, final_test_predict) test_roc_auc_score = roc_auc_score(final_test_gndtrth, final_test_probabe) final_result.append(test_f1_score) final_result.append(test_roc_auc_score) with open(result_path, "wb") as f: pickle.dump((final_test_gndtrth, final_test_predict, final_test_probabe), f) print("Saved Attack Test Ground Truth and Predict Sets") print("Test F1: %f\nAUC: %f" % (test_f1_score, test_roc_auc_score)) final_result.append(1.*correct/total) print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/(1.0*total), correct, total)) return final_result def delete_pickle(self): train_file = glob.glob(self.ATTACK_SETS +"train.p") for trf in train_file: os.remove(trf) test_file = glob.glob(self.ATTACK_SETS +"test.p") for tef in test_file: os.remove(tef) def saveModel(self, path): torch.save(self.attack_model.state_dict(), path) class attack_for_whitebox(): def __init__(self, TARGET_PATH, SHADOW_PATH, ATTACK_SETS, attack_train_loader, attack_test_loader, target_model, shadow_model, attack_model, device, class_num): self.device = device self.class_num = class_num self.ATTACK_SETS = ATTACK_SETS self.TARGET_PATH = TARGET_PATH self.target_model = target_model.to(self.device) self.target_model.load_state_dict(torch.load(self.TARGET_PATH)) self.target_model.eval() self.SHADOW_PATH = SHADOW_PATH self.shadow_model = shadow_model.to(self.device) self.shadow_model.load_state_dict(torch.load(self.SHADOW_PATH)) self.shadow_model.eval() self.attack_train_loader = attack_train_loader self.attack_test_loader = attack_test_loader self.attack_model = attack_model.to(self.device) torch.manual_seed(0) self.attack_model.apply(weights_init) self.target_criterion = nn.CrossEntropyLoss(reduction='none') self.attack_criterion = nn.CrossEntropyLoss() #self.optimizer = optim.SGD(self.attack_model.parameters(), lr=0.01, momentum=0.9, weight_decay=5e-4) self.optimizer = optim.Adam(self.attack_model.parameters(), lr=1e-5) self.attack_train_data = None self.attack_test_data = None def _get_data(self, model, inputs, targets): results = model(inputs) # outputs = F.softmax(outputs, dim=1) losses = self.target_criterion(results, targets) gradients = [] for loss in losses: loss.backward(retain_graph=True) gradient_list = reversed(list(model.named_parameters())) for name, parameter in gradient_list: if 'weight' in name: gradient = parameter.grad.clone() # [column[:, None], row].resize_(100,100) gradient = gradient.unsqueeze_(0) gradients.append(gradient.unsqueeze_(0)) break labels = [] for num in targets: label = [0 for i in range(self.class_num)] label[num.item()] = 1 labels.append(label) gradients = torch.cat(gradients, dim=0) losses = losses.unsqueeze_(1).detach() outputs, _ = torch.sort(results, descending=True) labels = torch.Tensor(labels) return outputs, losses, gradients, labels def prepare_dataset(self): with open(self.ATTACK_SETS + "train.p", "wb") as f: for inputs, targets, members in self.attack_train_loader: inputs, targets = inputs.to(self.device), targets.to(self.device) output, loss, gradient, label = self._get_data(self.shadow_model, inputs, targets) pickle.dump((output, loss, gradient, label, members), f) print("Finished Saving Train Dataset") with open(self.ATTACK_SETS + "test.p", "wb") as f: for inputs, targets, members in self.attack_test_loader: inputs, targets = inputs.to(self.device), targets.to(self.device) output, loss, gradient, label = self._get_data(self.target_model, inputs, targets) pickle.dump((output, loss, gradient, label, members), f) # pickle.dump((output, loss, gradient, label, members), open(self.ATTACK_PATH + "test.p", "wb")) print("Finished Saving Test Dataset") def train(self, epoch, result_path): self.attack_model.train() batch_idx = 1 train_loss = 0 correct = 0 total = 0 final_train_gndtrth = [] final_train_predict = [] final_train_probabe = [] final_result = [] with open(self.ATTACK_SETS + "train.p", "rb") as f: while(True): try: output, loss, gradient, label, members = pickle.load(f) output, loss, gradient, label, members = output.to(self.device), loss.to(self.device), gradient.to(self.device), label.to(self.device), members.to(self.device) results = self.attack_model(output, loss, gradient, label) # results = F.softmax(results, dim=1) losses = self.attack_criterion(results, members) losses.backward() self.optimizer.step() train_loss += losses.item() _, predicted = results.max(1) total += members.size(0) correct += predicted.eq(members).sum().item() if epoch: final_train_gndtrth.append(members) final_train_predict.append(predicted) final_train_probabe.append(results[:, 1]) batch_idx += 1 except EOFError: break if epoch: final_train_gndtrth = torch.cat(final_train_gndtrth, dim=0).cpu().detach().numpy() final_train_predict = torch.cat(final_train_predict, dim=0).cpu().detach().numpy() final_train_probabe = torch.cat(final_train_probabe, dim=0).cpu().detach().numpy() train_f1_score = f1_score(final_train_gndtrth, final_train_predict) train_roc_auc_score = roc_auc_score(final_train_gndtrth, final_train_probabe) final_result.append(train_f1_score) final_result.append(train_roc_auc_score) with open(result_path, "wb") as f: pickle.dump((final_train_gndtrth, final_train_predict, final_train_probabe), f) print("Saved Attack Train Ground Truth and Predict Sets") print("Train F1: %f\nAUC: %f" % (train_f1_score, train_roc_auc_score)) final_result.append(1.*correct/total) print( 'Train Acc: %.3f%% (%d/%d) | Loss: %.3f' % (100.*correct/total, correct, total, 1.*train_loss/batch_idx)) return final_result def test(self, epoch, result_path): self.attack_model.eval() batch_idx = 1 correct = 0 total = 0 final_test_gndtrth = [] final_test_predict = [] final_test_probabe = [] final_result = [] with torch.no_grad(): with open(self.ATTACK_SETS + "test.p", "rb") as f: while(True): try: output, loss, gradient, label, members = pickle.load(f) output, loss, gradient, label, members = output.to(self.device), loss.to(self.device), gradient.to(self.device), label.to(self.device), members.to(self.device) results = self.attack_model(output, loss, gradient, label) _, predicted = results.max(1) total += members.size(0) correct += predicted.eq(members).sum().item() results = F.softmax(results, dim=1) if epoch: final_test_gndtrth.append(members) final_test_predict.append(predicted) final_test_probabe.append(results[:, 1]) batch_idx += 1 except EOFError: break if epoch: final_test_gndtrth = torch.cat(final_test_gndtrth, dim=0).cpu().numpy() final_test_predict = torch.cat(final_test_predict, dim=0).cpu().numpy() final_test_probabe = torch.cat(final_test_probabe, dim=0).cpu().numpy() test_f1_score = f1_score(final_test_gndtrth, final_test_predict) test_roc_auc_score = roc_auc_score(final_test_gndtrth, final_test_probabe) final_result.append(test_f1_score) final_result.append(test_roc_auc_score) with open(result_path, "wb") as f: pickle.dump((final_test_gndtrth, final_test_predict, final_test_probabe), f) print("Saved Attack Test Ground Truth and Predict Sets") print("Test F1: %f\nAUC: %f" % (test_f1_score, test_roc_auc_score)) final_result.append(1.*correct/total) print( 'Test Acc: %.3f%% (%d/%d)' % (100.*correct/(1.0*total), correct, total)) return final_result def delete_pickle(self): train_file = glob.glob(self.ATTACK_SETS +"train.p") for trf in train_file: os.remove(trf) test_file = glob.glob(self.ATTACK_SETS +"test.p") for tef in test_file: os.remove(tef) def saveModel(self, path): torch.save(self.attack_model.state_dict(), path) def train_shadow_model(PATH, device, shadow_model, train_loader, test_loader, use_DP, noise, norm, loss, optimizer, delta): model = shadow(train_loader, test_loader, shadow_model, device, use_DP, noise, norm, loss, optimizer, delta) acc_train = 0 acc_test = 0 for i in range(100): print("<======================= Epoch " + str(i+1) + " =======================>") print("shadow training") acc_train = model.train() print("shadow testing") acc_test = model.test() overfitting = round(acc_train - acc_test, 6) print('The overfitting rate is %s' % overfitting) FILE_PATH = PATH + "_shadow.pth" model.saveModel(FILE_PATH) print("saved shadow model!!!") print("Finished training!!!") return acc_train, acc_test, overfitting def train_shadow_distillation(MODEL_PATH, DL_PATH, device, target_model, student_model, train_loader, test_loader): distillation = distillation_training(MODEL_PATH, train_loader, test_loader, student_model, target_model, device) for i in range(100): print("<======================= Epoch " + str(i+1) + " =======================>") print("shadow distillation training") acc_distillation_train = distillation.train() print("shadow distillation testing") acc_distillation_test = distillation.test() overfitting = round(acc_distillation_train - acc_distillation_test, 6) print('The overfitting rate is %s' % overfitting) result_path = DL_PATH + "_shadow.pth" distillation.saveModel(result_path) print("Saved shadow model!!!") print("Finished training!!!") return acc_distillation_train, acc_distillation_test, overfitting def get_attack_dataset_without_shadow(train_set, test_set, batch_size): mem_length = len(train_set)//3 nonmem_length = len(test_set)//3 mem_train, mem_test, _ = torch.utils.data.random_split(train_set, [mem_length, mem_length, len(train_set)-(mem_length*2)]) nonmem_train, nonmem_test, _ = torch.utils.data.random_split(test_set, [nonmem_length, nonmem_length, len(test_set)-(nonmem_length*2)]) mem_train, mem_test, nonmem_train, nonmem_test = list(mem_train), list(mem_test), list(nonmem_train), list(nonmem_test) for i in range(len(mem_train)): mem_train[i] = mem_train[i] + (1,) for i in range(len(nonmem_train)): nonmem_train[i] = nonmem_train[i] + (0,) for i in range(len(nonmem_test)): nonmem_test[i] = nonmem_test[i] + (0,) for i in range(len(mem_test)): mem_test[i] = mem_test[i] + (1,) attack_train = mem_train + nonmem_train attack_test = mem_test + nonmem_test attack_trainloader = torch.utils.data.DataLoader( attack_train, batch_size=batch_size, shuffle=True, num_workers=2) attack_testloader = torch.utils.data.DataLoader( attack_test, batch_size=batch_size, shuffle=True, num_workers=2) return attack_trainloader, attack_testloader def get_attack_dataset_with_shadow(target_train, target_test, shadow_train, shadow_test, batch_size): mem_train, nonmem_train, mem_test, nonmem_test = list(shadow_train), list(shadow_test), list(target_train), list(target_test) for i in range(len(mem_train)): mem_train[i] = mem_train[i] + (1,) for i in range(len(nonmem_train)): nonmem_train[i] = nonmem_train[i] + (0,) for i in range(len(nonmem_test)): nonmem_test[i] = nonmem_test[i] + (0,) for i in range(len(mem_test)): mem_test[i] = mem_test[i] + (1,) train_length = min(len(mem_train), len(nonmem_train)) test_length = min(len(mem_test), len(nonmem_test)) mem_train, _ = torch.utils.data.random_split(mem_train, [train_length, len(mem_train) - train_length]) non_mem_train, _ = torch.utils.data.random_split(nonmem_train, [train_length, len(nonmem_train) - train_length]) mem_test, _ = torch.utils.data.random_split(mem_test, [test_length, len(mem_test) - test_length]) non_mem_test, _ = torch.utils.data.random_split(nonmem_test, [test_length, len(nonmem_test) - test_length]) attack_train = mem_train + non_mem_train attack_test = mem_test + non_mem_test attack_trainloader = torch.utils.data.DataLoader( attack_train, batch_size=batch_size, shuffle=True, num_workers=2) attack_testloader = torch.utils.data.DataLoader( attack_test, batch_size=batch_size, shuffle=True, num_workers=2) return attack_trainloader, attack_testloader # black shadow def attack_mode0(TARGET_PATH, SHADOW_PATH, ATTACK_PATH, device, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, get_attack_set, num_classes): MODELS_PATH = ATTACK_PATH + "_meminf_attack0.pth" RESULT_PATH = ATTACK_PATH + "_meminf_attack0.p" ATTACK_SETS = ATTACK_PATH + "_meminf_attack_mode0_" attack = attack_for_blackbox(SHADOW_PATH, TARGET_PATH, ATTACK_SETS, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, device) if get_attack_set: attack.delete_pickle() attack.prepare_dataset() for i in range(50): flag = 1 if i == 49 else 0 print("Epoch %d :" % (i+1)) res_train = attack.train(flag, RESULT_PATH) res_test = attack.test(flag, RESULT_PATH) attack.saveModel(MODELS_PATH) print("Saved Attack Model") return res_train, res_test # black partial def attack_mode1(TARGET_PATH, ATTACK_PATH, device, attack_trainloader, attack_testloader, target_model, attack_model, get_attack_set, num_classes): MODELS_PATH = ATTACK_PATH + "_meminf_attack1.pth" RESULT_PATH = ATTACK_PATH + "_meminf_attack1.p" ATTACK_SETS = ATTACK_PATH + "_meminf_attack_mode1_" attack = attack_for_blackbox(TARGET_PATH, TARGET_PATH, ATTACK_SETS, attack_trainloader, attack_testloader, target_model, target_model, attack_model, device) if get_attack_set: attack.delete_pickle() attack.prepare_dataset() for i in range(50): flag = 1 if i == 49 else 0 print("Epoch %d :" % (i+1)) res_train = attack.train(flag, RESULT_PATH) res_test = attack.test(flag, RESULT_PATH) attack.saveModel(MODELS_PATH) print("Saved Attack Model") return res_train, res_test # white partial def attack_mode2(TARGET_PATH, ATTACK_PATH, device, attack_trainloader, attack_testloader, target_model, attack_model, get_attack_set, num_classes): MODELS_PATH = ATTACK_PATH + "_meminf_attack2.pth" RESULT_PATH = ATTACK_PATH + "_meminf_attack2.p" ATTACK_SETS = ATTACK_PATH + "_meminf_attack_mode2_" attack = attack_for_whitebox(TARGET_PATH, TARGET_PATH, ATTACK_SETS, attack_trainloader, attack_testloader, target_model, target_model, attack_model, device, num_classes) if get_attack_set: attack.delete_pickle() attack.prepare_dataset() for i in range(50): flag = 1 if i == 49 else 0 print("Epoch %d :" % (i+1)) res_train = attack.train(flag, RESULT_PATH) res_test = attack.test(flag, RESULT_PATH) attack.saveModel(MODELS_PATH) print("Saved Attack Model") return res_train, res_test # white shadow def attack_mode3(TARGET_PATH, SHADOW_PATH, ATTACK_PATH, device, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, get_attack_set, num_classes): MODELS_PATH = ATTACK_PATH + "_meminf_attack3.pth" RESULT_PATH = ATTACK_PATH + "_meminf_attack3.p" ATTACK_SETS = ATTACK_PATH + "_meminf_attack_mode3_" attack = attack_for_whitebox(TARGET_PATH, SHADOW_PATH, ATTACK_SETS, attack_trainloader, attack_testloader, target_model, shadow_model, attack_model, device, num_classes) if get_attack_set: attack.delete_pickle() attack.prepare_dataset() for i in range(50): flag = 1 if i == 49 else 0 print("Epoch %d :" % (i+1)) res_train = attack.train(flag, RESULT_PATH) res_test = attack.test(flag, RESULT_PATH) attack.saveModel(MODELS_PATH) print("Saved Attack Model") return res_train, res_test def get_gradient_size(model): gradient_size = [] gradient_list = reversed(list(model.named_parameters())) for name, parameter in gradient_list: if 'weight' in name: gradient_size.append(parameter.shape) return gradient_size
31,612
37.042118
183
py
IVOS-ATNet
IVOS-ATNet-master/eval_real-world.py
from davisinteractive.session import DavisInteractiveSession from davisinteractive import utils as interactive_utils from davisinteractive.dataset import Davis from davisinteractive.metrics import batched_jaccard from libs import custom_transforms as tr, davis2017_torchdataset import os import numpy as np from PIL import Image import csv from datetime import datetime import torch from torch.autograd import Variable from torchvision import transforms from torch.utils.data import DataLoader from libs import utils, utils_torch from libs.analyze_report import analyze_summary from config import Config from networks.atnet import ATnet class Main_tester(object): def __init__(self, config): self.config = config self.Davisclass = Davis(self.config.davis_dataset_dir) self.current_time = datetime.now().strftime('%Y%m%d-%H%M%S') self._palette = Image.open(self.config.palette_dir).getpalette() self.save_res_dir = str() self.save_log_dir = str() self.save_logger = None self.save_csvsummary_dir = str() self.net = ATnet() self.net.cuda() self.net.eval() self.net.load_state_dict(torch.load(self.config.test_load_state_dir)) # To implement ordered test self.scr_indices = [1, 2, 3] self.max_nb_interactions = 8 self.max_time = self.max_nb_interactions * 30 self.scr_samples = [] for v in sorted(self.Davisclass.sets[self.config.test_subset]): for idx in self.scr_indices: self.scr_samples.append((v, idx)) self.img_size, self.num_frames, self.n_objects, self.final_masks, self.tmpdict_siact = None, None, None, None, None self.pad_info, self.hpad1, self.wpad1, self.hpad2, self.wpad2 = None, None, None, None, None def run_for_diverse_metrics(self, ): with torch.no_grad(): for metric in self.config.test_metric_list: if metric == 'J': dir_name = os.path.split(os.path.split(__file__)[0])[1] + '[J]_' + self.current_time elif metric == 'J_AND_F': dir_name = os.path.split(os.path.split(__file__)[0])[1] + '[JF]_' + self.current_time else: dir_name = None print("Impossible metric is contained in config.test_metric_list!") raise NotImplementedError() self.save_res_dir = os.path.join(self.config.test_result_dir, dir_name) utils.mkdir(self.save_res_dir) self.save_csvsummary_dir = os.path.join(self.save_res_dir, 'summary_in_csv.csv') self.save_log_dir = os.path.join(self.save_res_dir, 'test_logs.txt') self.save_logger = utils.logger(self.save_log_dir) self.save_logger.printNlog(dir_name) curr_path = os.path.dirname(os.path.abspath(__file__)) os.system('cp {}/config.py {}/config.py'.format(curr_path, self.save_res_dir)) self.run_IVOS(metric) def run_IVOS(self, metric): seen_seq = {} numseq, tmpseq = 0, '' output_dict = dict() output_dict['average_objs_iou'] = dict() output_dict['average_iact_iou'] = np.zeros(self.max_nb_interactions) output_dict['annotated_frames'] = dict() with open(self.save_csvsummary_dir, mode='a') as csv_file: writer = csv.writer(csv_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL) writer.writerow(['sequence', 'obj_idx', 'scr_idx'] + ['round-' + str(i + 1) for i in range(self.max_nb_interactions)]) with DavisInteractiveSession(host=self.config.test_host, user_key=self.config.test_userkey, davis_root=self.config.davis_dataset_dir, subset=self.config.test_subset, report_save_dir=self.save_res_dir, max_nb_interactions=self.max_nb_interactions, max_time=self.max_time, metric_to_optimize=metric) as sess: sess.connector.service.robot.min_nb_nodes = self.config.test_min_nb_nodes sess.samples = self.scr_samples # sess.samples = [('dog', 3)] while sess.next(): # Get the current iteration scribbles self.sequence, scribbles, first_scribble = sess.get_scribbles(only_last=False) if first_scribble: anno_dict = {'frames': [], 'annotated_masks': [], 'masks_tobe_modified': []} n_interaction = 1 info = Davis.dataset[self.sequence] self.img_size = info['image_size'][::-1] self.num_frames = info['num_frames'] self.n_objects = info['num_objects'] info = None seen_seq[self.sequence] = 1 if self.sequence not in seen_seq.keys() else seen_seq[self.sequence] + 1 scr_id = seen_seq[self.sequence] self.final_masks = np.zeros([self.num_frames, self.img_size[0], self.img_size[1]]) self.pad_info = utils.apply_pad(self.final_masks[0])[1] self.hpad1, self.wpad1 = self.pad_info[0][0], self.pad_info[1][0] self.hpad2, self.wpad2 = self.pad_info[0][1], self.pad_info[1][1] self.h_ds, self.w_ds = int((self.img_size[0] + sum(self.pad_info[0])) / 4), int((self.img_size[1] + sum(self.pad_info[1])) / 4) self.anno_6chEnc_r5_list = [] self.anno_3chEnc_r5_list = [] self.prob_map_of_frames = torch.zeros((self.num_frames, self.n_objects, 4 * self.h_ds, 4 * self.w_ds)).cuda() self.gt_masks = self.Davisclass.load_annotations(self.sequence) IoU_over_eobj = [] else: n_interaction += 1 self.save_logger.printNlog('\nRunning sequence {} in (scribble index: {}) (round: {})' .format(self.sequence, sess.samples[sess.sample_idx][1], n_interaction)) annotated_now = interactive_utils.scribbles.annotated_frames(sess.sample_last_scribble)[0] anno_dict['frames'].append(annotated_now) # Where we save annotated frames anno_dict['masks_tobe_modified'].append(self.final_masks[annotated_now]) # mask before modefied at the annotated frame # Get Predicted mask & Mask decision from pred_mask self.final_masks = self.run_VOS_singleiact(n_interaction, scribbles, anno_dict['frames']) # self.final_mask changes if self.config.test_save_all_segs_option: utils.mkdir( os.path.join(self.save_res_dir, 'result_video', '{}-scr{:02d}/round{:02d}'.format(self.sequence, scr_id, n_interaction))) for fr in range(self.num_frames): savefname = os.path.join(self.save_res_dir, 'result_video', '{}-scr{:02d}/round{:02d}'.format(self.sequence, scr_id, n_interaction), '{:05d}.png'.format(fr)) tmpPIL = Image.fromarray(self.final_masks[fr].astype(np.uint8), 'P') tmpPIL.putpalette(self._palette) tmpPIL.save(savefname) # Submit your prediction sess.submit_masks(self.final_masks) # F, H, W # print sequence name if tmpseq != self.sequence: tmpseq, numseq = self.sequence, numseq + 1 print(str(numseq) + ':' + str(self.sequence) + '-' + str(seen_seq[self.sequence]) + '\n') ## Visualizers and Saver # IoU estimation jaccard = batched_jaccard(self.gt_masks, self.final_masks, average_over_objects=False, nb_objects=self.n_objects ) # frames, objid IoU_over_eobj.append(jaccard) anno_dict['annotated_masks'].append(self.final_masks[annotated_now]) # mask after modefied at the annotated frame if self.max_nb_interactions == len(anno_dict['frames']): # After Lastround -> total 90 iter seq_scrid_name = self.sequence + str(scr_id) # IoU manager IoU_over_eobj = np.stack(IoU_over_eobj, axis=0) # niact,frames,n_obj IoUeveryround_perobj = np.mean(IoU_over_eobj, axis=1) # niact,n_obj output_dict['average_iact_iou'] += np.sum(IoU_over_eobj[list(range(n_interaction)), anno_dict['frames']], axis=-1) output_dict['annotated_frames'][seq_scrid_name] = anno_dict['frames'] # write csv for obj_idx in range(self.n_objects): with open(self.save_csvsummary_dir, mode='a') as csv_file: writer = csv.writer(csv_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL) writer.writerow([self.sequence, str(obj_idx + 1), str(scr_id)] + list(IoUeveryround_perobj[:, obj_idx])) summary = sess.get_global_summary(save_file=self.save_res_dir + '/summary_' + sess.report_name[7:] + '.json') analyze_summary(self.save_res_dir + '/summary_' + sess.report_name[7:] + '.json', metric=metric) # final_IOU = summary['curve'][metric][-1] average_IoU_per_round = summary['curve'][metric][1:-1] torch.cuda.empty_cache() model = None return average_IoU_per_round def run_VOS_singleiact(self, n_interaction, scribbles_data, annotated_frames): annotated_frames_np = np.array(annotated_frames) num_workers = 4 annotated_now = annotated_frames[-1] scribbles_list = scribbles_data['scribbles'] seq_name = scribbles_data['sequence'] output_masks = self.final_masks.copy().astype(np.float64) prop_list = utils.get_prop_list(annotated_frames, annotated_now, self.num_frames, proportion=self.config.test_propagation_proportion) prop_fore = sorted(prop_list)[0] prop_rear = sorted(prop_list)[-1] # Interaction settings pm_ps_ns_3ch_t = [] # n_obj,3,h,w if n_interaction == 1: for obj_id in range(1, self.n_objects + 1): pos_scrimg = utils.scribble_to_image(scribbles_list, annotated_now, obj_id, dilation=self.config.scribble_dilation_param, prev_mask=self.final_masks[annotated_now]) pm_ps_ns_3ch_t.append(np.stack([np.ones_like(pos_scrimg) / 2, pos_scrimg, np.zeros_like(pos_scrimg)], axis=0)) pm_ps_ns_3ch_t = np.stack(pm_ps_ns_3ch_t, axis=0) # n_obj,3,h,w # Image.fromarray((scr_img[:, :, 1] * 255).astype(np.uint8)).save('/home/six/Desktop/CVPRW_figure/judo_obj1_scr.png') else: for obj_id in range(1, self.n_objects + 1): prev_round_input = (self.final_masks[annotated_now] == obj_id).astype(np.float32) # H,W pos_scrimg, neg_scrimg = utils.scribble_to_image(scribbles_list, annotated_now, obj_id, dilation=self.config.scribble_dilation_param, prev_mask=self.final_masks[annotated_now], blur=True, singleimg=False, seperate_pos_neg=True) pm_ps_ns_3ch_t.append(np.stack([prev_round_input, pos_scrimg, neg_scrimg], axis=0)) pm_ps_ns_3ch_t = np.stack(pm_ps_ns_3ch_t, axis=0) # n_obj,3,h,w pm_ps_ns_3ch_t = torch.from_numpy(pm_ps_ns_3ch_t).cuda() if (prop_list[0] != annotated_now) and (prop_list.count(annotated_now) != 2): print(str(prop_list)) raise NotImplementedError print(str(prop_list)) # we made our proplist first backward, and then forward composed_transforms = transforms.Compose([tr.Normalize_ApplymeanvarImage(self.config.mean, self.config.var), tr.ToTensor()]) db_test = davis2017_torchdataset.DAVIS2017(split='val', transform=composed_transforms, root=self.config.davis_dataset_dir, custom_frames=prop_list, seq_name=seq_name, rgb=True, obj_id=None, no_gt=True, retname=True, prev_round_masks=self.final_masks, ) testloader = DataLoader(db_test, batch_size=1, shuffle=False, num_workers=num_workers, pin_memory=True) flag = 0 # 1: propagating backward, 2: propagating forward print('[{:01d} round] processing...'.format(n_interaction)) for ii, batched in enumerate(testloader): # batched : image, scr_img, 0~fr, meta inpdict = dict() operating_frame = int(batched['meta']['frame_id'][0]) for inp in batched: if inp == 'meta': continue inpdict[inp] = Variable(batched[inp]).cuda() inpdict['image'] = inpdict['image'].expand(self.n_objects, -1, -1, -1) #################### Iaction ######################## if operating_frame == annotated_now: # Check the round is on interaction if flag == 0: flag += 1 adjacent_to_anno = True elif flag == 1: flag += 1 adjacent_to_anno = True continue else: raise NotImplementedError pm_ps_ns_3ch_t = torch.nn.ReflectionPad2d(self.pad_info[1] + self.pad_info[0])(pm_ps_ns_3ch_t) inputs = torch.cat([inpdict['image'], pm_ps_ns_3ch_t], dim=1) output_logit, anno_6chEnc_r5 = self.net.forward_ANet(inputs) # [nobj, 1, P_H, P_W], # [n_obj,2048,h/16,w/16] output_prob_anno = torch.sigmoid(output_logit) prob_onehot_t = output_prob_anno[:, 0].detach() anno_3chEnc_r5, _, _, r2_prev_fromanno = self.net.encoder_3ch.forward(inpdict['image']) self.anno_6chEnc_r5_list.append(anno_6chEnc_r5) self.anno_3chEnc_r5_list.append(anno_3chEnc_r5) if len(self.anno_6chEnc_r5_list) != len(annotated_frames): raise NotImplementedError #################### Propagation ######################## else: # Flag [1: propagating backward, 2: propagating forward] if adjacent_to_anno: r2_prev = r2_prev_fromanno predmask_prev = output_prob_anno else: predmask_prev = output_prob_prop adjacent_to_anno = False output_logit, r2_prev = self.net.forward_TNet( self.anno_3chEnc_r5_list, inpdict['image'], self.anno_6chEnc_r5_list, r2_prev, predmask_prev) # [nobj, 1, P_H, P_W] output_prob_prop = torch.sigmoid(output_logit) prob_onehot_t = output_prob_prop[:, 0].detach() smallest_alpha = 0.5 if flag == 1: sorted_frames = annotated_frames_np[annotated_frames_np < annotated_now] if len(sorted_frames) == 0: alpha = 1 else: closest_addianno_frame = np.max(sorted_frames) alpha = smallest_alpha + (1 - smallest_alpha) * ( (operating_frame - closest_addianno_frame) / (annotated_now - closest_addianno_frame)) else: sorted_frames = annotated_frames_np[annotated_frames_np > annotated_now] if len(sorted_frames) == 0: alpha = 1 else: closest_addianno_frame = np.min(sorted_frames) alpha = smallest_alpha + (1 - smallest_alpha) * ( (closest_addianno_frame - operating_frame) / (closest_addianno_frame - annotated_now)) prob_onehot_t = (alpha * prob_onehot_t) + ((1 - alpha) * self.prob_map_of_frames[operating_frame]) # Final mask indexing self.prob_map_of_frames[operating_frame] = prob_onehot_t output_masks[prop_fore:prop_rear + 1] = \ utils_torch.combine_masks_with_batch(self.prob_map_of_frames[prop_fore:prop_rear + 1], n_obj=self.n_objects, th=self.config.test_propth )[:, 0, self.hpad1:-self.hpad2, self.wpad1:-self.wpad2].cpu().numpy().astype(np.float) # f,h,w torch.cuda.empty_cache() return output_masks if __name__ == '__main__': config = Config() os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = str(config.test_gpu_id) tester = Main_tester(config) tester.run_for_diverse_metrics() # try:main_val(model, # Config, # min_nb_nodes= min_nb_nodes, # simplyfied_testset= simplyfied_test,tr(config.test_gpu_id) # metric = metric) # except: continue
17,873
50.2149
147
py
IVOS-ATNet
IVOS-ATNet-master/eval_davis-framework.py
from davisinteractive.session import DavisInteractiveSession from davisinteractive import utils as interactive_utils from davisinteractive.dataset import Davis from davisinteractive.metrics import batched_jaccard from libs import custom_transforms as tr, davis2017_torchdataset import os import numpy as np from PIL import Image import csv from datetime import datetime import torch from torch.autograd import Variable from torchvision import transforms from torch.utils.data import DataLoader from libs import utils, utils_torch from libs.analyze_report import analyze_summary from config import Config from networks.atnet import ATnet class Main_tester(object): def __init__(self, config): self.config = config self.Davisclass = Davis(self.config.davis_dataset_dir) self.current_time = datetime.now().strftime('%Y%m%d-%H%M%S') self._palette = Image.open(self.config.palette_dir).getpalette() self.save_res_dir = str() self.save_log_dir = str() self.save_logger = None self.save_csvsummary_dir = str() self.net = ATnet() self.net.cuda() self.net.eval() self.net.load_state_dict(torch.load(self.config.test_load_state_dir)) # To implement ordered test self.scr_indices = [1, 2, 3] self.max_nb_interactions = 8 self.max_time = self.max_nb_interactions * 30 self.scr_samples = [] for v in sorted(self.Davisclass.sets[self.config.test_subset]): for idx in self.scr_indices: self.scr_samples.append((v, idx)) self.img_size, self.num_frames, self.n_objects, self.final_masks, self.tmpdict_siact = None, None, None, None, None self.pad_info, self.hpad1, self.wpad1, self.hpad2, self.wpad2 = None, None, None, None, None def run_for_diverse_metrics(self, ): with torch.no_grad(): for metric in self.config.test_metric_list: if metric == 'J': dir_name = 'IVOS-ATNet_J_' + self.current_time elif metric == 'J_AND_F': dir_name = 'IVOS-ATNet_JF_' + self.current_time else: dir_name = None print("Impossible metric is contained in config.test_metric_list!") raise NotImplementedError() self.save_res_dir = os.path.join(self.config.test_result_df_dir, dir_name) utils.mkdir(self.save_res_dir) self.save_csvsummary_dir = os.path.join(self.save_res_dir, 'summary_in_csv.csv') self.save_log_dir = os.path.join(self.save_res_dir, 'test_logs.txt') self.save_logger = utils.logger(self.save_log_dir) self.save_logger.printNlog(dir_name) curr_path = os.path.dirname(os.path.abspath(__file__)) os.system('cp {}/config.py {}/config.py'.format(curr_path, self.save_res_dir)) self.run_IVOS(metric) def run_IVOS(self, metric): seen_seq = {} numseq, tmpseq = 0, '' output_dict = dict() output_dict['average_objs_iou'] = dict() output_dict['average_iact_iou'] = np.zeros(self.max_nb_interactions) output_dict['annotated_frames'] = dict() with open(self.save_csvsummary_dir, mode='a') as csv_file: writer = csv.writer(csv_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL) writer.writerow(['sequence', 'obj_idx', 'scr_idx'] + ['round-' + str(i + 1) for i in range(self.max_nb_interactions)]) with DavisInteractiveSession(host=self.config.test_host, user_key=self.config.test_userkey, davis_root=self.config.davis_dataset_dir, subset=self.config.test_subset, report_save_dir=self.save_res_dir, max_nb_interactions=self.max_nb_interactions, max_time=self.max_time, metric_to_optimize=metric) as sess: sess.connector.service.robot.min_nb_nodes = self.config.test_min_nb_nodes sess.samples = self.scr_samples # sess.samples = [('dog', 3)] while sess.next(): # Get the current iteration scribbles self.sequence, scribbles, first_scribble = sess.get_scribbles(only_last=False) if first_scribble: anno_dict = {'frames': [], 'annotated_masks': [], 'masks_tobe_modified': []} n_interaction = 1 info = Davis.dataset[self.sequence] self.img_size = info['image_size'][::-1] self.num_frames = info['num_frames'] self.n_objects = info['num_objects'] info = None seen_seq[self.sequence] = 1 if self.sequence not in seen_seq.keys() else seen_seq[self.sequence] + 1 scr_id = seen_seq[self.sequence] self.final_masks = np.zeros([self.num_frames, self.img_size[0], self.img_size[1]]) self.pad_info = utils.apply_pad(self.final_masks[0])[1] self.hpad1, self.wpad1 = self.pad_info[0][0], self.pad_info[1][0] self.hpad2, self.wpad2 = self.pad_info[0][1], self.pad_info[1][1] self.h_ds, self.w_ds = int((self.img_size[0] + sum(self.pad_info[0])) / 4), int((self.img_size[1] + sum(self.pad_info[1])) / 4) self.anno_6chEnc_r5_list = [] self.anno_3chEnc_r5_list = [] self.prob_map_of_frames = torch.zeros((self.num_frames, self.n_objects, 4 * self.h_ds, 4 * self.w_ds)).cuda() self.gt_masks = self.Davisclass.load_annotations(self.sequence) IoU_over_eobj = [] else: n_interaction += 1 self.save_logger.printNlog('\nRunning sequence {} in (scribble index: {}) (round: {})' .format(self.sequence, sess.samples[sess.sample_idx][1], n_interaction)) annotated_now = interactive_utils.scribbles.annotated_frames(sess.sample_last_scribble)[0] anno_dict['frames'].append(annotated_now) # Where we save annotated frames anno_dict['masks_tobe_modified'].append(self.final_masks[annotated_now]) # mask before modefied at the annotated frame # Get Predicted mask & Mask decision from pred_mask self.final_masks = self.run_VOS_singleiact(n_interaction, scribbles, anno_dict['frames']) # self.final_mask changes if self.config.test_save_all_segs_option: utils.mkdir( os.path.join(self.save_res_dir, 'result_video', '{}-scr{:02d}/round{:02d}'.format(self.sequence, scr_id, n_interaction))) for fr in range(self.num_frames): savefname = os.path.join(self.save_res_dir, 'result_video', '{}-scr{:02d}/round{:02d}'.format(self.sequence, scr_id, n_interaction), '{:05d}.png'.format(fr)) tmpPIL = Image.fromarray(self.final_masks[fr].astype(np.uint8), 'P') tmpPIL.putpalette(self._palette) tmpPIL.save(savefname) # Submit your prediction sess.submit_masks(self.final_masks) # F, H, W # print sequence name if tmpseq != self.sequence: tmpseq, numseq = self.sequence, numseq + 1 print(str(numseq) + ':' + str(self.sequence) + '-' + str(seen_seq[self.sequence]) + '\n') ## Visualizers and Saver # IoU estimation jaccard = batched_jaccard(self.gt_masks, self.final_masks, average_over_objects=False, nb_objects=self.n_objects ) # frames, objid IoU_over_eobj.append(jaccard) anno_dict['annotated_masks'].append(self.final_masks[annotated_now]) # mask after modefied at the annotated frame if self.max_nb_interactions == len(anno_dict['frames']): # After Lastround -> total 90 iter seq_scrid_name = self.sequence + str(scr_id) # IoU manager IoU_over_eobj = np.stack(IoU_over_eobj, axis=0) # niact,frames,n_obj IoUeveryround_perobj = np.mean(IoU_over_eobj, axis=1) # niact,n_obj output_dict['average_iact_iou'] += np.sum(IoU_over_eobj[list(range(n_interaction)), anno_dict['frames']], axis=-1) output_dict['annotated_frames'][seq_scrid_name] = anno_dict['frames'] # write csv for obj_idx in range(self.n_objects): with open(self.save_csvsummary_dir, mode='a') as csv_file: writer = csv.writer(csv_file, delimiter=',', quotechar='"', quoting=csv.QUOTE_MINIMAL) writer.writerow([self.sequence, str(obj_idx + 1), str(scr_id)] + list(IoUeveryround_perobj[:, obj_idx])) summary = sess.get_global_summary(save_file=self.save_res_dir + '/summary_' + sess.report_name[7:] + '.json') analyze_summary(self.save_res_dir + '/summary_' + sess.report_name[7:] + '.json', metric=metric) # final_IOU = summary['curve'][metric][-1] average_IoU_per_round = summary['curve'][metric][1:-1] torch.cuda.empty_cache() model = None return average_IoU_per_round def run_VOS_singleiact(self, n_interaction, scribbles_data, annotated_frames): annotated_frames_np = np.array(annotated_frames) num_workers = 4 annotated_now = annotated_frames[-1] scribbles_list = scribbles_data['scribbles'] seq_name = scribbles_data['sequence'] output_masks = self.final_masks.copy().astype(np.float64) prop_list = utils.get_prop_list(annotated_frames, annotated_now, self.num_frames, proportion=self.config.test_propagation_proportion) prop_fore = sorted(prop_list)[0] prop_rear = sorted(prop_list)[-1] # Interaction settings pm_ps_ns_3ch_t = [] # n_obj,3,h,w if n_interaction == 1: for obj_id in range(1, self.n_objects + 1): pos_scrimg = utils.scribble_to_image(scribbles_list, annotated_now, obj_id, dilation=self.config.scribble_dilation_param, prev_mask=self.final_masks[annotated_now]) pm_ps_ns_3ch_t.append(np.stack([np.ones_like(pos_scrimg) / 2, pos_scrimg, np.zeros_like(pos_scrimg)], axis=0)) pm_ps_ns_3ch_t = np.stack(pm_ps_ns_3ch_t, axis=0) # n_obj,3,h,w # Image.fromarray((scr_img[:, :, 1] * 255).astype(np.uint8)).save('/home/six/Desktop/CVPRW_figure/judo_obj1_scr.png') else: for obj_id in range(1, self.n_objects + 1): prev_round_input = (self.final_masks[annotated_now] == obj_id).astype(np.float32) # H,W pos_scrimg, neg_scrimg = utils.scribble_to_image(scribbles_list, annotated_now, obj_id, dilation=self.config.scribble_dilation_param, prev_mask=self.final_masks[annotated_now], blur=True, singleimg=False, seperate_pos_neg=True) pm_ps_ns_3ch_t.append(np.stack([prev_round_input, pos_scrimg, neg_scrimg], axis=0)) pm_ps_ns_3ch_t = np.stack(pm_ps_ns_3ch_t, axis=0) # n_obj,3,h,w pm_ps_ns_3ch_t = torch.from_numpy(pm_ps_ns_3ch_t).cuda() if (prop_list[0] != annotated_now) and (prop_list.count(annotated_now) != 2): print(str(prop_list)) raise NotImplementedError print(str(prop_list)) # we made our proplist first backward, and then forward composed_transforms = transforms.Compose([tr.Normalize_ApplymeanvarImage(self.config.mean, self.config.var), tr.ToTensor()]) db_test = davis2017_torchdataset.DAVIS2017(split='val', transform=composed_transforms, root=self.config.davis_dataset_dir, custom_frames=prop_list, seq_name=seq_name, rgb=True, obj_id=None, no_gt=True, retname=True, prev_round_masks=self.final_masks, ) testloader = DataLoader(db_test, batch_size=1, shuffle=False, num_workers=num_workers, pin_memory=True) flag = 0 # 1: propagating backward, 2: propagating forward print('[{:01d} round] processing...'.format(n_interaction)) for ii, batched in enumerate(testloader): # batched : image, scr_img, 0~fr, meta inpdict = dict() operating_frame = int(batched['meta']['frame_id'][0]) for inp in batched: if inp == 'meta': continue inpdict[inp] = Variable(batched[inp]).cuda() inpdict['image'] = inpdict['image'].expand(self.n_objects, -1, -1, -1) #################### Iaction ######################## if operating_frame == annotated_now: # Check the round is on interaction if flag == 0: flag += 1 adjacent_to_anno = True elif flag == 1: flag += 1 adjacent_to_anno = True continue else: raise NotImplementedError pm_ps_ns_3ch_t = torch.nn.ReflectionPad2d(self.pad_info[1] + self.pad_info[0])(pm_ps_ns_3ch_t) inputs = torch.cat([inpdict['image'], pm_ps_ns_3ch_t], dim=1) output_logit, anno_6chEnc_r5 = self.net.forward_ANet(inputs) # [nobj, 1, P_H, P_W], # [n_obj,2048,h/16,w/16] output_prob_anno = torch.sigmoid(output_logit) prob_onehot_t = output_prob_anno[:, 0].detach() anno_3chEnc_r5, _, _, r2_prev_fromanno = self.net.encoder_3ch.forward(inpdict['image']) self.anno_6chEnc_r5_list.append(anno_6chEnc_r5) self.anno_3chEnc_r5_list.append(anno_3chEnc_r5) if len(self.anno_6chEnc_r5_list) != len(annotated_frames): raise NotImplementedError #################### Propagation ######################## else: # Flag [1: propagating backward, 2: propagating forward] if adjacent_to_anno: r2_prev = r2_prev_fromanno predmask_prev = output_prob_anno else: predmask_prev = output_prob_prop adjacent_to_anno = False output_logit, r2_prev = self.net.forward_TNet( self.anno_3chEnc_r5_list, inpdict['image'], self.anno_6chEnc_r5_list, r2_prev, predmask_prev) # [nobj, 1, P_H, P_W] output_prob_prop = torch.sigmoid(output_logit) prob_onehot_t = output_prob_prop[:, 0].detach() smallest_alpha = 0.5 if flag == 1: sorted_frames = annotated_frames_np[annotated_frames_np < annotated_now] if len(sorted_frames) == 0: alpha = 1 else: closest_addianno_frame = np.max(sorted_frames) alpha = smallest_alpha + (1 - smallest_alpha) * ( (operating_frame - closest_addianno_frame) / (annotated_now - closest_addianno_frame)) else: sorted_frames = annotated_frames_np[annotated_frames_np > annotated_now] if len(sorted_frames) == 0: alpha = 1 else: closest_addianno_frame = np.min(sorted_frames) alpha = smallest_alpha + (1 - smallest_alpha) * ( (closest_addianno_frame - operating_frame) / (closest_addianno_frame - annotated_now)) prob_onehot_t = (alpha * prob_onehot_t) + ((1 - alpha) * self.prob_map_of_frames[operating_frame]) # Final mask indexing self.prob_map_of_frames[operating_frame] = prob_onehot_t output_masks[prop_fore:prop_rear + 1] = \ utils_torch.combine_masks_with_batch(self.prob_map_of_frames[prop_fore:prop_rear + 1], n_obj=self.n_objects, th=self.config.test_propth )[:, 0, self.hpad1:-self.hpad2, self.wpad1:-self.wpad2].cpu().numpy().astype(np.float) # f,h,w torch.cuda.empty_cache() return output_masks if __name__ == '__main__': config = Config() os.environ["CUDA_DEVICE_ORDER"] = "PCI_BUS_ID" os.environ["CUDA_VISIBLE_DEVICES"] = str(config.test_gpu_id) tester = Main_tester(config) tester.run_for_diverse_metrics() # try:main_val(model, # Config, # min_nb_nodes= min_nb_nodes, # simplyfied_testset= simplyfied_test,tr(config.test_gpu_id) # metric = metric) # except: continue
17,800
50.005731
147
py
IVOS-ATNet
IVOS-ATNet-master/networks/ltm_transfer.py
import torch import torch.nn as nn import torch.nn.functional as F class LTM_transfer(nn.Module): def __init__(self,md=4, stride=1): super(LTM_transfer, self).__init__() self.md = md #displacement (default = 4pixels) self.range = (md*2 + 1) ** 2 #(default = (4x2+1)**2 = 81) self.grid = None self.Channelwise_sum = None d_u = torch.linspace(-self.md * stride, self.md * stride, 2 * self.md + 1).view(1, -1).repeat((2 * self.md + 1, 1)).view(self.range, 1) # (25,1) d_v = torch.linspace(-self.md * stride, self.md * stride, 2 * self.md + 1).view(-1, 1).repeat((1, 2 * self.md + 1)).view(self.range, 1) # (25,1) self.d = torch.cat((d_u, d_v), dim=1).cuda() # (25,2) def L2normalize(self, x, d=1): eps = 1e-6 norm = x ** 2 norm = norm.sum(dim=d, keepdim=True) + eps norm = norm ** (0.5) return (x/norm) def UniformGrid(self, Input): ''' Make uniform grid :param Input: tensor(N,C,H,W) :return grid: (1,2,H,W) ''' # torchHorizontal = torch.linspace(-1.0, 1.0, W).view(1, 1, 1, W).expand(N, 1, H, W) # torchVertical = torch.linspace(-1.0, 1.0, H).view(1, 1, H, 1).expand(N, 1, H, W) # grid = torch.cat([torchHorizontal, torchVertical], 1).cuda() _, _, H, W = Input.size() # mesh grid xx = torch.arange(0, W).view(1, 1, 1, W).expand(1, 1, H, W) yy = torch.arange(0, H).view(1, 1, H, 1).expand(1, 1, H, W) grid = torch.cat((xx, yy), 1).float() if Input.is_cuda: grid = grid.cuda() return grid def warp(self, x, BM_d): vgrid = self.grid + BM_d # [N2HW] # [(2d+1)^2, 2, H, W] # scale grid to [-1,1] vgrid[:, 0, :, :] = 2.0 * vgrid[:, 0, :, :] / max(x.size(3) - 1, 1) - 1.0 vgrid[:, 1, :, :] = 2.0 * vgrid[:, 1, :, :] / max(x.size(2) - 1, 1) - 1.0 vgrid = vgrid.permute(0, 2, 3, 1) output = nn.functional.grid_sample(x, vgrid, mode='bilinear', padding_mode = 'border') #800MB memory occupied (d=2,C=64,H=256,W=256) mask = torch.autograd.Variable(torch.ones(x.size())).cuda() mask = nn.functional.grid_sample(mask, vgrid) #300MB memory occpied (d=2,C=64,H=256,W=256) mask = mask.masked_fill_(mask<0.999,0) mask = mask.masked_fill_(mask>0,1) return output * mask def forward(self,sim_feature, f_map, apply_softmax_on_simfeature = True): ''' Return bilateral cost volume(Set of bilateral correlation map) :param sim_feature: Correlation feature based on operating frame's HW (N,D2,H,W) :param f_map: Previous frame mask (N,1,H,W) :return Correlation Cost: (N,(2d+1)^2,H,W) ''' # feature1 = self.L2normalize(feature1) # feature2 = self.L2normalize(feature2) B_size,C_size,H_size,W_size = f_map.size() if self.grid is None: # Initialize first uniform grid self.grid = self.UniformGrid(f_map) if H_size != self.grid.size(2) or W_size != self.grid.size(3): # Update uniform grid to fit on input tensor shape self.grid = self.UniformGrid(f_map) # Displacement volume (N,(2d+1)^2,2,H,W) d = (i,j) , i in [-md,md] & j in [-md,md] D_vol = self.d.view(self.range, 2, 1, 1).expand(-1, -1, H_size, W_size) # [(2d+1)^2, 2, H, W] if apply_softmax_on_simfeature: sim_feature = F.softmax(sim_feature, dim=1) # B,D^2,H,W f_map = self.warp(f_map.transpose(0, 1).expand(self.range,-1,-1,-1), D_vol).transpose(0, 1) # B,D^2,H,W f_map = torch.sum(torch.mul(sim_feature, f_map),dim=1, keepdim=True) # B,1,H,W return f_map # B,1,H,W
3,767
38.25
153
py
IVOS-ATNet
IVOS-ATNet-master/networks/atnet.py
import torch import torch.nn as nn import torch.nn.functional as F from networks.deeplab.aspp import ASPP from networks.deeplab.backbone.resnet import SEResNet50 from networks.correlation_package.correlation import Correlation from networks.ltm_transfer import LTM_transfer class ATnet(nn.Module): def __init__(self, pretrained=1, resfix=False, corr_displacement=4, corr_stride=2): super(ATnet, self).__init__() print("Constructing ATnet architecture..") self.encoder_6ch = Encoder_6ch(resfix) self.encoder_3ch = Encoder_3ch(resfix) self.indicator_encoder = ConverterEncoder() # self.decoder_iact = Decoder() self.decoder_prop = Decoder_prop() self.ltm_local_affinity = Correlation(pad_size=corr_displacement * corr_stride, kernel_size=1, max_displacement=corr_displacement * corr_stride, stride1=1, stride2=corr_stride, corr_multiply=1) self.ltm_transfer = LTM_transfer(md=corr_displacement, stride=corr_stride) self.prev_conv1x1 = nn.Conv2d(256, 256, kernel_size=1, padding=0) # 1/4, 256 self.conv1x1 = nn.Conv2d(2048*2, 2048, kernel_size=1, padding=0) # 1/16, 2048 self.refer_weight = None self._initialize_weights(pretrained) def forward_ANet(self, x): # Bx4xHxW to Bx1xHxW r5, r4, r3, r2 = self.encoder_6ch(x) estimated_mask, m2 = self.decoder_iact(r5, r3, r2, only_return_feature=False) r5_indicator = self.indicator_encoder(r5, m2) return estimated_mask, r5_indicator def forward_TNet(self, anno_propEnc_r5_list, targframe_3ch, anno_iactEnc_r5_list, r2_prev, predmask_prev, debug_f_mask = False): #1/16, 2048 f_targ, _, r3_targ, r2_targ = self.encoder_3ch(targframe_3ch) f_mask_r5 = self.correlation_global_transfer(anno_propEnc_r5_list, f_targ, anno_iactEnc_r5_list) # 1/16, 2048 r2_targ_c = self.prev_conv1x1(r2_targ) r2_prev = self.prev_conv1x1(r2_prev) f_mask_r2 = self.correlation_local_transfer(r2_prev, r2_targ_c, predmask_prev) # 1/4, 1 [B,1,H/4,W/4] r5_concat = torch.cat([f_targ, f_mask_r5], dim=1) # 1/16, 2048*2 r5_concat = self.conv1x1(r5_concat) estimated_mask, m2 = self.decoder_prop(r5_concat, r3_targ, r2_targ, f_mask_r2) if not debug_f_mask: return estimated_mask, r2_targ else: return estimated_mask, r2_targ, f_mask_r2 def correlation_global_transfer(self, anno_feature_list, targ_feature, anno_indicator_feature_list ): ''' :param anno_feature_list: [B,C,H,W] x list (N values in list) :param targ_feature: [B,C,H,W] :param anno_indicator_feature_list: [B,C,H,W] x list (N values in list) :return targ_mask_feature: [B,C,H,W] ''' b, c, h, w = anno_indicator_feature_list[0].size() # b means n_objs targ_feature = targ_feature.view(b, c, h * w) # [B, C, HxW] n_features = len(anno_feature_list) anno_feature = [] for f_idx in range(n_features): anno_feature.append(anno_feature_list[f_idx].view(b, c, h * w).transpose(1, 2)) # [B, HxW', C] anno_feature = torch.cat(anno_feature, dim=1) # [B, NxHxW', C] sim_feature = torch.bmm(anno_feature, targ_feature) # [B, NxHxW', HxW] sim_feature = F.softmax(sim_feature, dim=2) / n_features # [B, NxHxW', HxW] anno_indicator_feature = [] for f_idx in range(n_features): anno_indicator_feature.append(anno_indicator_feature_list[f_idx].view(b, c, h * w)) # [B, C, HxW'] anno_indicator_feature = torch.cat(anno_indicator_feature, dim=-1) # [B, C, NxHxW'] targ_mask_feature = torch.bmm(anno_indicator_feature, sim_feature) # [B, C, HxW] targ_mask_feature = targ_mask_feature.view(b, c, h, w) return targ_mask_feature def correlation_local_transfer(self, r2_prev, r2_targ, predmask_prev): ''' :param r2_prev: [B,C,H,W] :param r2_targ: [B,C,H,W] :param predmask_prev: [B,1,4*H,4*W] :return targ_mask_feature_r2: [B,1,H,W] ''' predmask_prev = F.interpolate(predmask_prev, scale_factor=0.25, mode='bilinear',align_corners=True) # B,1,H,W sim_feature = self.ltm_local_affinity.forward(r2_targ,r2_prev,) # B,D^2,H,W sim_feature = F.softmax(sim_feature, dim=2) # B,D^2,H,W predmask_targ = self.ltm_transfer.forward(sim_feature, predmask_prev, apply_softmax_on_simfeature = False) # B,1,H,W return predmask_targ def _initialize_weights(self, pretrained): for m in self.modules(): if pretrained: break else: if isinstance(m, nn.Conv2d): m.weight.data.normal_(0, 0.001) if m.bias is not None: m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.Linear): m.weight.data.normal_(0, 0.01) m.bias.data.zero_() class Encoder_3ch(nn.Module): # T-Net Encoder def __init__(self, resfix): super(Encoder_3ch, self).__init__() self.conv0_3ch = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=True) resnet = SEResNet50(output_stride=16, BatchNorm=nn.BatchNorm2d, pretrained=True) self.bn1 = resnet.bn1 self.relu = resnet.relu # 1/2, 64 self.maxpool = resnet.maxpool self.res2 = resnet.layer1 # 1/4, 256 self.res3 = resnet.layer2 # 1/8, 512 self.res4 = resnet.layer3 # 1/16, 1024 self.res5 = resnet.layer4 # 1/16, 2048 # freeze BNs if resfix: for m in self.modules(): if isinstance(m, nn.BatchNorm2d): for p in m.parameters(): p.requires_grad = False def forward(self, x): x = self.conv0_3ch(x) # 1/2, 64 x = self.bn1(x) c1 = self.relu(x) # 1/2, 64 x = self.maxpool(c1) # 1/4, 64 r2 = self.res2(x) # 1/4, 256 r3 = self.res3(r2) # 1/8, 512 r4 = self.res4(r3) # 1/16, 1024 r5 = self.res5(r4) # 1/16, 2048 return r5, r4, r3, r2 def forward_r2(self,x): x = self.conv0_3ch(x) # 1/2, 64 x = self.bn1(x) c1 = self.relu(x) # 1/2, 64 x = self.maxpool(c1) # 1/4, 64 r2 = self.res2(x) # 1/4, 256 return r2 class Encoder_6ch(nn.Module): # A-Net Encoder def __init__(self, resfix): super(Encoder_6ch, self).__init__() self.conv0_6ch = nn.Conv2d(6, 64, kernel_size=7, stride=2, padding=3, bias=True) resnet = SEResNet50(output_stride=16, BatchNorm=nn.BatchNorm2d, pretrained=True) self.bn1 = resnet.bn1 self.relu = resnet.relu # 1/2, 64 self.maxpool = resnet.maxpool self.res2 = resnet.layer1 # 1/4, 256 self.res3 = resnet.layer2 # 1/8, 512 self.res4 = resnet.layer3 # 1/16, 1024 self.res5 = resnet.layer4 # 1/16, 2048 # freeze BNs if resfix: for m in self.modules(): if isinstance(m, nn.BatchNorm2d): for p in m.parameters(): p.requires_grad = False def forward(self, x): x = self.conv0_6ch(x) # 1/2, 64 x = self.bn1(x) c1 = self.relu(x) # 1/2, 64 x = self.maxpool(c1) # 1/4, 64 r2 = self.res2(x) # 1/4, 256 r3 = self.res3(r2) # 1/8, 512 r4 = self.res4(r3) # 1/16, 1024 r5 = self.res5(r4) # 1/16, 2048 return r5, r4, r3, r2 class Decoder(nn.Module): # A-Net Decoder def __init__(self): super(Decoder, self).__init__() mdim = 256 self.aspp_decoder = ASPP(backbone='res', output_stride=16, BatchNorm=nn.BatchNorm2d, pretrained=1) self.convG0 = nn.Conv2d(2048, mdim, kernel_size=3, padding=1) self.convG1 = nn.Conv2d(mdim, mdim, kernel_size=3, padding=1) self.convG2 = nn.Conv2d(mdim, mdim, kernel_size=3, padding=1) self.RF3 = Refine(512, mdim) # 1/16 -> 1/8 self.RF2 = Refine(256, mdim) # 1/8 -> 1/4 self.lastconv = nn.Sequential(nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=1, bias=False), nn.BatchNorm2d(256), nn.ReLU(), nn.Dropout(0.5), nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False), nn.BatchNorm2d(256), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(256, 1, kernel_size=1, stride=1)) def forward(self, r5, r3_targ, r2_targ, only_return_feature = False): aspp_out = self.aspp_decoder(r5) #1/16 mdim aspp_out = F.interpolate(aspp_out, scale_factor=4, mode='bilinear',align_corners=True) #1/4 mdim m4 = self.convG0(F.relu(r5)) # out: # 1/16, mdim m4 = self.convG1(F.relu(m4)) # out: # 1/16, mdim m4 = self.convG2(F.relu(m4)) # out: # 1/16, mdim m3 = self.RF3(r3_targ, m4) # out: 1/8, mdim m2 = self.RF2(r2_targ, m3) # out: 1/4, mdim m2 = torch.cat((m2, aspp_out), dim=1) # out: 1/4, mdim*2 if only_return_feature: return m2 x = self.lastconv(m2) x = F.interpolate(x, scale_factor=4, mode='bilinear', align_corners=True) return x, m2 class Decoder_prop(nn.Module): # T-Net Decoder def __init__(self): super(Decoder_prop, self).__init__() mdim = 256 self.aspp_decoder = ASPP(backbone='res', output_stride=16, BatchNorm=nn.BatchNorm2d, pretrained=1) self.convG0 = nn.Conv2d(2048, mdim, kernel_size=3, padding=1) self.convG1 = nn.Conv2d(mdim, mdim, kernel_size=3, padding=1) self.convG2 = nn.Conv2d(mdim, mdim, kernel_size=3, padding=1) self.RF3 = Refine(512, mdim) # 1/16 -> 1/8 self.RF2 = Refine(256, mdim) # 1/8 -> 1/4 self.lastconv = nn.Sequential(nn.Conv2d(512, 256, kernel_size=3, stride=1, padding=1, bias=False), nn.BatchNorm2d(256), nn.ReLU(), nn.Dropout(0.5), nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False), nn.BatchNorm2d(256), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(256, 1, kernel_size=1, stride=1)) def forward(self, r5, r3_targ, r2_targ, f_mask_r2): aspp_out = self.aspp_decoder(r5) #1/16 mdim aspp_out = F.interpolate(aspp_out, scale_factor=4, mode='bilinear',align_corners=True) #1/4 mdim m4 = self.convG0(F.relu(r5)) # out: # 1/16, mdim m4 = self.convG1(F.relu(m4)) # out: # 1/16, mdim m4 = self.convG2(F.relu(m4)) # out: # 1/16, mdim m3 = self.RF3(r3_targ, m4) # out: 1/8, mdim m3 = m3 + 0.5 * F.interpolate(f_mask_r2, scale_factor=0.5, mode='bilinear',align_corners=True) #1/4 mdim m2 = self.RF2(r2_targ, m3) # out: 1/4, mdim m2 = m2 + 0.5 * f_mask_r2 m2 = torch.cat((m2, aspp_out), dim=1) # out: 1/4, mdim*2 x = self.lastconv(m2) x = F.interpolate(x, scale_factor=4, mode='bilinear', align_corners=True) return x, m2 class ConverterEncoder(nn.Module): def __init__(self): super(ConverterEncoder, self).__init__() # [1/4, 512] to [1/8, 1024] downsample1 = nn.Sequential(nn.Conv2d(512, 1024, kernel_size=1, stride=2, bias=False), nn.BatchNorm2d(1024), ) self.block1 = SEBottleneck(512, 256, stride = 2, downsample = downsample1) # [1/8, 1024] to [1/16, 2048] downsample2 = nn.Sequential(nn.Conv2d(1024, 2048, kernel_size=1, stride=2, bias=False), nn.BatchNorm2d(2048), ) self.block2 = SEBottleneck(1024, 512, stride = 2, downsample=downsample2) self.conv1x1 = nn.Conv2d(2048 * 2, 2048, kernel_size=1, padding=0) # 1/16, 2048 def forward(self, r5, m2): ''' :param r5: 1/16, 2048 :param m2: 1/4, 512 :return: ''' x = self.block1(m2) x = self.block2(x) x = torch.cat((x,r5),dim=1) x = self.conv1x1(x) return x class SEBottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, BatchNorm=nn.BatchNorm2d): super(SEBottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = BatchNorm(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, dilation=dilation, padding=dilation, bias=False) self.bn2 = BatchNorm(planes) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = BatchNorm(planes * self.expansion) self.relu = nn.ReLU(inplace=True) # SE self.global_pool = nn.AdaptiveAvgPool2d(1) self.conv_down = nn.Conv2d( planes * 4, planes // 4, kernel_size=1, bias=False) self.conv_up = nn.Conv2d( planes // 4, planes * 4, kernel_size=1, bias=False) self.sig = nn.Sigmoid() self.downsample = downsample self.stride = stride self.dilation = dilation 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) out1 = self.global_pool(out) out1 = self.conv_down(out1) out1 = self.relu(out1) out1 = self.conv_up(out1) out1 = self.sig(out1) if self.downsample is not None: residual = self.downsample(x) res = out1 * out + residual res = self.relu(res) return res class SELayer(nn.Module): def __init__(self, channel, reduction=16): super(SELayer, self).__init__() self.avg_pool = nn.AdaptiveAvgPool2d(1) self.fc = nn.Sequential( nn.Linear(channel, channel // reduction, bias=False), nn.ReLU(inplace=True), nn.Linear(channel // reduction, channel, bias=False), nn.Sigmoid() ) def forward(self, x): b, c, _, _ = x.size() y = self.avg_pool(x).view(b, c) y = self.fc(y).view(b, c, 1, 1) return x * y.expand_as(x) class Refine(nn.Module): def __init__(self, inplanes, planes, scale_factor=2): super(Refine, self).__init__() self.convFS1 = nn.Conv2d(inplanes, planes, kernel_size=3, padding=1) self.convFS2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1) self.convFS3 = nn.Conv2d(planes, planes, kernel_size=3, padding=1) self.convMM1 = nn.Conv2d(planes, planes, kernel_size=3, padding=1) self.convMM2 = nn.Conv2d(planes, planes, kernel_size=3, padding=1) self.scale_factor = scale_factor def forward(self, f, pm): s = self.convFS1(f) sr = self.convFS2(F.relu(s)) sr = self.convFS3(F.relu(sr)) s = s + sr m = s + F.interpolate(pm, scale_factor=self.scale_factor, mode='bilinear',align_corners=True) mr = self.convMM1(F.relu(m)) mr = self.convMM2(F.relu(mr)) m = m + mr return m
16,091
38.153285
144
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/aspp.py
import math import torch import torch.nn as nn import torch.nn.functional as F from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d class _ASPPModule(nn.Module): def __init__(self, inplanes, planes, kernel_size, padding, dilation, BatchNorm, pretrained): super(_ASPPModule, self).__init__() self.atrous_conv = nn.Conv2d(inplanes, planes, kernel_size=kernel_size, stride=1, padding=padding, dilation=dilation, bias=False) self.bn = BatchNorm(planes) self.relu = nn.ReLU() self._init_weight(pretrained) def forward(self, x): x = self.atrous_conv(x) x = self.bn(x) return self.relu(x) def _init_weight(self,pretrained): for m in self.modules(): if pretrained: break else: if isinstance(m, nn.Conv2d): torch.nn.init.kaiming_normal_(m.weight) elif isinstance(m, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() class ASPP(nn.Module): def __init__(self, backbone, output_stride, BatchNorm, pretrained): super(ASPP, self).__init__() if backbone == 'drn': inplanes = 512 elif backbone == 'mobilenet': inplanes = 320 else: inplanes = 2048 if output_stride == 16: dilations = [1, 6, 12, 18] elif output_stride == 8: dilations = [1, 12, 24, 36] else: raise NotImplementedError self.aspp1 = _ASPPModule(inplanes, 256, 1, padding=0, dilation=dilations[0], BatchNorm=BatchNorm, pretrained=pretrained) self.aspp2 = _ASPPModule(inplanes, 256, 3, padding=dilations[1], dilation=dilations[1], BatchNorm=BatchNorm, pretrained=pretrained) self.aspp3 = _ASPPModule(inplanes, 256, 3, padding=dilations[2], dilation=dilations[2], BatchNorm=BatchNorm, pretrained=pretrained) self.aspp4 = _ASPPModule(inplanes, 256, 3, padding=dilations[3], dilation=dilations[3], BatchNorm=BatchNorm, pretrained=pretrained) self.global_avg_pool = nn.Sequential(nn.AdaptiveAvgPool2d((1, 1)), nn.Conv2d(inplanes, 256, 1, stride=1, bias=False), BatchNorm(256), nn.ReLU()) self.conv1 = nn.Conv2d(1280, 256, 1, bias=False) self.bn1 = BatchNorm(256) self.relu = nn.ReLU() self.dropout = nn.Dropout(0.5) self._init_weight(pretrained) def forward(self, x): x1 = self.aspp1(x) x2 = self.aspp2(x) x3 = self.aspp3(x) x4 = self.aspp4(x) x5 = self.global_avg_pool(x) # if type(x4.size()[2]) != int: # tmpsize = (x4.size()[2].item(),x4.size()[3].item()) # else: # tmpsize = (x4.size()[2],x4.size()[3]) # x5 = F.interpolate(x5, size=(14,14), mode='bilinear', align_corners=True) x5 = F.interpolate(x5, size=x4.size()[2:], mode='bilinear', align_corners=True) x = torch.cat((x1, x2, x3, x4, x5), dim=1) x = self.conv1(x) x = self.bn1(x) x = self.relu(x) return self.dropout(x) def _init_weight(self,pretrained): for m in self.modules(): if pretrained: break else: 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)) torch.nn.init.kaiming_normal_(m.weight) elif isinstance(m, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def build_aspp(backbone, output_stride, BatchNorm,pretrained): return ASPP(backbone, output_stride, BatchNorm, pretrained)
4,257
38.425926
139
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/decoder.py
import math import torch import torch.nn as nn import torch.nn.functional as F from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d class Decoder(nn.Module): def __init__(self, num_classes, backbone, BatchNorm): super(Decoder, self).__init__() if backbone == 'resnet' or backbone == 'drn': low_level_inplanes = 256 elif backbone == 'xception': low_level_inplanes = 128 elif backbone == 'mobilenet': low_level_inplanes = 24 else: raise NotImplementedError self.conv1 = nn.Conv2d(low_level_inplanes, 48, 1, bias=False) self.bn1 = BatchNorm(48) self.relu = nn.ReLU() self.last_conv = nn.Sequential(nn.Conv2d(304, 256, kernel_size=3, stride=1, padding=1, bias=False), BatchNorm(256), nn.ReLU(), nn.Dropout(0.5), nn.Conv2d(256, 256, kernel_size=3, stride=1, padding=1, bias=False), BatchNorm(256), nn.ReLU(), nn.Dropout(0.1), nn.Conv2d(256, num_classes, kernel_size=1, stride=1)) self._init_weight() def forward(self, x, low_level_feat): low_level_feat = self.conv1(low_level_feat) low_level_feat = self.bn1(low_level_feat) low_level_feat = self.relu(low_level_feat) x = F.interpolate(x, size=low_level_feat.size()[2:], mode='bilinear', align_corners=True) x = torch.cat((x, low_level_feat), dim=1) x = self.last_conv(x) return x def _init_weight(self): for m in self.modules(): if isinstance(m, nn.Conv2d): torch.nn.init.kaiming_normal_(m.weight) elif isinstance(m, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def build_decoder(num_classes, backbone, BatchNorm): return Decoder(num_classes, backbone, BatchNorm)
2,280
39.017544
107
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/deeplab.py
import torch import torch.nn as nn import torch.nn.functional as F from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d from networks.deeplab.aspp import build_aspp from networks.deeplab.decoder import build_decoder from networks.deeplab.backbone import build_backbone class DeepLab(nn.Module): def __init__(self, backbone='resnet', output_stride=16, num_classes=21, sync_bn=True, freeze_bn=False): super(DeepLab, self).__init__() if backbone == 'drn': output_stride = 8 if sync_bn == True: BatchNorm = SynchronizedBatchNorm2d else: BatchNorm = nn.BatchNorm2d self.backbone = build_backbone(backbone, output_stride, BatchNorm) self.aspp = build_aspp(backbone, output_stride, BatchNorm) self.decoder = build_decoder(num_classes, backbone, BatchNorm) if freeze_bn: self.freeze_bn() def forward(self, input): x, low_level_feat = self.backbone(input) x = self.aspp(x) x = self.decoder(x, low_level_feat) x = F.interpolate(x, size=input.size()[2:], mode='bilinear', align_corners=True) return x def freeze_bn(self): for m in self.modules(): if isinstance(m, SynchronizedBatchNorm2d): m.eval() elif isinstance(m, nn.BatchNorm2d): m.eval() def get_1x_lr_params(self): modules = [self.backbone] for i in range(len(modules)): for m in modules[i].named_modules(): if isinstance(m[1], nn.Conv2d) or isinstance(m[1], SynchronizedBatchNorm2d) \ or isinstance(m[1], nn.BatchNorm2d): for p in m[1].parameters(): if p.requires_grad: yield p def get_10x_lr_params(self): modules = [self.aspp, self.decoder] for i in range(len(modules)): for m in modules[i].named_modules(): if isinstance(m[1], nn.Conv2d) or isinstance(m[1], SynchronizedBatchNorm2d) \ or isinstance(m[1], nn.BatchNorm2d): for p in m[1].parameters(): if p.requires_grad: yield p if __name__ == "__main__": model = DeepLab(backbone='mobilenet', output_stride=16) model.eval() input = torch.rand(1, 3, 513, 513) output = model(input) print(output.size())
2,493
33.638889
93
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/backbone/resnet.py
import math import torch.nn as nn import torch.utils.model_zoo as model_zoo from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, BatchNorm=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = BatchNorm(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, dilation=dilation, padding=dilation, bias=False) self.bn2 = BatchNorm(planes) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = BatchNorm(planes * self.expansion) self.relu = nn.ReLU(inplace=True) self.downsample = downsample self.stride = stride self.dilation = dilation 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 SEBottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, dilation=1, downsample=None, BatchNorm=None): super(SEBottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = BatchNorm(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, dilation=dilation, padding=dilation, bias=False) self.bn2 = BatchNorm(planes) self.conv3 = nn.Conv2d(planes, planes * self.expansion, kernel_size=1, bias=False) self.bn3 = BatchNorm(planes * self.expansion) self.relu = nn.ReLU(inplace=True) # SE self.global_pool = nn.AdaptiveAvgPool2d(1) self.conv_down = nn.Conv2d( planes * 4, planes // 4, kernel_size=1, bias=False) self.conv_up = nn.Conv2d( planes // 4, planes * 4, kernel_size=1, bias=False) self.sig = nn.Sigmoid() self.downsample = downsample self.stride = stride self.dilation = dilation 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) out1 = self.global_pool(out) out1 = self.conv_down(out1) out1 = self.relu(out1) out1 = self.conv_up(out1) out1 = self.sig(out1) if self.downsample is not None: residual = self.downsample(x) res = out1 * out + residual res = self.relu(res) return res class ResNet(nn.Module): def __init__(self, block, layers, output_stride, BatchNorm, pretrained=True, modelname = 'res101'): self.inplanes = 64 self.modelname = modelname super(ResNet, self).__init__() blocks = [1, 2, 4] if output_stride == 16: strides = [1, 2, 2, 1] dilations = [1, 1, 1, 2] elif output_stride == 8: strides = [1, 2, 1, 1] dilations = [1, 1, 2, 4] else: raise NotImplementedError # Modules self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = BatchNorm(64) 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], stride=strides[0], dilation=dilations[0], BatchNorm=BatchNorm) self.layer2 = self._make_layer(block, 128, layers[1], stride=strides[1], dilation=dilations[1], BatchNorm=BatchNorm) self.layer3 = self._make_layer(block, 256, layers[2], stride=strides[2], dilation=dilations[2], BatchNorm=BatchNorm) self.layer4 = self._make_MG_unit(block, 512, blocks=blocks, stride=strides[3], dilation=dilations[3], BatchNorm=BatchNorm) # self.layer4 = self._make_layer(block, 512, layers[3], stride=strides[3], dilation=dilations[3], BatchNorm=BatchNorm) self._init_weight() if pretrained: self._load_pretrained_model() def _make_layer(self, block, planes, blocks, stride=1, dilation=1, BatchNorm=None): 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), BatchNorm(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, dilation, downsample, BatchNorm)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, dilation=dilation, BatchNorm=BatchNorm)) return nn.Sequential(*layers) def _make_MG_unit(self, block, planes, blocks, stride=1, dilation=1, BatchNorm=None): 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), BatchNorm(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, dilation=blocks[0]*dilation, downsample=downsample, BatchNorm=BatchNorm)) self.inplanes = planes * block.expansion for i in range(1, len(blocks)): layers.append(block(self.inplanes, planes, stride=1, dilation=blocks[i]*dilation, BatchNorm=BatchNorm)) return nn.Sequential(*layers) def forward(self, input): x = self.conv1(input) x = self.bn1(x) x = self.relu(x) x = self.maxpool(x) x = self.layer1(x) #256 128 128 low_level_feat = x x = self.layer2(x) #512 64 64 x = self.layer3(x) #1024 32 32 x = self.layer4(x) #2048 32 32 return x, low_level_feat def _init_weight(self): 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, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _load_pretrained_model(self): if self.modelname =='res101': pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/resnet101-5d3b4d8f.pth') elif self.modelname == 'res50': pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/resnet50-19c8e357.pth') elif self.modelname == 'SEres50': pretrain_dict = model_zoo.load_url('https://download.pytorch.org/models/resnet50-19c8e357.pth') else: raise NotImplementedError model_dict = {} state_dict = self.state_dict() for k, v in pretrain_dict.items(): if k in state_dict: model_dict[k] = v state_dict.update(model_dict) self.load_state_dict(state_dict) def ResNet101(output_stride, BatchNorm, pretrained=True,): """Constructs a ResNet-101 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(Bottleneck, [3, 4, 23, 3], output_stride, BatchNorm, pretrained=pretrained, modelname='res101') return model def ResNet50(output_stride, BatchNorm, pretrained=True): """Constructs a ResNet-50 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(Bottleneck, [3, 4, 6, 3], output_stride, BatchNorm, pretrained=pretrained, modelname='res50') return model def SEResNet50(output_stride, BatchNorm, pretrained=True): """Constructs a ResNet-50 model. Args: pretrained (bool): If True, returns a model pre-trained on ImageNet """ model = ResNet(SEBottleneck, [3, 4, 6, 3], output_stride, BatchNorm, pretrained=pretrained, modelname='SEres50') return model if __name__ == "__main__": import torch model = ResNet50(BatchNorm=nn.BatchNorm2d, pretrained=True, output_stride=16) input = torch.rand(1, 3, 512, 512) output, low_level_feat = model(input) print(output.size()) print(low_level_feat.size())
9,076
36.979079
130
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/backbone/drn.py
import torch.nn as nn import math import torch.utils.model_zoo as model_zoo from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d webroot = 'https://tigress-web.princeton.edu/~fy/drn/models/' model_urls = { 'resnet50': 'https://download.pytorch.org/models/resnet50-19c8e357.pth', 'drn-c-26': webroot + 'drn_c_26-ddedf421.pth', 'drn-c-42': webroot + 'drn_c_42-9d336e8c.pth', 'drn-c-58': webroot + 'drn_c_58-0a53a92c.pth', 'drn-d-22': webroot + 'drn_d_22-4bd2f8ea.pth', 'drn-d-38': webroot + 'drn_d_38-eebb45f0.pth', 'drn-d-54': webroot + 'drn_d_54-0e0534ff.pth', 'drn-d-105': webroot + 'drn_d_105-12b40979.pth' } def conv3x3(in_planes, out_planes, stride=1, padding=1, dilation=1): return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride, padding=padding, bias=False, dilation=dilation) class BasicBlock(nn.Module): expansion = 1 def __init__(self, inplanes, planes, stride=1, downsample=None, dilation=(1, 1), residual=True, BatchNorm=None): super(BasicBlock, self).__init__() self.conv1 = conv3x3(inplanes, planes, stride, padding=dilation[0], dilation=dilation[0]) self.bn1 = BatchNorm(planes) self.relu = nn.ReLU(inplace=True) self.conv2 = conv3x3(planes, planes, padding=dilation[1], dilation=dilation[1]) self.bn2 = BatchNorm(planes) self.downsample = downsample self.stride = stride self.residual = residual 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) if self.residual: out += residual out = self.relu(out) return out class Bottleneck(nn.Module): expansion = 4 def __init__(self, inplanes, planes, stride=1, downsample=None, dilation=(1, 1), residual=True, BatchNorm=None): super(Bottleneck, self).__init__() self.conv1 = nn.Conv2d(inplanes, planes, kernel_size=1, bias=False) self.bn1 = BatchNorm(planes) self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=dilation[1], bias=False, dilation=dilation[1]) self.bn2 = BatchNorm(planes) self.conv3 = nn.Conv2d(planes, planes * 4, kernel_size=1, bias=False) self.bn3 = BatchNorm(planes * 4) 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 DRN(nn.Module): def __init__(self, block, layers, arch='D', channels=(16, 32, 64, 128, 256, 512, 512, 512), BatchNorm=None): super(DRN, self).__init__() self.inplanes = channels[0] self.out_dim = channels[-1] self.arch = arch if arch == 'C': self.conv1 = nn.Conv2d(3, channels[0], kernel_size=7, stride=1, padding=3, bias=False) self.bn1 = BatchNorm(channels[0]) self.relu = nn.ReLU(inplace=True) self.layer1 = self._make_layer( BasicBlock, channels[0], layers[0], stride=1, BatchNorm=BatchNorm) self.layer2 = self._make_layer( BasicBlock, channels[1], layers[1], stride=2, BatchNorm=BatchNorm) elif arch == 'D': self.layer0 = nn.Sequential( nn.Conv2d(3, channels[0], kernel_size=7, stride=1, padding=3, bias=False), BatchNorm(channels[0]), nn.ReLU(inplace=True) ) self.layer1 = self._make_conv_layers( channels[0], layers[0], stride=1, BatchNorm=BatchNorm) self.layer2 = self._make_conv_layers( channels[1], layers[1], stride=2, BatchNorm=BatchNorm) self.layer3 = self._make_layer(block, channels[2], layers[2], stride=2, BatchNorm=BatchNorm) self.layer4 = self._make_layer(block, channels[3], layers[3], stride=2, BatchNorm=BatchNorm) self.layer5 = self._make_layer(block, channels[4], layers[4], dilation=2, new_level=False, BatchNorm=BatchNorm) self.layer6 = None if layers[5] == 0 else \ self._make_layer(block, channels[5], layers[5], dilation=4, new_level=False, BatchNorm=BatchNorm) if arch == 'C': self.layer7 = None if layers[6] == 0 else \ self._make_layer(BasicBlock, channels[6], layers[6], dilation=2, new_level=False, residual=False, BatchNorm=BatchNorm) self.layer8 = None if layers[7] == 0 else \ self._make_layer(BasicBlock, channels[7], layers[7], dilation=1, new_level=False, residual=False, BatchNorm=BatchNorm) elif arch == 'D': self.layer7 = None if layers[6] == 0 else \ self._make_conv_layers(channels[6], layers[6], dilation=2, BatchNorm=BatchNorm) self.layer8 = None if layers[7] == 0 else \ self._make_conv_layers(channels[7], layers[7], dilation=1, BatchNorm=BatchNorm) self._init_weight() def _init_weight(self): 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, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1, dilation=1, new_level=True, residual=True, BatchNorm=None): assert dilation == 1 or dilation % 2 == 0 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), BatchNorm(planes * block.expansion), ) layers = list() layers.append(block( self.inplanes, planes, stride, downsample, dilation=(1, 1) if dilation == 1 else ( dilation // 2 if new_level else dilation, dilation), residual=residual, BatchNorm=BatchNorm)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, residual=residual, dilation=(dilation, dilation), BatchNorm=BatchNorm)) return nn.Sequential(*layers) def _make_conv_layers(self, channels, convs, stride=1, dilation=1, BatchNorm=None): modules = [] for i in range(convs): modules.extend([ nn.Conv2d(self.inplanes, channels, kernel_size=3, stride=stride if i == 0 else 1, padding=dilation, bias=False, dilation=dilation), BatchNorm(channels), nn.ReLU(inplace=True)]) self.inplanes = channels return nn.Sequential(*modules) def forward(self, x): if self.arch == 'C': x = self.conv1(x) x = self.bn1(x) x = self.relu(x) elif self.arch == 'D': x = self.layer0(x) x = self.layer1(x) x = self.layer2(x) x = self.layer3(x) low_level_feat = x x = self.layer4(x) x = self.layer5(x) if self.layer6 is not None: x = self.layer6(x) if self.layer7 is not None: x = self.layer7(x) if self.layer8 is not None: x = self.layer8(x) return x, low_level_feat class DRN_A(nn.Module): def __init__(self, block, layers, BatchNorm=None): self.inplanes = 64 super(DRN_A, self).__init__() self.out_dim = 512 * block.expansion self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False) self.bn1 = BatchNorm(64) 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], BatchNorm=BatchNorm) self.layer2 = self._make_layer(block, 128, layers[1], stride=2, BatchNorm=BatchNorm) self.layer3 = self._make_layer(block, 256, layers[2], stride=1, dilation=2, BatchNorm=BatchNorm) self.layer4 = self._make_layer(block, 512, layers[3], stride=1, dilation=4, BatchNorm=BatchNorm) self._init_weight() def _init_weight(self): 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, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _make_layer(self, block, planes, blocks, stride=1, dilation=1, BatchNorm=None): 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), BatchNorm(planes * block.expansion), ) layers = [] layers.append(block(self.inplanes, planes, stride, downsample, BatchNorm=BatchNorm)) self.inplanes = planes * block.expansion for i in range(1, blocks): layers.append(block(self.inplanes, planes, dilation=(dilation, dilation, ), BatchNorm=BatchNorm)) 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) return x def drn_a_50(BatchNorm, pretrained=True): model = DRN_A(Bottleneck, [3, 4, 6, 3], BatchNorm=BatchNorm) if pretrained: model.load_state_dict(model_zoo.load_url(model_urls['resnet50'])) return model def drn_c_26(BatchNorm, pretrained=True): model = DRN(BasicBlock, [1, 1, 2, 2, 2, 2, 1, 1], arch='C', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-c-26']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_c_42(BatchNorm, pretrained=True): model = DRN(BasicBlock, [1, 1, 3, 4, 6, 3, 1, 1], arch='C', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-c-42']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_c_58(BatchNorm, pretrained=True): model = DRN(Bottleneck, [1, 1, 3, 4, 6, 3, 1, 1], arch='C', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-c-58']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_d_22(BatchNorm, pretrained=True): model = DRN(BasicBlock, [1, 1, 2, 2, 2, 2, 1, 1], arch='D', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-d-22']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_d_24(BatchNorm, pretrained=True): model = DRN(BasicBlock, [1, 1, 2, 2, 2, 2, 2, 2], arch='D', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-d-24']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_d_38(BatchNorm, pretrained=True): model = DRN(BasicBlock, [1, 1, 3, 4, 6, 3, 1, 1], arch='D', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-d-38']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_d_40(BatchNorm, pretrained=True): model = DRN(BasicBlock, [1, 1, 3, 4, 6, 3, 2, 2], arch='D', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-d-40']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_d_54(BatchNorm, pretrained=True): model = DRN(Bottleneck, [1, 1, 3, 4, 6, 3, 1, 1], arch='D', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-d-54']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model def drn_d_105(BatchNorm, pretrained=True): model = DRN(Bottleneck, [1, 1, 3, 4, 23, 3, 1, 1], arch='D', BatchNorm=BatchNorm) if pretrained: pretrained = model_zoo.load_url(model_urls['drn-d-105']) del pretrained['fc.weight'] del pretrained['fc.bias'] model.load_state_dict(pretrained) return model if __name__ == "__main__": import torch model = drn_a_50(BatchNorm=nn.BatchNorm2d, pretrained=True) input = torch.rand(1, 3, 512, 512) output, low_level_feat = model(input) print(output.size()) print(low_level_feat.size())
14,657
35.372208
100
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/backbone/xception.py
import math import torch import torch.nn as nn import torch.nn.functional as F import torch.utils.model_zoo as model_zoo from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d def fixed_padding(inputs, kernel_size, dilation): kernel_size_effective = kernel_size + (kernel_size - 1) * (dilation - 1) pad_total = kernel_size_effective - 1 pad_beg = pad_total // 2 pad_end = pad_total - pad_beg padded_inputs = F.pad(inputs, (pad_beg, pad_end, pad_beg, pad_end)) return padded_inputs class SeparableConv2d(nn.Module): def __init__(self, inplanes, planes, kernel_size=3, stride=1, dilation=1, bias=False, BatchNorm=None): super(SeparableConv2d, self).__init__() self.conv1 = nn.Conv2d(inplanes, inplanes, kernel_size, stride, 0, dilation, groups=inplanes, bias=bias) self.bn = BatchNorm(inplanes) self.pointwise = nn.Conv2d(inplanes, planes, 1, 1, 0, 1, 1, bias=bias) def forward(self, x): x = fixed_padding(x, self.conv1.kernel_size[0], dilation=self.conv1.dilation[0]) x = self.conv1(x) x = self.bn(x) x = self.pointwise(x) return x class Block(nn.Module): def __init__(self, inplanes, planes, reps, stride=1, dilation=1, BatchNorm=None, start_with_relu=True, grow_first=True, is_last=False): super(Block, self).__init__() if planes != inplanes or stride != 1: self.skip = nn.Conv2d(inplanes, planes, 1, stride=stride, bias=False) self.skipbn = BatchNorm(planes) else: self.skip = None self.relu = nn.ReLU(inplace=True) rep = [] filters = inplanes if grow_first: rep.append(self.relu) rep.append(SeparableConv2d(inplanes, planes, 3, 1, dilation, BatchNorm=BatchNorm)) rep.append(BatchNorm(planes)) filters = planes for i in range(reps - 1): rep.append(self.relu) rep.append(SeparableConv2d(filters, filters, 3, 1, dilation, BatchNorm=BatchNorm)) rep.append(BatchNorm(filters)) if not grow_first: rep.append(self.relu) rep.append(SeparableConv2d(inplanes, planes, 3, 1, dilation, BatchNorm=BatchNorm)) rep.append(BatchNorm(planes)) if stride != 1: rep.append(self.relu) rep.append(SeparableConv2d(planes, planes, 3, 2, BatchNorm=BatchNorm)) rep.append(BatchNorm(planes)) if stride == 1 and is_last: rep.append(self.relu) rep.append(SeparableConv2d(planes, planes, 3, 1, BatchNorm=BatchNorm)) rep.append(BatchNorm(planes)) if not start_with_relu: rep = rep[1:] self.rep = nn.Sequential(*rep) def forward(self, inp): x = self.rep(inp) if self.skip is not None: skip = self.skip(inp) skip = self.skipbn(skip) else: skip = inp x = x + skip return x class AlignedXception(nn.Module): """ Modified Alighed Xception """ def __init__(self, output_stride, BatchNorm, pretrained=True): super(AlignedXception, self).__init__() if output_stride == 16: entry_block3_stride = 2 middle_block_dilation = 1 exit_block_dilations = (1, 2) elif output_stride == 8: entry_block3_stride = 1 middle_block_dilation = 2 exit_block_dilations = (2, 4) else: raise NotImplementedError # Entry flow self.conv1 = nn.Conv2d(3, 32, 3, stride=2, padding=1, bias=False) self.bn1 = BatchNorm(32) self.relu = nn.ReLU(inplace=True) self.conv2 = nn.Conv2d(32, 64, 3, stride=1, padding=1, bias=False) self.bn2 = BatchNorm(64) self.block1 = Block(64, 128, reps=2, stride=2, BatchNorm=BatchNorm, start_with_relu=False) self.block2 = Block(128, 256, reps=2, stride=2, BatchNorm=BatchNorm, start_with_relu=False, grow_first=True) self.block3 = Block(256, 728, reps=2, stride=entry_block3_stride, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True, is_last=True) # Middle flow self.block4 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block5 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block6 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block7 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block8 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block9 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block10 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block11 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block12 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block13 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block14 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block15 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block16 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block17 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block18 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) self.block19 = Block(728, 728, reps=3, stride=1, dilation=middle_block_dilation, BatchNorm=BatchNorm, start_with_relu=True, grow_first=True) # Exit flow self.block20 = Block(728, 1024, reps=2, stride=1, dilation=exit_block_dilations[0], BatchNorm=BatchNorm, start_with_relu=True, grow_first=False, is_last=True) self.conv3 = SeparableConv2d(1024, 1536, 3, stride=1, dilation=exit_block_dilations[1], BatchNorm=BatchNorm) self.bn3 = BatchNorm(1536) self.conv4 = SeparableConv2d(1536, 1536, 3, stride=1, dilation=exit_block_dilations[1], BatchNorm=BatchNorm) self.bn4 = BatchNorm(1536) self.conv5 = SeparableConv2d(1536, 2048, 3, stride=1, dilation=exit_block_dilations[1], BatchNorm=BatchNorm) self.bn5 = BatchNorm(2048) # Init weights self._init_weight() # Load pretrained model if pretrained: self._load_pretrained_model() def forward(self, x): # Entry flow x = self.conv1(x) x = self.bn1(x) x = self.relu(x) x = self.conv2(x) x = self.bn2(x) x = self.relu(x) x = self.block1(x) # add relu here x = self.relu(x) low_level_feat = x x = self.block2(x) x = self.block3(x) # Middle flow x = self.block4(x) x = self.block5(x) x = self.block6(x) x = self.block7(x) x = self.block8(x) x = self.block9(x) x = self.block10(x) x = self.block11(x) x = self.block12(x) x = self.block13(x) x = self.block14(x) x = self.block15(x) x = self.block16(x) x = self.block17(x) x = self.block18(x) x = self.block19(x) # Exit flow x = self.block20(x) x = self.relu(x) x = self.conv3(x) x = self.bn3(x) x = self.relu(x) x = self.conv4(x) x = self.bn4(x) x = self.relu(x) x = self.conv5(x) x = self.bn5(x) x = self.relu(x) return x, low_level_feat def _init_weight(self): 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, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() def _load_pretrained_model(self): pretrain_dict = model_zoo.load_url('http://data.lip6.fr/cadene/pretrainedmodels/xception-b5690688.pth') model_dict = {} state_dict = self.state_dict() for k, v in pretrain_dict.items(): if k in model_dict: if 'pointwise' in k: v = v.unsqueeze(-1).unsqueeze(-1) if k.startswith('block11'): model_dict[k] = v model_dict[k.replace('block11', 'block12')] = v model_dict[k.replace('block11', 'block13')] = v model_dict[k.replace('block11', 'block14')] = v model_dict[k.replace('block11', 'block15')] = v model_dict[k.replace('block11', 'block16')] = v model_dict[k.replace('block11', 'block17')] = v model_dict[k.replace('block11', 'block18')] = v model_dict[k.replace('block11', 'block19')] = v elif k.startswith('block12'): model_dict[k.replace('block12', 'block20')] = v elif k.startswith('bn3'): model_dict[k] = v model_dict[k.replace('bn3', 'bn4')] = v elif k.startswith('conv4'): model_dict[k.replace('conv4', 'conv5')] = v elif k.startswith('bn4'): model_dict[k.replace('bn4', 'bn5')] = v else: model_dict[k] = v state_dict.update(model_dict) self.load_state_dict(state_dict) if __name__ == "__main__": import torch model = AlignedXception(BatchNorm=nn.BatchNorm2d, pretrained=True, output_stride=16) input = torch.rand(1, 3, 512, 512) output, low_level_feat = model(input) print(output.size()) print(low_level_feat.size())
11,561
39.145833
116
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/backbone/mobilenet.py
import torch import torch.nn.functional as F import torch.nn as nn import math from networks.deeplab.sync_batchnorm.batchnorm import SynchronizedBatchNorm2d import torch.utils.model_zoo as model_zoo def conv_bn(inp, oup, stride, BatchNorm): return nn.Sequential( nn.Conv2d(inp, oup, 3, stride, 1, bias=False), BatchNorm(oup), nn.ReLU6(inplace=True) ) def fixed_padding(inputs, kernel_size, dilation): kernel_size_effective = kernel_size + (kernel_size - 1) * (dilation - 1) pad_total = kernel_size_effective - 1 pad_beg = pad_total // 2 pad_end = pad_total - pad_beg padded_inputs = F.pad(inputs, (pad_beg, pad_end, pad_beg, pad_end)) return padded_inputs class InvertedResidual(nn.Module): def __init__(self, inp, oup, stride, dilation, expand_ratio, BatchNorm): super(InvertedResidual, self).__init__() self.stride = stride assert stride in [1, 2] hidden_dim = round(inp * expand_ratio) self.use_res_connect = self.stride == 1 and inp == oup self.kernel_size = 3 self.dilation = dilation if expand_ratio == 1: self.conv = nn.Sequential( # dw nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 0, dilation, groups=hidden_dim, bias=False), BatchNorm(hidden_dim), nn.ReLU6(inplace=True), # pw-linear nn.Conv2d(hidden_dim, oup, 1, 1, 0, 1, 1, bias=False), BatchNorm(oup), ) else: self.conv = nn.Sequential( # pw nn.Conv2d(inp, hidden_dim, 1, 1, 0, 1, bias=False), BatchNorm(hidden_dim), nn.ReLU6(inplace=True), # dw nn.Conv2d(hidden_dim, hidden_dim, 3, stride, 0, dilation, groups=hidden_dim, bias=False), BatchNorm(hidden_dim), nn.ReLU6(inplace=True), # pw-linear nn.Conv2d(hidden_dim, oup, 1, 1, 0, 1, bias=False), BatchNorm(oup), ) def forward(self, x): x_pad = fixed_padding(x, self.kernel_size, dilation=self.dilation) if self.use_res_connect: x = x + self.conv(x_pad) else: x = self.conv(x_pad) return x class MobileNetV2(nn.Module): def __init__(self, output_stride=8, BatchNorm=None, width_mult=1., pretrained=True): super(MobileNetV2, self).__init__() block = InvertedResidual input_channel = 32 current_stride = 1 rate = 1 interverted_residual_setting = [ # t, c, n, s [1, 16, 1, 1], [6, 24, 2, 2], [6, 32, 3, 2], [6, 64, 4, 2], [6, 96, 3, 1], [6, 160, 3, 2], [6, 320, 1, 1], ] # building first layer input_channel = int(input_channel * width_mult) self.features = [conv_bn(3, input_channel, 2, BatchNorm)] current_stride *= 2 # building inverted residual blocks for t, c, n, s in interverted_residual_setting: if current_stride == output_stride: stride = 1 dilation = rate rate *= s else: stride = s dilation = 1 current_stride *= s output_channel = int(c * width_mult) for i in range(n): if i == 0: self.features.append(block(input_channel, output_channel, stride, dilation, t, BatchNorm)) else: self.features.append(block(input_channel, output_channel, 1, dilation, t, BatchNorm)) input_channel = output_channel self.features = nn.Sequential(*self.features) self._initialize_weights() if pretrained: self._load_pretrained_model() self.low_level_features = self.features[0:4] self.high_level_features = self.features[4:] def forward(self, x): low_level_feat = self.low_level_features(x) x = self.high_level_features(low_level_feat) return x, low_level_feat def _load_pretrained_model(self): pretrain_dict = model_zoo.load_url('http://jeff95.me/models/mobilenet_v2-6a65762b.pth') model_dict = {} state_dict = self.state_dict() for k, v in pretrain_dict.items(): if k in state_dict: model_dict[k] = v state_dict.update(model_dict) self.load_state_dict(state_dict) def _initialize_weights(self): 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)) torch.nn.init.kaiming_normal_(m.weight) elif isinstance(m, SynchronizedBatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() elif isinstance(m, nn.BatchNorm2d): m.weight.data.fill_(1) m.bias.data.zero_() if __name__ == "__main__": input = torch.rand(1, 3, 512, 512) model = MobileNetV2(output_stride=16, BatchNorm=nn.BatchNorm2d) output, low_level_feat = model(input) print(output.size()) print(low_level_feat.size())
5,398
34.519737
110
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/sync_batchnorm/replicate.py
# -*- coding: utf-8 -*- # File : replicate.py # Author : Jiayuan Mao # Email : maojiayuan@gmail.com # Date : 27/01/2018 # # This file is part of Synchronized-BatchNorm-PyTorch. # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch # Distributed under MIT License. import functools from torch.nn.parallel.data_parallel import DataParallel __all__ = [ 'CallbackContext', 'execute_replication_callbacks', 'DataParallelWithCallback', 'patch_replication_callback' ] class CallbackContext(object): pass def execute_replication_callbacks(modules): """ Execute an replication callback `__data_parallel_replicate__` on each module created by original replication. The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)` Note that, as all modules are isomorphism, we assign each sub-module with a context (shared among multiple copies of this module on different devices). Through this context, different copies can share some information. We guarantee that the callback on the master copy (the first copy) will be called ahead of calling the callback of any slave copies. """ master_copy = modules[0] nr_modules = len(list(master_copy.modules())) ctxs = [CallbackContext() for _ in range(nr_modules)] for i, module in enumerate(modules): for j, m in enumerate(module.modules()): if hasattr(m, '__data_parallel_replicate__'): m.__data_parallel_replicate__(ctxs[j], i) class DataParallelWithCallback(DataParallel): """ Data Parallel with a replication callback. An replication callback `__data_parallel_replicate__` of each module will be invoked after being created by original `replicate` function. The callback will be invoked with arguments `__data_parallel_replicate__(ctx, copy_id)` Examples: > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False) > sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1]) # sync_bn.__data_parallel_replicate__ will be invoked. """ def replicate(self, module, device_ids): modules = super(DataParallelWithCallback, self).replicate(module, device_ids) execute_replication_callbacks(modules) return modules def patch_replication_callback(data_parallel): """ Monkey-patch an existing `DataParallel` object. Add the replication callback. Useful when you have customized `DataParallel` implementation. Examples: > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False) > sync_bn = DataParallel(sync_bn, device_ids=[0, 1]) > patch_replication_callback(sync_bn) # this is equivalent to > sync_bn = SynchronizedBatchNorm1d(10, eps=1e-5, affine=False) > sync_bn = DataParallelWithCallback(sync_bn, device_ids=[0, 1]) """ assert isinstance(data_parallel, DataParallel) old_replicate = data_parallel.replicate @functools.wraps(old_replicate) def new_replicate(module, device_ids): modules = old_replicate(module, device_ids) execute_replication_callbacks(modules) return modules data_parallel.replicate = new_replicate
3,218
35.579545
115
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/sync_batchnorm/unittest.py
# -*- coding: utf-8 -*- # File : unittest.py # Author : Jiayuan Mao # Email : maojiayuan@gmail.com # Date : 27/01/2018 # # This file is part of Synchronized-BatchNorm-PyTorch. # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch # Distributed under MIT License. import unittest import numpy as np from torch.autograd import Variable def as_numpy(v): if isinstance(v, Variable): v = v.data return v.cpu().numpy() class TorchTestCase(unittest.TestCase): def assertTensorClose(self, a, b, atol=1e-3, rtol=1e-3): npa, npb = as_numpy(a), as_numpy(b) self.assertTrue( np.allclose(npa, npb, atol=atol), 'Tensor close check failed\n{}\n{}\nadiff={}, rdiff={}'.format(a, b, np.abs(npa - npb).max(), np.abs((npa - npb) / np.fmax(npa, 1e-5)).max()) )
834
26.833333
157
py
IVOS-ATNet
IVOS-ATNet-master/networks/deeplab/sync_batchnorm/batchnorm.py
# -*- coding: utf-8 -*- # File : batchnorm.py # Author : Jiayuan Mao # Email : maojiayuan@gmail.com # Date : 27/01/2018 # # This file is part of Synchronized-BatchNorm-PyTorch. # https://github.com/vacancy/Synchronized-BatchNorm-PyTorch # Distributed under MIT License. import collections import torch import torch.nn.functional as F from torch.nn.modules.batchnorm import _BatchNorm from torch.nn.parallel._functions import ReduceAddCoalesced, Broadcast from .comm import SyncMaster __all__ = ['SynchronizedBatchNorm1d', 'SynchronizedBatchNorm2d', 'SynchronizedBatchNorm3d'] def _sum_ft(tensor): """sum over the first and last dimention""" return tensor.sum(dim=0).sum(dim=-1) def _unsqueeze_ft(tensor): """add new dementions at the front and the tail""" return tensor.unsqueeze(0).unsqueeze(-1) _ChildMessage = collections.namedtuple('_ChildMessage', ['sum', 'ssum', 'sum_size']) _MasterMessage = collections.namedtuple('_MasterMessage', ['sum', 'inv_std']) class _SynchronizedBatchNorm(_BatchNorm): def __init__(self, num_features, eps=1e-5, momentum=0.1, affine=True): super(_SynchronizedBatchNorm, self).__init__(num_features, eps=eps, momentum=momentum, affine=affine) self._sync_master = SyncMaster(self._data_parallel_master) self._is_parallel = False self._parallel_id = None self._slave_pipe = None def forward(self, input): # If it is not parallel computation or is in evaluation mode, use PyTorch's implementation. if not (self._is_parallel and self.training): return F.batch_norm( input, self.running_mean, self.running_var, self.weight, self.bias, self.training, self.momentum, self.eps) # Resize the input to (B, C, -1). input_shape = input.size() input = input.view(input.size(0), self.num_features, -1) # Compute the sum and square-sum. sum_size = input.size(0) * input.size(2) input_sum = _sum_ft(input) input_ssum = _sum_ft(input ** 2) # Reduce-and-broadcast the statistics. if self._parallel_id == 0: mean, inv_std = self._sync_master.run_master(_ChildMessage(input_sum, input_ssum, sum_size)) else: mean, inv_std = self._slave_pipe.run_slave(_ChildMessage(input_sum, input_ssum, sum_size)) # Compute the output. if self.affine: # MJY:: Fuse the multiplication for speed. output = (input - _unsqueeze_ft(mean)) * _unsqueeze_ft(inv_std * self.weight) + _unsqueeze_ft(self.bias) else: output = (input - _unsqueeze_ft(mean)) * _unsqueeze_ft(inv_std) # Reshape it. return output.view(input_shape) def __data_parallel_replicate__(self, ctx, copy_id): self._is_parallel = True self._parallel_id = copy_id # parallel_id == 0 means master device. if self._parallel_id == 0: ctx.sync_master = self._sync_master else: self._slave_pipe = ctx.sync_master.register_slave(copy_id) def _data_parallel_master(self, intermediates): """Reduce the sum and square-sum, compute the statistics, and broadcast it.""" # Always using same "device order" makes the ReduceAdd operation faster. # Thanks to:: Tete Xiao (http://tetexiao.com/) intermediates = sorted(intermediates, key=lambda i: i[1].sum.get_device()) to_reduce = [i[1][:2] for i in intermediates] to_reduce = [j for i in to_reduce for j in i] # flatten target_gpus = [i[1].sum.get_device() for i in intermediates] sum_size = sum([i[1].sum_size for i in intermediates]) sum_, ssum = ReduceAddCoalesced.apply(target_gpus[0], 2, *to_reduce) mean, inv_std = self._compute_mean_std(sum_, ssum, sum_size) broadcasted = Broadcast.apply(target_gpus, mean, inv_std) outputs = [] for i, rec in enumerate(intermediates): outputs.append((rec[0], _MasterMessage(*broadcasted[i * 2:i * 2 + 2]))) return outputs def _compute_mean_std(self, sum_, ssum, size): """Compute the mean and standard-deviation with sum and square-sum. This method also maintains the moving average on the master device.""" assert size > 1, 'BatchNorm computes unbiased standard-deviation, which requires size > 1.' mean = sum_ / size sumvar = ssum - sum_ * mean unbias_var = sumvar / (size - 1) bias_var = sumvar / size self.running_mean = (1 - self.momentum) * self.running_mean + self.momentum * mean.data self.running_var = (1 - self.momentum) * self.running_var + self.momentum * unbias_var.data return mean, bias_var.clamp(self.eps) ** -0.5 class SynchronizedBatchNorm1d(_SynchronizedBatchNorm): r"""Applies Synchronized Batch Normalization over a 2d or 3d input that is seen as a mini-batch. .. math:: y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta This module differs from the built-in PyTorch BatchNorm1d as the mean and standard-deviation are reduced across all devices during training. For example, when one uses `nn.DataParallel` to wrap the network during training, PyTorch's implementation normalize the tensor on each device using the statistics only on that device, which accelerated the computation and is also easy to implement, but the statistics might be inaccurate. Instead, in this synchronized version, the statistics will be computed over all training samples distributed on multiple devices. Note that, for one-GPU or CPU-only case, this module behaves exactly same as the built-in PyTorch implementation. The mean and standard-deviation are calculated per-dimension over the mini-batches and gamma and beta are learnable parameter vectors of size C (where C is the input size). During training, this layer keeps a running estimate of its computed mean and variance. The running sum is kept with a default momentum of 0.1. During evaluation, this running mean/variance is used for normalization. Because the BatchNorm is done over the `C` dimension, computing statistics on `(N, L)` slices, it's common terminology to call this Temporal BatchNorm Args: num_features: num_features from an expected input of size `batch_size x num_features [x width]` eps: a value added to the denominator for numerical stability. Default: 1e-5 momentum: the value used for the running_mean and running_var computation. Default: 0.1 affine: a boolean value that when set to ``True``, gives the layer learnable affine parameters. Default: ``True`` Shape: - Input: :math:`(N, C)` or :math:`(N, C, L)` - Output: :math:`(N, C)` or :math:`(N, C, L)` (same shape as input) Examples: >>> # With Learnable Parameters >>> m = SynchronizedBatchNorm1d(100) >>> # Without Learnable Parameters >>> m = SynchronizedBatchNorm1d(100, affine=False) >>> input = torch.autograd.Variable(torch.randn(20, 100)) >>> output = m(input) """ def _check_input_dim(self, input): if input.dim() != 2 and input.dim() != 3: raise ValueError('expected 2D or 3D input (got {}D input)' .format(input.dim())) super(SynchronizedBatchNorm1d, self)._check_input_dim(input) class SynchronizedBatchNorm2d(_SynchronizedBatchNorm): r"""Applies Batch Normalization over a 4d input that is seen as a mini-batch of 3d inputs .. math:: y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta This module differs from the built-in PyTorch BatchNorm2d as the mean and standard-deviation are reduced across all devices during training. For example, when one uses `nn.DataParallel` to wrap the network during training, PyTorch's implementation normalize the tensor on each device using the statistics only on that device, which accelerated the computation and is also easy to implement, but the statistics might be inaccurate. Instead, in this synchronized version, the statistics will be computed over all training samples distributed on multiple devices. Note that, for one-GPU or CPU-only case, this module behaves exactly same as the built-in PyTorch implementation. The mean and standard-deviation are calculated per-dimension over the mini-batches and gamma and beta are learnable parameter vectors of size C (where C is the input size). During training, this layer keeps a running estimate of its computed mean and variance. The running sum is kept with a default momentum of 0.1. During evaluation, this running mean/variance is used for normalization. Because the BatchNorm is done over the `C` dimension, computing statistics on `(N, H, W)` slices, it's common terminology to call this Spatial BatchNorm Args: num_features: num_features from an expected input of size batch_size x num_features x height x width eps: a value added to the denominator for numerical stability. Default: 1e-5 momentum: the value used for the running_mean and running_var computation. Default: 0.1 affine: a boolean value that when set to ``True``, gives the layer learnable affine parameters. Default: ``True`` Shape: - Input: :math:`(N, C, H, W)` - Output: :math:`(N, C, H, W)` (same shape as input) Examples: >>> # With Learnable Parameters >>> m = SynchronizedBatchNorm2d(100) >>> # Without Learnable Parameters >>> m = SynchronizedBatchNorm2d(100, affine=False) >>> input = torch.autograd.Variable(torch.randn(20, 100, 35, 45)) >>> output = m(input) """ def _check_input_dim(self, input): if input.dim() != 4: raise ValueError('expected 4D input (got {}D input)' .format(input.dim())) super(SynchronizedBatchNorm2d, self)._check_input_dim(input) class SynchronizedBatchNorm3d(_SynchronizedBatchNorm): r"""Applies Batch Normalization over a 5d input that is seen as a mini-batch of 4d inputs .. math:: y = \frac{x - mean[x]}{ \sqrt{Var[x] + \epsilon}} * gamma + beta This module differs from the built-in PyTorch BatchNorm3d as the mean and standard-deviation are reduced across all devices during training. For example, when one uses `nn.DataParallel` to wrap the network during training, PyTorch's implementation normalize the tensor on each device using the statistics only on that device, which accelerated the computation and is also easy to implement, but the statistics might be inaccurate. Instead, in this synchronized version, the statistics will be computed over all training samples distributed on multiple devices. Note that, for one-GPU or CPU-only case, this module behaves exactly same as the built-in PyTorch implementation. The mean and standard-deviation are calculated per-dimension over the mini-batches and gamma and beta are learnable parameter vectors of size C (where C is the input size). During training, this layer keeps a running estimate of its computed mean and variance. The running sum is kept with a default momentum of 0.1. During evaluation, this running mean/variance is used for normalization. Because the BatchNorm is done over the `C` dimension, computing statistics on `(N, D, H, W)` slices, it's common terminology to call this Volumetric BatchNorm or Spatio-temporal BatchNorm Args: num_features: num_features from an expected input of size batch_size x num_features x depth x height x width eps: a value added to the denominator for numerical stability. Default: 1e-5 momentum: the value used for the running_mean and running_var computation. Default: 0.1 affine: a boolean value that when set to ``True``, gives the layer learnable affine parameters. Default: ``True`` Shape: - Input: :math:`(N, C, D, H, W)` - Output: :math:`(N, C, D, H, W)` (same shape as input) Examples: >>> # With Learnable Parameters >>> m = SynchronizedBatchNorm3d(100) >>> # Without Learnable Parameters >>> m = SynchronizedBatchNorm3d(100, affine=False) >>> input = torch.autograd.Variable(torch.randn(20, 100, 35, 45, 10)) >>> output = m(input) """ def _check_input_dim(self, input): if input.dim() != 5: raise ValueError('expected 5D input (got {}D input)' .format(input.dim())) super(SynchronizedBatchNorm3d, self)._check_input_dim(input)
12,932
44.861702
116
py
IVOS-ATNet
IVOS-ATNet-master/libs/custom_transforms.py
import numpy as np import torch class Normalize_ApplymeanvarImage(object): def __init__(self, mean, var, change_channels=False): self.mean = mean self.var = var self.change_channels = change_channels def __call__(self, sample): for elem in sample.keys(): if 'image' in elem: if self.change_channels: sample[elem] = sample[elem][:, :, [2, 1, 0]] sample[elem] = sample[elem].astype(np.float32)/255.0 sample[elem] = np.subtract(sample[elem], np.array(self.mean, dtype=np.float32))/np.array(self.var, dtype=np.float32) return sample def __str__(self): return 'SubtractMeanImage'+str(self.mean) class ToTensor(object): """Convert ndarrays in sample to Tensors.""" def __call__(self, sample): for elem in sample.keys(): if 'meta' in elem: continue tmp = sample[elem] if tmp.ndim == 2: tmp = tmp[:, :, np.newaxis] # swap color axis because # numpy image: H x W x C # torch image: C X H X W tmp = tmp.transpose((2, 0, 1)) sample[elem] = torch.from_numpy(tmp) return sample
1,272
25.520833
132
py
IVOS-ATNet
IVOS-ATNet-master/libs/utils_torch.py
import torch def combine_masks_with_batch(masks, n_obj, th=0.5, return_as_onehot = False): """ Combine mask for different objects. Different methods are the following: * `max_per_pixel`: Computes the final mask taking the pixel with the highest probability for every object. # Arguments masks: Tensor with shape[B, nobj, H, W]. H, W on batches must be same method: String. Method that specifies how the masks are fused. # Returns [B, 1, H, W] """ # masks : B, nobj, h, w # output : h,w marker = torch.argmax(masks, dim=1, keepdim=True) # if not return_as_onehot: out_mask = torch.unsqueeze(torch.zeros_like(masks)[:,0],1) #[B, 1, H, W] for obj_id in range(n_obj): try :tmp_mask = (marker == obj_id) * (masks[:,obj_id].unsqueeze(1) > th) except: raise NotImplementedError out_mask[tmp_mask] = obj_id + 1 # [B, 1, H, W] if return_as_onehot: out_mask = torch.zeros_like(masks) # [B, nobj, H, W] for obj_id in range(n_obj): try :tmp_mask = (marker == obj_id) * (masks[:,obj_id].unsqueeze(1) > th) except: raise NotImplementedError out_mask[:, obj_id] = tmp_mask[:,0].type(torch.cuda.FloatTensor) return out_mask
1,315
34.567568
84
py
IVOS-ATNet
IVOS-ATNet-master/libs/davis2017_torchdataset.py
from __future__ import division import os import numpy as np import cv2 from libs import utils from torch.utils.data import Dataset import json from PIL import Image class DAVIS2017(Dataset): """DAVIS 2017 dataset constructed using the PyTorch built-in functionalities""" def __init__(self, split='val', root='', num_frames=None, custom_frames=None, transform=None, retname=False, seq_name=None, obj_id=None, gt_only_first_frame=False, no_gt=False, batch_gt=False, rgb=False, effective_batch=None, prev_round_masks = None,#f,h,w ): """Loads image to label pairs for tool pose estimation split: Split or list of splits of the dataset root: dataset directory with subfolders "JPEGImages" and "Annotations" num_frames: Select number of frames of the sequence (None for all frames) custom_frames: List or Tuple with the number of the frames to include transform: Data transformations retname: Retrieve meta data in the sample key 'meta' seq_name: Use a specific sequence obj_id: Use a specific object of a sequence (If None and sequence is specified, the batch_gt is True) gt_only_first_frame: Provide the GT only in the first frame no_gt: No GT is provided batch_gt: For every frame sequence batch all the different objects gt rgb: Use RGB channel order in the image """ if isinstance(split, str): self.split = [split] else: split.sort() self.split = split self.db_root_dir = root self.transform = transform self.seq_name = seq_name self.obj_id = obj_id self.num_frames = num_frames self.custom_frames = custom_frames self.retname = retname self.rgb = rgb if seq_name is not None and obj_id is None: batch_gt = True self.batch_gt = batch_gt self.all_seqs_list = [] self.seqs = [] for splt in self.split: with open(os.path.join(self.db_root_dir, 'ImageSets', '2017', splt + '.txt')) as f: seqs_tmp = f.readlines() seqs_tmp = list(map(lambda elem: elem.strip(), seqs_tmp)) self.seqs.extend(seqs_tmp) self.seq_list_file = os.path.join(self.db_root_dir, 'ImageSets', '2017', '_'.join(self.split) + '_instances.txt') # Precompute the dictionary with the objects per sequence if not self._check_preprocess(): self._preprocess() if self.seq_name is None: img_list = [] labels = [] prevmask_list= [] for seq in self.seqs: images = np.sort(os.listdir(os.path.join(self.db_root_dir, 'JPEGImages/480p/', seq.strip()))) images_path = list(map(lambda x: os.path.join('JPEGImages/480p/', seq.strip(), x), images)) lab = np.sort(os.listdir(os.path.join(self.db_root_dir, 'Annotations/480p/', seq.strip()))) lab_path = list(map(lambda x: os.path.join('Annotations/480p/', seq.strip(), x), lab)) if num_frames is not None: seq_len = len(images_path) num_frames = min(num_frames, seq_len) frame_vector = np.arange(num_frames) frames_ids = list(np.round(frame_vector*seq_len/float(num_frames)).astype(np.int)) frames_ids[-1] = min(frames_ids[-1], seq_len) images_path = [images_path[x] for x in frames_ids] if no_gt: lab_path = [None] * len(images_path) else: lab_path = [lab_path[x] for x in frames_ids] elif isinstance(custom_frames, tuple) or isinstance(custom_frames, list): assert min(custom_frames) >= 0 and max(custom_frames) <= len(images_path) images_path = [images_path[x] for x in custom_frames] prevmask_list = [prev_round_masks[x] for x in custom_frames] if no_gt: lab_path = [None] * len(images_path) else: lab_path = [lab_path[x] for x in custom_frames] if gt_only_first_frame: lab_path = [lab_path[0]] lab_path.extend([None] * (len(images_path) - 1)) elif no_gt: lab_path = [None] * len(images_path) if self.batch_gt: obj = self.seq_dict[seq] if -1 in obj: obj.remove(-1) for ii in range(len(img_list), len(images_path)+len(img_list)): self.all_seqs_list.append([obj, ii]) else: for obj in self.seq_dict[seq]: if obj != -1: for ii in range(len(img_list), len(images_path)+len(img_list)): self.all_seqs_list.append([obj, ii]) img_list.extend(images_path) labels.extend(lab_path) else: # Initialize the per sequence images for online training assert self.seq_name in self.seq_dict.keys(), '{} not in {} set.'.format(self.seq_name, '_'.join(self.split)) names_img = np.sort(os.listdir(os.path.join(self.db_root_dir, 'JPEGImages/480p/', str(seq_name)))) img_list = list(map(lambda x: os.path.join('JPEGImages/480p/', str(seq_name), x), names_img)) name_label = np.sort(os.listdir(os.path.join(self.db_root_dir, 'Annotations/480p/', str(seq_name)))) labels = list(map(lambda x: os.path.join('Annotations/480p/', str(seq_name), x), name_label)) prevmask_list = [] if num_frames is not None: seq_len = len(img_list) num_frames = min(num_frames, seq_len) frame_vector = np.arange(num_frames) frames_ids = list(np.round(frame_vector * seq_len / float(num_frames)).astype(np.int)) frames_ids[-1] = min(frames_ids[-1], seq_len) img_list = [img_list[x] for x in frames_ids] if no_gt: labels = [None] * len(img_list) else: labels = [labels[x] for x in frames_ids] elif isinstance(custom_frames, tuple) or isinstance(custom_frames, list): assert min(custom_frames) >= 0 and max(custom_frames) <= len(img_list) img_list = [img_list[x] for x in custom_frames] prevmask_list = [prev_round_masks[x] for x in custom_frames] if no_gt: labels = [None] * len(img_list) else: labels = [labels[x] for x in custom_frames] if gt_only_first_frame: labels = [labels[0]] labels.extend([None]*(len(img_list)-1)) elif no_gt: labels = [None] * len(img_list) if obj_id is not None: assert obj_id in self.seq_dict[self.seq_name], \ "{} doesn't have this object id {}.".format(self.seq_name, str(obj_id)) if self.batch_gt: self.obj_id = self.seq_dict[self.seq_name] if -1 in self.obj_id: self.obj_id.remove(-1) self.obj_id = [0]+self.obj_id assert (len(labels) == len(img_list)) if effective_batch: self.img_list = img_list * effective_batch self.labels = labels * effective_batch else: self.img_list = img_list self.labels = labels self.prevmasks_list = prevmask_list # print('Done initializing DAVIS2017 '+'_'.join(self.split)+' Dataset') # print('Number of images: {}'.format(len(self.img_list))) # if self.seq_name is None: # print('Number of elements {}'.format(len(self.all_seqs_list))) def _check_preprocess(self): _seq_list_file = self.seq_list_file if not os.path.isfile(_seq_list_file): return False else: self.seq_dict = json.load(open(self.seq_list_file, 'r')) return True def _preprocess(self): self.seq_dict = {} for seq in self.seqs: # Read object masks and get number of objects name_label = np.sort(os.listdir(os.path.join(self.db_root_dir, 'Annotations/480p/', seq))) label_path = os.path.join(self.db_root_dir, 'Annotations/480p/', seq, name_label[0]) _mask = np.array(Image.open(label_path)) _mask_ids = np.unique(_mask) n_obj = _mask_ids[-1] self.seq_dict[seq] = list(range(1, n_obj+1)) with open(self.seq_list_file, 'w') as outfile: outfile.write('{{\n\t"{:s}": {:s}'.format(self.seqs[0], json.dumps(self.seq_dict[self.seqs[0]]))) for ii in range(1, len(self.seqs)): outfile.write(',\n\t"{:s}": {:s}'.format(self.seqs[ii], json.dumps(self.seq_dict[self.seqs[ii]]))) outfile.write('\n}\n') print('Preprocessing finished') def __len__(self): if self.seq_name is None: return len(self.all_seqs_list) else: return len(self.img_list) def __getitem__(self, idx): # print(idx) img, gt, prev_round_mask = self.make_img_gt_mask_pair(idx) pad_img, pad_info = utils.apply_pad(img) pad_gt= utils.apply_pad(gt, padinfo = pad_info)#h,w,n sample = {'image': pad_img, 'gt': pad_gt} if self.retname: if self.seq_name is None: obj_id = self.all_seqs_list[idx][0] img_path = self.img_list[self.all_seqs_list[idx][1]] else: obj_id = self.obj_id img_path = self.img_list[idx] seq_name = img_path.split('/')[-2] frame_id = img_path.split('/')[-1].split('.')[-2] sample['meta'] = {'seq_name': seq_name, 'frame_id': frame_id, 'obj_id': obj_id, 'im_size': (img.shape[0], img.shape[1]), 'pad_size': (pad_img.shape[0], pad_img.shape[1]), 'pad_info': pad_info} if self.transform is not None: sample = self.transform(sample) return sample def make_img_gt_mask_pair(self, idx): """ Make the image-ground-truth pair """ prev_round_mask_tmp = self.prevmasks_list[idx] if self.seq_name is None: obj_id = self.all_seqs_list[idx][0] img_path = self.img_list[self.all_seqs_list[idx][1]] label_path = self.labels[self.all_seqs_list[idx][1]] else: obj_id = self.obj_id img_path = self.img_list[idx] label_path = self.labels[idx] seq_name = img_path.split('/')[-2] n_obj = 1 if isinstance(obj_id, int) else len(obj_id) img = cv2.imread(os.path.join(self.db_root_dir, img_path)) img = np.array(img, dtype=np.float32) if self.rgb: img = img[:, :, [2, 1, 0]] if label_path is not None: label = Image.open(os.path.join(self.db_root_dir, label_path)) else: if self.batch_gt: gt = np.zeros(np.append(img.shape[:-1], n_obj), dtype=np.float32) else: gt = np.zeros(img.shape[:-1], dtype=np.float32) if label_path is not None: gt_tmp = np.array(label, dtype=np.uint8) if self.batch_gt: gt = np.zeros(np.append(n_obj, gt_tmp.shape), dtype=np.float32) for ii, k in enumerate(obj_id): gt[ii, :, :] = gt_tmp == k gt = gt.transpose((1, 2, 0)) else: gt = (gt_tmp == obj_id).astype(np.float32) if self.batch_gt: prev_round_mask = np.zeros(np.append(img.shape[:-1], n_obj), dtype=np.float32) for ii, k in enumerate(obj_id): prev_round_mask[:, :, ii] = prev_round_mask_tmp == k else: prev_round_mask = (prev_round_mask_tmp == obj_id).astype(np.float32) return img, gt, prev_round_mask def get_img_size(self): img = cv2.imread(os.path.join(self.db_root_dir, self.img_list[0])) return list(img.shape[:2]) def __str__(self): return 'DAVIS2017' if __name__ =='__main__': a = DAVIS2017(split='val', custom_frames=[21,22], seq_name='gold-fish', rgb=True, no_gt=False, retname=True,prev_round_masks=np.zeros([40,480,854])) c= a.__getitem__(0) b=1
13,111
42.417219
153
py
GraphLoG
GraphLoG-main/pretrain_graphlog.py
import argparse from loader import MoleculeDataset import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from tqdm import tqdm import numpy as np import os, sys import pdb import copy import random from model import GNN, ProjectNet from sklearn.metrics import roc_auc_score from splitters import scaffold_split, random_split, random_scaffold_split import pandas as pd from util import ExtractSubstructureContextPair from torch_geometric.data import DataLoader from dataloader import DataLoaderSubstructContext from torch_geometric.nn import global_add_pool, global_mean_pool, global_max_pool from tensorboardX import SummaryWriter # Graph pooling functions def pool_func(x, batch, mode = "mean"): if mode == "sum": return global_add_pool(x, batch) elif mode == "mean": return global_mean_pool(x, batch) elif mode == "max": return global_max_pool(x, batch) # Mask some nodes in a graph def mask_nodes(batch, args, num_atom_type=119): masked_node_indices = list() # select indices of masked nodes for i in range(batch.batch[-1] + 1): idx = torch.nonzero((batch.batch == i).float()).squeeze(-1) num_node = idx.shape[0] if args.mask_num == 0: sample_size = int(num_node * args.mask_rate + 1) else: sample_size = min(args.mask_num, int(num_node * 0.5)) masked_node_idx = random.sample(idx.tolist(), sample_size) masked_node_idx.sort() masked_node_indices += masked_node_idx batch.masked_node_indices = torch.tensor(masked_node_indices) # mask nodes' features for node_idx in masked_node_indices: batch.x[node_idx] = torch.tensor([num_atom_type, 0]) return batch # NCE loss within a graph def intra_NCE_loss(node_reps, node_modify_reps, batch, tau=0.1, epsilon=1e-6): node_reps_norm = torch.norm(node_reps, dim = 1).unsqueeze(-1) node_modify_reps_norm = torch.norm(node_modify_reps, dim = 1).unsqueeze(-1) sim = torch.mm(node_reps, node_modify_reps.t()) / ( torch.mm(node_reps_norm, node_modify_reps_norm.t()) + epsilon) exp_sim = torch.exp(sim / tau) mask = torch.stack([(batch.batch == i).float() for i in batch.batch.tolist()], dim = 1) exp_sim_mask = exp_sim * mask exp_sim_all = torch.index_select(exp_sim_mask, 1, batch.masked_node_indices) exp_sim_positive = torch.index_select(exp_sim_all, 0, batch.masked_node_indices) positive_ratio = exp_sim_positive.sum(0) / (exp_sim_all.sum(0) + epsilon) NCE_loss = -torch.log(positive_ratio).sum() / batch.masked_node_indices.shape[0] mask_select = torch.index_select(mask, 1, batch.masked_node_indices) thr = 1. / mask_select.sum(0) correct_cnt = (positive_ratio > thr).float().sum() return NCE_loss, correct_cnt # NCE loss across different graphs def inter_NCE_loss(graph_reps, graph_modify_reps, device, tau=0.1, epsilon=1e-6): graph_reps_norm = torch.norm(graph_reps, dim = 1).unsqueeze(-1) graph_modify_reps_norm = torch.norm(graph_modify_reps, dim = 1).unsqueeze(-1) sim = torch.mm(graph_reps, graph_modify_reps.t()) / ( torch.mm(graph_reps_norm, graph_modify_reps_norm.t()) + epsilon) exp_sim = torch.exp(sim / tau) mask = torch.eye(graph_reps.shape[0]).to(device) positive = (exp_sim * mask).sum(0) negative = (exp_sim * (1 - mask)).sum(0) positive_ratio = positive / (positive + negative + epsilon) NCE_loss = -torch.log(positive_ratio).sum() / graph_reps.shape[0] thr = 1. / ((1 - mask).sum(0) + 1.) correct_cnt = (positive_ratio > thr).float().sum() return NCE_loss, correct_cnt # NCE loss for global-local mutual information maximization def gl_NCE_loss(node_reps, graph_reps, batch, tau=0.1, epsilon=1e-6): node_reps_norm = torch.norm(node_reps, dim = 1).unsqueeze(-1) graph_reps_norm = torch.norm(graph_reps, dim = 1).unsqueeze(-1) sim = torch.mm(node_reps, graph_reps.t()) / ( torch.mm(node_reps_norm, graph_reps_norm.t()) + epsilon) exp_sim = torch.exp(sim / tau) mask = torch.stack([(batch == i).float() for i in range(graph_reps.shape[0])], dim = 1) positive = exp_sim * mask negative = exp_sim * (1 - mask) positive_ratio = positive / (positive + negative.sum(0).unsqueeze(0) + epsilon) NCE_loss = -torch.log(positive_ratio + (1 - mask)).sum() / node_reps.shape[0] thr = 1. / ((1 - mask).sum(0) + 1.).unsqueeze(0) correct_cnt = (positive_ratio > thr).float().sum() return NCE_loss, correct_cnt # NCE loss between graphs and prototypes def proto_NCE_loss(graph_reps, tau=0.1, epsilon=1e-6): global proto, proto_connection # similarity for original and modified graphs graph_reps_norm = torch.norm(graph_reps, dim=1).unsqueeze(-1) exp_sim_list = [] mask_list = [] NCE_loss = 0 for i in range(len(proto)-1, -1, -1): tmp_proto = proto[i] proto_norm = torch.norm(tmp_proto, dim=1).unsqueeze(-1) sim = torch.mm(graph_reps, tmp_proto.t()) / ( torch.mm(graph_reps_norm, proto_norm.t()) + epsilon) exp_sim = torch.exp(sim / tau) if i != (len(proto) - 1): # apply the connection mask exp_sim_last = exp_sim_list[-1] idx_last = torch.argmax(exp_sim_last, dim = 1).unsqueeze(-1) connection = proto_connection[i] connection_mask = (connection.unsqueeze(0) == idx_last.float()).float() exp_sim = exp_sim * connection_mask # define NCE loss between prototypes from consecutive layers upper_proto = proto[i+1] upper_proto_norm = torch.norm(upper_proto, dim=1).unsqueeze(-1) proto_sim = torch.mm(tmp_proto, upper_proto.t()) / ( torch.mm(proto_norm, upper_proto_norm.t()) + epsilon) proto_exp_sim = torch.exp(proto_sim / tau) proto_positive_list = [proto_exp_sim[j, connection[j].long()] for j in range(proto_exp_sim.shape[0])] proto_positive = torch.stack(proto_positive_list, dim=0) proto_positive_ratio = proto_positive / (proto_exp_sim.sum(1) + epsilon) NCE_loss += -torch.log(proto_positive_ratio).mean() mask = (exp_sim == exp_sim.max(1)[0].unsqueeze(-1)).float() exp_sim_list.append(exp_sim) mask_list.append(mask) # define NCE loss between graph embedding and prototypes for i in range(len(proto)): exp_sim = exp_sim_list[i] mask = mask_list[i] positive = exp_sim * mask negative = exp_sim * (1 - mask) positive_ratio = positive.sum(1) / (positive.sum(1) + negative.sum(1) + epsilon) NCE_loss += -torch.log(positive_ratio).mean() return NCE_loss # Update prototypes with batch information def update_proto_lowest(graph_reps, decay_ratio=0.7, epsilon=1e-6): global proto, proto_state graph_reps_norm = torch.norm(graph_reps, dim=1).unsqueeze(-1) proto_norm = torch.norm(proto[0], dim=1).unsqueeze(-1) sim = torch.mm(graph_reps, proto[0].t()) / ( torch.mm(graph_reps_norm, proto_norm.t()) + epsilon) # update states of prototypes mask = (sim == sim.max(1)[0].unsqueeze(-1)).float() cnt = mask.sum(0) proto_state[0].data = proto_state[0].data + cnt.data # update prototypes batch_cnt = mask.t() / (cnt.unsqueeze(-1) + epsilon) batch_mean = torch.mm(batch_cnt, graph_reps) proto[0].data = proto[0].data * (cnt == 0).float().unsqueeze(-1).data + ( proto[0].data * decay_ratio + batch_mean.data * (1 - decay_ratio)) * (cnt != 0).float().unsqueeze(-1).data return # Initialze prototypes and their state def init_proto_lowest(args, model, proj, loader, device, num_iter = 5): model.eval() proj.eval() for iter in range(num_iter): for step, batch in enumerate(tqdm(loader, desc="Iteration")): batch = batch.to(device) # get node and graph representations node_reps = model(batch.x, batch.edge_index, batch.edge_attr) graph_reps = pool_func(node_reps, batch.batch, mode=args.graph_pooling) # feature projection graph_reps_proj = proj(graph_reps) # update prototypes update_proto_lowest(graph_reps_proj, decay_ratio = args.decay_ratio) global proto, proto_state idx = torch.nonzero((proto_state[0] >= 2).float()).squeeze(-1) proto_selected = torch.index_select(proto[0], 0, idx) proto_selected.requires_grad = True return proto_selected # Initialze prototypes and their state def init_proto(args, index, device, num_iter = 20): global proto, proto_state proto_connection = torch.zeros(proto[index-1].shape[0]).to(device) for iter in range(num_iter): for i in range(proto[index-1].shape[0]): # update the closest prototype sim = torch.mm(proto[index], proto[index-1][i,:].unsqueeze(-1)).squeeze(-1) idx = torch.argmax(sim) if iter == (num_iter - 1): proto_state[index][idx] = 1 proto_connection[i] = idx proto[index].data[idx, :] = proto[index].data[idx, :] * args.decay_ratio + \ proto[index-1].data[i, :] * (1 - args.decay_ratio) # penalize rival sim[idx] = 0 rival_idx = torch.argmax(sim) proto[index].data[rival_idx, :] = proto[index].data[rival_idx, :] * (2 - args.decay_ratio) - \ proto[index-1].data[i, :] * (1 - args.decay_ratio) indices = torch.nonzero(proto_state[index]).squeeze(-1) proto_selected = torch.index_select(proto[index], 0, indices) proto_selected.requires_grad = True for i in range(indices.shape[0]): idx = indices[i] idx_connection = torch.nonzero((proto_connection == idx.float()).float()).squeeze(-1) proto_connection[idx_connection] = i return proto_selected, proto_connection # For one epoch pretraining def pretrain(args, model, proj, loader, optimizer, device): model.train() proj.train() NCE_loss_intra_cnt = 0 NCE_loss_inter_cnt = 0 correct_intra_cnt = 0 correct_inter_cnt = 0 total_intra_cnt = 0 total_inter_cnt = 0 for step, batch in enumerate(tqdm(loader, desc="Iteration")): batch_modify = copy.deepcopy(batch) batch_modify = mask_nodes(batch_modify, args) batch, batch_modify = batch.to(device), batch_modify.to(device) # get node and graph representations node_reps = model(batch.x, batch.edge_index, batch.edge_attr) node_modify_reps = model(batch_modify.x, batch_modify.edge_index, batch_modify.edge_attr) graph_reps = pool_func(node_reps, batch.batch, mode=args.graph_pooling) graph_modify_reps = pool_func(node_modify_reps, batch_modify.batch, mode=args.graph_pooling) # feature projection node_reps_proj = proj(node_reps) node_modify_reps_proj = proj(node_modify_reps) graph_reps_proj = proj(graph_reps) graph_modify_reps_proj = proj(graph_modify_reps) # NCE loss NCE_loss_intra, correct_intra = intra_NCE_loss(node_reps_proj, node_modify_reps_proj, batch_modify, tau=args.tau) NCE_loss_inter, correct_inter = inter_NCE_loss(graph_reps_proj, graph_modify_reps_proj, device, tau=args.tau) NCE_loss_intra_cnt += NCE_loss_intra.item() NCE_loss_inter_cnt += NCE_loss_inter.item() correct_intra_cnt += correct_intra correct_inter_cnt += correct_inter total_intra_cnt += batch_modify.masked_node_indices.shape[0] total_inter_cnt += graph_reps.shape[0] # optimization optimizer.zero_grad() NCE_loss = args.alpha * NCE_loss_intra + args.beta * NCE_loss_inter NCE_loss.backward() optimizer.step() if (step + 1) % args.disp_interval == 0: print( 'iteration: %d, intra NCE loss: %f, intra acc: %f, inter NCE loss: %f, inter acc: %f' % ( step + 1, NCE_loss_intra.item(), float(correct_intra_cnt) / float(total_intra_cnt), NCE_loss_inter.item(), float(correct_inter_cnt) / float(total_inter_cnt))) return NCE_loss_intra_cnt / step, float(correct_intra_cnt) / float( total_intra_cnt), NCE_loss_inter_cnt / step, float(correct_inter_cnt) / float(total_inter_cnt) # For every epoch training def train(args, model, proj, loader, optimizer, device): global proto, proto_connection model.train() proj.train() NCE_loss_intra_cnt = 0 NCE_loss_inter_cnt = 0 NCE_loss_proto_cnt = 0 correct_intra_cnt = 0 correct_inter_cnt = 0 total_intra_cnt = 0 total_inter_cnt = 0 for step, batch in enumerate(tqdm(loader, desc="Iteration")): batch_modify = copy.deepcopy(batch) batch_modify = mask_nodes(batch_modify, args) batch, batch_modify = batch.to(device), batch_modify.to(device) # get node and graph representations node_reps = model(batch.x, batch.edge_index, batch.edge_attr) node_modify_reps = model(batch_modify.x, batch_modify.edge_index, batch_modify.edge_attr) graph_reps = pool_func(node_reps, batch.batch, mode=args.graph_pooling) graph_modify_reps = pool_func(node_modify_reps, batch_modify.batch, mode=args.graph_pooling) # feature projection node_reps_proj = proj(node_reps) node_modify_reps_proj = proj(node_modify_reps) graph_reps_proj = proj(graph_reps) graph_modify_reps_proj = proj(graph_modify_reps) # NCE loss NCE_loss_intra, correct_intra = intra_NCE_loss(node_reps_proj, node_modify_reps_proj, batch_modify, tau=args.tau) NCE_loss_inter, correct_inter = inter_NCE_loss(graph_reps_proj, graph_modify_reps_proj, device, tau=args.tau) NCE_loss_proto = proto_NCE_loss(graph_reps_proj, tau=args.tau) NCE_loss_intra_cnt += NCE_loss_intra.item() NCE_loss_inter_cnt += NCE_loss_inter.item() NCE_loss_proto_cnt += NCE_loss_proto.item() correct_intra_cnt += correct_intra correct_inter_cnt += correct_inter total_intra_cnt += batch_modify.masked_node_indices.shape[0] total_inter_cnt += graph_reps.shape[0] # optimization optimizer.zero_grad() NCE_loss = args.alpha * NCE_loss_intra + args.beta * NCE_loss_inter + \ args.gamma * NCE_loss_proto NCE_loss.backward() optimizer.step() if (step + 1) % args.disp_interval == 0: print( 'iteration: %d, intra NCE loss: %f, intra acc: %f, inter NCE loss: %f, inter acc: %f' % ( step + 1, NCE_loss_intra.item(), float(correct_intra_cnt) / float(total_intra_cnt), NCE_loss_inter.item(), float(correct_inter_cnt) / float(total_inter_cnt))) template = 'iteration: %d, proto NCE loss: %f' value_list = [step + 1, NCE_loss_proto.item()] for i in range(args.hierarchy): template += (', active num ' + str(i+1) + ': %d') value_list.append(proto[i].shape[0]) print (template % tuple(value_list)) return NCE_loss_intra_cnt / step, float(correct_intra_cnt) / float( total_intra_cnt), NCE_loss_inter_cnt / step, float(correct_inter_cnt) / float( total_inter_cnt), NCE_loss_proto_cnt / step def main(): # Training settings parser = argparse.ArgumentParser(description='GraphLoG for GNN pre-training') parser.add_argument('--device', type=int, default=0, help='which gpu to use if any (default: 0)') parser.add_argument('--batch_size', type=int, default=512, help='input batch size for training (default: 512)') parser.add_argument('--local_epochs', type=int, default=1, help='number of epochs for local learning (default: 1)') parser.add_argument('--global_epochs', type=int, default=10, help='number of epochs for global learning (default: 10)') parser.add_argument('--lr', type=float, default=0.001, help='learning rate (default: 0.001)') parser.add_argument('--decay', type=float, default=0, help='weight decay (default: 0)') parser.add_argument('--num_layer', type=int, default=5, help='number of GNN message passing layers (default: 5).') parser.add_argument('--emb_dim', type=int, default=300, help='embedding dimensions (default: 300)') parser.add_argument('--dropout_ratio', type=float, default=0, help='dropout ratio (default: 0)') parser.add_argument('--mask_rate', type=float, default=0.3, help='dropout ratio (default: 0.3)') parser.add_argument('--mask_num', type=int, default=0, help='the number of modified nodes (default: 0)') parser.add_argument('--JK', type=str, default="last", help='how the node features are combined across layers. last, sum, max or concat') parser.add_argument('--graph_pooling', type=str, default="mean", help='graph level pooling (sum, mean, max)') parser.add_argument('--dataset', type=str, default='zinc_standard_agent', help='root directory of dataset for pretraining') parser.add_argument('--output_model_file', type=str, default='', help='filename to output the model') parser.add_argument('--gnn_type', type=str, default="gin") parser.add_argument('--seed', type=int, default=0, help="Seed for splitting dataset.") parser.add_argument('--num_workers', type=int, default=1, help='number of workers for dataset loading') parser.add_argument('--tau', type=float, default=0.04, help='the temperature parameter for softmax') parser.add_argument('--decay_ratio', type=float, default=0.95, help='the decay ratio for moving average') parser.add_argument('--num_proto', type=int, default=50, help='the number of initial prototypes') parser.add_argument('--hierarchy', type=int, default=3, help='the number of hierarchy') parser.add_argument('--alpha', type=float, default=1, help='the weight of intra-graph NCE loss') parser.add_argument('--beta', type=float, default=1, help='the weight of inter-graph NCE loss') parser.add_argument('--gamma', type=float, default=0.1, help='the weight of prototype NCE loss') parser.add_argument('--disp_interval', type=int, default=10, help='the display interval') args = parser.parse_args() torch.manual_seed(args.seed) np.random.seed(args.seed) device = torch.device("cuda:" + str(args.device)) if torch.cuda.is_available() else torch.device("cpu") if torch.cuda.is_available(): torch.cuda.manual_seed_all(args.seed) print("num GNN layer: %d" % (args.num_layer)) # set up dataset and transform function. dataset = MoleculeDataset("./dataset/" + args.dataset, dataset=args.dataset) loader = DataLoader(dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers) # set up pretraining models and feature projector model = GNN(args.num_layer, args.emb_dim, JK=args.JK, drop_ratio=args.dropout_ratio, gnn_type=args.gnn_type).to(device) if args.JK == 'concat': proj = ProjectNet((args.num_layer + 1) * args.emb_dim).to(device) else: proj = ProjectNet(args.emb_dim).to(device) # set up the optimizer for pretraining model_param_group = [{"params": model.parameters(), "lr": args.lr}, {"params": proj.parameters(), "lr": args.lr}] optimizer_pretrain = optim.Adam(model_param_group, lr=args.lr, weight_decay=args.decay) # initialize prototypes and their state global proto, proto_state, proto_connection if args.JK == 'concat': proto = [torch.rand((args.num_proto, (args.num_layer + 1) * args.emb_dim)).to(device) for i in range(args.hierarchy)] else: proto = [torch.rand((args.num_proto, args.emb_dim)).to(device) for i in range(args.hierarchy)] proto_state = [torch.zeros(args.num_proto).to(device) for i in range(args.hierarchy)] proto_connection = [] # pre-training with only local objective for epoch in range(1, args.local_epochs + 1): print("====epoch " + str(epoch)) train_intra_loss, train_intra_acc, train_inter_loss, train_inter_acc = pretrain( args, model, proj, loader, optimizer_pretrain, device) print(train_intra_loss, train_intra_acc, train_inter_loss, train_inter_acc) print("") # initialize prototypes and their state according to pretrained representations print("Initalize prototypes: layer 1") tmp_proto = init_proto_lowest(args, model, proj, loader, device) proto[0] = tmp_proto for i in range(1, args.hierarchy): print ("Initialize prototypes: layer ", i + 1) tmp_proto, tmp_proto_connection = init_proto(args, i, device) proto[i] = tmp_proto proto_connection.append(tmp_proto_connection) # set up the optimizer model_param_group = [{"params": model.parameters(), "lr": args.lr}, {"params": proj.parameters(), "lr": args.lr}] for i in range(args.hierarchy): model_param_group += [{'params': proto[i], 'lr': args.lr, 'weight_decay': 0}] optimizer = optim.Adam(model_param_group, lr=args.lr, weight_decay=args.decay) # Training with local and global objectives for epoch in range(1, args.global_epochs + 1): print("====epoch " + str(epoch)) train_intra_loss, train_intra_acc, train_inter_loss, train_inter_acc, train_proto_loss = train( args, model, proj, loader, optimizer, device) print(train_intra_loss, train_intra_acc, train_inter_loss, train_inter_acc, train_proto_loss) if not args.output_model_file == "": torch.save(model.state_dict(), args.output_model_file + ".pth") os.system('watch nvidia-smi') if __name__ == "__main__": main()
22,536
43.364173
118
py
GraphLoG
GraphLoG-main/batch.py
import torch from torch_geometric.data import Data, Batch class BatchMasking(Data): r"""A plain old python object modeling a batch of graphs as one big (dicconnected) graph. With :class:`torch_geometric.data.Data` being the base class, all its methods can also be used here. In addition, single graphs can be reconstructed via the assignment vector :obj:`batch`, which maps each node to its respective graph identifier. """ def __init__(self, batch=None, **kwargs): super(BatchMasking, self).__init__(**kwargs) self.batch = batch @staticmethod def from_data_list(data_list): r"""Constructs a batch object from a python list holding :class:`torch_geometric.data.Data` objects. The assignment vector :obj:`batch` is created on the fly.""" keys = [set(data.keys) for data in data_list] keys = list(set.union(*keys)) assert 'batch' not in keys batch = BatchMasking() for key in keys: batch[key] = [] batch.batch = [] cumsum_node = 0 cumsum_edge = 0 for i, data in enumerate(data_list): num_nodes = data.num_nodes batch.batch.append(torch.full((num_nodes, ), i, dtype=torch.long)) for key in data.keys: item = data[key] if key in ['edge_index', 'masked_atom_indices']: item = item + cumsum_node elif key == 'connected_edge_indices': item = item + cumsum_edge batch[key].append(item) cumsum_node += num_nodes cumsum_edge += data.edge_index.shape[1] for key in keys: batch[key] = torch.cat( batch[key], dim=data_list[0].cat_dim(key, batch[key][0])) batch.batch = torch.cat(batch.batch, dim=-1) return batch.contiguous() def cumsum(self, key, item): r"""If :obj:`True`, the attribute :obj:`key` with content :obj:`item` should be added up cumulatively before concatenated together. .. note:: This method is for internal use only, and should only be overridden if the batch concatenation process is corrupted for a specific data attribute. """ return key in ['edge_index', 'face', 'masked_atom_indices', 'connected_edge_indices'] @property def num_graphs(self): """Returns the number of graphs in the batch.""" return self.batch[-1].item() + 1 class BatchAE(Data): r"""A plain old python object modeling a batch of graphs as one big (dicconnected) graph. With :class:`torch_geometric.data.Data` being the base class, all its methods can also be used here. In addition, single graphs can be reconstructed via the assignment vector :obj:`batch`, which maps each node to its respective graph identifier. """ def __init__(self, batch=None, **kwargs): super(BatchAE, self).__init__(**kwargs) self.batch = batch @staticmethod def from_data_list(data_list): r"""Constructs a batch object from a python list holding :class:`torch_geometric.data.Data` objects. The assignment vector :obj:`batch` is created on the fly.""" keys = [set(data.keys) for data in data_list] keys = list(set.union(*keys)) assert 'batch' not in keys batch = BatchAE() for key in keys: batch[key] = [] batch.batch = [] cumsum_node = 0 for i, data in enumerate(data_list): num_nodes = data.num_nodes batch.batch.append(torch.full((num_nodes, ), i, dtype=torch.long)) for key in data.keys: item = data[key] if key in ['edge_index', 'negative_edge_index']: item = item + cumsum_node batch[key].append(item) cumsum_node += num_nodes for key in keys: batch[key] = torch.cat( batch[key], dim=batch.cat_dim(key)) batch.batch = torch.cat(batch.batch, dim=-1) return batch.contiguous() @property def num_graphs(self): """Returns the number of graphs in the batch.""" return self.batch[-1].item() + 1 def cat_dim(self, key): return -1 if key in ["edge_index", "negative_edge_index"] else 0 class BatchSubstructContext(Data): r"""A plain old python object modeling a batch of graphs as one big (dicconnected) graph. With :class:`torch_geometric.data.Data` being the base class, all its methods can also be used here. In addition, single graphs can be reconstructed via the assignment vector :obj:`batch`, which maps each node to its respective graph identifier. """ """ Specialized batching for substructure context pair! """ def __init__(self, batch=None, **kwargs): super(BatchSubstructContext, self).__init__(**kwargs) self.batch = batch @staticmethod def from_data_list(data_list): r"""Constructs a batch object from a python list holding :class:`torch_geometric.data.Data` objects. The assignment vector :obj:`batch` is created on the fly.""" #keys = [set(data.keys) for data in data_list] #keys = list(set.union(*keys)) #assert 'batch' not in keys batch = BatchSubstructContext() keys = ["center_substruct_idx", "edge_attr_substruct", "edge_index_substruct", "x_substruct", "overlap_context_substruct_idx", "edge_attr_context", "edge_index_context", "x_context"] for key in keys: #print(key) batch[key] = [] #batch.batch = [] #used for pooling the context batch.batch_overlapped_context = [] batch.overlapped_context_size = [] cumsum_main = 0 cumsum_substruct = 0 cumsum_context = 0 i = 0 for data in data_list: #If there is no context, just skip!! if hasattr(data, "x_context"): num_nodes = data.num_nodes num_nodes_substruct = len(data.x_substruct) num_nodes_context = len(data.x_context) #batch.batch.append(torch.full((num_nodes, ), i, dtype=torch.long)) batch.batch_overlapped_context.append(torch.full((len(data.overlap_context_substruct_idx), ), i, dtype=torch.long)) batch.overlapped_context_size.append(len(data.overlap_context_substruct_idx)) ###batching for the main graph #for key in data.keys: # if not "context" in key and not "substruct" in key: # item = data[key] # item = item + cumsum_main if batch.cumsum(key, item) else item # batch[key].append(item) ###batching for the substructure graph for key in ["center_substruct_idx", "edge_attr_substruct", "edge_index_substruct", "x_substruct"]: item = data[key] item = item + cumsum_substruct if batch.cumsum(key, item) else item batch[key].append(item) ###batching for the context graph for key in ["overlap_context_substruct_idx", "edge_attr_context", "edge_index_context", "x_context"]: item = data[key] item = item + cumsum_context if batch.cumsum(key, item) else item batch[key].append(item) cumsum_main += num_nodes cumsum_substruct += num_nodes_substruct cumsum_context += num_nodes_context i += 1 for key in keys: batch[key] = torch.cat( batch[key], dim=batch.cat_dim(key)) #batch.batch = torch.cat(batch.batch, dim=-1) batch.batch_overlapped_context = torch.cat(batch.batch_overlapped_context, dim=-1) batch.overlapped_context_size = torch.LongTensor(batch.overlapped_context_size) return batch.contiguous() def cat_dim(self, key): return -1 if key in ["edge_index", "edge_index_substruct", "edge_index_context"] else 0 def cumsum(self, key, item): r"""If :obj:`True`, the attribute :obj:`key` with content :obj:`item` should be added up cumulatively before concatenated together. .. note:: This method is for internal use only, and should only be overridden if the batch concatenation process is corrupted for a specific data attribute. """ return key in ["edge_index", "edge_index_substruct", "edge_index_context", "overlap_context_substruct_idx", "center_substruct_idx"] @property def num_graphs(self): """Returns the number of graphs in the batch.""" return self.batch[-1].item() + 1
8,940
38.043668
190
py
GraphLoG
GraphLoG-main/dataloader.py
import torch.utils.data from torch.utils.data.dataloader import default_collate from batch import BatchSubstructContext, BatchMasking, BatchAE class DataLoaderSubstructContext(torch.utils.data.DataLoader): r"""Data loader which merges data objects from a :class:`torch_geometric.data.dataset` to a mini-batch. Args: dataset (Dataset): The dataset from which to load the data. batch_size (int, optional): How may samples per batch to load. (default: :obj:`1`) shuffle (bool, optional): If set to :obj:`True`, the data will be reshuffled at every epoch (default: :obj:`True`) """ def __init__(self, dataset, batch_size=1, shuffle=True, **kwargs): super(DataLoaderSubstructContext, self).__init__( dataset, batch_size, shuffle, collate_fn=lambda data_list: BatchSubstructContext.from_data_list(data_list), **kwargs) class DataLoaderMasking(torch.utils.data.DataLoader): r"""Data loader which merges data objects from a :class:`torch_geometric.data.dataset` to a mini-batch. Args: dataset (Dataset): The dataset from which to load the data. batch_size (int, optional): How may samples per batch to load. (default: :obj:`1`) shuffle (bool, optional): If set to :obj:`True`, the data will be reshuffled at every epoch (default: :obj:`True`) """ def __init__(self, dataset, batch_size=1, shuffle=True, **kwargs): super(DataLoaderMasking, self).__init__( dataset, batch_size, shuffle, collate_fn=lambda data_list: BatchMasking.from_data_list(data_list), **kwargs) class DataLoaderAE(torch.utils.data.DataLoader): r"""Data loader which merges data objects from a :class:`torch_geometric.data.dataset` to a mini-batch. Args: dataset (Dataset): The dataset from which to load the data. batch_size (int, optional): How may samples per batch to load. (default: :obj:`1`) shuffle (bool, optional): If set to :obj:`True`, the data will be reshuffled at every epoch (default: :obj:`True`) """ def __init__(self, dataset, batch_size=1, shuffle=True, **kwargs): super(DataLoaderAE, self).__init__( dataset, batch_size, shuffle, collate_fn=lambda data_list: BatchAE.from_data_list(data_list), **kwargs)
2,503
36.939394
89
py
GraphLoG
GraphLoG-main/model.py
import torch from torch_geometric.nn import MessagePassing from torch_geometric.utils import add_self_loops, degree, softmax from torch_geometric.nn import global_add_pool, global_mean_pool, global_max_pool, GlobalAttention, Set2Set import torch.nn.functional as F from torch_scatter import scatter_add from torch_geometric.nn.inits import glorot, zeros num_atom_type = 120 #including the extra mask tokens num_chirality_tag = 3 num_bond_type = 6 #including aromatic and self-loop edge, and extra masked tokens num_bond_direction = 3 class GINConv(MessagePassing): """ Extension of GIN aggregation to incorporate edge information by concatenation. Args: emb_dim (int): dimensionality of embeddings for nodes and edges. embed_input (bool): whether to embed input or not. See https://arxiv.org/abs/1810.00826 """ def __init__(self, emb_dim, aggr = "add"): super(GINConv, self).__init__() #multi-layer perceptron self.mlp = torch.nn.Sequential(torch.nn.Linear(emb_dim, 2*emb_dim), torch.nn.ReLU(), torch.nn.Linear(2*emb_dim, emb_dim)) self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim) self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim) torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data) torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data) self.aggr = aggr def forward(self, x, edge_index, edge_attr): #add self loops in the edge space edge_index = add_self_loops(edge_index, num_nodes = x.size(0)) #add features corresponding to self-loop edges. self_loop_attr = torch.zeros(x.size(0), 2) self_loop_attr[:,0] = 4 #bond type for self-loop edge self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype) edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0) edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1]) return self.propagate(self.aggr, edge_index, x=x, edge_attr=edge_embeddings) def message(self, x_j, edge_attr): return x_j + edge_attr def update(self, aggr_out): return self.mlp(aggr_out) class GCNConv(MessagePassing): def __init__(self, emb_dim, aggr = "add"): super(GCNConv, self).__init__() self.emb_dim = emb_dim self.linear = torch.nn.Linear(emb_dim, emb_dim) self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim) self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim) torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data) torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data) self.aggr = aggr def norm(self, edge_index, num_nodes, dtype): ### assuming that self-loops have been already added in edge_index edge_weight = torch.ones((edge_index.size(1), ), dtype=dtype, device=edge_index.device) row, col = edge_index deg = scatter_add(edge_weight, row, dim=0, dim_size=num_nodes) deg_inv_sqrt = deg.pow(-0.5) deg_inv_sqrt[deg_inv_sqrt == float('inf')] = 0 return deg_inv_sqrt[row] * edge_weight * deg_inv_sqrt[col] def forward(self, x, edge_index, edge_attr): #add self loops in the edge space edge_index = add_self_loops(edge_index, num_nodes = x.size(0)) #add features corresponding to self-loop edges. self_loop_attr = torch.zeros(x.size(0), 2) self_loop_attr[:,0] = 4 #bond type for self-loop edge self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype) edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0) edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1]) norm = self.norm(edge_index, x.size(0), x.dtype) x = self.linear(x) return self.propagate(self.aggr, edge_index, x=x, edge_attr=edge_embeddings, norm = norm) def message(self, x_j, edge_attr, norm): return norm.view(-1, 1) * (x_j + edge_attr) class GATConv(MessagePassing): def __init__(self, emb_dim, heads=2, negative_slope=0.2, aggr = "add"): super(GATConv, self).__init__() self.aggr = aggr self.emb_dim = emb_dim self.heads = heads self.negative_slope = negative_slope self.weight_linear = torch.nn.Linear(emb_dim, heads * emb_dim) self.att = torch.nn.Parameter(torch.Tensor(1, heads, 2 * emb_dim)) self.bias = torch.nn.Parameter(torch.Tensor(emb_dim)) self.edge_embedding1 = torch.nn.Embedding(num_bond_type, heads * emb_dim) self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, heads * emb_dim) torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data) torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data) self.reset_parameters() def reset_parameters(self): glorot(self.att) zeros(self.bias) def forward(self, x, edge_index, edge_attr): #add self loops in the edge space edge_index = add_self_loops(edge_index, num_nodes = x.size(0)) #add features corresponding to self-loop edges. self_loop_attr = torch.zeros(x.size(0), 2) self_loop_attr[:,0] = 4 #bond type for self-loop edge self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype) edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0) edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1]) x = self.weight_linear(x).view(-1, self.heads, self.emb_dim) return self.propagate(self.aggr, edge_index, x=x, edge_attr=edge_embeddings) def message(self, edge_index, x_i, x_j, edge_attr): edge_attr = edge_attr.view(-1, self.heads, self.emb_dim) x_j += edge_attr alpha = (torch.cat([x_i, x_j], dim=-1) * self.att).sum(dim=-1) alpha = F.leaky_relu(alpha, self.negative_slope) alpha = softmax(alpha, edge_index[0]) return x_j * alpha.view(-1, self.heads, 1) def update(self, aggr_out): aggr_out = aggr_out.mean(dim=1) aggr_out = aggr_out + self.bias return aggr_out class GraphSAGEConv(MessagePassing): def __init__(self, emb_dim, aggr = "mean"): super(GraphSAGEConv, self).__init__() self.emb_dim = emb_dim self.linear = torch.nn.Linear(emb_dim, emb_dim) self.edge_embedding1 = torch.nn.Embedding(num_bond_type, emb_dim) self.edge_embedding2 = torch.nn.Embedding(num_bond_direction, emb_dim) torch.nn.init.xavier_uniform_(self.edge_embedding1.weight.data) torch.nn.init.xavier_uniform_(self.edge_embedding2.weight.data) self.aggr = aggr def forward(self, x, edge_index, edge_attr): #add self loops in the edge space edge_index = add_self_loops(edge_index, num_nodes = x.size(0)) #add features corresponding to self-loop edges. self_loop_attr = torch.zeros(x.size(0), 2) self_loop_attr[:,0] = 4 #bond type for self-loop edge self_loop_attr = self_loop_attr.to(edge_attr.device).to(edge_attr.dtype) edge_attr = torch.cat((edge_attr, self_loop_attr), dim = 0) edge_embeddings = self.edge_embedding1(edge_attr[:,0]) + self.edge_embedding2(edge_attr[:,1]) x = self.linear(x) return self.propagate(self.aggr, edge_index, x=x, edge_attr=edge_embeddings) def message(self, x_j, edge_attr): return x_j + edge_attr def update(self, aggr_out): return F.normalize(aggr_out, p = 2, dim = -1) class GNN(torch.nn.Module): """ Args: num_layer (int): the number of GNN layers emb_dim (int): dimensionality of embeddings JK (str): last, concat, max or sum. max_pool_layer (int): the layer from which we use max pool rather than add pool for neighbor aggregation drop_ratio (float): dropout rate gnn_type: gin, gcn, graphsage, gat Output: node representations """ def __init__(self, num_layer, emb_dim, JK = "last", drop_ratio = 0, gnn_type = "gin"): super(GNN, self).__init__() self.num_layer = num_layer self.drop_ratio = drop_ratio self.JK = JK if self.num_layer < 2: raise ValueError("Number of GNN layers must be greater than 1.") self.x_embedding1 = torch.nn.Embedding(num_atom_type, emb_dim) self.x_embedding2 = torch.nn.Embedding(num_chirality_tag, emb_dim) torch.nn.init.xavier_uniform_(self.x_embedding1.weight.data) torch.nn.init.xavier_uniform_(self.x_embedding2.weight.data) ###List of MLPs self.gnns = torch.nn.ModuleList() for layer in range(num_layer): if gnn_type == "gin": self.gnns.append(GINConv(emb_dim, aggr = "add")) elif gnn_type == "gcn": self.gnns.append(GCNConv(emb_dim)) elif gnn_type == "gat": self.gnns.append(GATConv(emb_dim)) elif gnn_type == "graphsage": self.gnns.append(GraphSAGEConv(emb_dim)) ###List of batchnorms self.batch_norms = torch.nn.ModuleList() for layer in range(num_layer): self.batch_norms.append(torch.nn.BatchNorm1d(emb_dim)) #def forward(self, x, edge_index, edge_attr): def forward(self, *argv): if len(argv) == 3: x, edge_index, edge_attr = argv[0], argv[1], argv[2] elif len(argv) == 1: data = argv[0] x, edge_index, edge_attr = data.x, data.edge_index, data.edge_attr else: raise ValueError("unmatched number of arguments.") x = self.x_embedding1(x[:,0]) + self.x_embedding2(x[:,1]) h_list = [x] for layer in range(self.num_layer): h = self.gnns[layer](h_list[layer], edge_index, edge_attr) h = self.batch_norms[layer](h) #h = F.dropout(F.relu(h), self.drop_ratio, training = self.training) if layer == self.num_layer - 1: #remove relu for the last layer h = F.dropout(h, self.drop_ratio, training = self.training) else: h = F.dropout(F.relu(h), self.drop_ratio, training = self.training) h_list.append(h) ### Different implementations of Jk-concat if self.JK == "concat": node_representation = torch.cat(h_list, dim = 1) elif self.JK == "last": node_representation = h_list[-1] elif self.JK == "max": h_list = [h.unsqueeze_(0) for h in h_list] node_representation = torch.max(torch.cat(h_list, dim = 0), dim = 0)[0] elif self.JK == "sum": h_list = [h.unsqueeze_(0) for h in h_list] node_representation = torch.sum(torch.cat(h_list, dim = 0), dim = 0)[0] return node_representation class GNN_graphpred(torch.nn.Module): """ Extension of GIN to incorporate edge information by concatenation. Args: num_layer (int): the number of GNN layers emb_dim (int): dimensionality of embeddings num_tasks (int): number of tasks in multi-task learning scenario drop_ratio (float): dropout rate JK (str): last, concat, max or sum. graph_pooling (str): sum, mean, max, attention, set2set gnn_type: gin, gcn, graphsage, gat See https://arxiv.org/abs/1810.00826 JK-net: https://arxiv.org/abs/1806.03536 """ def __init__(self, num_layer, emb_dim, num_tasks, JK = "last", drop_ratio = 0, graph_pooling = "mean", gnn_type = "gin"): super(GNN_graphpred, self).__init__() self.num_layer = num_layer self.drop_ratio = drop_ratio self.JK = JK self.emb_dim = emb_dim self.num_tasks = num_tasks if self.num_layer < 2: raise ValueError("Number of GNN layers must be greater than 1.") self.gnn = GNN(num_layer, emb_dim, JK, drop_ratio, gnn_type = gnn_type) #Different kind of graph pooling if graph_pooling == "sum": self.pool = global_add_pool elif graph_pooling == "mean": self.pool = global_mean_pool elif graph_pooling == "max": self.pool = global_max_pool elif graph_pooling == "attention": if self.JK == "concat": self.pool = GlobalAttention(gate_nn = torch.nn.Linear((self.num_layer + 1) * emb_dim, 1)) else: self.pool = GlobalAttention(gate_nn = torch.nn.Linear(emb_dim, 1)) elif graph_pooling[:-1] == "set2set": set2set_iter = int(graph_pooling[-1]) if self.JK == "concat": self.pool = Set2Set((self.num_layer + 1) * emb_dim, set2set_iter) else: self.pool = Set2Set(emb_dim, set2set_iter) else: raise ValueError("Invalid graph pooling type.") #For graph-level binary classification if graph_pooling[:-1] == "set2set": self.mult = 2 else: self.mult = 1 if self.JK == "concat": rep_dim = self.mult * (self.num_layer + 1) * self.emb_dim self.graph_pred_linear = torch.nn.Linear(rep_dim, self.num_tasks) # self.graph_pred_linear = torch.nn.Sequential( # torch.nn.Linear(rep_dim, rep_dim), # torch.nn.ReLU(), # torch.nn.Linear(rep_dim, self.num_tasks) # ) else: rep_dim = self.mult * self.emb_dim self.graph_pred_linear = torch.nn.Linear(rep_dim, self.num_tasks) # self.graph_pred_linear = torch.nn.Sequential( # torch.nn.Linear(rep_dim, rep_dim), # torch.nn.ReLU(), # torch.nn.Linear(rep_dim, self.num_tasks) # ) def from_pretrained(self, model_file): #self.gnn = GNN(self.num_layer, self.emb_dim, JK = self.JK, drop_ratio = self.drop_ratio) self.gnn.load_state_dict(torch.load(model_file)) def forward(self, *argv): if len(argv) == 4: x, edge_index, edge_attr, batch = argv[0], argv[1], argv[2], argv[3] elif len(argv) == 1: data = argv[0] x, edge_index, edge_attr, batch = data.x, data.edge_index, data.edge_attr, data.batch else: raise ValueError("unmatched number of arguments.") node_representation = self.gnn(x, edge_index, edge_attr) return self.graph_pred_linear(self.pool(node_representation, batch)) class ProjectNet(torch.nn.Module): def __init__(self, rep_dim): super(ProjectNet, self).__init__() self.rep_dim = rep_dim self.proj = torch.nn.Sequential( torch.nn.Linear(self.rep_dim, self.rep_dim), torch.nn.ReLU(), torch.nn.Linear(self.rep_dim, self.rep_dim) ) def forward(self, x): x_proj = self.proj(x) return x_proj if __name__ == "__main__": pass
15,224
36.967581
129
py
GraphLoG
GraphLoG-main/finetune.py
import argparse from loader import MoleculeDataset from torch_geometric.data import DataLoader import torch import torch.nn as nn import torch.nn.functional as F import torch.optim as optim from torch.optim.lr_scheduler import StepLR from tqdm import tqdm import os, sys import numpy as np import random from model import GNN, GNN_graphpred from sklearn.metrics import roc_auc_score from splitters import scaffold_split import pandas as pd import os import shutil from tensorboardX import SummaryWriter criterion = nn.BCEWithLogitsLoss(reduction = "none") def train(args, model, device, loader, optimizer, scheduler): model.train() scheduler.step() for step, batch in enumerate(tqdm(loader, desc="Iteration")): batch = batch.to(device) pred = model(batch.x, batch.edge_index, batch.edge_attr, batch.batch) y = batch.y.view(pred.shape).to(torch.float64) #Whether y is non-null or not. is_valid = y**2 > 0 #Loss matrix loss_mat = criterion(pred.double(), (y+1)/2) #loss matrix after removing null target loss_mat = torch.where(is_valid, loss_mat, torch.zeros(loss_mat.shape).to(loss_mat.device).to(loss_mat.dtype)) optimizer.zero_grad() loss = torch.sum(loss_mat)/torch.sum(is_valid) loss.backward() optimizer.step() def eval(args, model, device, loader): model.eval() y_true = [] y_scores = [] for step, batch in enumerate(tqdm(loader, desc="Iteration")): batch = batch.to(device) with torch.no_grad(): pred = model(batch.x, batch.edge_index, batch.edge_attr, batch.batch) y_true.append(batch.y.view(pred.shape)) y_scores.append(pred) y_true = torch.cat(y_true, dim = 0).cpu().numpy() y_scores = torch.cat(y_scores, dim = 0).cpu().numpy() roc_list = [] for i in range(y_true.shape[1]): #AUC is only defined when there is at least one positive data. if np.sum(y_true[:,i] == 1) > 0 and np.sum(y_true[:,i] == -1) > 0: is_valid = y_true[:,i]**2 > 0 roc_list.append(roc_auc_score((y_true[is_valid,i] + 1)/2, y_scores[is_valid,i])) if len(roc_list) < y_true.shape[1]: print("Some target is missing!") print("Missing ratio: %f" %(1 - float(len(roc_list))/y_true.shape[1])) return sum(roc_list)/len(roc_list) #y_true.shape[1] def main(): # Training settings parser = argparse.ArgumentParser(description='PyTorch implementation of pre-training of graph neural networks') parser.add_argument('--device', type=int, default=0, help='which gpu to use if any (default: 0)') parser.add_argument('--batch_size', type=int, default=32, help='input batch size for training (default: 32)') parser.add_argument('--epochs', type=int, default=100, help='number of epochs to train (default: 100)') parser.add_argument('--num_run', type=int, default=5, help='number of independent runs (default: 5)') parser.add_argument('--lr', type=float, default=0.001, help='learning rate (default: 0.001)') parser.add_argument('--lr_scale', type=float, default=1, help='relative learning rate for the feature extraction layer (default: 1)') parser.add_argument('--frozen', action='store_true', default=False, help='whether to freeze gnn extractor') parser.add_argument('--decay', type=float, default=0, help='weight decay (default: 0)') parser.add_argument('--num_layer', type=int, default=5, help='number of GNN message passing layers (default: 5).') parser.add_argument('--emb_dim', type=int, default=300, help='embedding dimensions (default: 300)') parser.add_argument('--dropout_ratio', type=float, default=0.5, help='dropout ratio (default: 0.5)') parser.add_argument('--graph_pooling', type=str, default="mean", help='graph level pooling (sum, mean, max, set2set, attention)') parser.add_argument('--JK', type=str, default="last", help='how the node features across layers are combined. last, sum, max or concat') parser.add_argument('--gnn_type', type=str, default="gin") parser.add_argument('--dataset', type=str, default = 'bbbp', help='root directory of dataset. For now, only classification.') parser.add_argument('--input_model_file', type=str, default = '', help='filename to read the model (if there is any)') parser.add_argument('--filename', type=str, default = '', help='output filename') parser.add_argument('--seed', type=int, default=None, help = "Seed for splitting the dataset.") parser.add_argument('--runseed', type=int, default=None, help = "Seed for minibatch selection, random initialization.") parser.add_argument('--split', type = str, default="scaffold", help = "random or scaffold or random_scaffold") parser.add_argument('--eval_train', type=int, default = 0, help='evaluating training or not') parser.add_argument('--num_workers', type=int, default = 1, help='number of workers for dataset loading') args = parser.parse_args() if args.seed: seed = args.seed print ('Manual seed: ', seed) else: seed = random.randint(0, 10000) print ('Random seed: ', seed) # Bunch of classification tasks if args.dataset == "tox21": num_tasks = 12 elif args.dataset == "hiv": num_tasks = 1 elif args.dataset == "pcba": num_tasks = 128 elif args.dataset == "muv": num_tasks = 17 elif args.dataset == "bace": num_tasks = 1 elif args.dataset == "bbbp": num_tasks = 1 elif args.dataset == "toxcast": num_tasks = 617 elif args.dataset == "sider": num_tasks = 27 elif args.dataset == "clintox": num_tasks = 2 else: raise ValueError("Invalid dataset name.") # set up dataset dataset = MoleculeDataset("./dataset/" + args.dataset, dataset=args.dataset) print(dataset) if args.split == "scaffold": smiles_list = pd.read_csv('./dataset/' + args.dataset + '/processed/smiles.csv', header=None)[0].tolist() train_dataset, valid_dataset, test_dataset = scaffold_split(dataset, smiles_list, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1) print("scaffold") elif args.split == "random": train_dataset, valid_dataset, test_dataset = random_split(dataset, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1, seed=seed) print("random") elif args.split == "random_scaffold": smiles_list = pd.read_csv('./dataset/' + args.dataset + '/processed/smiles.csv', header=None)[0].tolist() train_dataset, valid_dataset, test_dataset = random_scaffold_split(dataset, smiles_list, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1, seed=seed) print("random scaffold") else: raise ValueError("Invalid split option.") print(train_dataset[0]) # run multiple times best_valid_auc_list = [] last_epoch_auc_list = [] for run_idx in range(args.num_run): print ('\nRun ', run_idx + 1) if args.runseed: runseed = args.runseed print('Manual runseed: ', runseed) else: runseed = random.randint(0, 10000) print('Random runseed: ', runseed) torch.manual_seed(runseed) np.random.seed(runseed) device = torch.device("cuda:" + str(args.device)) if torch.cuda.is_available() else torch.device("cpu") if torch.cuda.is_available(): torch.cuda.manual_seed_all(runseed) train_loader = DataLoader(train_dataset, batch_size=args.batch_size, shuffle=True, num_workers=args.num_workers) val_loader = DataLoader(valid_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers) test_loader = DataLoader(test_dataset, batch_size=args.batch_size, shuffle=False, num_workers=args.num_workers) # set up model model = GNN_graphpred(args.num_layer, args.emb_dim, num_tasks, JK=args.JK, drop_ratio=args.dropout_ratio, graph_pooling=args.graph_pooling, gnn_type=args.gnn_type) if not args.input_model_file == "": model.from_pretrained(args.input_model_file) model.to(device) # set up optimizer # different learning rate for different part of GNN model_param_group = [] if args.frozen: model_param_group.append({"params": model.gnn.parameters(), "lr": 0}) else: model_param_group.append({"params": model.gnn.parameters()}) if args.graph_pooling == "attention": model_param_group.append({"params": model.pool.parameters(), "lr": args.lr * args.lr_scale}) model_param_group.append({"params": model.graph_pred_linear.parameters(), "lr": args.lr * args.lr_scale}) optimizer = optim.Adam(model_param_group, lr=args.lr, weight_decay=args.decay) print(optimizer) scheduler = StepLR(optimizer, step_size=30, gamma=0.3) # run fine-tuning best_valid = 0 best_valid_test = 0 last_epoch_test = 0 for epoch in range(1, args.epochs + 1): print("====epoch " + str(epoch), " lr: ", optimizer.param_groups[-1]['lr']) train(args, model, device, train_loader, optimizer, scheduler) print("====Evaluation") if args.eval_train: train_acc = eval(args, model, device, train_loader) else: print("omit the training accuracy computation") train_acc = 0 val_acc = eval(args, model, device, val_loader) test_acc = eval(args, model, device, test_loader) if val_acc > best_valid: best_valid = val_acc best_valid_test = test_acc if epoch == args.epochs: last_epoch_test = test_acc print("train: %f val: %f test: %f" % (train_acc, val_acc, test_acc)) print("") best_valid_auc_list.append(best_valid_test) last_epoch_auc_list.append(last_epoch_test) # summarize results best_valid_auc_list = np.array(best_valid_auc_list) last_epoch_auc_list = np.array(last_epoch_auc_list) if args.dataset in ["muv", "hiv"]: print('Best validation epoch:') print('Mean: {}\tStd: {}'.format(np.mean(best_valid_auc_list), np.std(best_valid_auc_list))) else: print('Last epoch:') print('Mean: {}\tStd: {}'.format(np.mean(last_epoch_auc_list), np.std(last_epoch_auc_list))) os.system('watch nvidia-smi') if __name__ == "__main__": main()
11,223
40.724907
129
py
GraphLoG
GraphLoG-main/splitters.py
import torch import random import numpy as np from itertools import compress from rdkit.Chem.Scaffolds import MurckoScaffold from collections import defaultdict from sklearn.model_selection import StratifiedKFold # splitter function def generate_scaffold(smiles, include_chirality=False): """ Obtain Bemis-Murcko scaffold from smiles :param smiles: :param include_chirality: :return: smiles of scaffold """ scaffold = MurckoScaffold.MurckoScaffoldSmiles( smiles=smiles, includeChirality=include_chirality) return scaffold # # test generate_scaffold # s = 'Cc1cc(Oc2nccc(CCC)c2)ccc1' # scaffold = generate_scaffold(s) # assert scaffold == 'c1ccc(Oc2ccccn2)cc1' def scaffold_split(dataset, smiles_list, task_idx=None, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1, return_smiles=False): """ Adapted from https://github.com/deepchem/deepchem/blob/master/deepchem/splits/splitters.py Split dataset by Bemis-Murcko scaffolds This function can also ignore examples containing null values for a selected task when splitting. Deterministic split :param dataset: pytorch geometric dataset obj :param smiles_list: list of smiles corresponding to the dataset obj :param task_idx: column idx of the data.y tensor. Will filter out examples with null value in specified task column of the data.y tensor prior to splitting. If None, then no filtering :param null_value: float that specifies null value in data.y to filter if task_idx is provided :param frac_train: :param frac_valid: :param frac_test: :param return_smiles: :return: train, valid, test slices of the input dataset obj. If return_smiles = True, also returns ([train_smiles_list], [valid_smiles_list], [test_smiles_list]) """ np.testing.assert_almost_equal(frac_train + frac_valid + frac_test, 1.0) if task_idx != None: # filter based on null values in task_idx # get task array y_task = np.array([data.y[task_idx].item() for data in dataset]) # boolean array that correspond to non null values non_null = y_task != null_value smiles_list = list(compress(enumerate(smiles_list), non_null)) else: non_null = np.ones(len(dataset)) == 1 smiles_list = list(compress(enumerate(smiles_list), non_null)) # create dict of the form {scaffold_i: [idx1, idx....]} all_scaffolds = {} for i, smiles in smiles_list: scaffold = generate_scaffold(smiles, include_chirality=True) if scaffold not in all_scaffolds: all_scaffolds[scaffold] = [i] else: all_scaffolds[scaffold].append(i) # sort from largest to smallest sets all_scaffolds = {key: sorted(value) for key, value in all_scaffolds.items()} all_scaffold_sets = [ scaffold_set for (scaffold, scaffold_set) in sorted( all_scaffolds.items(), key=lambda x: (len(x[1]), x[1][0]), reverse=True) ] # get train, valid test indices train_cutoff = frac_train * len(smiles_list) valid_cutoff = (frac_train + frac_valid) * len(smiles_list) train_idx, valid_idx, test_idx = [], [], [] for scaffold_set in all_scaffold_sets: if len(train_idx) + len(scaffold_set) > train_cutoff: if len(train_idx) + len(valid_idx) + len(scaffold_set) > valid_cutoff: test_idx.extend(scaffold_set) else: valid_idx.extend(scaffold_set) else: train_idx.extend(scaffold_set) assert len(set(train_idx).intersection(set(valid_idx))) == 0 assert len(set(test_idx).intersection(set(valid_idx))) == 0 train_dataset = dataset[torch.tensor(train_idx)] valid_dataset = dataset[torch.tensor(valid_idx)] test_dataset = dataset[torch.tensor(test_idx)] if not return_smiles: return train_dataset, valid_dataset, test_dataset else: train_smiles = [smiles_list[i][1] for i in train_idx] valid_smiles = [smiles_list[i][1] for i in valid_idx] test_smiles = [smiles_list[i][1] for i in test_idx] return train_dataset, valid_dataset, test_dataset, (train_smiles, valid_smiles, test_smiles) def random_scaffold_split(dataset, smiles_list, task_idx=None, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1, seed=0): """ Adapted from https://github.com/pfnet-research/chainer-chemistry/blob/master/chainer_chemistry/dataset/splitters/scaffold_splitter.py Split dataset by Bemis-Murcko scaffolds This function can also ignore examples containing null values for a selected task when splitting. Deterministic split :param dataset: pytorch geometric dataset obj :param smiles_list: list of smiles corresponding to the dataset obj :param task_idx: column idx of the data.y tensor. Will filter out examples with null value in specified task column of the data.y tensor prior to splitting. If None, then no filtering :param null_value: float that specifies null value in data.y to filter if task_idx is provided :param frac_train: :param frac_valid: :param frac_test: :param seed; :return: train, valid, test slices of the input dataset obj """ np.testing.assert_almost_equal(frac_train + frac_valid + frac_test, 1.0) if task_idx != None: # filter based on null values in task_idx # get task array y_task = np.array([data.y[task_idx].item() for data in dataset]) # boolean array that correspond to non null values non_null = y_task != null_value smiles_list = list(compress(enumerate(smiles_list), non_null)) else: non_null = np.ones(len(dataset)) == 1 smiles_list = list(compress(enumerate(smiles_list), non_null)) rng = np.random.RandomState(seed) scaffolds = defaultdict(list) for ind, smiles in smiles_list: scaffold = generate_scaffold(smiles, include_chirality=True) scaffolds[scaffold].append(ind) scaffold_sets = rng.permutation(list(scaffolds.values())) n_total_valid = int(np.floor(frac_valid * len(dataset))) n_total_test = int(np.floor(frac_test * len(dataset))) train_idx = [] valid_idx = [] test_idx = [] for scaffold_set in scaffold_sets: if len(valid_idx) + len(scaffold_set) <= n_total_valid: valid_idx.extend(scaffold_set) elif len(test_idx) + len(scaffold_set) <= n_total_test: test_idx.extend(scaffold_set) else: train_idx.extend(scaffold_set) train_dataset = dataset[torch.tensor(train_idx)] valid_dataset = dataset[torch.tensor(valid_idx)] test_dataset = dataset[torch.tensor(test_idx)] return train_dataset, valid_dataset, test_dataset def random_split(dataset, task_idx=None, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1, seed=0, smiles_list=None): """ :param dataset: :param task_idx: :param null_value: :param frac_train: :param frac_valid: :param frac_test: :param seed: :param smiles_list: list of smiles corresponding to the dataset obj, or None :return: train, valid, test slices of the input dataset obj. If smiles_list != None, also returns ([train_smiles_list], [valid_smiles_list], [test_smiles_list]) """ np.testing.assert_almost_equal(frac_train + frac_valid + frac_test, 1.0) if task_idx != None: # filter based on null values in task_idx # get task array y_task = np.array([data.y[task_idx].item() for data in dataset]) non_null = y_task != null_value # boolean array that correspond to non null values idx_array = np.where(non_null)[0] dataset = dataset[torch.tensor(idx_array)] # examples containing non # null labels in the specified task_idx else: pass num_mols = len(dataset) random.seed(seed) all_idx = list(range(num_mols)) random.shuffle(all_idx) train_idx = all_idx[:int(frac_train * num_mols)] valid_idx = all_idx[int(frac_train * num_mols):int(frac_valid * num_mols) + int(frac_train * num_mols)] test_idx = all_idx[int(frac_valid * num_mols) + int(frac_train * num_mols):] assert len(set(train_idx).intersection(set(valid_idx))) == 0 assert len(set(valid_idx).intersection(set(test_idx))) == 0 assert len(train_idx) + len(valid_idx) + len(test_idx) == num_mols train_dataset = dataset[torch.tensor(train_idx)] valid_dataset = dataset[torch.tensor(valid_idx)] test_dataset = dataset[torch.tensor(test_idx)] if not smiles_list: return train_dataset, valid_dataset, test_dataset else: train_smiles = [smiles_list[i] for i in train_idx] valid_smiles = [smiles_list[i] for i in valid_idx] test_smiles = [smiles_list[i] for i in test_idx] return train_dataset, valid_dataset, test_dataset, (train_smiles, valid_smiles, test_smiles) def cv_random_split(dataset, fold_idx = 0, frac_train=0.9, frac_valid=0.1, seed=0, smiles_list=None): """ :param dataset: :param task_idx: :param null_value: :param frac_train: :param frac_valid: :param frac_test: :param seed: :param smiles_list: list of smiles corresponding to the dataset obj, or None :return: train, valid, test slices of the input dataset obj. If smiles_list != None, also returns ([train_smiles_list], [valid_smiles_list], [test_smiles_list]) """ np.testing.assert_almost_equal(frac_train + frac_valid, 1.0) skf = StratifiedKFold(n_splits=10, shuffle = True, random_state = seed) labels = [data.y.item() for data in dataset] idx_list = [] for idx in skf.split(np.zeros(len(labels)), labels): idx_list.append(idx) train_idx, val_idx = idx_list[fold_idx] train_dataset = dataset[torch.tensor(train_idx)] valid_dataset = dataset[torch.tensor(val_idx)] return train_dataset, valid_dataset if __name__ == "__main__": from loader import MoleculeDataset from rdkit import Chem import pandas as pd # # test scaffold_split dataset = MoleculeDataset('dataset/tox21', dataset='tox21') smiles_list = pd.read_csv('dataset/tox21/processed/smiles.csv', header=None)[0].tolist() train_dataset, valid_dataset, test_dataset = scaffold_split(dataset, smiles_list, task_idx=None, null_value=0, frac_train=0.8,frac_valid=0.1, frac_test=0.1) # train_dataset, valid_dataset, test_dataset = random_scaffold_split(dataset, smiles_list, task_idx=None, null_value=0, frac_train=0.8,frac_valid=0.1, frac_test=0.1, seed = 0) unique_ids = set(train_dataset.data.id.tolist() + valid_dataset.data.id.tolist() + test_dataset.data.id.tolist()) assert len(unique_ids) == len(dataset) # check that we did not have any # missing or overlapping examples # test scaffold_split with smiles returned dataset = MoleculeDataset('dataset/bbbp', dataset='bbbp') smiles_list = pd.read_csv('dataset/bbbp/processed/smiles.csv', header=None)[ 0].tolist() train_dataset, valid_dataset, test_dataset, (train_smiles, valid_smiles, test_smiles) = \ scaffold_split(dataset, smiles_list, task_idx=None, null_value=0, frac_train=0.8,frac_valid=0.1, frac_test=0.1, return_smiles=True) assert len(train_dataset) == len(train_smiles) for i in range(len(train_dataset)): data_obj_n_atoms = train_dataset[i].x.size()[0] smiles_n_atoms = len(list(Chem.MolFromSmiles(train_smiles[ i]).GetAtoms())) assert data_obj_n_atoms == smiles_n_atoms assert len(valid_dataset) == len(valid_smiles) for i in range(len(valid_dataset)): data_obj_n_atoms = valid_dataset[i].x.size()[0] smiles_n_atoms = len(list(Chem.MolFromSmiles(valid_smiles[ i]).GetAtoms())) assert data_obj_n_atoms == smiles_n_atoms assert len(test_dataset) == len(test_smiles) for i in range(len(test_dataset)): data_obj_n_atoms = test_dataset[i].x.size()[0] smiles_n_atoms = len(list(Chem.MolFromSmiles(test_smiles[ i]).GetAtoms())) assert data_obj_n_atoms == smiles_n_atoms # test random_split from loader import MoleculeDataset dataset = MoleculeDataset('dataset/tox21', dataset='tox21') train_dataset, valid_dataset, test_dataset = random_split(dataset, task_idx=None, null_value=0, frac_train=0.8,frac_valid=0.1, frac_test=0.1) unique_ids = set(train_dataset.data.id.tolist() + valid_dataset.data.id.tolist() + test_dataset.data.id.tolist()) assert len(unique_ids) == len(dataset) # check that we did not have any # missing or overlapping examples # test random_split with smiles returned dataset = MoleculeDataset('dataset/bbbp', dataset='bbbp') smiles_list = pd.read_csv('dataset/bbbp/processed/smiles.csv', header=None)[ 0].tolist() train_dataset, valid_dataset, test_dataset, (train_smiles, valid_smiles, test_smiles) = \ random_split(dataset, task_idx=None, null_value=0, frac_train=0.8, frac_valid=0.1, frac_test=0.1, seed=42, smiles_list=smiles_list) assert len(train_dataset) == len(train_smiles) for i in range(len(train_dataset)): data_obj_n_atoms = train_dataset[i].x.size()[0] smiles_n_atoms = len(list(Chem.MolFromSmiles(train_smiles[ i]).GetAtoms())) assert data_obj_n_atoms == smiles_n_atoms assert len(valid_dataset) == len(valid_smiles) for i in range(len(valid_dataset)): data_obj_n_atoms = valid_dataset[i].x.size()[0] smiles_n_atoms = len(list(Chem.MolFromSmiles(valid_smiles[ i]).GetAtoms())) assert data_obj_n_atoms == smiles_n_atoms assert len(test_dataset) == len(test_smiles) for i in range(len(test_dataset)): data_obj_n_atoms = test_dataset[i].x.size()[0] smiles_n_atoms = len(list(Chem.MolFromSmiles(test_smiles[ i]).GetAtoms())) assert data_obj_n_atoms == smiles_n_atoms
14,949
41.351275
179
py
GraphLoG
GraphLoG-main/util.py
import torch import copy import random import networkx as nx import numpy as np from torch_geometric.utils import convert from loader import graph_data_obj_to_nx_simple, nx_to_graph_data_obj_simple from rdkit import Chem from rdkit.Chem import AllChem from loader import mol_to_graph_data_obj_simple, \ graph_data_obj_to_mol_simple from loader import MoleculeDataset def check_same_molecules(s1, s2): mol1 = AllChem.MolFromSmiles(s1) mol2 = AllChem.MolFromSmiles(s2) return AllChem.MolToInchi(mol1) == AllChem.MolToInchi(mol2) class NegativeEdge: def __init__(self): """ Randomly sample negative edges """ pass def __call__(self, data): num_nodes = data.num_nodes num_edges = data.num_edges edge_set = set([str(data.edge_index[0, i].cpu().item()) + "," + str( data.edge_index[1, i].cpu().item()) for i in range(data.edge_index.shape[1])]) redandunt_sample = torch.randint(0, num_nodes, (2, 5 * num_edges)) sampled_ind = [] sampled_edge_set = set([]) for i in range(5 * num_edges): node1 = redandunt_sample[0, i].cpu().item() node2 = redandunt_sample[1, i].cpu().item() edge_str = str(node1) + "," + str(node2) if not edge_str in edge_set and not edge_str in sampled_edge_set and not node1 == node2: sampled_edge_set.add(edge_str) sampled_ind.append(i) if len(sampled_ind) == num_edges / 2: break data.negative_edge_index = redandunt_sample[:, sampled_ind] return data class ExtractSubstructureContextPair: def __init__(self, k, l1, l2): """ Randomly selects a node from the data object, and adds attributes that contain the substructure that corresponds to k hop neighbours rooted at the node, and the context substructures that corresponds to the subgraph that is between l1 and l2 hops away from the root node. :param k: :param l1: :param l2: """ self.k = k self.l1 = l1 self.l2 = l2 # for the special case of 0, addresses the quirk with # single_source_shortest_path_length if self.k == 0: self.k = -1 if self.l1 == 0: self.l1 = -1 if self.l2 == 0: self.l2 = -1 def __call__(self, data, root_idx=None): """ :param data: pytorch geometric data object :param root_idx: If None, then randomly samples an atom idx. Otherwise sets atom idx of root (for debugging only) :return: None. Creates new attributes in original data object: data.center_substruct_idx data.x_substruct data.edge_attr_substruct data.edge_index_substruct data.x_context data.edge_attr_context data.edge_index_context data.overlap_context_substruct_idx """ num_atoms = data.x.size()[0] if root_idx == None: root_idx = random.sample(range(num_atoms), 1)[0] G = graph_data_obj_to_nx_simple(data) # same ordering as input data obj # Get k-hop subgraph rooted at specified atom idx substruct_node_idxes = nx.single_source_shortest_path_length(G, root_idx, self.k).keys() if len(substruct_node_idxes) > 0: substruct_G = G.subgraph(substruct_node_idxes) substruct_G, substruct_node_map = reset_idxes(substruct_G) # need # to reset node idx to 0 -> num_nodes - 1, otherwise data obj does not # make sense, since the node indices in data obj must start at 0 substruct_data = nx_to_graph_data_obj_simple(substruct_G) data.x_substruct = substruct_data.x data.edge_attr_substruct = substruct_data.edge_attr data.edge_index_substruct = substruct_data.edge_index data.center_substruct_idx = torch.tensor([substruct_node_map[ root_idx]]) # need # to convert center idx from original graph node ordering to the # new substruct node ordering # Get subgraphs that is between l1 and l2 hops away from the root node l1_node_idxes = nx.single_source_shortest_path_length(G, root_idx, self.l1).keys() l2_node_idxes = nx.single_source_shortest_path_length(G, root_idx, self.l2).keys() context_node_idxes = set(l1_node_idxes).symmetric_difference( set(l2_node_idxes)) if len(context_node_idxes) > 0: context_G = G.subgraph(context_node_idxes) context_G, context_node_map = reset_idxes(context_G) # need to # reset node idx to 0 -> num_nodes - 1, otherwise data obj does not # make sense, since the node indices in data obj must start at 0 context_data = nx_to_graph_data_obj_simple(context_G) data.x_context = context_data.x data.edge_attr_context = context_data.edge_attr data.edge_index_context = context_data.edge_index # Get indices of overlapping nodes between substruct and context, # WRT context ordering context_substruct_overlap_idxes = list(set( context_node_idxes).intersection(set(substruct_node_idxes))) if len(context_substruct_overlap_idxes) > 0: context_substruct_overlap_idxes_reorder = [context_node_map[old_idx] for old_idx in context_substruct_overlap_idxes] # need to convert the overlap node idxes, which is from the # original graph node ordering to the new context node ordering data.overlap_context_substruct_idx = \ torch.tensor(context_substruct_overlap_idxes_reorder) return data def __repr__(self): return '{}(k={},l1={}, l2={})'.format(self.__class__.__name__, self.k, self.l1, self.l2) def reset_idxes(G): """ Resets node indices such that they are numbered from 0 to num_nodes - 1 :param G: :return: copy of G with relabelled node indices, mapping """ mapping = {} for new_idx, old_idx in enumerate(G.nodes()): mapping[old_idx] = new_idx new_G = nx.relabel_nodes(G, mapping, copy=True) return new_G, mapping class MaskAtom: def __init__(self, num_atom_type, num_edge_type, mask_rate, mask_num=0, mask_edge=True): """ Randomly masks an atom, and optionally masks edges connecting to it. The mask atom type index is num_possible_atom_type The mask edge type index in num_possible_edge_type :param num_atom_type: :param num_edge_type: :param mask_rate: % of atoms to be masked :param mask_num: number of atoms to be masked :param mask_edge: If True, also mask the edges that connect to the masked atoms """ self.num_atom_type = num_atom_type self.num_edge_type = num_edge_type self.mask_rate = mask_rate self.mask_num = mask_num self.mask_edge = mask_edge def __call__(self, data, masked_atom_indices=None): """ :param data: pytorch geometric data object. Assume that the edge ordering is the default pytorch geometric ordering, where the two directions of a single edge occur in pairs. Eg. data.edge_index = tensor([[0, 1, 1, 2, 2, 3], [1, 0, 2, 1, 3, 2]]) :param masked_atom_indices: If None, then randomly samples num_atoms * mask rate number of atom indices Otherwise a list of atom idx that sets the atoms to be masked (for debugging only) :return: None, Creates new attributes in original data object: data.mask_node_idx data.mask_node_label data.mask_edge_idx data.mask_edge_label """ if masked_atom_indices == None: # sample x distinct atoms to be masked, based on mask rate. But # will sample at least 1 atom num_atoms = data.x.size()[0] if self.mask_num == 0: sample_size = int(num_atoms * self.mask_rate + 1) else: sample_size = self.mask_num masked_atom_indices = random.sample(range(num_atoms), sample_size) # create mask node label by copying atom feature of mask atom mask_node_labels_list = [] for atom_idx in masked_atom_indices: mask_node_labels_list.append(data.x[atom_idx].view(1, -1)) data.mask_node_label = torch.cat(mask_node_labels_list, dim=0) data.masked_atom_indices = torch.tensor(masked_atom_indices) # modify the original node feature of the masked node for atom_idx in masked_atom_indices: data.x[atom_idx] = torch.tensor([self.num_atom_type, 0]) if self.mask_edge: # create mask edge labels by copying edge features of edges that are bonded to # mask atoms connected_edge_indices = [] for bond_idx, (u, v) in enumerate(data.edge_index.cpu().numpy().T): for atom_idx in masked_atom_indices: if atom_idx in set((u, v)) and \ bond_idx not in connected_edge_indices: connected_edge_indices.append(bond_idx) if len(connected_edge_indices) > 0: # create mask edge labels by copying bond features of the bonds connected to # the mask atoms mask_edge_labels_list = [] for bond_idx in connected_edge_indices[::2]: # because the # edge ordering is such that two directions of a single # edge occur in pairs, so to get the unique undirected # edge indices, we take every 2nd edge index from list mask_edge_labels_list.append( data.edge_attr[bond_idx].view(1, -1)) data.mask_edge_label = torch.cat(mask_edge_labels_list, dim=0) # modify the original bond features of the bonds connected to the mask atoms for bond_idx in connected_edge_indices: data.edge_attr[bond_idx] = torch.tensor( [self.num_edge_type, 0]) data.connected_edge_indices = torch.tensor( connected_edge_indices[::2]) else: data.mask_edge_label = torch.empty((0, 2)).to(torch.int64) data.connected_edge_indices = torch.tensor( connected_edge_indices).to(torch.int64) return data def __repr__(self): if self.mask_num == 0: return '{}(num_atom_type={}, num_edge_type={}, mask_rate={}, mask_edge={})'.format( self.__class__.__name__, self.num_atom_type, self.num_edge_type, self.mask_rate, self.mask_edge) else: return '{}(num_atom_type={}, num_edge_type={}, mask_num={}, mask_edge={})'.format( self.__class__.__name__, self.num_atom_type, self.num_edge_type, self.mask_num, self.mask_edge) if __name__ == "__main__": transform = NegativeEdge() dataset = MoleculeDataset("dataset/tox21", dataset="tox21") transform(dataset[0])
11,865
41.378571
100
py
GraphLoG
GraphLoG-main/loader.py
import os import torch import pickle import collections import math import pandas as pd import numpy as np import networkx as nx from rdkit import Chem from rdkit.Chem import Descriptors from rdkit.Chem import AllChem from rdkit import DataStructs from rdkit.Chem.rdMolDescriptors import GetMorganFingerprintAsBitVect from torch.utils import data from torch_geometric.data import Data from torch_geometric.data import InMemoryDataset from torch_geometric.data import Batch from itertools import repeat, product, chain # allowable node and edge features allowable_features = { 'possible_atomic_num_list' : list(range(1, 119)), 'possible_formal_charge_list' : [-5, -4, -3, -2, -1, 0, 1, 2, 3, 4, 5], 'possible_chirality_list' : [ Chem.rdchem.ChiralType.CHI_UNSPECIFIED, Chem.rdchem.ChiralType.CHI_TETRAHEDRAL_CW, Chem.rdchem.ChiralType.CHI_TETRAHEDRAL_CCW, Chem.rdchem.ChiralType.CHI_OTHER ], 'possible_hybridization_list' : [ Chem.rdchem.HybridizationType.S, Chem.rdchem.HybridizationType.SP, Chem.rdchem.HybridizationType.SP2, Chem.rdchem.HybridizationType.SP3, Chem.rdchem.HybridizationType.SP3D, Chem.rdchem.HybridizationType.SP3D2, Chem.rdchem.HybridizationType.UNSPECIFIED ], 'possible_numH_list' : [0, 1, 2, 3, 4, 5, 6, 7, 8], 'possible_implicit_valence_list' : [0, 1, 2, 3, 4, 5, 6], 'possible_degree_list' : [0, 1, 2, 3, 4, 5, 6, 7, 8, 9, 10], 'possible_bonds' : [ Chem.rdchem.BondType.SINGLE, Chem.rdchem.BondType.DOUBLE, Chem.rdchem.BondType.TRIPLE, Chem.rdchem.BondType.AROMATIC ], 'possible_bond_dirs' : [ # only for double bond stereo information Chem.rdchem.BondDir.NONE, Chem.rdchem.BondDir.ENDUPRIGHT, Chem.rdchem.BondDir.ENDDOWNRIGHT ] } def mol_to_graph_data_obj_simple(mol): """ Converts rdkit mol object to graph Data object required by the pytorch geometric package. NB: Uses simplified atom and bond features, and represent as indices :param mol: rdkit mol object :return: graph data object with the attributes: x, edge_index, edge_attr """ # atoms num_atom_features = 2 # atom type, chirality tag atom_features_list = [] for atom in mol.GetAtoms(): atom_feature = [allowable_features['possible_atomic_num_list'].index( atom.GetAtomicNum())] + [allowable_features[ 'possible_chirality_list'].index(atom.GetChiralTag())] atom_features_list.append(atom_feature) x = torch.tensor(np.array(atom_features_list), dtype=torch.long) # bonds num_bond_features = 2 # bond type, bond direction if len(mol.GetBonds()) > 0: # mol has bonds edges_list = [] edge_features_list = [] for bond in mol.GetBonds(): i = bond.GetBeginAtomIdx() j = bond.GetEndAtomIdx() edge_feature = [allowable_features['possible_bonds'].index( bond.GetBondType())] + [allowable_features[ 'possible_bond_dirs'].index( bond.GetBondDir())] edges_list.append((i, j)) edge_features_list.append(edge_feature) edges_list.append((j, i)) edge_features_list.append(edge_feature) # data.edge_index: Graph connectivity in COO format with shape [2, num_edges] edge_index = torch.tensor(np.array(edges_list).T, dtype=torch.long) # data.edge_attr: Edge feature matrix with shape [num_edges, num_edge_features] edge_attr = torch.tensor(np.array(edge_features_list), dtype=torch.long) else: # mol has no bonds edge_index = torch.empty((2, 0), dtype=torch.long) edge_attr = torch.empty((0, num_bond_features), dtype=torch.long) data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr) return data def graph_data_obj_to_mol_simple(data_x, data_edge_index, data_edge_attr): """ Convert pytorch geometric data obj to rdkit mol object. NB: Uses simplified atom and bond features, and represent as indices. :param: data_x: :param: data_edge_index: :param: data_edge_attr :return: """ mol = Chem.RWMol() # atoms atom_features = data_x.cpu().numpy() num_atoms = atom_features.shape[0] for i in range(num_atoms): atomic_num_idx, chirality_tag_idx = atom_features[i] atomic_num = allowable_features['possible_atomic_num_list'][atomic_num_idx] chirality_tag = allowable_features['possible_chirality_list'][chirality_tag_idx] atom = Chem.Atom(atomic_num) atom.SetChiralTag(chirality_tag) mol.AddAtom(atom) # bonds edge_index = data_edge_index.cpu().numpy() edge_attr = data_edge_attr.cpu().numpy() num_bonds = edge_index.shape[1] for j in range(0, num_bonds, 2): begin_idx = int(edge_index[0, j]) end_idx = int(edge_index[1, j]) bond_type_idx, bond_dir_idx = edge_attr[j] bond_type = allowable_features['possible_bonds'][bond_type_idx] bond_dir = allowable_features['possible_bond_dirs'][bond_dir_idx] mol.AddBond(begin_idx, end_idx, bond_type) # set bond direction new_bond = mol.GetBondBetweenAtoms(begin_idx, end_idx) new_bond.SetBondDir(bond_dir) # Chem.SanitizeMol(mol) # fails for COC1=CC2=C(NC(=N2)[S@@](=O)CC2=NC=C( # C)C(OC)=C2C)C=C1, when aromatic bond is possible # when we do not have aromatic bonds # Chem.SanitizeMol(mol, sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) return mol def graph_data_obj_to_nx_simple(data): """ Converts graph Data object required by the pytorch geometric package to network x data object. NB: Uses simplified atom and bond features, and represent as indices. NB: possible issues with recapitulating relative stereochemistry since the edges in the nx object are unordered. :param data: pytorch geometric Data object :return: network x object """ G = nx.Graph() # atoms atom_features = data.x.cpu().numpy() num_atoms = atom_features.shape[0] for i in range(num_atoms): atomic_num_idx, chirality_tag_idx = atom_features[i] G.add_node(i, atom_num_idx=atomic_num_idx, chirality_tag_idx=chirality_tag_idx) pass # bonds edge_index = data.edge_index.cpu().numpy() edge_attr = data.edge_attr.cpu().numpy() num_bonds = edge_index.shape[1] for j in range(0, num_bonds, 2): begin_idx = int(edge_index[0, j]) end_idx = int(edge_index[1, j]) bond_type_idx, bond_dir_idx = edge_attr[j] if not G.has_edge(begin_idx, end_idx): G.add_edge(begin_idx, end_idx, bond_type_idx=bond_type_idx, bond_dir_idx=bond_dir_idx) return G def nx_to_graph_data_obj_simple(G): """ Converts nx graph to pytorch geometric Data object. Assume node indices are numbered from 0 to num_nodes - 1. NB: Uses simplified atom and bond features, and represent as indices. NB: possible issues with recapitulating relative stereochemistry since the edges in the nx object are unordered. :param G: nx graph obj :return: pytorch geometric Data object """ # atoms num_atom_features = 2 # atom type, chirality tag atom_features_list = [] for _, node in G.nodes(data=True): atom_feature = [node['atom_num_idx'], node['chirality_tag_idx']] atom_features_list.append(atom_feature) x = torch.tensor(np.array(atom_features_list), dtype=torch.long) # bonds num_bond_features = 2 # bond type, bond direction if len(G.edges()) > 0: # mol has bonds edges_list = [] edge_features_list = [] for i, j, edge in G.edges(data=True): edge_feature = [edge['bond_type_idx'], edge['bond_dir_idx']] edges_list.append((i, j)) edge_features_list.append(edge_feature) edges_list.append((j, i)) edge_features_list.append(edge_feature) # data.edge_index: Graph connectivity in COO format with shape [2, num_edges] edge_index = torch.tensor(np.array(edges_list).T, dtype=torch.long) # data.edge_attr: Edge feature matrix with shape [num_edges, num_edge_features] edge_attr = torch.tensor(np.array(edge_features_list), dtype=torch.long) else: # mol has no bonds edge_index = torch.empty((2, 0), dtype=torch.long) edge_attr = torch.empty((0, num_bond_features), dtype=torch.long) data = Data(x=x, edge_index=edge_index, edge_attr=edge_attr) return data def get_gasteiger_partial_charges(mol, n_iter=12): """ Calculates list of gasteiger partial charges for each atom in mol object. :param mol: rdkit mol object :param n_iter: number of iterations. Default 12 :return: list of computed partial charges for each atom. """ Chem.rdPartialCharges.ComputeGasteigerCharges(mol, nIter=n_iter, throwOnParamFailure=True) partial_charges = [float(a.GetProp('_GasteigerCharge')) for a in mol.GetAtoms()] return partial_charges def create_standardized_mol_id(smiles): """ :param smiles: :return: inchi """ if check_smiles_validity(smiles): # remove stereochemistry smiles = AllChem.MolToSmiles(AllChem.MolFromSmiles(smiles), isomericSmiles=False) mol = AllChem.MolFromSmiles(smiles) if mol != None: # to catch weird issue with O=C1O[al]2oc(=O)c3ccc(cn3)c3ccccc3c3cccc(c3)c3ccccc3c3cc(C(F)(F)F)c(cc3o2)-c2ccccc2-c2cccc(c2)-c2ccccc2-c2cccnc21 if '.' in smiles: # if multiple species, pick largest molecule mol_species_list = split_rdkit_mol_obj(mol) largest_mol = get_largest_mol(mol_species_list) inchi = AllChem.MolToInchi(largest_mol) else: inchi = AllChem.MolToInchi(mol) return inchi else: return else: return class MoleculeDataset(InMemoryDataset): def __init__(self, root, #data = None, #slices = None, transform=None, pre_transform=None, pre_filter=None, dataset='zinc250k', empty=False): """ Adapted from qm9.py. Disabled the download functionality :param root: directory of the dataset, containing a raw and processed dir. The raw dir should contain the file containing the smiles, and the processed dir can either empty or a previously processed file :param dataset: name of the dataset. Currently only implemented for zinc250k, chembl_with_labels, tox21, hiv, bace, bbbp, clintox, esol, freesolv, lipophilicity, muv, pcba, sider, toxcast :param empty: if True, then will not load any data obj. For initializing empty dataset """ self.dataset = dataset self.root = root super(MoleculeDataset, self).__init__(root, transform, pre_transform, pre_filter) self.transform, self.pre_transform, self.pre_filter = transform, pre_transform, pre_filter if not empty: self.data, self.slices = torch.load(self.processed_paths[0]) def get(self, idx): data = Data() for key in self.data.keys: item, slices = self.data[key], self.slices[key] s = list(repeat(slice(None), item.dim())) s[data.cat_dim(key, item)] = slice(slices[idx], slices[idx + 1]) data[key] = item[s] return data @property def raw_file_names(self): file_name_list = os.listdir(self.raw_dir) # assert len(file_name_list) == 1 # currently assume we have a # # single raw file return file_name_list @property def processed_file_names(self): return 'geometric_data_processed.pt' def download(self): raise NotImplementedError('Must indicate valid location of raw data. ' 'No download allowed') def process(self): data_smiles_list = [] data_list = [] if self.dataset == 'zinc_standard_agent': input_path = self.raw_paths[0] input_df = pd.read_csv(input_path, sep=',', compression='gzip', dtype='str') smiles_list = list(input_df['smiles']) zinc_id_list = list(input_df['zinc_id']) for i in range(len(smiles_list)): print(i) s = smiles_list[i] # each example contains a single species try: rdkit_mol = AllChem.MolFromSmiles(s) if rdkit_mol != None: # ignore invalid mol objects # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id id = int(zinc_id_list[i].split('ZINC')[1].lstrip('0')) data.id = torch.tensor( [id]) # id here is zinc id value, stripped of # leading zeros data_list.append(data) data_smiles_list.append(smiles_list[i]) except: continue elif self.dataset == 'chembl_filtered': ### get downstream test molecules. from splitters import scaffold_split ### downstream_dir = [ 'dataset/bace', 'dataset/bbbp', 'dataset/clintox', 'dataset/esol', 'dataset/freesolv', 'dataset/hiv', 'dataset/lipophilicity', 'dataset/muv', # 'dataset/pcba/processed/smiles.csv', 'dataset/sider', 'dataset/tox21', 'dataset/toxcast' ] downstream_inchi_set = set() for d_path in downstream_dir: print(d_path) dataset_name = d_path.split('/')[1] downstream_dataset = MoleculeDataset(d_path, dataset=dataset_name) downstream_smiles = pd.read_csv(os.path.join(d_path, 'processed', 'smiles.csv'), header=None)[0].tolist() assert len(downstream_dataset) == len(downstream_smiles) _, _, _, (train_smiles, valid_smiles, test_smiles) = scaffold_split(downstream_dataset, downstream_smiles, task_idx=None, null_value=0, frac_train=0.8,frac_valid=0.1, frac_test=0.1, return_smiles=True) ### remove both test and validation molecules remove_smiles = test_smiles + valid_smiles downstream_inchis = [] for smiles in remove_smiles: species_list = smiles.split('.') for s in species_list: # record inchi for all species, not just # largest (by default in create_standardized_mol_id if input has # multiple species) inchi = create_standardized_mol_id(s) downstream_inchis.append(inchi) downstream_inchi_set.update(downstream_inchis) smiles_list, rdkit_mol_objs, folds, labels = \ _load_chembl_with_labels_dataset(os.path.join(self.root, 'raw')) print('processing') for i in range(len(rdkit_mol_objs)): print(i) rdkit_mol = rdkit_mol_objs[i] if rdkit_mol != None: # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) mw = Descriptors.MolWt(rdkit_mol) if 50 <= mw <= 900: inchi = create_standardized_mol_id(smiles_list[i]) if inchi != None and inchi not in downstream_inchi_set: data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) # fold information if i in folds[0]: data.fold = torch.tensor([0]) elif i in folds[1]: data.fold = torch.tensor([1]) else: data.fold = torch.tensor([2]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'tox21': smiles_list, rdkit_mol_objs, labels = \ _load_tox21_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] ## convert aromatic bonds to double bonds #Chem.SanitizeMol(rdkit_mol, #sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'hiv': smiles_list, rdkit_mol_objs, labels = \ _load_hiv_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'bace': smiles_list, rdkit_mol_objs, folds, labels = \ _load_bace_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor([labels[i]]) data.fold = torch.tensor([folds[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'bbbp': smiles_list, rdkit_mol_objs, labels = \ _load_bbbp_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] if rdkit_mol != None: # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'clintox': smiles_list, rdkit_mol_objs, labels = \ _load_clintox_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] if rdkit_mol != None: # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'esol': smiles_list, rdkit_mol_objs, labels = \ _load_esol_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'freesolv': smiles_list, rdkit_mol_objs, labels = \ _load_freesolv_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'lipophilicity': smiles_list, rdkit_mol_objs, labels = \ _load_lipophilicity_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'muv': smiles_list, rdkit_mol_objs, labels = \ _load_muv_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'pcba': smiles_list, rdkit_mol_objs, labels = \ _load_pcba_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'pcba_pretrain': smiles_list, rdkit_mol_objs, labels = \ _load_pcba_dataset(self.raw_paths[0]) downstream_inchi = set(pd.read_csv(os.path.join(self.root, 'downstream_mol_inchi_may_24_2019'), sep=',', header=None)[0]) for i in range(len(smiles_list)): print(i) if '.' not in smiles_list[i]: # remove examples with # multiples species rdkit_mol = rdkit_mol_objs[i] mw = Descriptors.MolWt(rdkit_mol) if 50 <= mw <= 900: inchi = create_standardized_mol_id(smiles_list[i]) if inchi != None and inchi not in downstream_inchi: # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) # elif self.dataset == '' elif self.dataset == 'sider': smiles_list, rdkit_mol_objs, labels = \ _load_sider_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'toxcast': smiles_list, rdkit_mol_objs, labels = \ _load_toxcast_dataset(self.raw_paths[0]) for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] if rdkit_mol != None: # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) # id here is the index of the mol in # the dataset data.y = torch.tensor(labels[i, :]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'ptc_mr': input_path = self.raw_paths[0] input_df = pd.read_csv(input_path, sep=',', header=None, names=['id', 'label', 'smiles']) smiles_list = input_df['smiles'] labels = input_df['label'].values for i in range(len(smiles_list)): print(i) s = smiles_list[i] rdkit_mol = AllChem.MolFromSmiles(s) if rdkit_mol != None: # ignore invalid mol objects # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'mutag': smiles_path = os.path.join(self.root, 'raw', 'mutag_188_data.can') # smiles_path = 'dataset/mutag/raw/mutag_188_data.can' labels_path = os.path.join(self.root, 'raw', 'mutag_188_target.txt') # labels_path = 'dataset/mutag/raw/mutag_188_target.txt' smiles_list = pd.read_csv(smiles_path, sep=' ', header=None)[0] labels = pd.read_csv(labels_path, header=None)[0].values for i in range(len(smiles_list)): print(i) s = smiles_list[i] rdkit_mol = AllChem.MolFromSmiles(s) if rdkit_mol != None: # ignore invalid mol objects # # convert aromatic bonds to double bonds # Chem.SanitizeMol(rdkit_mol, # sanitizeOps=Chem.SanitizeFlags.SANITIZE_KEKULIZE) data = mol_to_graph_data_obj_simple(rdkit_mol) # manually add mol id data.id = torch.tensor( [i]) data.y = torch.tensor([labels[i]]) data_list.append(data) data_smiles_list.append(smiles_list[i]) else: raise ValueError('Invalid dataset name') if self.pre_filter is not None: data_list = [data for data in data_list if self.pre_filter(data)] if self.pre_transform is not None: data_list = [self.pre_transform(data) for data in data_list] # write data_smiles_list in processed paths data_smiles_series = pd.Series(data_smiles_list) data_smiles_series.to_csv(os.path.join(self.processed_dir, 'smiles.csv'), index=False, header=False) data, slices = self.collate(data_list) torch.save((data, slices), self.processed_paths[0]) # NB: only properly tested when dataset_1 is chembl_with_labels and dataset_2 # is pcba_pretrain def merge_dataset_objs(dataset_1, dataset_2): """ Naively merge 2 molecule dataset objects, and ignore identities of molecules. Assumes both datasets have multiple y labels, and will pad accordingly. ie if dataset_1 has obj_1 with y dim 1310 and dataset_2 has obj_2 with y dim 128, then the resulting obj_1 and obj_2 will have dim 1438, where obj_1 have the last 128 cols with 0, and obj_2 have the first 1310 cols with 0. :return: pytorch geometric dataset obj, with the x, edge_attr, edge_index, new y attributes only """ d_1_y_dim = dataset_1[0].y.size()[0] d_2_y_dim = dataset_2[0].y.size()[0] data_list = [] # keep only x, edge_attr, edge_index, padded_y then append for d in dataset_1: old_y = d.y new_y = torch.cat([old_y, torch.zeros(d_2_y_dim, dtype=torch.long)]) data_list.append(Data(x=d.x, edge_index=d.edge_index, edge_attr=d.edge_attr, y=new_y)) for d in dataset_2: old_y = d.y new_y = torch.cat([torch.zeros(d_1_y_dim, dtype=torch.long), old_y.long()]) data_list.append(Data(x=d.x, edge_index=d.edge_index, edge_attr=d.edge_attr, y=new_y)) # create 'empty' dataset obj. Just randomly pick a dataset and root path # that has already been processed new_dataset = MoleculeDataset(root='dataset/chembl_with_labels', dataset='chembl_with_labels', empty=True) # collate manually new_dataset.data, new_dataset.slices = new_dataset.collate(data_list) return new_dataset def create_circular_fingerprint(mol, radius, size, chirality): """ :param mol: :param radius: :param size: :param chirality: :return: np array of morgan fingerprint """ fp = GetMorganFingerprintAsBitVect(mol, radius, nBits=size, useChirality=chirality) return np.array(fp) class MoleculeFingerprintDataset(data.Dataset): def __init__(self, root, dataset, radius, size, chirality=True): """ Create dataset object containing list of dicts, where each dict contains the circular fingerprint of the molecule, label, id, and possibly precomputed fold information :param root: directory of the dataset, containing a raw and processed_fp dir. The raw dir should contain the file containing the smiles, and the processed_fp dir can either be empty or a previously processed file :param dataset: name of dataset. Currently only implemented for tox21, hiv, chembl_with_labels :param radius: radius of the circular fingerprints :param size: size of the folded fingerprint vector :param chirality: if True, fingerprint includes chirality information """ self.dataset = dataset self.root = root self.radius = radius self.size = size self.chirality = chirality self._load() def _process(self): data_smiles_list = [] data_list = [] if self.dataset == 'chembl_with_labels': smiles_list, rdkit_mol_objs, folds, labels = \ _load_chembl_with_labels_dataset(os.path.join(self.root, 'raw')) print('processing') for i in range(len(rdkit_mol_objs)): print(i) rdkit_mol = rdkit_mol_objs[i] if rdkit_mol != None: # # convert aromatic bonds to double bonds fp_arr = create_circular_fingerprint(rdkit_mol, self.radius, self.size, self.chirality) fp_arr = torch.tensor(fp_arr) # manually add mol id id = torch.tensor([i]) # id here is the index of the mol in # the dataset y = torch.tensor(labels[i, :]) # fold information if i in folds[0]: fold = torch.tensor([0]) elif i in folds[1]: fold = torch.tensor([1]) else: fold = torch.tensor([2]) data_list.append({'fp_arr': fp_arr, 'id': id, 'y': y, 'fold': fold}) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'tox21': smiles_list, rdkit_mol_objs, labels = \ _load_tox21_dataset(os.path.join(self.root, 'raw/tox21.csv')) print('processing') for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] ## convert aromatic bonds to double bonds fp_arr = create_circular_fingerprint(rdkit_mol, self.radius, self.size, self.chirality) fp_arr = torch.tensor(fp_arr) # manually add mol id id = torch.tensor([i]) # id here is the index of the mol in # the dataset y = torch.tensor(labels[i, :]) data_list.append({'fp_arr': fp_arr, 'id': id, 'y': y}) data_smiles_list.append(smiles_list[i]) elif self.dataset == 'hiv': smiles_list, rdkit_mol_objs, labels = \ _load_hiv_dataset(os.path.join(self.root, 'raw/HIV.csv')) print('processing') for i in range(len(smiles_list)): print(i) rdkit_mol = rdkit_mol_objs[i] # # convert aromatic bonds to double bonds fp_arr = create_circular_fingerprint(rdkit_mol, self.radius, self.size, self.chirality) fp_arr = torch.tensor(fp_arr) # manually add mol id id = torch.tensor([i]) # id here is the index of the mol in # the dataset y = torch.tensor([labels[i]]) data_list.append({'fp_arr': fp_arr, 'id': id, 'y': y}) data_smiles_list.append(smiles_list[i]) else: raise ValueError('Invalid dataset name') # save processed data objects and smiles processed_dir = os.path.join(self.root, 'processed_fp') data_smiles_series = pd.Series(data_smiles_list) data_smiles_series.to_csv(os.path.join(processed_dir, 'smiles.csv'), index=False, header=False) with open(os.path.join(processed_dir, 'fingerprint_data_processed.pkl'), 'wb') as f: pickle.dump(data_list, f) def _load(self): processed_dir = os.path.join(self.root, 'processed_fp') # check if saved file exist. If so, then load from save file_name_list = os.listdir(processed_dir) if 'fingerprint_data_processed.pkl' in file_name_list: with open(os.path.join(processed_dir, 'fingerprint_data_processed.pkl'), 'rb') as f: self.data_list = pickle.load(f) # if no saved file exist, then perform processing steps, save then # reload else: self._process() self._load() def __len__(self): return len(self.data_list) def __getitem__(self, index): ## if iterable class is passed, return dataset objection if hasattr(index, "__iter__"): dataset = MoleculeFingerprintDataset(self.root, self.dataset, self.radius, self.size, chirality=self.chirality) dataset.data_list = [self.data_list[i] for i in index] return dataset else: return self.data_list[index] def _load_tox21_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] tasks = ['NR-AR', 'NR-AR-LBD', 'NR-AhR', 'NR-Aromatase', 'NR-ER', 'NR-ER-LBD', 'NR-PPAR-gamma', 'SR-ARE', 'SR-ATAD5', 'SR-HSE', 'SR-MMP', 'SR-p53'] labels = input_df[tasks] # convert 0 to -1 labels = labels.replace(0, -1) # convert nan to 0 labels = labels.fillna(0) assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.values def _load_hiv_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] labels = input_df['HIV_active'] # convert 0 to -1 labels = labels.replace(0, -1) # there are no nans assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.values def _load_bace_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing indices for each of the 3 folds, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['mol'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] labels = input_df['Class'] # convert 0 to -1 labels = labels.replace(0, -1) # there are no nans folds = input_df['Model'] folds = folds.replace('Train', 0) # 0 -> train folds = folds.replace('Valid', 1) # 1 -> valid folds = folds.replace('Test', 2) # 2 -> test assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) assert len(smiles_list) == len(folds) return smiles_list, rdkit_mol_objs_list, folds.values, labels.values def _load_bbbp_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] preprocessed_rdkit_mol_objs_list = [m if m != None else None for m in rdkit_mol_objs_list] preprocessed_smiles_list = [AllChem.MolToSmiles(m) if m != None else None for m in preprocessed_rdkit_mol_objs_list] labels = input_df['p_np'] # convert 0 to -1 labels = labels.replace(0, -1) # there are no nans assert len(smiles_list) == len(preprocessed_rdkit_mol_objs_list) assert len(smiles_list) == len(preprocessed_smiles_list) assert len(smiles_list) == len(labels) return preprocessed_smiles_list, preprocessed_rdkit_mol_objs_list, \ labels.values def _load_clintox_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] preprocessed_rdkit_mol_objs_list = [m if m != None else None for m in rdkit_mol_objs_list] preprocessed_smiles_list = [AllChem.MolToSmiles(m) if m != None else None for m in preprocessed_rdkit_mol_objs_list] tasks = ['FDA_APPROVED', 'CT_TOX'] labels = input_df[tasks] # convert 0 to -1 labels = labels.replace(0, -1) # there are no nans assert len(smiles_list) == len(preprocessed_rdkit_mol_objs_list) assert len(smiles_list) == len(preprocessed_smiles_list) assert len(smiles_list) == len(labels) return preprocessed_smiles_list, preprocessed_rdkit_mol_objs_list, \ labels.values # input_path = 'dataset/clintox/raw/clintox.csv' # smiles_list, rdkit_mol_objs_list, labels = _load_clintox_dataset(input_path) def _load_esol_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels (regression task) """ # NB: some examples have multiple species input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] labels = input_df['measured log solubility in mols per litre'] assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.values # input_path = 'dataset/esol/raw/delaney-processed.csv' # smiles_list, rdkit_mol_objs_list, labels = _load_esol_dataset(input_path) def _load_freesolv_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels (regression task) """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] labels = input_df['expt'] assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.values def _load_lipophilicity_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels (regression task) """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] labels = input_df['exp'] assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.values def _load_muv_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] tasks = ['MUV-466', 'MUV-548', 'MUV-600', 'MUV-644', 'MUV-652', 'MUV-689', 'MUV-692', 'MUV-712', 'MUV-713', 'MUV-733', 'MUV-737', 'MUV-810', 'MUV-832', 'MUV-846', 'MUV-852', 'MUV-858', 'MUV-859'] labels = input_df[tasks] # convert 0 to -1 labels = labels.replace(0, -1) # convert nan to 0 labels = labels.fillna(0) assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.values def _load_sider_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] tasks = ['Hepatobiliary disorders', 'Metabolism and nutrition disorders', 'Product issues', 'Eye disorders', 'Investigations', 'Musculoskeletal and connective tissue disorders', 'Gastrointestinal disorders', 'Social circumstances', 'Immune system disorders', 'Reproductive system and breast disorders', 'Neoplasms benign, malignant and unspecified (incl cysts and polyps)', 'General disorders and administration site conditions', 'Endocrine disorders', 'Surgical and medical procedures', 'Vascular disorders', 'Blood and lymphatic system disorders', 'Skin and subcutaneous tissue disorders', 'Congenital, familial and genetic disorders', 'Infections and infestations', 'Respiratory, thoracic and mediastinal disorders', 'Psychiatric disorders', 'Renal and urinary disorders', 'Pregnancy, puerperium and perinatal conditions', 'Ear and labyrinth disorders', 'Cardiac disorders', 'Nervous system disorders', 'Injury, poisoning and procedural complications'] labels = input_df[tasks] # convert 0 to -1 labels = labels.replace(0, -1) assert len(smiles_list) == len(rdkit_mol_objs_list) assert len(smiles_list) == len(labels) return smiles_list, rdkit_mol_objs_list, labels.value def _load_toxcast_dataset(input_path): """ :param input_path: :return: list of smiles, list of rdkit mol obj, np.array containing the labels """ # NB: some examples have multiple species, some example smiles are invalid input_df = pd.read_csv(input_path, sep=',') smiles_list = input_df['smiles'] rdkit_mol_objs_list = [AllChem.MolFromSmiles(s) for s in smiles_list] # Some smiles could not be successfully converted # to rdkit mol object so them to None preprocessed_rdkit_mol_objs_list = [m if m != None else None for m in rdkit_mol_objs_list] preprocessed_smiles_list = [AllChem.MolToSmiles(m) if m != None else None for m in preprocessed_rdkit_mol_objs_list] tasks = list(input_df.columns)[1:] labels = input_df[tasks] # convert 0 to -1 labels = labels.replace(0, -1) # convert nan to 0 labels = labels.fillna(0) assert len(smiles_list) == len(preprocessed_rdkit_mol_objs_list) assert len(smiles_list) == len(preprocessed_smiles_list) assert len(smiles_list) == len(labels) return preprocessed_smiles_list, preprocessed_rdkit_mol_objs_list, \ labels.values def _load_chembl_with_labels_dataset(root_path): """ Data from 'Large-scale comparison of machine learning methods for drug target prediction on ChEMBL' :param root_path: path to the folder containing the reduced chembl dataset :return: list of smiles, preprocessed rdkit mol obj list, list of np.array containing indices for each of the 3 folds, np.array containing the labels """ # adapted from https://github.com/ml-jku/lsc/blob/master/pythonCode/lstm/loadData.py # first need to download the files and unzip: # wget http://bioinf.jku.at/research/lsc/chembl20/dataPythonReduced.zip # unzip and rename to chembl_with_labels # wget http://bioinf.jku.at/research/lsc/chembl20/dataPythonReduced/chembl20Smiles.pckl # into the dataPythonReduced directory # wget http://bioinf.jku.at/research/lsc/chembl20/dataPythonReduced/chembl20LSTM.pckl # 1. load folds and labels f=open(os.path.join(root_path, 'folds0.pckl'), 'rb') folds=pickle.load(f) f.close() f=open(os.path.join(root_path, 'labelsHard.pckl'), 'rb') targetMat=pickle.load(f) sampleAnnInd=pickle.load(f) targetAnnInd=pickle.load(f) f.close() targetMat=targetMat targetMat=targetMat.copy().tocsr() targetMat.sort_indices() targetAnnInd=targetAnnInd targetAnnInd=targetAnnInd-targetAnnInd.min() folds=[np.intersect1d(fold, sampleAnnInd.index.values).tolist() for fold in folds] targetMatTransposed=targetMat[sampleAnnInd[list(chain(*folds))]].T.tocsr() targetMatTransposed.sort_indices() # # num positive examples in each of the 1310 targets trainPosOverall=np.array([np.sum(targetMatTransposed[x].data > 0.5) for x in range(targetMatTransposed.shape[0])]) # # num negative examples in each of the 1310 targets trainNegOverall=np.array([np.sum(targetMatTransposed[x].data < -0.5) for x in range(targetMatTransposed.shape[0])]) # dense array containing the labels for the 456331 molecules and 1310 targets denseOutputData=targetMat.A # possible values are {-1, 0, 1} # 2. load structures f=open(os.path.join(root_path, 'chembl20LSTM.pckl'), 'rb') rdkitArr=pickle.load(f) f.close() assert len(rdkitArr) == denseOutputData.shape[0] assert len(rdkitArr) == len(folds[0]) + len(folds[1]) + len(folds[2]) preprocessed_rdkitArr = [] print('preprocessing') for i in range(len(rdkitArr)): print(i) m = rdkitArr[i] if m == None: preprocessed_rdkitArr.append(None) else: mol_species_list = split_rdkit_mol_obj(m) if len(mol_species_list) == 0: preprocessed_rdkitArr.append(None) else: largest_mol = get_largest_mol(mol_species_list) if len(largest_mol.GetAtoms()) <= 2: preprocessed_rdkitArr.append(None) else: preprocessed_rdkitArr.append(largest_mol) assert len(preprocessed_rdkitArr) == denseOutputData.shape[0] smiles_list = [AllChem.MolToSmiles(m) if m != None else None for m in preprocessed_rdkitArr] # bc some empty mol in the # rdkitArr zzz... assert len(preprocessed_rdkitArr) == len(smiles_list) return smiles_list, preprocessed_rdkitArr, folds, denseOutputData # root_path = 'dataset/chembl_with_labels' def check_smiles_validity(smiles): try: m = Chem.MolFromSmiles(smiles) if m: return True else: return False except: return False def split_rdkit_mol_obj(mol): """ Split rdkit mol object containing multiple species or one species into a list of mol objects or a list containing a single object respectively :param mol: :return: """ smiles = AllChem.MolToSmiles(mol, isomericSmiles=True) smiles_list = smiles.split('.') mol_species_list = [] for s in smiles_list: if check_smiles_validity(s): mol_species_list.append(AllChem.MolFromSmiles(s)) return mol_species_list def get_largest_mol(mol_list): """ Given a list of rdkit mol objects, returns mol object containing the largest num of atoms. If multiple containing largest num of atoms, picks the first one :param mol_list: :return: """ num_atoms_list = [len(m.GetAtoms()) for m in mol_list] largest_mol_idx = num_atoms_list.index(max(num_atoms_list)) return mol_list[largest_mol_idx] def create_all_datasets(): #### create dataset downstream_dir = [ 'bace', 'bbbp', 'clintox', 'esol', 'freesolv', 'hiv', 'lipophilicity', 'muv', 'sider', 'tox21', 'toxcast' ] for dataset_name in downstream_dir: print(dataset_name) root = "dataset/" + dataset_name os.makedirs(root + "/processed", exist_ok=True) dataset = MoleculeDataset(root, dataset=dataset_name) print(dataset) dataset = MoleculeDataset(root = "dataset/chembl_filtered", dataset="chembl_filtered") print(dataset) dataset = MoleculeDataset(root = "dataset/zinc_standard_agent", dataset="zinc_standard_agent") print(dataset) # test MoleculeDataset object if __name__ == "__main__": create_all_datasets()
56,150
41.250564
165
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/run_test_ensemble.py
"""Script to launch ensemble on test set results.""" import argparse import json from collections import defaultdict from functools import partial from pathlib import Path import jax import jax.numpy as jnp import numpy as np import pandas as pd import SimpleITK as sitk # noqa: N813 from absl import logging from omegaconf import OmegaConf from tqdm import tqdm from imgx.datasets import ( DIR_TFDS_PROCESSED_MAP, IMAGE_SPACING_MAP, NUM_CLASSES_MAP, ) from imgx.exp.eval import ( get_jit_segmentation_metrics, get_non_jit_segmentation_metrics_per_step, ) logging.set_verbosity(logging.INFO) def parse_args() -> argparse.Namespace: """Parse arguments.""" parser = argparse.ArgumentParser(description=__doc__) parser.add_argument( "--log_dir", type=Path, help="Folder of wandb.", default=None, ) args = parser.parse_args() return args def vote_ensemble(test_dir: Path, dir_tfds: Path, num_classes: int) -> None: """Ensemble prediction via voting. Args: test_dir: path having predictions. dir_tfds: path of tfds data, having ground truth. num_classes: number of classes in labels. """ # get seed dirs and sort by seeds lst_seed_dir = sorted( test_dir.glob("seed_*/"), key=lambda x: int(x.stem.split("_")[-1]) ) num_seeds = len(lst_seed_dir) # map relative_path to list of full path, corresponding to seeds path_dict = defaultdict(list) for seed_dir in lst_seed_dir: for x in seed_dir.glob("**/*.nii.gz"): rel_path = x.relative_to(seed_dir) path_dict[rel_path].append(x) # vote to ensemble logging.info("Calculating ensemble predictions.") for rel_path, pred_paths in path_dict.items(): # a list of shape (D, W, H) mask_preds = [ sitk.GetArrayFromImage(sitk.ReadImage(x)) for x in pred_paths ] # (D, W, H, num_classes, num_seeds) mask_onehot = jax.nn.one_hot( jnp.stack(mask_preds, axis=-1), num_classes=num_classes, axis=-2 ) # vote (D, W, H) mask_pred = jnp.argmax(jnp.sum(mask_onehot, axis=-1), axis=-1).astype( "uint16" ) # copy meta data uid = pred_paths[0].stem.split("_")[0] volume_mask_true = sitk.ReadImage( dir_tfds / f"{uid}_mask_preprocessed.nii.gz" ) volume_mask_pred = sitk.GetImageFromArray(mask_pred) volume_mask_pred.CopyInformation(volume_mask_true) # save out_path = test_dir / f"ensemble_{num_seeds}" / rel_path out_path.parent.mkdir(parents=True, exist_ok=True) sitk.WriteImage( image=volume_mask_pred, fileName=out_path, useCompression=True, ) def evaluate_ensemble_prediction( dir_path: Path, dir_tfds: Path, num_classes: int, spacing: jnp.ndarray ) -> None: """Evaluate the saved predictions from ensemble. Args: dir_path: path having predictions. dir_tfds: path of tfds data, having ground truth. num_classes: number of classes in labels. spacing: spacing for voxels. """ num_steps = int(dir_path.name.split("_")[1]) uids = [ x.name.split("_")[0] for x in (dir_path / "step_0").glob("*.nii.gz") ] lst_df_scalar = [] for uid in tqdm(uids, total=len(uids)): # (D, W, H) mask_true = sitk.GetArrayFromImage( sitk.ReadImage(dir_tfds / f"{uid}_mask_preprocessed.nii.gz") ) # (D, W, H, num_classes) mask_true = jax.nn.one_hot(mask_true, num_classes=num_classes, axis=-1) # (1, W, H, D, num_classes) mask_true = jnp.transpose(mask_true, axes=(2, 1, 0, 3))[None, ...] pred_paths = [ dir_path / f"step_{i}" / f"{uid}_mask_pred.nii.gz" for i in range(num_steps) ] # a list of shape (D, W, H) mask_preds = [ sitk.GetArrayFromImage(sitk.ReadImage(x)) for x in pred_paths ] # (D, W, H, num_classes, num_steps) mask_pred = jax.nn.one_hot( jnp.stack(mask_preds, axis=-1), num_classes=num_classes, axis=-2 ) # (1, W, H, D, num_classes, num_steps) mask_pred = jnp.transpose(mask_pred, axes=(2, 1, 0, 3, 4))[None, ...] # metrics scalars_jit = jax.vmap( partial( get_jit_segmentation_metrics, mask_true=mask_true, spacing=spacing, ), in_axes=-1, out_axes=-1, )(mask_pred) scalars_nonjit = get_non_jit_segmentation_metrics_per_step( mask_pred=mask_pred, mask_true=mask_true, spacing=spacing, ) scalars = {**scalars_jit, **scalars_nonjit} # flatten per step scalars_flatten = {} for k, v in scalars.items(): for i in range(v.shape[-1]): scalars_flatten[f"{k}_step_{i}"] = v[..., i] scalars_flatten[k] = v[..., -1] scalars = scalars_flatten scalars = jax.tree_map(lambda x: np.asarray(x).tolist(), scalars) scalars["uid"] = [uid] lst_df_scalar.append(pd.DataFrame(scalars)) # assemble metrics df_scalar = pd.concat(lst_df_scalar) df_scalar = df_scalar.sort_values("uid") df_scalar.to_csv(dir_path / "metrics_per_sample.csv", index=False) # average over samples in the dataset scalars = df_scalar.drop("uid", axis=1).mean().to_dict() scalars = {"test_" + k: v for k, v in scalars.items()} scalars["num_images_in_total"] = len(df_scalar) with open(dir_path / "mean_metrics.json", "w", encoding="utf-8") as f: json.dump(scalars, f, sort_keys=True, indent=4) def main() -> None: # pylint:disable=R0915 """Main function.""" args = parse_args() config = OmegaConf.load(args.log_dir / "files" / "config_backup.yaml") if config.task.name != "diffusion": raise ValueError("Ensemble is only for diffusion.") data_config = config.data dir_tfds = DIR_TFDS_PROCESSED_MAP[data_config.name] spacing = jnp.array(IMAGE_SPACING_MAP[data_config.name]) num_classes = NUM_CLASSES_MAP[data_config["name"]] test_dir = args.log_dir / "files" / "test_evaluation" # no ensemble if 1 seed only lst_seed_dir = sorted( test_dir.glob("seed_*/"), key=lambda x: int(x.stem.split("_")[-1]) ) if len(lst_seed_dir) == 1: logging.info("Ensemble not performed as there is one seed only.") return # ensemble vote_ensemble(test_dir=test_dir, dir_tfds=dir_tfds, num_classes=num_classes) # evaluate for dir_path in test_dir.glob("ensemble_*/sample_*_steps"): logging.info(f"Evaluating ensemble predictions metrics for {dir_path}.") evaluate_ensemble_prediction( dir_path=dir_path, dir_tfds=dir_tfds, num_classes=num_classes, spacing=spacing, ) if __name__ == "__main__": main()
7,072
31.296804
80
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/run_test.py
"""Script to launch evaluation on test sets.""" import argparse import json from pathlib import Path import jax import numpy as np from absl import logging from omegaconf import OmegaConf from imgx import TEST_SPLIT from imgx.device import broadcast_to_local_devices from imgx.exp import Experiment from imgx.exp.train_state import get_eval_params_and_state_from_ckpt logging.set_verbosity(logging.INFO) def get_checkpoint_dir( log_dir: Path, num_batch: int, metric: str, max_metric: bool ) -> Path: """Get the checkpoint directory. Args: log_dir: Directory of entire log. num_batch: number of batches to select checkpoint. -1 means select the latest one. metric: metric to maximise or minimise. max_metric: maximise the metric or not. Returns: A directory having arrays.npy and tree.pkl. Raises: ValueError: if any file not found. """ ckpt_dir = log_dir / "files" / "ckpt" if num_batch < 0: # take the one having the best metrics best_metric_scalar = -np.inf if max_metric else np.inf for ckpt_i_dir in ckpt_dir.glob("batch_*/"): if not ckpt_i_dir.is_dir(): continue # load metric num_batch_i = int(ckpt_i_dir.stem.split("_")[-1]) metric_path = ckpt_i_dir / "mean_metrics.json" if not metric_path.exists(): continue with open(metric_path, encoding="utf-8") as f: scalars = json.load(f) if metric not in scalars: raise ValueError(f"Metrics {metric} not found in {ckpt_i_dir}") metric_scalar = scalars[metric] # use the ckpt if it's the first or its metric is better # if same performance, prefer being trained for longer if ( (num_batch < 0) or (max_metric and (best_metric_scalar <= metric_scalar)) or ((not max_metric) and (best_metric_scalar >= metric_scalar)) ): best_metric_scalar = metric_scalar num_batch = num_batch_i if num_batch < 0: raise ValueError(f"Checkpoint not found under {ckpt_dir}") ckpt_dir = ckpt_dir / f"batch_{num_batch}" # sanity check if not ckpt_dir.exists(): raise ValueError(f"Checkpoint directory {ckpt_dir} does not exist.") array_path = ckpt_dir / "arrays.npy" if not array_path.exists(): raise ValueError(f"Checkpoint {array_path} does not exist.") tree_path = ckpt_dir / "tree.pkl" if not tree_path.exists(): raise ValueError(f"Checkpoint {tree_path} does not exist.") return ckpt_dir def parse_args() -> argparse.Namespace: """Parse arguments.""" parser = argparse.ArgumentParser(description=__doc__) parser.add_argument( "--log_dir", type=Path, help="Folder of wandb.", default=None, ) parser.add_argument( "--num_batch", type=int, help="Number of batches for identify checkpoint.", default=-1, ) parser.add_argument( "--num_timesteps", type=int, help="Number of sampling steps for diffusion.", default=-1, ) parser.add_argument( "--num_seeds", type=int, help="Number of seeds for inference.", default=1, ) parser.add_argument( "--metric", type=str, help="Metric to select template.", default="mean_binary_dice_score", ) parser.add_argument("--max_metric", dest="max_metric", action="store_true") parser.add_argument("--min_metric", dest="max_metric", action="store_false") parser.set_defaults(max_metric=True) args = parser.parse_args() return args def main() -> None: """Main function.""" args = parse_args() config = OmegaConf.load(args.log_dir / "files" / "config_backup.yaml") out_dir = args.log_dir / "files" / "test_evaluation" if config.task.name == "diffusion": if args.num_timesteps <= 0: raise ValueError("num_timesteps required for diffusion.") config.task.diffusion.num_timesteps = args.num_timesteps logging.info(f"Sampling {args.num_timesteps} steps.") ckpt_dir = get_checkpoint_dir( log_dir=args.log_dir, num_batch=args.num_batch, metric=args.metric, max_metric=args.max_metric, ) logging.info(f"Using checkpoint {ckpt_dir}.") # load checkpoint params, state = get_eval_params_and_state_from_ckpt( ckpt_dir=ckpt_dir, use_ema=config.training.ema.use, ) # prevent any gradient related actions params = jax.lax.stop_gradient(params) state = jax.lax.stop_gradient(state) # inference per seed for seed in range(args.num_seeds): logging.info(f"Starting test split evaluation for seed {seed}.") out_dir_seed = out_dir / f"seed_{seed}" out_dir_seed.mkdir(parents=True, exist_ok=True) if config.task.name == "diffusion": out_dir_seed = out_dir_seed / f"sample_{args.num_timesteps}_steps" # init exp # necessary as data set will be exhausted run = Experiment(config=config) run.eval_init() rng = jax.random.PRNGKey(seed) rng = broadcast_to_local_devices(rng) run.eval_step( split=TEST_SPLIT, params=params, state=state, rng=rng, out_dir=out_dir_seed, save_predictions=True, ) logging.info(f"Finished test split evaluation for seed {seed}.") if __name__ == "__main__": main()
5,684
30.236264
80
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/math_util.py
"""Module for math functions.""" import jax import jax.numpy as jnp def logits_to_mask(x: jnp.ndarray, axis: int) -> jnp.ndarray: """Transform logits to one hot mask. The one will be on the class having largest logit. Args: x: logits. axis: axis of num_classes. Returns: One hot probabilities. """ return jax.nn.one_hot( x=jnp.argmax(x, axis=axis), num_classes=x.shape[axis], axis=axis, )
471
18.666667
61
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/run_valid.py
"""Script to launch evaluation on validation tests.""" import argparse from pathlib import Path from typing import List import jax from absl import logging from omegaconf import OmegaConf from imgx import VALID_SPLIT from imgx.device import broadcast_to_local_devices from imgx.exp import Experiment from imgx.exp.train_state import get_eval_params_and_state_from_ckpt logging.set_verbosity(logging.INFO) def get_checkpoint_dirs( log_dir: Path, ) -> List[Path]: """Get the directory of all checkpoints. Args: log_dir: Directory of entire log. Returns: A list of directories having arrays.npy and tree.pkl. Raises: ValueError: if any file not found. """ ckpt_dir = log_dir / "files" / "ckpt" ckpt_dirs = [] num_batches = [] for ckpt_i_dir in ckpt_dir.glob("batch_*/"): if not ckpt_i_dir.is_dir(): continue array_path = ckpt_i_dir / "arrays.npy" if not array_path.exists(): continue tree_path = ckpt_i_dir / "tree.pkl" if not tree_path.exists(): continue num_batches.append(int(ckpt_i_dir.stem.split("_")[-1])) ckpt_dirs.append(ckpt_i_dir) ckpt_dirs = [ x[1] for x in sorted(zip(num_batches, ckpt_dirs), reverse=True) ] return ckpt_dirs def parse_args() -> argparse.Namespace: """Parse arguments.""" parser = argparse.ArgumentParser(description=__doc__) parser.add_argument( "--log_dir", type=Path, help="Folder of wandb.", default=None, ) parser.add_argument( "--num_timesteps", type=int, help="Number of sampling steps for diffusion.", default=-1, ) args = parser.parse_args() return args def main() -> None: """Main function.""" args = parse_args() config = OmegaConf.load(args.log_dir / "files" / "config_backup.yaml") if config.task.name == "diffusion": if args.num_timesteps <= 0: raise ValueError("num_timesteps required for diffusion.") config.task.diffusion.num_timesteps = args.num_timesteps ckpt_dirs = get_checkpoint_dirs( log_dir=args.log_dir, ) # init exp run = Experiment(config=config) run.eval_init() for ckpt_dir in ckpt_dirs: # load checkpoint params, state = get_eval_params_and_state_from_ckpt( ckpt_dir=ckpt_dir, use_ema=config.training.ema.use, ) # prevent any gradient related actions params = jax.lax.stop_gradient(params) state = jax.lax.stop_gradient(state) # inference logging.info(f"Starting valid split evaluation for {ckpt_dir}.") rng = jax.random.PRNGKey(config.seed) rng = broadcast_to_local_devices(rng) run.eval_step( split=VALID_SPLIT, params=params, state=state, rng=rng, out_dir=ckpt_dir, save_predictions=False, ) # clean up del params del state if __name__ == "__main__": main()
3,106
24.891667
74
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/__init__.py
"""A Jax-based DL toolkit for biomedical and bioinformatics applications.""" from pathlib import Path # machine error EPS = 1.0e-5 NAN_MASK = "nan_mask" # path for all non-tensorflow-dataset data sets DIR_DATA = Path("datasets") # splits TRAIN_SPLIT = "train" VALID_SPLIT = "valid" TEST_SPLIT = "test" # jax device # one model can be stored across multiple shards/slices # given 8 devices, it can be grouped into 4x2 # if num_devices_per_replica = 2, then one model is stored across 2 devices # so the replica_axis would be of size 4 SHARD_AXIS = "shard_axis" REPLICA_AXIS = "replica_axis" # data dict keys UID = "uid" IMAGE = "image" LABEL = "label"
657
21.689655
76
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/run_train.py
"""Script to launch training.""" from pathlib import Path import hydra import jax import wandb from absl import logging from omegaconf import DictConfig, OmegaConf from imgx import VALID_SPLIT from imgx.config import flatten_dict from imgx.exp import Experiment from imgx.exp.train_state import get_eval_params_and_state, save_ckpt logging.set_verbosity(logging.INFO) def set_debug_config(config: DictConfig) -> DictConfig: """Modify config for debugging purpose. Args: config: original config. Returns: modified config. Raises: ValueError: if data set is unknown. """ # reduce all model size config.model.unet3d.num_channels = (1, 2, 4, 8) config.model.unet3d_slice.num_channels = (1, 2, 4, 8) config.model.unet3d_time.num_channels = (1, 2, 4, 8) config.model.unet3d_slice_time.num_channels = (1, 2, 4, 8) # make training shorter n_devices = jax.local_device_count() config.data.max_num_samples = 11 config.training.batch_size_per_replica = 2 config.training.batch_size = ( n_devices * config.training.batch_size_per_replica ) config.training.max_num_samples = 100 # make logging more frequent config.logging.eval_freq = 1 config.logging.save_freq = 4 return config def get_batch_size_per_step(config: DictConfig) -> int: """Return the actual number of samples per step. Args: config: total config. Returns: Number of samples across all devices. """ if "batch_size_per_replica" not in config["training"]: logging.warning("Batch size per step is not accurate.") return 1 num_devices = jax.local_device_count() num_devices_per_replica = config["training"]["num_devices_per_replica"] num_models = num_devices // num_devices_per_replica batch_size = config["training"]["batch_size_per_replica"] * num_models return batch_size @hydra.main( version_base=None, config_path="conf", config_name="config_segmentation" ) def main( # pylint:disable=too-many-statements config: DictConfig, ) -> None: """Main function. Args: config: config loaded from yaml. """ # update config if config.debug: config = set_debug_config(config) # init wandb files_dir = None if config.logging.wandb.project: wandb_run = wandb.init( project=config.logging.wandb.project, entity=config.logging.wandb.entity, config=flatten_dict(dict(config)), ) files_dir = Path(wandb_run.settings.files_dir) # backup config OmegaConf.save(config=config, f=files_dir / "config_backup.yaml") # init devices devices = jax.local_devices() if config.training.num_devices_per_replica != 1: raise ValueError("Distributed training not supported.") logging.info(f"Local devices are: {devices}") # init exp run = Experiment(config=config) train_state = run.train_init() run.eval_init() logging.info("Start training.") batch_size_per_step = get_batch_size_per_step(config) max_num_steps = config.training.max_num_samples // batch_size_per_step for i in range(1, 1 + max_num_steps): # train step train_state, train_scalars = run.train_step( train_state=train_state, ) train_scalars = {"train_" + k: v for k, v in train_scalars.items()} scalars = { "num_samples": i * batch_size_per_step, **train_scalars, } to_save_ckpt = ( (i > 0) and (i % config.logging.save_freq == 0) and (files_dir is not None) ) # evaluate if saving ckpt or time to evaluate to_eval = to_save_ckpt or (i % config.logging.eval_freq == 0) if to_save_ckpt and (files_dir is not None): ckpt_dir = files_dir / "ckpt" / f"batch_{i}" else: ckpt_dir = None if to_eval and config.eval: # TODO on TPU evaluation causes OOM params, state = get_eval_params_and_state(train_state) val_scalars = run.eval_step( split=VALID_SPLIT, params=params, state=state, rng=jax.random.PRNGKey(config.seed), out_dir=ckpt_dir, save_predictions=False, ) val_scalars = {"valid_" + k: v for k, v in val_scalars.items()} scalars = { **scalars, **val_scalars, } if config.logging.wandb.project: wandb.log(scalars) scalars = { k: v if isinstance(v, int) else f"{v:.2e}" for k, v in scalars.items() } logging.info(f"Batch {i}: {scalars}") # save checkpoint and metrics if ckpt_dir is not None: save_ckpt( train_state=train_state, ckpt_dir=ckpt_dir, ) # backup config every time OmegaConf.save(config=config, f=ckpt_dir / "config.yaml") logging.info(f"Checkpoint saved at {ckpt_dir}") if __name__ == "__main__": main() # pylint: disable=no-value-for-parameter
5,235
29.619883
76
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/device.py
"""Module to handle multi-devices.""" from typing import Optional, Tuple, Union import chex import jax import jax.numpy as jnp def broadcast_to_local_devices(value: chex.ArrayTree) -> chex.ArrayTree: """Broadcasts an object to all local devices. Args: value: value to be broadcast. Returns: broadcast value. """ devices = jax.local_devices() return jax.tree_map( lambda v: jax.device_put_sharded(len(devices) * [v], devices), value ) def get_first_replica_values(value: chex.ArrayTree) -> chex.ArrayTree: """Gets values from the first replica. Args: value: broadcast value. Returns: value of the first replica. """ return jax.tree_map(lambda x: x[0], value) def bind_rng_to_host_or_device( rng: jnp.ndarray, bind_to: Optional[str] = None, axis_name: Optional[Union[str, Tuple[str, ...]]] = None, ) -> jnp.ndarray: """Binds a rng to the host or device. https://github.com/google-research/scenic/blob/main/scenic/train_lib/train_utils.py#L577 Must be called from within a pmapped function. Note that when binding to "device", we also bind the rng to hosts, as we fold_in the rng with axis_index, which is unique for devices across all hosts. Args: rng: A jax.random.PRNGKey. bind_to: Must be one of the 'host' or 'device'. None means no binding. axis_name: The axis of the devices we are binding rng across, necessary if bind_to is device. Returns: jax.random.PRNGKey specialized to host/device. """ if bind_to is None: return rng if bind_to == "host": return jax.random.fold_in(rng, jax.process_index()) if bind_to == "device": return jax.random.fold_in(rng, jax.lax.axis_index(axis_name)) raise ValueError( "`bind_to` should be one of the `[None, 'host', 'device']`" ) def shard( pytree: chex.ArrayTree, num_replicas: int, ) -> chex.ArrayTree: """Reshapes all arrays in the pytree to add a leading shard dimension. We assume that all arrays in the pytree have leading dimension divisible by num_devices_per_replica. Args: pytree: A pytree of arrays to be sharded. num_replicas: number of model replicas. Returns: Sharded data. """ def _shard_array(array: jnp.ndarray) -> jnp.ndarray: return array.reshape((num_replicas, -1) + array.shape[1:]) return jax.tree_map(_shard_array, pytree) def unshard(pytree: chex.ArrayTree) -> chex.ArrayTree: """Reshapes arrays from [ndev, bs, ...] to [host_bs, ...]. Args: pytree: A pytree of arrays to be sharded. Returns: Sharded data. """ def _unshard_array(array: jnp.ndarray) -> jnp.ndarray: ndev, bs = array.shape[:2] return array.reshape((ndev * bs,) + array.shape[2:]) return jax.tree_map(_unshard_array, pytree) def is_tpu() -> bool: """Return true if the device is tpu. Returns: True if tpu. """ return jax.local_devices()[0].platform == "tpu"
3,095
25.689655
92
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/metric/area.py
"""Metrics to measure foreground area.""" import jax.numpy as jnp def class_proportion(mask: jnp.ndarray) -> jnp.ndarray: """Calculate proportion per class. Args: mask: shape = (batch, d1, ..., dn, num_classes). Returns: Proportion, shape = (batch, num_classes). """ reduce_axes = tuple(range(1, mask.ndim - 1)) volume = jnp.float32(jnp.prod(jnp.array(mask.shape[1:-1]))) sqrt_volume = jnp.sqrt(volume) mask = jnp.float32(mask) # attempt to avoid over/underflow return jnp.sum(mask / sqrt_volume, axis=reduce_axes) / sqrt_volume
590
27.142857
70
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/metric/distribution.py
"""Metric functions for probability distributions.""" import jax.numpy as jnp def normal_kl( p_mean: jnp.ndarray, p_log_variance: jnp.ndarray, q_mean: jnp.ndarray, q_log_variance: jnp.ndarray, ) -> jnp.ndarray: """Compute the KL divergence between two 1D normal distributions. Although the inputs are arrays, each value is considered independently. This function is not symmetric. Input array shapes should be broadcastable. Args: p_mean: mean of distribution p. p_log_variance: log variance of distribution p. q_mean: mean of distribution q. q_log_variance: log variance of distribution q. Returns: KL divergence. """ return 0.5 * ( -1.0 + q_log_variance - p_log_variance + jnp.exp(p_log_variance - q_log_variance) + ((p_mean - q_mean) ** 2) * jnp.exp(-q_log_variance) ) def approx_standard_normal_cdf(x: jnp.ndarray) -> jnp.ndarray: """Approximate cumulative distribution function of standard normal. if x ~ Normal(mean, var), then cdf(z) = p(x <= z) https://www.aimspress.com/article/doi/10.3934/math.2022648#b13 https://www.jstor.org/stable/2346872 Args: x: array of any shape with any float values. Returns: CDF estimation. """ return 0.5 * ( 1.0 + jnp.tanh(jnp.sqrt(2.0 / jnp.pi) * (x + 0.044715 * x**3)) ) def discretized_gaussian_log_likelihood( x: jnp.ndarray, mean: jnp.ndarray, log_variance: jnp.ndarray, x_delta: float = 1.0 / 255.0, x_bound: float = 0.999, ) -> jnp.ndarray: """Log-likelihood of a normal distribution discretizing to an image. Args: x: target image, with value inside normalized in [-1, 1]. mean: normal distribution mean. log_variance: log of distribution variance. x_delta: discretization step, used to estimate probability. x_bound: values with abs > x_bound are calculated differently. Returns: Discretized log likelihood over 2*delta. """ log_scales = 0.5 * log_variance centered_x = x - mean inv_stdv = jnp.exp(-log_scales) # let y be a variable # cdf(z+delta) = p(y <= z+delta) plus_in = inv_stdv * (centered_x + x_delta) cdf_plus = approx_standard_normal_cdf(plus_in) # log( p(y <= z+delta) ) log_cdf_plus = jnp.log(cdf_plus.clip(min=1e-12)) # cdf(z-delta) = p(y <= z-delta) minus_in = inv_stdv * (centered_x - x_delta) cdf_minus = approx_standard_normal_cdf(minus_in) # log( 1-p(y <= z-delta) ) = log( p(y > z-delta) ) log_one_minus_cdf_minus = jnp.log((1.0 - cdf_minus).clip(min=1e-12)) # p(z-delta < y <= z+delta) cdf_delta = cdf_plus - cdf_minus log_cdf_delta = jnp.log(cdf_delta.clip(min=1e-12)) # if x < -0.999, log( p(y <= z+delta) ) # if x > 0.999, log( p(y > z-delta) ) # if -0.999 <= x <= 0.999, log( p(z-delta < y <= z+delta) ) return jnp.where( x < -x_bound, log_cdf_plus, jnp.where(x > x_bound, log_one_minus_cdf_minus, log_cdf_delta), )
3,097
29.07767
75
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/metric/dice.py
"""Metric functions for image segmentation.""" import jax.numpy as jnp def dice_score( mask_pred: jnp.ndarray, mask_true: jnp.ndarray, ) -> jnp.ndarray: """Soft Dice score, larger is better. Args: mask_pred: soft mask with probabilities, (batch, ..., num_classes). mask_true: one hot targets, (batch, ..., num_classes). Returns: Dice score of shape (batch, num_classes). """ reduce_axis = tuple(range(mask_pred.ndim))[1:-1] numerator = 2.0 * jnp.sum(mask_pred * mask_true, axis=reduce_axis) denominator = jnp.sum(mask_pred + mask_true, axis=reduce_axis) return jnp.where( condition=denominator > 0, x=numerator / denominator, y=jnp.nan, ) def iou( mask_pred: jnp.ndarray, mask_true: jnp.ndarray, ) -> jnp.ndarray: """IOU (Intersection Over Union), or Jaccard index. Args: mask_pred: binary mask of predictions, (batch, ..., num_classes). mask_true: one hot targets, (batch, ..., num_classes). Returns: IoU of shape (batch, num_classes). """ reduce_axis = tuple(range(mask_pred.ndim))[1:-1] numerator = jnp.sum(mask_pred * mask_true, axis=reduce_axis) sum_mask = jnp.sum(mask_pred + mask_true, axis=reduce_axis) denominator = sum_mask - numerator return jnp.where( condition=sum_mask > 0, x=numerator / denominator, y=jnp.nan )
1,407
28.333333
75
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/metric/centroid.py
"""Metric centroid distance.""" from typing import Optional, Tuple import jax.numpy as jnp def get_coordinate_grid(shape: Tuple[int, ...]) -> jnp.ndarray: """Generate a grid with given shape. This function is not jittable as the output depends on the value of shapes. Args: shape: shape of the grid, (d1, ..., dn). Returns: grid: grid coordinates, of shape (n, d1, ..., dn). grid[:, i1, ..., in] = [i1, ..., in] """ return jnp.stack( jnp.meshgrid( *(jnp.arange(d) for d in shape), indexing="ij", ), axis=0, dtype=jnp.float32, ) def get_centroid( mask: jnp.ndarray, grid: jnp.ndarray, ) -> Tuple[jnp.ndarray, jnp.ndarray]: """Calculate the centroid of the mask. Args: mask: boolean mask of shape = (batch, d1, ..., dn, num_classes) grid: shape = (n, d1, ..., dn) Returns: centroid of shape = (batch, n, num_classes). nan mask of shape = (batch, num_classes). """ mask_reduce_axes = tuple(range(1, mask.ndim - 1)) grid_reduce_axes = tuple(range(2, mask.ndim)) # (batch, n, d1, ..., dn, num_classes) masked_grid = jnp.expand_dims(mask, axis=1) * jnp.expand_dims( grid, axis=(0, -1) ) # (batch, n, num_classes) numerator = jnp.sum(masked_grid, axis=grid_reduce_axes) # (batch, num_classes) summed_mask = jnp.sum(mask, axis=mask_reduce_axes) # (batch, 1, num_classes) denominator = summed_mask[:, None, :] # if mask is not empty return real centroid, else nan centroid = jnp.where( condition=denominator > 0, x=numerator / denominator, y=jnp.nan ) return centroid, summed_mask == 0 def centroid_distance( mask_true: jnp.ndarray, mask_pred: jnp.ndarray, grid: jnp.ndarray, spacing: Optional[jnp.ndarray] = None, ) -> jnp.ndarray: """Calculate the L2-distance between two centroids. Args: mask_true: shape = (batch, d1, ..., dn, num_classes). mask_pred: shape = (batch, d1, ..., dn, num_classes). grid: shape = (n, d1, ..., dn). spacing: spacing of pixel/voxels along each dimension, (n,). Returns: distance, shape = (batch, num_classes). """ # centroid (batch, n, num_classes) nan_mask (batch, num_classes) centroid_true, nan_mask_true = get_centroid( mask=mask_true, grid=grid, ) centroid_pred, nan_mask_pred = get_centroid( mask=mask_pred, grid=grid, ) nan_mask = nan_mask_true | nan_mask_pred if spacing is not None: centroid_true = jnp.where( condition=nan_mask[:, None, :], x=jnp.nan, y=centroid_true * spacing[None, :, None], ) centroid_pred = jnp.where( condition=nan_mask[:, None, :], x=jnp.nan, y=centroid_pred * spacing[None, :, None], ) # return nan if the centroid cannot be defined for one sample with one class return jnp.where( condition=nan_mask, x=jnp.nan, y=jnp.linalg.norm(centroid_true - centroid_pred, axis=1), )
3,160
28.542056
80
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/datasets/augmentation.py
"""Image augmentation functions.""" from functools import partial from typing import Callable, Dict, Sequence import jax import numpy as np from jax import numpy as jnp from jax.scipy.ndimage import map_coordinates from omegaconf import DictConfig from imgx import IMAGE, LABEL from imgx.datasets import FOREGROUND_RANGE, IMAGE_SHAPE_MAP from imgx.metric.centroid import get_coordinate_grid def get_2d_rotation_matrix( radians: jnp.ndarray, ) -> jnp.ndarray: """Return 2d rotation matrix given radians. The affine transformation applies as following: [x, = [[* * 0] * [x, y, [* * 0] y, 1] [0 0 1]] 1] Args: radians: tuple of one values, correspond to xy planes. Returns: Rotation matrix of shape (3, 3). """ sin, cos = jnp.sin(radians[0]), jnp.cos(radians[0]) return jnp.array( [ [cos, -sin, 0.0], [sin, cos, 0.0], [0.0, 0.0, 1.0], ] ) def get_3d_rotation_matrix( radians: jnp.ndarray, ) -> jnp.ndarray: """Return 3d rotation matrix given radians. The affine transformation applies as following: [x, = [[* * * 0] * [x, y, [* * * 0] y, z, [* * * 0] z, 1] [0 0 0 1]] 1] Args: radians: tuple of three values, correspond to yz, xz, xy planes. Returns: Rotation matrix of shape (4, 4). """ affine = jnp.eye(4) # rotation of yz around x-axis sin, cos = jnp.sin(radians[0]), jnp.cos(radians[0]) affine_ax = jnp.array( [ [1.0, 0.0, 0.0, 0.0], [0.0, cos, -sin, 0.0], [0.0, sin, cos, 0.0], [0.0, 0.0, 0.0, 1.0], ] ) affine = jnp.matmul(affine_ax, affine) # rotation of zx around y-axis sin, cos = jnp.sin(radians[1]), jnp.cos(radians[1]) affine_ax = jnp.array( [ [cos, 0.0, sin, 0.0], [0.0, 1.0, 0.0, 0.0], [-sin, 0.0, cos, 0.0], [0.0, 0.0, 0.0, 1.0], ] ) affine = jnp.matmul(affine_ax, affine) # rotation of xy around z-axis sin, cos = jnp.sin(radians[2]), jnp.cos(radians[2]) affine_ax = jnp.array( [ [cos, -sin, 0.0, 0.0], [sin, cos, 0.0, 0.0], [0.0, 0.0, 1.0, 0.0], [0.0, 0.0, 0.0, 1.0], ] ) affine = jnp.matmul(affine_ax, affine) return affine def get_rotation_matrix( radians: jnp.ndarray, ) -> jnp.ndarray: """Return rotation matrix given radians. Args: radians: correspond to rotate around each axis. Returns: Rotation matrix of shape (n+1, n+1). Raises: ValueError: if not 2D or 3D. """ if radians.size == 1: return get_2d_rotation_matrix(radians) if radians.size == 3: return get_3d_rotation_matrix(radians) raise ValueError("Only support 2D/3D rotations.") def get_translation_matrix( shifts: jnp.ndarray, ) -> jnp.ndarray: """Return 3d translation matrix given shifts. For example, the 3D affine transformation applies as following: [x, = [[1 0 0 *] * [x, y, [0 1 0 *] y, z, [0 0 1 *] z, 1] [0 0 0 1]] 1] Args: shifts: correspond to each axis shift. Returns: Translation matrix of shape (n+1, n+1). """ ndims = shifts.size shifts = jnp.concatenate([shifts, jnp.array([1.0])]) return jnp.concatenate( [ jnp.eye(ndims + 1, ndims), shifts[:, None], ], axis=1, ) def get_scaling_matrix( scales: jnp.ndarray, ) -> jnp.ndarray: """Return scaling matrix given scales. For example, the 3D affine transformation applies as following: [x, = [[* 0 0 0] * [x, y, [0 * 0 0] y, z, [0 0 * 0] z, 1] [0 0 0 1]] 1] Args: scales: correspond to each axis scaling. Returns: Affine matrix of shape (n+1, n+1). """ scales = jnp.concatenate([scales, jnp.array([1.0])]) return jnp.diag(scales) def get_affine_matrix( radians: jnp.ndarray, shifts: jnp.ndarray, scales: jnp.ndarray, ) -> jnp.ndarray: """Return an affine matrix from parameters. The matrix is not squared, as the last row is not needed. For rotation, translation, and scaling matrix, they are kept for composition purpose. For example, the 3D affine transformation applies as following: [x, = [[* * * *] * [x, y, [* * * *] y, z, [* * * *] z, 1] [0 0 0 1]] 1] Args: radians: correspond to rotate around each axis. shifts: correspond to each axis shift. scales: correspond to each axis scaling. Returns: Affine matrix of shape (n+1, n+1). """ affine_rot = get_rotation_matrix(radians) affine_shift = get_translation_matrix(shifts) affine_scale = get_scaling_matrix(scales) return jnp.matmul(affine_shift, jnp.matmul(affine_scale, affine_rot)) def batch_get_random_affine_matrix( key: jax.random.PRNGKey, max_rotation: jnp.ndarray, min_translation: jnp.ndarray, max_translation: jnp.ndarray, max_scaling: jnp.ndarray, ) -> jnp.ndarray: """Get a batch of random affine matrices. Args: key: jax random key. max_rotation: maximum rotation in radians. min_translation: minimum translation in pixel/voxels. max_translation: maximum translation in pixel/voxels. max_scaling: maximum scaling difference in pixel/voxels. Returns: Affine matrix of shape (batch, n+1, n+1). """ key_radian, key_shift, key_scale = jax.random.split(key, num=3) radians = jax.random.uniform( key=key_radian, shape=max_rotation.shape, minval=-max_rotation, maxval=max_rotation, ) shifts = jax.random.uniform( key=key_shift, shape=max_translation.shape, minval=min_translation, maxval=max_translation, ) scales = jax.random.uniform( key=key_scale, shape=max_scaling.shape, minval=1.0 - max_scaling, maxval=1.0 + max_scaling, ) # vmap on first axis, which is a batch return jax.vmap(get_affine_matrix)(radians, shifts, scales) def apply_affine_to_grid( grid: jnp.ndarray, affine_matrix: jnp.ndarray ) -> jnp.ndarray: """Apply affine matrix to grid. The grid has non-negative coordinates, means the origin is at a corner. Need to shift the grid such that the origin is at center, then apply affine, then shift the origin back. Args: grid: grid coordinates, of shape (n, d1, ..., dn). grid[:, i1, ..., in] = [i1, ..., in] affine_matrix: shape (n+1, n+1) Returns: Grid with updated coordinates. """ # (n+1, d1, ..., dn) extended_grid = jnp.concatenate( [grid, jnp.ones((1,) + grid.shape[1:])], axis=0 ) # shift to center shift = (jnp.array(grid.shape[1:]) - 1) / 2 shift_matrix = get_translation_matrix(-shift) # (n+1, n+1) # (n+1, n+1) * (n+1, d1, ..., dn) = (n+1, d1, ..., dn) extended_grid = jnp.einsum("ji,i...->j...", shift_matrix, extended_grid) # affine # (n+1, n+1) * (n+1, d1, ..., dn) = (n+1, d1, ..., dn) extended_grid = jnp.einsum("ji,i...->j...", affine_matrix, extended_grid) # shift to corner shift_matrix = get_translation_matrix(shift)[:-1, :] # (n, n+1) # (n, n+1) * (n+1, d1, ..., dn) = (n, d1, ..., dn) extended_grid = jnp.einsum("ji,i...->j...", shift_matrix, extended_grid) return extended_grid def batch_apply_affine_to_grid( grid: jnp.ndarray, affine_matrix: jnp.ndarray ) -> jnp.ndarray: """Apply batch of affine matrix to grid. Args: grid: grid coordinates, of shape (n, d1, ..., dn). grid[:, i1, ..., in] = [i1, ..., in] affine_matrix: shape (batch, n+1, n+1). Returns: Grid with updated coordinates, shape (batch, n, d1, ..., dn). """ return jax.vmap(apply_affine_to_grid, in_axes=(None, 0))( grid, affine_matrix ) def batch_resample_image_label( input_dict: Dict[str, jnp.ndarray], grid: jnp.ndarray, ) -> Dict[str, jnp.ndarray]: """Apply batch of affine matrix to image and label. Args: input_dict: dict having image and label. image shape (batch, d1, ..., dn) grid: grid coordinates, of shape (batch, n, d1, ..., dn). Returns: Updated image and label, of same shape. """ resample_image_vmap = jax.vmap( partial( map_coordinates, order=1, mode="constant", cval=0.0, ), in_axes=(0, 0), ) resample_label_vmap = jax.vmap( partial( map_coordinates, order=0, mode="constant", cval=0.0, ), in_axes=(0, 0), ) image = resample_image_vmap( input_dict[IMAGE], grid, ) label = resample_label_vmap( input_dict[LABEL], grid, ) return {IMAGE: image, LABEL: label} def batch_random_affine_transform( key: jax.random.PRNGKey, input_dict: Dict[str, jnp.ndarray], grid: jnp.ndarray, max_rotation: jnp.ndarray, max_translation: jnp.ndarray, max_scaling: jnp.ndarray, ) -> Dict[str, jnp.ndarray]: """Keep image and label only. TODO: image does not have channel. Args: key: jax random key. input_dict: dict having image and label. image shape (batch, d1, ..., dn) grid: grid coordinates, of shape (n, d1, ..., dn). grid[:, i1, ..., in] = [i1, ..., in] max_rotation: maximum rotation in radians, shape = (batch, ...). max_translation: maximum translation in pixel/voxels, shape = (batch, ...). max_scaling: maximum scaling difference in pixel/voxels, shape = (batch, ...). Returns: Augmented dict having image and label. image and label all have shapes (batch, d1, ..., dn). """ # (batch, ...) batch_size = input_dict[IMAGE].shape[0] max_rotation = jnp.tile(max_rotation[None, ...], (batch_size, 1)) max_translation = jnp.tile(max_translation[None, ...], (batch_size, 1)) min_translation = -max_translation max_scaling = jnp.tile(max_scaling[None, ...], (batch_size, 1)) # refine translation to avoid remove classes shape = jnp.array(input_dict[LABEL].shape[1:]) shape = jnp.tile(shape[None, ...], (batch_size, 1)) max_translation = jnp.minimum( max_translation, shape - 1 - input_dict[FOREGROUND_RANGE][..., -1] ) min_translation = jnp.maximum( min_translation, -input_dict[FOREGROUND_RANGE][..., 0] ) # (batch, n+1, n+1) affine_matrix = batch_get_random_affine_matrix( key=key, max_rotation=max_rotation, min_translation=min_translation, max_translation=max_translation, max_scaling=max_scaling, ) # (batch, n, d1, ..., dn) grid = batch_apply_affine_to_grid(grid=grid, affine_matrix=affine_matrix) return batch_resample_image_label( input_dict=input_dict, grid=grid, ) def build_aug_fn_from_fns( fns: Sequence[Callable], ) -> Callable: """Combine a list of data augmentation functions. Args: fns: entire config. Returns: A data augmentation function. """ def aug_fn( key: jax.random.PRNGKey, input_dict: Dict[str, jnp.ndarray] ) -> Dict[str, jnp.ndarray]: keys = jax.random.split(key, num=len(fns)) for k, fn in zip(keys, fns): input_dict = fn(k, input_dict) return input_dict return aug_fn def build_aug_fn_from_config( config: DictConfig, ) -> Callable: """Return a data augmentation function. Args: config: entire config. Returns: A data augmentation function. """ data_config = config.data dataset_name = data_config["name"] image_shape = IMAGE_SHAPE_MAP[dataset_name] da_config = data_config[dataset_name]["data_augmentation"] grid = get_coordinate_grid(shape=image_shape) max_rotation = np.array(da_config["max_rotation"]) max_translation = np.array(da_config["max_translation"]) max_scaling = np.array(da_config["max_scaling"]) aug_fns = [ partial( batch_random_affine_transform, grid=grid, max_rotation=max_rotation, max_translation=max_translation, max_scaling=max_scaling, ) ] if len(aug_fns) == 1: return aug_fns[0] return build_aug_fn_from_fns(aug_fns)
12,812
26.793926
77
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/datasets/iterator.py
"""Dataset related classes and functions.""" from functools import partial from typing import Callable, Dict, Iterator, Optional, Tuple import jax import jax.numpy as jnp import jax.scipy import jmp import tensorflow as tf import tensorflow_datasets as tfds from absl import logging from omegaconf import DictConfig from imgx import IMAGE, LABEL, TEST_SPLIT, TRAIN_SPLIT, VALID_SPLIT from imgx.datasets import FOREGROUND_RANGE, Dataset from imgx.datasets.util import ( get_foreground_range, maybe_pad_batch, tf_to_numpy, ) from imgx.device import shard def create_image_label_dict_from_dict( x: Dict[str, tf.Tensor], ) -> Dict[str, tf.Tensor]: """Create a dict from inputs. Args: x: dict having image, label, and other tensors. Returns: Dict having image and label. """ return { IMAGE: x[IMAGE], LABEL: x[LABEL], FOREGROUND_RANGE: get_foreground_range(x[LABEL]), } def load_split_from_image_tfds_builder( builder: tfds.core.DatasetBuilder, batch_size: int, split: str, augment_train_example_fn: Optional[Callable] = None, shuffle_buffer_size: Optional[int] = None, shuffle_seed: int = 0, max_num_samples: int = -1, dtype: jnp.dtype = jnp.float32, ) -> Tuple[tf.data.Dataset, int]: """Loads a split from a TensorFlow Dataset compatible builder. https://github.com/google-research/scenic/blob/main/scenic/dataset_lib/dataset_utils.py Args: builder: A TFDS compatible dataset builder. batch_size: The batch size returned by the data pipeline. split: Name of the split to be loaded. augment_train_example_fn: A function that given a train example returns the augmented example. Note that this function is applied AFTER caching and repeat to get true randomness. shuffle_buffer_size: Size of the tf.data.dataset shuffle buffer. shuffle_seed: Seed for shuffling the training data. max_num_samples: maximum number of samples to consider. dtype: data type for images. Returns: - A repeated dataset. - Number of steps after batch if the dataset is not repeated, returns -1 for training. """ is_train = split == TRAIN_SPLIT # Prepare arguments. shuffle_buffer_size = shuffle_buffer_size or (8 * batch_size) # Download data. builder.download_and_prepare() # Each host is responsible for a fixed subset of data. if is_train: split = tfds.even_splits(split, jax.process_count())[ jax.process_index() ] dataset = builder.as_dataset( split=split, shuffle_files=False, ) # Shrink data set if required if max_num_samples > 0: logging.info( f"Taking first {max_num_samples} data samples for split {split}." ) dataset = dataset.take(max_num_samples) # Caching. dataset = dataset.cache() num_steps = -1 # not set for training if is_train: # First repeat then batch. dataset = dataset.repeat() # Augmentation should be done after repeat for true randomness. if augment_train_example_fn: dataset = dataset.map( augment_train_example_fn, num_parallel_calls=tf.data.experimental.AUTOTUNE, ) # Shuffle after augmentation to avoid loading non-augmented images into # buffer. dataset = dataset.shuffle(shuffle_buffer_size, seed=shuffle_seed) dataset = dataset.batch(batch_size, drop_remainder=True) else: # First batch then repeat. dataset = dataset.batch(batch_size, drop_remainder=False) num_steps = tf.data.experimental.cardinality(dataset).numpy() if split == VALID_SPLIT: # repeat dataset for validation dataset = dataset.repeat() # NOTE: You may be tempted to move the casting earlier on in the pipeline, # but for bf16 some operations will end up silently placed on the TPU and # this causes stalls while TF and JAX battle for the accelerator. if dtype != jnp.float32: def cast_fn(batch: Dict[str, jnp.ndarray]) -> Dict[str, jnp.ndarray]: batch[IMAGE] = tf.cast(batch[IMAGE], tf.dtypes.as_dtype(dtype)) return batch dataset = dataset.map(cast_fn) dataset = dataset.prefetch(tf.data.experimental.AUTOTUNE) return dataset, num_steps def get_image_iterator( builder: tfds.core.DatasetBuilder, split: str, is_train: bool, batch_size_per_replica: int, num_replicas: int, shuffle_seed: int, max_num_samples: int, dtype: jnp.dtype = jnp.float32, ) -> Tuple[Iterator, int]: """Returns iterator from builder. Args: builder: data set builder. split: split name. is_train: if the split is for training. batch_size_per_replica: Number of samples consumed per model per step. num_replicas: number of model replicas. shuffle_seed: Seed for shuffling the training data. max_num_samples: maximum number of samples in iterator. dtype: data type for images. Returns: - Batch iterator. - Number of steps after batch if the dataset is not repeated, returns -1 for training. """ batch_size = batch_size_per_replica * num_replicas dataset, num_steps = load_split_from_image_tfds_builder( builder=builder, batch_size=batch_size, split=split, shuffle_seed=shuffle_seed, augment_train_example_fn=create_image_label_dict_from_dict, max_num_samples=max_num_samples, dtype=dtype, ) maybe_pad_batches = partial( maybe_pad_batch, is_train=is_train, batch_size=batch_size ) dataset_iter = iter(dataset) dataset_iter = map(tf_to_numpy, dataset_iter) dataset_iter = map(maybe_pad_batches, dataset_iter) shard_batches = partial(shard, num_replicas=num_replicas) dataset_iter = map(shard_batches, dataset_iter) return dataset_iter, num_steps def get_image_tfds_dataset( dataset_name: str, config: DictConfig, ) -> Dataset: """Returns generators for the dataset train, valid, and test sets. Args: dataset_name: Data set name. config: entire config. Returns: A Dataset() which includes train_iter, valid_iter, and test_iter. """ batch_size_per_replica = config["training"]["batch_size_per_replica"] num_devices_per_replica = config["training"]["num_devices_per_replica"] num_replicas = jax.local_device_count() // num_devices_per_replica shuffle_seed = config["seed"] max_num_samples = config["data"]["max_num_samples"] dtype = jnp.float32 if config["training"]["mixed_precision"]["use"]: dtype = jmp.half_dtype() builder = tfds.builder(dataset_name) train_iter, _ = get_image_iterator( builder=builder, split=TRAIN_SPLIT, is_train=True, batch_size_per_replica=batch_size_per_replica, num_replicas=num_replicas, shuffle_seed=shuffle_seed, max_num_samples=max_num_samples, dtype=dtype, ) valid_iter, num_valid_steps = get_image_iterator( builder=builder, split=VALID_SPLIT, is_train=False, batch_size_per_replica=batch_size_per_replica, num_replicas=num_replicas, shuffle_seed=shuffle_seed, max_num_samples=max_num_samples, dtype=dtype, ) test_iter, num_test_steps = get_image_iterator( builder=builder, split=TEST_SPLIT, is_train=False, batch_size_per_replica=batch_size_per_replica, num_replicas=num_replicas, shuffle_seed=shuffle_seed, max_num_samples=max_num_samples, dtype=dtype, ) return Dataset( train_iter=train_iter, valid_iter=valid_iter, test_iter=test_iter, num_valid_steps=num_valid_steps, num_test_steps=num_test_steps, )
8,041
31.297189
91
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/datasets/util.py
"""Util functions for image. Some are adapted from https://github.com/google-research/scenic/blob/03735eb81f64fd1241c4efdb946ea6de3d326fe1/scenic/dataset_lib/dataset_utils.py """ import functools import queue import threading from typing import Any, Callable, Dict, Generator, Iterable, Tuple import chex import jax import jax.numpy as jnp import numpy as np import tensorflow as tf import tensorflow.experimental.numpy as tnp from absl import logging from imgx import IMAGE def maybe_pad_batch( batch: Dict[str, chex.ArrayTree], is_train: bool, batch_size: int, batch_dim: int = 0, ) -> Dict[str, chex.ArrayTree]: """Zero pad the batch on the right to the batch_size. All leave tensors in the batch pytree will be padded. This function expects the root structure of the batch pytree to be a dictionary and returns a dictionary with the same structure (and substructures), additionally with the key 'batch_mask' added to the root dict, with 1.0 indicating indices which are true data and 0.0 indicating a padded index. `batch_mask` will be used for calculating the weighted cross entropy, or weighted accuracy. Note that in this codebase, we assume we drop the last partial batch from the training set, so if the batch is from the training set (i.e. `train=True`), or when the batch is from the test/validation set, but it is a complete batch, we *modify* the batch dict by adding an array of ones as the `batch_mask` of all examples in the batch. Otherwise, we create a new dict that has the padded patch and its corresponding `batch_mask` array. Note that batch_mask can be also used as the label mask (not input mask), for task that are pixel/token level. This is simply done by applying the mask we make for padding the partial batches on top of the existing label mask. Args: batch: A dictionary containing a pytree. If `inputs_key` is not set, we use the first leave to get the current batch size. Otherwise, the tensor mapped with `inputs_key` at the root dictionary is used. is_train: if the batch is from the training data. In that case, we drop the last (incomplete) batch and thus don't do any padding. batch_size: All arrays in the dict will be padded to have first dimension equal to desired_batch_size. batch_dim: Batch dimension. The default is 0, but it can be different if a sharded batch is given. Returns: A dictionary mapping the same keys to the padded batches. Additionally, we add a key representing weights, to indicate how the batch was padded. Raises: ValueError: if configs are conflicting. """ sample_tensor = batch[IMAGE] batch_pad = batch_size - sample_tensor.shape[batch_dim] if is_train and batch_pad != 0: raise ValueError( "In this codebase, we assumed that we always drop the " "last partial batch of the train set. Please use " "` drop_remainder=True` for the training set." ) # Most batches do not need padding, so we quickly return to avoid slowdown. if is_train or batch_pad == 0: return batch def zero_pad(array: np.ndarray) -> np.ndarray: pad_with = ( [(0, 0)] * batch_dim + [(0, batch_pad)] + [(0, 0)] * (array.ndim - batch_dim - 1) ) return np.pad(array, pad_with, mode="constant") padded_batch = jax.tree_map(zero_pad, batch) return padded_batch def unpad( pytree: chex.ArrayTree, num_samples: int, ) -> chex.ArrayTree: """Remove padded data for all arrays in the pytree. We assume that all arrays in the pytree have the same leading dimension. Args: pytree: A pytree of arrays to be sharded. num_samples: number of samples to keep. Returns: Data without padding """ def _unpad_array(x: jnp.ndarray) -> jnp.ndarray: return x[:num_samples, ...] return jax.tree_map(_unpad_array, pytree) def tf_to_numpy(batch: Dict) -> np.ndarray: """Convert an input batch from tf Tensors to numpy arrays. Args: batch: A dictionary that has items in a batch: image and labels. Returns: Numpy arrays of the given tf Tensors. """ def convert_data(x: tf.Tensor) -> np.ndarray: """Use _numpy() for zero-copy conversion between TF and NumPy. Args: x: tf tensor. Returns: numpy array. """ return x._numpy() # pylint: disable=protected-access return jax.tree_map(convert_data, batch) def get_center_pad_shape( current_shape: Tuple[int, ...], target_shape: Tuple[int, ...] ) -> Tuple[Tuple[int, ...], Tuple[int, ...]]: """Get pad sizes for sitk.ConstantPad. The padding is added symmetrically. Args: current_shape: current shape of the image. target_shape: target shape of the image. Returns: pad_lower: shape to pad on the lower side. pad_upper: shape to pad on the upper side. """ pad_lower = [] pad_upper = [] for i, size_i in enumerate(current_shape): pad_i = max(target_shape[i] - size_i, 0) pad_lower_i = pad_i // 2 pad_upper_i = pad_i - pad_lower_i pad_lower.append(pad_lower_i) pad_upper.append(pad_upper_i) return tuple(pad_lower), tuple(pad_upper) def get_center_crop_shape( current_shape: Tuple[int, ...], target_shape: Tuple[int, ...] ) -> Tuple[Tuple[int, ...], Tuple[int, ...]]: """Get crop sizes for sitk.Crop. The crop is performed symmetrically. Args: current_shape: current shape of the image. target_shape: target shape of the image. Returns: crop_lower: shape to pad on the lower side. crop_upper: shape to pad on the upper side. """ crop_lower = [] crop_upper = [] for i, size_i in enumerate(current_shape): crop_i = max(size_i - target_shape[i], 0) crop_lower_i = crop_i // 2 crop_upper_i = crop_i - crop_lower_i crop_lower.append(crop_lower_i) crop_upper.append(crop_upper_i) return tuple(crop_lower), tuple(crop_upper) def try_to_get_center_crop_shape( label_min: int, label_max: int, current_length: int, target_length: int ) -> Tuple[int, int]: """Try to crop at the center of label, 1D. Args: label_min: label index minimum, inclusive. label_max: label index maximum, exclusive. current_length: current image length. target_length: target image length. Returns: crop_lower: shape to pad on the lower side. crop_upper: shape to pad on the upper side. Raises: ValueError: if label min max is out of range. """ if label_min < 0 or label_max > current_length: raise ValueError("Label index out of range.") if current_length <= target_length: # no need of crop return 0, 0 # attend to perform crop centered at label center label_center = (label_max - 1 + label_min) / 2.0 bbox_lower = int(np.ceil(label_center - target_length / 2.0)) bbox_upper = bbox_lower + target_length # if lower is negative, then have to shift the window to right bbox_lower = max(bbox_lower, 0) # if upper is too large, then have to shift the window to left if bbox_upper > current_length: bbox_lower -= bbox_upper - current_length # calculate crop crop_lower = bbox_lower # bbox index starts at 0 crop_upper = current_length - target_length - crop_lower return crop_lower, crop_upper def get_center_crop_shape_from_bbox( bbox_min: Tuple[int, ...], bbox_max: Tuple[int, ...], current_shape: Tuple[int, ...], target_shape: Tuple[int, ...], ) -> Tuple[Tuple[int, ...], Tuple[int, ...]]: """Get crop sizes for sitk.Crop from label bounding box. The crop is not necessarily performed symmetrically. Args: bbox_min: [start_in_1st_spatial_dim, ...], inclusive, starts at zero. bbox_max: [end_in_1st_spatial_dim, ...], exclusive, starts at zero. current_shape: current shape of the image. target_shape: target shape of the image. Returns: crop_lower: shape to pad on the lower side. crop_upper: shape to pad on the upper side. """ crop_lower = [] crop_upper = [] for i, current_length in enumerate(current_shape): crop_lower_i, crop_upper_i = try_to_get_center_crop_shape( label_min=bbox_min[i], label_max=bbox_max[i], current_length=current_length, target_length=target_shape[i], ) crop_lower.append(crop_lower_i) crop_upper.append(crop_upper_i) return tuple(crop_lower), tuple(crop_upper) def get_foreground_range(label: tf.Tensor) -> tf.Tensor: """Get the foreground range for a given label. This function is not defined in jax for augmentation because, nonzero is not jittable as the number of nonzero elements is unknown. Args: label: shape (d1, ..., dn), here n = ndim below. Returns: shape (ndim, 2), for each dimension, it's [min, max]. """ # (ndim, num_nonzero) nonzero_indices = tnp.stack(tnp.nonzero(label)) # (ndim, 2) return tnp.stack( [tnp.min(nonzero_indices, axis=-1), tnp.max(nonzero_indices, axis=-1)], axis=-1, ) def get_function_name(function: Callable[..., Any]) -> str: """Get name of any function. Args: function: function to query. Returns: function name. """ if isinstance(function, functools.partial): return f"partial({function.func.__name__})" return function.__name__ def py_prefetch( iterable_function: Callable[[], Iterable[chex.ArrayTree]], buffer_size: int = 5, ) -> Generator[chex.ArrayTree, None, None]: """Performs prefetching of elements from an iterable in a separate thread. Args: iterable_function: A python function that when called with no arguments returns an iterable. This is used to build a fresh iterable for each thread (crucial if working with tensorflow datasets because `tf.graph` objects are thread local). buffer_size (int): Number of elements to keep in the prefetch buffer. Yields: Prefetched elements from the original iterable. Raises: ValueError: if the buffer_size <= 1. Any error thrown by the iterable_function. Note this is not raised inside the producer, but after it finishes executing. """ if buffer_size <= 1: raise ValueError("the buffer_size should be > 1") buffer: queue.Queue = queue.Queue(maxsize=(buffer_size - 1)) producer_error = [] end = object() def producer() -> None: """Enqueues items from iterable on a given thread.""" try: # Build a new iterable for each thread. This is crucial if # working with tensorflow datasets # because tf.graph objects are thread local. iterable = iterable_function() for item in iterable: buffer.put(item) except Exception as err: # pylint: disable=broad-except logging.exception( "Error in producer thread for %s", get_function_name(iterable_function), ) producer_error.append(err) finally: buffer.put(end) threading.Thread(target=producer, daemon=True).start() # Consumer. while True: value = buffer.get() if value is end: break yield value if producer_error: raise producer_error[0]
11,835
32.247191
123
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/diffusion/variance_schedule.py
"""Variance schedule for diffusion models.""" from __future__ import annotations import enum import numpy as np from jax import numpy as jnp class DiffusionBetaSchedule(enum.Enum): """Class to define beta schedule.""" LINEAR = enum.auto() QUADRADIC = enum.auto() COSINE = enum.auto() WARMUP10 = enum.auto() WARMUP50 = enum.auto() def get_beta_schedule( num_timesteps: int, beta_schedule: DiffusionBetaSchedule, beta_start: float, beta_end: float, ) -> jnp.ndarray: """Get variance (beta) schedule for q(x_t | x_{t-1}). Args: num_timesteps: number of time steps in total, T. beta_schedule: schedule for beta. beta_start: beta for t=0. beta_end: beta for t=T-1. Returns: Shape (num_timesteps,) array of beta values, for t=0, ..., T-1. Values are in ascending order. Raises: ValueError: for unknown schedule. """ if beta_schedule == DiffusionBetaSchedule.LINEAR: return jnp.linspace( beta_start, beta_end, num_timesteps, ) if beta_schedule == DiffusionBetaSchedule.QUADRADIC: return ( jnp.linspace( beta_start**0.5, beta_end**0.5, num_timesteps, ) ** 2 ) if beta_schedule == DiffusionBetaSchedule.COSINE: def f(t: float) -> float: """Eq 17 in https://arxiv.org/abs/2102.09672. Args: t: time step with values in [0, 1]. Returns: Cumulative product of alpha. """ return np.cos((t + 0.008) / 1.008 * np.pi / 2) ** 2 betas = [0.0] alphas_cumprod_prev = 1.0 for i in range(1, num_timesteps): t = i / (num_timesteps - 1) alphas_cumprod = f(t) beta = 1 - alphas_cumprod / alphas_cumprod_prev betas.append(beta) return jnp.array(betas) * (beta_end - beta_start) + beta_start if beta_schedule == DiffusionBetaSchedule.WARMUP10: num_timesteps_warmup = max(num_timesteps // 10, 1) betas_warmup = ( jnp.linspace( beta_start**0.5, beta_end**0.5, num_timesteps_warmup, ) ** 2 ) return jnp.concatenate( [ betas_warmup, jnp.ones((num_timesteps - num_timesteps_warmup,)) * beta_end, ] ) if beta_schedule == DiffusionBetaSchedule.WARMUP50: num_timesteps_warmup = max(num_timesteps // 2, 1) betas_warmup = ( jnp.linspace( beta_start**0.5, beta_end**0.5, num_timesteps_warmup, ) ** 2 ) return jnp.concatenate( [ betas_warmup, jnp.ones((num_timesteps - num_timesteps_warmup,)) * beta_end, ] ) raise ValueError(f"Unknown beta_schedule {beta_schedule}.") def downsample_beta_schedule( betas: jnp.ndarray, num_timesteps: int, num_timesteps_to_keep: int, ) -> jnp.ndarray: """Downsample beta schedule. Args: betas: beta schedule, shape (num_timesteps,). Values are in ascending order. num_timesteps: number of time steps in total, T. num_timesteps_to_keep: number of time steps to keep. Returns: Downsampled beta schedule, shape (num_timesteps_to_keep,). """ if betas.shape != (num_timesteps,): raise ValueError( f"betas.shape ({betas.shape}) must be equal to " f"(num_timesteps,)=({num_timesteps},)" ) if (num_timesteps - 1) % (num_timesteps_to_keep - 1) != 0: raise ValueError( f"num_timesteps-1={num_timesteps-1} can't be evenly divided by " f"num_timesteps_to_keep-1={num_timesteps_to_keep-1}." ) if num_timesteps_to_keep < 2: raise ValueError( f"num_timesteps_to_keep ({num_timesteps_to_keep}) must be >= 2." ) if num_timesteps_to_keep == num_timesteps: return betas if num_timesteps_to_keep < num_timesteps: step_scale = (num_timesteps - 1) // (num_timesteps_to_keep - 1) beta0 = betas[0] alphas = 1.0 - betas alphas_cumprod = jnp.cumprod(alphas) # (num_timesteps_to_keep,) alphas_cumprod = alphas_cumprod[::step_scale] # (num_timesteps_to_keep-1,) betas = 1.0 - alphas_cumprod[1:] / alphas_cumprod[:-1] # (num_timesteps_to_keep,) betas = jnp.append(beta0, betas) return betas raise ValueError( f"num_timesteps_to_keep ({num_timesteps_to_keep}) " f"must be <= num_timesteps ({num_timesteps})" )
4,844
29.093168
77
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/diffusion/gaussian_diffusion.py
"""Gaussian diffusion related functions. https://github.com/WuJunde/MedSegDiff/blob/master/guided_diffusion/gaussian_diffusion.py https://github.com/hojonathanho/diffusion/blob/master/diffusion_tf/diffusion_utils_2.py """ import dataclasses import enum from typing import Callable, Iterator, Sequence, Tuple, Union import haiku as hk import jax.numpy as jnp import jax.random from imgx import EPS from imgx.diffusion.variance_schedule import ( DiffusionBetaSchedule, downsample_beta_schedule, get_beta_schedule, ) from imgx.metric.distribution import ( discretized_gaussian_log_likelihood, normal_kl, ) class DiffusionModelOutputType(enum.Enum): """Class to define model's output meaning. - X_START: model predicts x_0. - X_PREVIOUS: model predicts x_{t-1}. - EPSILON: model predicts noise epsilon. """ X_START = enum.auto() X_PREVIOUS = enum.auto() EPSILON = enum.auto() class DiffusionModelVarianceType(enum.Enum): r"""Class to define p(x_{t-1} | x_t) variance. - FIXED_SMALL: a smaller variance, \tilde{beta}_t = (1-\bar{alpha}_{t-1})/(1-\bar{alpha}_{t})*beta_t. - FIXED_LARGE: a larger variance, beta_t. - LEARNED: model outputs an array with channel=2, for mean and variance. - LEARNED_RANGE: model outputs an array with channel=2, for mean and variance. But the variance is not raw values, it's a coefficient to control the value between FIXED_SMALL and FIXED_LARGE. """ FIXED_SMALL = enum.auto() FIXED_LARGE = enum.auto() LEARNED = enum.auto() LEARNED_RANGE = enum.auto() class DiffusionSpace(enum.Enum): """Class to define the meaning of x. Model always outputs logits. """ SCALED_PROBS = enum.auto() # values will be [-1, 1] LOGITS = enum.auto() def extract_and_expand( arr: jnp.ndarray, t: jnp.ndarray, ndim: int ) -> jnp.ndarray: """Extract values from a 1D array and expand. This function is not jittable. Args: arr: 1D of shape (num_timesteps, ). t: storing index values < self.num_timesteps, shape (batch, ). ndim: number of dimensions for an array of shape (batch, ...). Returns: Expanded array of shape (batch, ...), expanded axes have dim 1. """ return jnp.expand_dims(arr[t], axis=tuple(range(1, ndim))) @dataclasses.dataclass class GaussianDiffusion(hk.Module): """Class for Gaussian diffusion sampling. https://github.com/WuJunde/MedSegDiff/blob/master/guided_diffusion/gaussian_diffusion.py TODO: split segmentation related functions to a sub-class. """ def __init__( self, model: hk.Module, num_timesteps: int, # T num_timesteps_beta: int, beta_schedule: DiffusionBetaSchedule, beta_start: float, beta_end: float, model_out_type: DiffusionModelOutputType, model_var_type: DiffusionModelVarianceType, x_space: DiffusionSpace, x_limit: float, use_ddim: bool, noise_fn: Callable = jax.random.normal, ) -> None: """Init. q(x_t | x_{t-1}) ~ Normal(sqrt(1-beta_t)*x_{t-1}, beta_t*I) Args: model: haiku model. num_timesteps: number of diffusion steps. num_timesteps_beta: number of steps when defining beta schedule. beta_schedule: schedule for betas. beta_start: beta for t=0. beta_end: beta for t=T. model_out_type: type of model output. model_var_type: type of variance for p(x_{t-1} | x_t). x_space: x is logits or scaled_probs. x_limit: x_t has values in [-x_limit, x_limit], the range has to be symmetric, as for T, the distribution is centered at zero. use_ddim: use ddim_sample. noise_fn: a function that gets noise of the same shape as x_t. """ super().__init__() self.model = model self.num_timesteps = num_timesteps self.num_timesteps_beta = num_timesteps_beta self.use_ddim = use_ddim self.model_out_type = model_out_type self.model_var_type = model_var_type self.x_space = x_space self.x_limit = x_limit self.noise_fn = noise_fn # shape are all (T,) # corresponding to 0, ..., T-1, where 0 means one step betas = get_beta_schedule( num_timesteps=num_timesteps_beta, beta_schedule=beta_schedule, beta_start=beta_start, beta_end=beta_end, ) self.betas = downsample_beta_schedule( betas=betas, num_timesteps=num_timesteps_beta, num_timesteps_to_keep=num_timesteps, ) alphas = 1.0 - self.betas # alpha_t self.alphas_cumprod = jnp.cumprod(alphas) # \bar{alpha}_t self.alphas_cumprod_prev = jnp.append(1.0, self.alphas_cumprod[:-1]) self.alphas_cumprod_next = jnp.append(self.alphas_cumprod[1:], 0.0) self.sqrt_alphas_cumprod = jnp.sqrt(self.alphas_cumprod) self.sqrt_one_minus_alphas_cumprod = jnp.sqrt(1.0 - self.alphas_cumprod) self.log_one_minus_alphas_cumprod = jnp.log(1.0 - self.alphas_cumprod) # last value is inf as last value of alphas_cumprod is zero self.sqrt_recip_alphas_cumprod = jnp.sqrt(1.0 / self.alphas_cumprod) self.sqrt_recip_alphas_cumprod_minus_one = jnp.sqrt( 1.0 / self.alphas_cumprod - 1 ) # q(x_{t-1} | x_t, x_0) # mean = coeff_start * x_0 + coeff_t * x_t # first values are nan self.posterior_mean_coeff_start = ( self.betas * jnp.sqrt(self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) ) self.posterior_mean_coeff_t = ( jnp.sqrt(alphas) * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) ) # variance # log calculation clipped because the posterior variance is 0 at t=0 # alphas_cumprod_prev has 1.0 appended in front self.posterior_variance = ( self.betas * (1.0 - self.alphas_cumprod_prev) / (1.0 - self.alphas_cumprod) ) # posterior_variance first value is zero self.posterior_log_variance_clipped = jnp.log( jnp.append(self.posterior_variance[1], self.posterior_variance[1:]) ) def q_mean_log_variance( self, x_start: jnp.ndarray, t: jnp.ndarray ) -> Tuple[jnp.ndarray, jnp.ndarray]: """Get the distribution q(x_t | x_0). Args: x_start: noiseless input, shape (batch, ...). t: storing index values < self.num_timesteps, shape (batch, ). Returns: mean: shape (batch, ...), expanded axes have dim 1. log_variance: shape (batch, ...), expanded axes have dim 1. """ mean = ( extract_and_expand(self.sqrt_alphas_cumprod, t, x_start.ndim) * x_start ) log_variance = extract_and_expand( self.log_one_minus_alphas_cumprod, t, x_start.ndim ) return mean, log_variance def q_sample( self, x_start: jnp.ndarray, noise: jnp.ndarray, t: jnp.ndarray, ) -> jnp.ndarray: """Sample from q(x_t | x_0). Args: x_start: noiseless input, shape (batch, ...). noise: same shape as x_start. t: storing index values < self.num_timesteps, shape (batch, ). Returns: Noisy array with same shape as x_start. """ mean = ( extract_and_expand(self.sqrt_alphas_cumprod, t, x_start.ndim) * x_start ) var = extract_and_expand( self.sqrt_one_minus_alphas_cumprod, t, x_start.ndim ) x_t = mean + var * noise x_t = self.clip_x(x_t) return x_t def q_posterior_mean( self, x_start: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray ) -> jnp.ndarray: """Get mean of the distribution q(x_{t-1} | x_t, x_0). Args: x_start: noiseless input, shape (batch, ...). x_t: noisy input, same shape as x_start. t: storing index values < self.num_timesteps, shape (batch, ). Returns: mean: same shape as x_start. """ return ( extract_and_expand(self.posterior_mean_coeff_start, t, x_start.ndim) * x_start + extract_and_expand(self.posterior_mean_coeff_t, t, x_start.ndim) * x_t ) def q_posterior_mean_variance( self, x_start: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray ) -> Tuple[jnp.ndarray, jnp.ndarray]: """Get the distribution q(x_{t-1} | x_t, x_0). Args: x_start: noiseless input, shape (batch, ...). x_t: noisy input, same shape as x_start. t: storing index values < self.num_timesteps, shape (batch, ). Returns: mean: same shape as x_start. log_variance: shape (batch, ...), expanded axes have dim 1. """ mean = self.q_posterior_mean(x_start, x_t, t) log_variance = extract_and_expand( self.posterior_log_variance_clipped, t, x_start.ndim ) return mean, log_variance def p_mean_variance( # pylint:disable=R0912 self, model_out: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray, ) -> Tuple[jnp.ndarray, jnp.ndarray, jnp.ndarray]: """Get the distribution p(x_{t-1} | x_t). Args: model_out: model predicted output. If model estimates variance, shape (batch, ..., 2*num_classes), else shape (batch, ..., num_classes). x_t: noisy input, shape (batch, ..., num_classes). t: storing index values < self.num_timesteps, shape (batch, ). Returns: x_start: predicted, same shape as x_t, values are clipped. mean: same shape as x_t. log_variance: compatible shape to (batch, ..., num_classes). """ # variance if self.model_var_type == DiffusionModelVarianceType.FIXED_SMALL: log_variance = self.posterior_log_variance_clipped # extend shape log_variance = extract_and_expand(log_variance, t, x_t.ndim) elif self.model_var_type == DiffusionModelVarianceType.FIXED_LARGE: # TODO why appending? variance = jnp.append(self.posterior_variance[1], self.betas[1:]) log_variance = jnp.log(variance) # extend shape log_variance = extract_and_expand(log_variance, t, x_t.ndim) elif self.model_var_type == DiffusionModelVarianceType.LEARNED: # model_out (batch, ..., num_classes*2) model_out, log_variance = jnp.split( model_out, indices_or_sections=2, axis=-1 ) elif self.model_var_type == DiffusionModelVarianceType.LEARNED_RANGE: # model_out (batch, ..., num_classes*2) model_out, var_coeff = jnp.split( model_out, indices_or_sections=2, axis=-1 ) log_min_variance = self.posterior_log_variance_clipped log_max_variance = jnp.log(self.betas) log_min_variance = extract_and_expand(log_min_variance, t, x_t.ndim) log_max_variance = extract_and_expand(log_max_variance, t, x_t.ndim) # var_coeff values are in [-1, 1] for [min_var, max_var]. var_coeff = jnp.clip(var_coeff, -1.0, 1.0) var_coeff = (var_coeff + 1) / 2 log_variance = ( var_coeff * log_max_variance + (1 - var_coeff) * log_min_variance ) else: raise ValueError( f"Unknown DiffusionModelVarianceType {self.model_var_type}." ) # mean if self.model_out_type == DiffusionModelOutputType.X_START: # q(x_{t-1} | x_t, x_0) x_start = self.logits_to_x(model_out) x_start = self.clip_x(x_start) mean = self.q_posterior_mean(x_start=x_start, x_t=x_t, t=t) elif self.model_out_type == DiffusionModelOutputType.X_PREVIOUS: # x_{t-1} x_prev = self.logits_to_x(model_out) x_prev = self.clip_x(x_prev) mean = x_prev x_start = self.predict_xstart_from_xprev_xt( x_prev=x_prev, x_t=x_t, t=t ) x_start = self.clip_x(x_start) elif self.model_out_type == DiffusionModelOutputType.EPSILON: x_start = self.predict_xstart_from_epsilon_xt( x_t=x_t, epsilon=model_out, t=t ) x_start = self.clip_x(x_start) mean = self.q_posterior_mean(x_start=x_start, x_t=x_t, t=t) else: raise ValueError( f"Unknown DiffusionModelOutputType {self.model_out_type}." ) return x_start, mean, log_variance def p_sample( self, model_out: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray, ) -> Tuple[jnp.ndarray, jnp.ndarray]: """Sample x_{t-1} ~ p(x_{t-1} | x_t). Args: model_out: model predicted output. If model estimates variance, shape (batch, ..., 2*num_classes), else shape (batch, ..., num_classes). x_t: noisy input, shape (batch, ..., num_classes). t: storing index values < self.num_timesteps, shape (batch, ). Returns: sample: x_{t-1}, same shape as x_t. x_start_pred: same shape as x_t. """ x_start_pred, mean, log_variance = self.p_mean_variance( model_out=model_out, x_t=x_t, t=t, ) noise = self.noise_sample(shape=x_t.shape, dtype=x_t.dtype) # no noise when t=0 # mean + exp(log(sigma**2)/2) * noise = mean + sigma * noise nonzero_mask = jnp.expand_dims( jnp.array(t != 0, dtype=noise.dtype), axis=tuple(range(1, noise.ndim)), ) sample = mean + nonzero_mask * jnp.exp(0.5 * log_variance) * noise # clip as the value may be out of range sample = self.clip_x(sample) return sample, x_start_pred def ddim_sample( self, model_out: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray, eta: float = 0.0, ) -> Tuple[jnp.ndarray, jnp.ndarray]: """Sample x_{t-1} ~ p(x_{t-1} | x_t). TODO: what are the differences between p_sample / ddim_sample? Args: model_out: model predicted output. If model estimates variance, shape (batch, ..., 2*num_classes), else shape (batch, ..., num_classes). x_t: noisy input, shape (batch, ..., num_classes). t: storing index values < self.num_timesteps, shape (batch, ). eta: control the noise level in sampling. Returns: sample: x_{t-1}, same shape as x_t. x_start_pred: same shape as x_t. """ # TODO why not using log variance output here? x_start_pred, _, _ = self.p_mean_variance( model_out=model_out, x_t=x_t, t=t, ) noise = self.noise_sample(shape=x_t.shape, dtype=x_t.dtype) epsilon = self.predict_epsilon_from_xstart_xt( x_t=x_t, x_start=x_start_pred, t=t ) alphas_cumprod_prev = extract_and_expand( self.alphas_cumprod_prev, t, x_t.ndim ) coeff_start = jnp.sqrt(alphas_cumprod_prev) log_variance = ( extract_and_expand(self.posterior_log_variance_clipped, t, x_t.ndim) * eta ) coeff_epsilon = jnp.sqrt(1.0 - alphas_cumprod_prev - log_variance**2) mean = coeff_start * x_start_pred + coeff_epsilon * epsilon nonzero_mask = jnp.expand_dims( jnp.array(t != 0, dtype=noise.dtype), axis=tuple(range(1, noise.ndim)), ) sample = mean + nonzero_mask * log_variance * noise # clip as the value may be out of range sample = self.clip_x(sample) return sample, x_start_pred def sample_mask( self, image: jnp.ndarray, x_t: jnp.ndarray, ) -> jnp.ndarray: """Generate segmentation mask from the model conditioned on image. The noise here is defined on segmentation mask. x_t is considered as logits. Args: image: image to be segmented, shape = (batch, ..., C). x_t: segmentation logits to be refined, shape = (batch, ..., num_classes). Returns: Sampled segmentation logits, shape = (batch, ..., num_classes). """ for x_start_t in self.sample_mask_progressive(image=image, x_t=x_t): x_start = x_start_t return x_start def sample_mask_progressive( self, image: jnp.ndarray, x_t: jnp.ndarray, ) -> Iterator[jnp.ndarray]: """Generate segmentation mask from the model conditioned on image. The noise here is defined on segmentation mask. x_t is considered as logits. Args: image: image to be segmented, shape = (batch, ..., C). x_t: segmentation logits to be refined, shape = (batch, ..., num_classes). Yields: x_start of shape = (batch, ..., num_classes, T). """ for t in reversed(range(self.num_timesteps)): # (batch, ) t_batch = jnp.array( [t] * x_t.shape[0], dtype=jnp.int16, ) # (batch, ..., ch_input + num_classes) model_in = jnp.concatenate([image, x_t], axis=-1) # (batch, ..., num_classes) or (batch, ..., 2*num_classes) model_out = self.model(model_in, t_batch) if self.use_ddim: x_t, x_start = self.ddim_sample( model_out=model_out, x_t=x_t, t=t_batch, ) else: x_t, x_start = self.p_sample( model_out=model_out, x_t=x_t, t=t_batch, ) yield x_start def predict_xstart_from_xprev_xt( self, x_prev: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray ) -> jnp.ndarray: """Get x_0 from x_{t-1} and x_t. The mean of q(x_{t-1} | x_t, x_0) is coeff_start * x_0 + coeff_t * x_t. So x_{t-1} = coeff_start * x_0 + coeff_t * x_t. x_0 = (x_{t-1} - coeff_t * x_t) / coeff_start = 1/coeff_start * x_{t-1} - coeff_t/coeff_start * x_t Args: x_prev: noisy input at t-1, shape (batch, ...). x_t: noisy input, same shape as x_prev. t: storing index values < self.num_timesteps, shape (batch, ). Returns: predicted x_0, same shape as x_prev. """ coeff_prev = extract_and_expand( 1.0 / self.posterior_mean_coeff_start, t, x_t.ndim ) coeff_t = extract_and_expand( self.posterior_mean_coeff_t / self.posterior_mean_coeff_start, t, x_t.ndim, ) return coeff_prev * x_prev - coeff_t * x_t def predict_xprev_from_xstart_xt( self, x_start: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray ) -> jnp.ndarray: """Get x_{t-1} from x_0 and x_t. The mean of q(x_{t-1} | x_t, x_0) is coeff_start * x_0 + coeff_t * x_t. So x_{t-1} = coeff_start * x_0 + coeff_t * x_t. Args: x_start: noisy input at t, shape (batch, ...). x_t: noisy input, same shape as x_start. t: storing index values < self.num_timesteps, shape (batch, ). Returns: predicted x_0, same shape as x_prev. """ coeff_start = extract_and_expand( self.posterior_mean_coeff_start, t, x_t.ndim ) coeff_t = extract_and_expand( self.posterior_mean_coeff_t, t, x_t.ndim, ) return coeff_start * x_start + coeff_t * x_t def sample_xprev_from_xstart_xt( self, x_start: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray ) -> jnp.ndarray: """Sample x_{t-1} from q(x_{t-1} | x_0, x_t). The mean of q(x_{t-1} | x_t, x_0) is coeff_start * x_0 + coeff_t * x_t. So x_{t-1} = coeff_start * x_0 + coeff_t * x_t. Args: x_start: noisy input at t, shape (batch, ...). x_t: noisy input, same shape as x_start. t: storing index values < self.num_timesteps, shape (batch, ). Returns: predicted x_0, same shape as x_prev. """ x_prev = self.predict_xprev_from_xstart_xt( x_start=x_start, x_t=x_t, t=t, ) noise = self.noise_sample(shape=x_t.shape, dtype=x_t.dtype) log_variance = extract_and_expand( self.posterior_log_variance_clipped, t, x_t.ndim ) sample = x_prev + noise * log_variance return self.clip_x(sample) def predict_xstart_from_epsilon_xt( self, x_t: jnp.ndarray, epsilon: jnp.ndarray, t: jnp.ndarray ) -> jnp.ndarray: """Get x_0 from epsilon. The reparameterization gives: x_t = sqrt(alphas_cumprod) * x_0 + sqrt(1-alphas_cumprod) * epsilon so, x_0 = 1/sqrt(alphas_cumprod) * x_t - sqrt(1-alphas_cumprod)/sqrt(alphas_cumprod) * epsilon = 1/sqrt(alphas_cumprod) * x_t - sqrt(1/alphas_cumprod - 1) * epsilon Args: x_t: noisy input at t-1, shape (batch, ...). epsilon: noise, shape (batch, ...), expanded axes have dim 1. t: storing index values < self.num_timesteps, shape (batch, ). Returns: predicted x_0, same shape as x_t. """ coeff_t = extract_and_expand( self.sqrt_recip_alphas_cumprod, t, x_t.ndim ) coeff_epsilon = extract_and_expand( self.sqrt_recip_alphas_cumprod_minus_one, t, x_t.ndim ) return coeff_t * x_t - coeff_epsilon * epsilon def predict_epsilon_from_xstart_xt( self, x_t: jnp.ndarray, x_start: jnp.ndarray, t: jnp.ndarray ) -> jnp.ndarray: """Get epsilon from x_0 and x_t. The reparameterization gives: x_t = sqrt(alphas_cumprod) * x_0 + sqrt(1-alphas_cumprod) * epsilon so, epsilon = (x_t - sqrt(alphas_cumprod) * x_0) / sqrt(1-alphas_cumprod) = (1/sqrt(alphas_cumprod) * x_t - x_0) /sqrt(1/alphas_cumprod-1) Args: x_t: noisy input at t-1, shape (batch, ...). x_start: predicted x_0, same shape as x_t. t: storing index values < self.num_timesteps, shape (batch, ). Returns: predicted x_0, same shape as x_t. """ coeff_t = extract_and_expand( self.sqrt_recip_alphas_cumprod, t, x_t.ndim ) denominator = extract_and_expand( self.sqrt_recip_alphas_cumprod_minus_one, t, x_t.ndim ) return (coeff_t * x_t - x_start) / denominator def sample_timestep( self, batch_size: int, min_val: Union[int, jnp.ndarray] = 0 ) -> jnp.ndarray: """Sample t of shape (batch, ). Define this function to avoid defining randon key. Args: batch_size: number of steps. min_val: minimum value, inclusive. Returns: Time steps with value between 0 and T-1, both sides inclusive. """ min_val = jnp.minimum(min_val, self.num_timesteps - 1) return jax.random.randint( hk.next_rng_key(), shape=(batch_size,), minval=min_val, # inclusive maxval=self.num_timesteps, # exclusive ) def noise_sample( self, shape: Sequence[int], dtype: jnp.dtype ) -> jnp.ndarray: """Return a noise of the same shape as input. Define this function to avoid defining randon key. Args: shape: array shape. dtype: data type. Returns: Noise of the same shape and dtype as x. """ return self.noise_fn(key=hk.next_rng_key(), shape=shape, dtype=dtype) def clip_x(self, x: jnp.ndarray) -> jnp.ndarray: """Clip the x_start/x_t to desired range. TODO: where should clip be used? Args: x: any array. Returns: Clipped array. """ if self.x_limit <= 0: return x return jnp.clip(x, -self.x_limit, self.x_limit) def logits_to_x(self, logits: jnp.ndarray) -> jnp.ndarray: """Map logits to x space. Args: logits: unnormalised logits. Returns: Array in the same space as x_start. """ if self.x_space == DiffusionSpace.LOGITS: return logits if self.x_space == DiffusionSpace.SCALED_PROBS: x = jax.nn.softmax(logits, axis=-1) x = x * 2.0 - 1.0 return x raise ValueError(f"Unknown x space {self.x_space}.") def x_to_logits(self, x: jnp.ndarray) -> jnp.ndarray: """Map x to logits. Args: x: in the same space as x_start. Returns: Logits. """ if self.x_space == DiffusionSpace.LOGITS: return x if self.x_space == DiffusionSpace.SCALED_PROBS: probs = (x + 1) / 2 probs = jnp.clip(probs, EPS, 1.0) return jnp.log(probs) raise ValueError(f"Unknown x space {self.x_space}.") def variational_lower_bound( self, model_out: jnp.ndarray, x_start: jnp.ndarray, x_t: jnp.ndarray, t: jnp.ndarray, ) -> jnp.ndarray: """Variational lower-bound, smaller is better. The resulting units are bits (rather than nats, as one might expect). This allows for comparison to other papers. Args: model_out: model predicted output, may present different things, shape (batch, ...). x_start: cleaned, same shape as x_t. x_t: noisy input, shape (batch, ...). t: storing index values < self.num_timesteps, shape (batch, ). Returns: lower bounds of shape (batch, ). """ reduce_axis = tuple(range(x_t.ndim))[1:] # q(x_{t-1} | x_t, x_0) q_mean, q_log_variance = self.q_posterior_mean_variance( x_start=x_start, x_t=x_t, t=t ) # p(x_{t-1} | x_t) _, p_mean, p_log_variance = self.p_mean_variance( model_out=model_out, x_t=x_t, t=t, ) kl = normal_kl( q_mean=q_mean, q_log_variance=q_log_variance, p_mean=p_mean, p_log_variance=p_log_variance, ) nll = -discretized_gaussian_log_likelihood( x_start, mean=q_mean, log_variance=q_log_variance ) # (batch, ) kl = jnp.mean(kl, axis=reduce_axis) / jnp.log(2.0) nll = jnp.mean(nll, axis=reduce_axis) / jnp.log(2.0) # return neg-log-likelihood for t = 0 return jnp.where(t == 0, nll, kl)
27,736
33.413151
92
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/loss/cross_entropy.py
"""Loss functions for classification.""" import jax import jax.numpy as jnp import optax def mean_cross_entropy( logits: jnp.ndarray, mask_true: jnp.ndarray, ) -> jnp.ndarray: """Cross entropy. Args: logits: unscaled prediction, (batch, ..., num_classes). mask_true: one hot targets, (batch, ..., num_classes). Returns: Cross entropy loss value of shape (1, ). """ # (batch, ...) loss = optax.softmax_cross_entropy(logits=logits, labels=mask_true) return jnp.mean(loss) def mean_focal_loss( logits: jnp.ndarray, mask_true: jnp.ndarray, gamma: float = 2.0, ) -> jnp.ndarray: """Focal loss. https://arxiv.org/abs/1708.02002 Args: logits: unscaled prediction, (batch, ..., num_classes). mask_true: one hot targets, (batch, ..., num_classes). gamma: adjust class imbalance, 0 is equivalent to cross entropy. Returns: Focal loss value of shape (1, ). """ # normalise logits to be the log of probabilities logits = jax.nn.log_softmax(logits, axis=-1) probs = jnp.exp(logits) focal_loss = -((1 - probs) ** gamma) * logits * mask_true # (batch, ..., num_classes) -> (batch, ...) # label are one hot, just sum over class axis focal_loss = jnp.sum(focal_loss, axis=-1) return jnp.mean(focal_loss)
1,352
26.06
72
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/loss/dice.py
"""Loss functions for image segmentation.""" import jax import jax.numpy as jnp def mean_with_background(batch_cls_loss: jnp.ndarray) -> jnp.ndarray: """Return average with background class. Args: batch_cls_loss: shape (batch, num_classes). Returns: Mean loss of shape (1,). """ return jnp.nanmean(batch_cls_loss) def mean_without_background(batch_cls_loss: jnp.ndarray) -> jnp.ndarray: """Return average without background class. Args: batch_cls_loss: shape (batch, num_classes). Returns: Mean loss of shape (1,). """ return jnp.nanmean(batch_cls_loss[:, 1:]) def dice_loss( logits: jnp.ndarray, mask_true: jnp.ndarray, ) -> jnp.ndarray: """Mean dice loss, smaller is better. Losses are not calculated on instance-classes, where there is no label. This is to avoid the need of smoothing and potentially nan gradients. Args: logits: unscaled prediction, (batch, ..., num_classes). mask_true: one hot targets, (batch, ..., num_classes). Returns: Dice loss value of shape (batch, num_classes). """ mask_pred = jax.nn.softmax(logits) reduce_axis = tuple(range(mask_pred.ndim))[1:-1] # (batch, num_classes) numerator = 2.0 * jnp.sum(mask_pred * mask_true, axis=reduce_axis) denominator = jnp.sum(mask_pred + mask_true, axis=reduce_axis) not_nan_mask = jnp.sum(mask_true, axis=reduce_axis) > 0 # nan loss are replaced by 0.0 return jnp.where( condition=not_nan_mask, x=1.0 - numerator / denominator, y=jnp.nan, ) def mean_dice_loss( logits: jnp.ndarray, mask_true: jnp.ndarray, include_background: bool, ) -> jnp.ndarray: """Mean dice loss, smaller is better. Losses are not calculated on instance-classes, where there is no label. This is to avoid the need of smoothing and potentially nan gradients. Args: logits: unscaled prediction, (batch, ..., num_classes). mask_true: one hot targets, (batch, ..., num_classes). include_background: include background as a separate class. Returns: Dice loss value of shape (1, ). """ loss = dice_loss(logits=logits, mask_true=mask_true) return jax.lax.cond( include_background, mean_with_background, mean_without_background, loss, )
2,386
27.082353
75
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/model/unet_3d_slice_time.py
"""UNet for segmentation.""" import dataclasses from typing import Callable, List, Tuple import haiku as hk import jax from jax import numpy as jnp from imgx.model.basic import instance_norm, sinusoidal_positional_embedding from imgx.model.unet_3d_slice import Conv2dNormAct, Conv2dPool @dataclasses.dataclass class TimeConv2dResBlock(hk.Module): """Conv2dResBlock with time embedding input. This class is defined separately to use remat, as remat does not allow condition loop (if / else). https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/model.py """ out_channels: int kernel_size: int activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.gelu def __call__( self, x: jnp.ndarray, t: jnp.ndarray, ) -> jnp.ndarray: """Forward pass. Args: x: tensor to be up-sampled, (batch, w, h, in_channels). t: time embedding, (batch, t_channels). Returns: Tensor. """ res = x x = hk.Conv2D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x) x = hk.Conv2D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) t = self.activation(t[:, None, None, :]) t = hk.Linear(output_size=self.out_channels)(t) x += t x = instance_norm(x) x = self.activation(x + res) return x @dataclasses.dataclass class Unet3dSliceTime(hk.Module): """2D UNet for 3D images. https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/nets/basic_unet.py """ in_shape: Tuple[int, int, int] # spatial shape in_channels: int # input channels out_channels: int num_channels: Tuple[int, ...] # channel at each depth, including the bottom num_timesteps: int # T kernel_size: int = 3 scale_factor: int = 2 # spatial down-sampling/up-sampling remat: bool = False # remat reduces memory cost at cost of compute speed def encoder( self, image: jnp.ndarray, t: jnp.ndarray, ) -> List[jnp.ndarray]: """Encoder the image. Args: image: image tensor of shape (batch, H, W, in_channels). t: time embedding of shape (batch, t_channels). Returns: List of embeddings from each layer. """ conv = Conv2dNormAct( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ) conv = hk.remat(conv) if self.remat else conv emb = conv(image) conv_t = TimeConv2dResBlock( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ) conv_t = hk.remat(conv_t) if self.remat else conv_t emb = conv_t(x=emb, t=t) embeddings = [emb] for ch in self.num_channels: conv = Conv2dPool(out_channels=ch, scale_factor=self.scale_factor) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) conv_t = TimeConv2dResBlock( out_channels=ch, kernel_size=self.kernel_size, ) conv_t = hk.remat(conv_t) if self.remat else conv_t emb = conv_t(x=emb, t=t) embeddings.append(emb) return embeddings def decoder( self, embeddings: List[jnp.ndarray], t: jnp.ndarray, ) -> jnp.ndarray: """Decode the embedding and perform prediction. Args: embeddings: list of embeddings from each layer. Starting with the first layer. t: time embedding of shape (batch, t_channels). Returns: Unnormalized logits. """ if len(embeddings) != len(self.num_channels) + 1: raise ValueError("UNet decoder input length does not match") emb = embeddings[-1] # calculate up-sampled channel # [32, 64, 128, 256] -> [32, 32, 64, 128] channels = self.num_channels[:1] + self.num_channels[:-1] for i, ch in enumerate(channels[::-1]): # skipped.shape <= up-scaled shape # as padding may be added when down-sampling skipped = embeddings[-i - 2] skipped_shape = skipped.shape[-3:-1] # deconv and pad to make emb of same shape as skipped conv = hk.Conv2DTranspose( output_channels=ch, kernel_shape=self.scale_factor, stride=self.scale_factor, ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) emb = emb[ ..., : skipped_shape[0], : skipped_shape[1], :, ] # add skipped emb += skipped # conv conv_t = TimeConv2dResBlock( out_channels=ch, kernel_size=self.kernel_size, ) conv_t = hk.remat(conv_t) if self.remat else conv_t emb = conv_t(emb, t) conv = hk.Conv2D(output_channels=self.out_channels, kernel_shape=1) conv = hk.remat(conv) if self.remat else conv out = conv(emb) return out def __call__( # type: ignore[no-untyped-def] self, image: jnp.ndarray, t: jnp.ndarray, **kwargs, # noqa: ARG002 ) -> jnp.ndarray: """Forward pass. Args: image: (batch, h, w, d, in_channels). t: (batch, ). kwargs: unused arguments. Returns: Predictions (batch, h, w, d, out_channels). Raises: ValueError: if input shape does not match. """ if image.shape[-4:] != (*self.in_shape, self.in_channels): raise ValueError( f"Input shape {image.shape[-4:]} does not match" f" configs {(*self.in_shape, self.in_channels)}" ) # (batch, h, w, d, in_channels) -> (batch, d, h, w, in_channels) image = jnp.transpose(image, (0, 3, 1, 2, 4)) # (batch, d, h, w, in_channels) -> (batch*d, h, w, in_channels) image = jnp.reshape(image, (-1, *self.in_shape[:2], self.in_channels)) # (batch, ) -> (batch*d,) t = jnp.repeat(t, repeats=self.in_shape[2], axis=0) dim_t = self.num_channels[0] * 4 t = sinusoidal_positional_embedding(x=t, dim=dim_t) embeddings = self.encoder(image=image, t=t) out = self.decoder(embeddings=embeddings, t=t) # (batch*d, h, w, out_channels) -> (batch, d, h, w, out_channels) out = jnp.reshape( out, (-1, self.in_shape[2], *self.in_shape[:2], self.out_channels) ) # (batch, d, h, w, out_channels) -> (batch, h, w, d, out_channels) out = jnp.transpose(out, (0, 2, 3, 1, 4)) return out
7,105
31.153846
95
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/model/unet_3d_time.py
"""UNet for segmentation.""" import dataclasses from typing import Callable, List, Tuple import haiku as hk import jax from jax import numpy as jnp from imgx.model.basic import instance_norm, sinusoidal_positional_embedding from imgx.model.unet_3d import Conv3dNormAct, Conv3dPool @dataclasses.dataclass class TimeConv3dResBlock(hk.Module): """Conv3dResBlock with time embedding input. This class is defined separately to use remat, as remat does not allow condition loop (if / else). https://github.com/CompVis/stable-diffusion/blob/main/ldm/modules/diffusionmodules/model.py """ out_channels: int kernel_size: int activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.gelu def __call__( self, x: jnp.ndarray, t: jnp.ndarray, ) -> jnp.ndarray: """Forward pass. Args: x: tensor to be up-sampled, (batch, w, h, d, in_channels). t: time embedding, (batch, t_channels). Returns: Tensor. """ res = x x = hk.Conv3D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x) x = hk.Conv3D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) t = self.activation(t[:, None, None, None, :]) t = hk.Linear(output_size=self.out_channels)(t) x += t x = instance_norm(x) x = self.activation(x + res) return x @dataclasses.dataclass class Unet3dTime(hk.Module): """3D UNet. https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/nets/basic_unet.py """ in_shape: Tuple[int, int, int] # spatial shape in_channels: int # input channels out_channels: int num_channels: Tuple[int, ...] # channel at each depth, including the bottom num_timesteps: int # T kernel_size: int = 3 scale_factor: int = 2 # spatial down-sampling/up-sampling remat: bool = False # remat reduces memory cost at cost of compute speed def encoder( self, image: jnp.ndarray, t: jnp.ndarray, ) -> List[jnp.ndarray]: """Encoder the image. Args: image: image tensor of shape (batch, H, W, D, in_channels). t: time embedding of shape (batch, t_channels). Returns: List of embeddings from each layer. """ conv = Conv3dNormAct( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ) conv = hk.remat(conv) if self.remat else conv emb = conv(image) conv_t = TimeConv3dResBlock( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ) conv_t = hk.remat(conv_t) if self.remat else conv_t emb = conv_t(x=emb, t=t) embeddings = [emb] for ch in self.num_channels: conv = Conv3dPool(out_channels=ch, scale_factor=self.scale_factor) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) conv_t = TimeConv3dResBlock( out_channels=ch, kernel_size=self.kernel_size, ) conv_t = hk.remat(conv_t) if self.remat else conv_t emb = conv_t(x=emb, t=t) embeddings.append(emb) return embeddings def decoder( self, embeddings: List[jnp.ndarray], t: jnp.ndarray, ) -> jnp.ndarray: """Decode the embedding and perform prediction. Args: embeddings: list of embeddings from each layer. Starting with the first layer. t: time embedding of shape (batch, t_channels). Returns: Unnormalized logits. """ if len(embeddings) != len(self.num_channels) + 1: raise ValueError("UNet decoder input length does not match") emb = embeddings[-1] # calculate up-sampled channel # [32, 64, 128, 256] -> [32, 32, 64, 128] channels = self.num_channels[:1] + self.num_channels[:-1] for i, ch in enumerate(channels[::-1]): # skipped.shape <= up-scaled shape # as padding may be added when down-sampling skipped = embeddings[-i - 2] skipped_shape = skipped.shape[-4:-1] # deconv and pad to make emb of same shape as skipped conv = hk.Conv3DTranspose( output_channels=ch, kernel_shape=self.scale_factor, stride=self.scale_factor, ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) emb = emb[ ..., : skipped_shape[0], : skipped_shape[1], : skipped_shape[2], :, ] # add skipped emb += skipped # conv conv_t = TimeConv3dResBlock( out_channels=ch, kernel_size=self.kernel_size, ) conv_t = hk.remat(conv_t) if self.remat else conv_t emb = conv_t(emb, t) conv = hk.Conv3D(output_channels=self.out_channels, kernel_shape=1) conv = hk.remat(conv) if self.remat else conv out = conv(emb) return out def __call__( # type: ignore[no-untyped-def] self, image: jnp.ndarray, t: jnp.ndarray, **kwargs, # noqa: ARG002 ) -> jnp.ndarray: """Forward pass. Args: image: (batch, h, w, d, in_channels). t: (batch, ). kwargs: unused arguments. Returns: Predictions (batch, h, w, d, out_channels). Raises: ValueError: if input shape does not match. """ if image.shape[-4:] != (*self.in_shape, self.in_channels): raise ValueError( f"Input shape {image.shape[-4:]} does not match" f" configs {(*self.in_shape, self.in_channels)}" ) dim_t = self.num_channels[0] * 4 t = sinusoidal_positional_embedding(x=t, dim=dim_t) embeddings = self.encoder(image=image, t=t) out = self.decoder(embeddings=embeddings, t=t) return out
6,438
30.10628
95
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/model/unet_3d_slice.py
"""UNet for segmentation.""" import dataclasses from typing import Callable, List, Tuple import haiku as hk import jax from jax import numpy as jnp from imgx.model.basic import instance_norm @dataclasses.dataclass class Conv2dNormAct(hk.Module): """Block with conv2d-norm-act.""" out_channels: int kernel_size: int activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.gelu def __call__( self, x: jnp.ndarray, ) -> jnp.ndarray: """Forward pass. Args: x: tensor to be up-sampled. Returns: Tensor. """ x = hk.Conv2D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x) return x @dataclasses.dataclass class Conv2dResBlock(hk.Module): """Block with two conv2d-norm-act layers and residual link.""" out_channels: int kernel_size: int activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.gelu def __call__( self, x: jnp.ndarray, ) -> jnp.ndarray: """Forward pass. Args: x: tensor to be up-sampled. Returns: Tensor. """ res = x x = hk.Conv2D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x) x = hk.Conv2D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x + res) return x @dataclasses.dataclass class Conv2dPool(hk.Module): """Patch merging, a down-sample layer.""" out_channels: int scale_factor: int def __call__( self, x: jnp.ndarray, ) -> jnp.ndarray: """Forward. Args: x: shape (batch, h, w, d, in_channels). Returns: Down-sampled array. """ x = hk.Conv2D( output_channels=self.out_channels, kernel_shape=self.scale_factor, stride=self.scale_factor, with_bias=False, )(x) x = instance_norm(x) return x @dataclasses.dataclass class Unet3dSlice(hk.Module): """2D UNet for 3D images. https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/nets/basic_unet.py """ in_shape: Tuple[int, int, int] # spatial shape in_channels: int # input channels out_channels: int num_channels: Tuple[int, ...] # channel at each depth, including the bottom kernel_size: int = 3 scale_factor: int = 2 # spatial down-sampling/up-sampling remat: bool = False # remat reduces memory cost at cost of compute speed def encoder( self, image: jnp.ndarray, ) -> List[jnp.ndarray]: """Encoder the image. Args: image: image tensor of shape (batch, H, W, C). Returns: List of embeddings from each layer. """ conv = hk.Sequential( [ Conv2dNormAct( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ), Conv2dResBlock( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ), ] ) conv = hk.remat(conv) if self.remat else conv emb = conv(image) embeddings = [emb] for ch in self.num_channels: conv = hk.Sequential( [ Conv2dPool(out_channels=ch, scale_factor=self.scale_factor), Conv2dResBlock( out_channels=ch, kernel_size=self.kernel_size, ), ] ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) embeddings.append(emb) return embeddings def decoder( self, embeddings: List[jnp.ndarray], ) -> jnp.ndarray: """Decode the embedding and perform prediction. Args: embeddings: list of embeddings from each layer. Starting with the first layer. Returns: Unnormalized logits. """ if len(embeddings) != len(self.num_channels) + 1: raise ValueError("UNet decoder input length does not match") emb = embeddings[-1] # calculate up-sampled channel # [32, 64, 128, 256] -> [32, 32, 64, 128] channels = self.num_channels[:1] + self.num_channels[:-1] for i, ch in enumerate(channels[::-1]): # skipped.shape <= up-scaled shape # as padding may be added when down-sampling skipped = embeddings[-i - 2] skipped_shape = skipped.shape[-3:-1] # deconv and pad to make emb of same shape as skipped conv = hk.Conv2DTranspose( output_channels=ch, kernel_shape=self.scale_factor, stride=self.scale_factor, ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) emb = emb[ ..., : skipped_shape[0], : skipped_shape[1], :, ] # add skipped emb += skipped # conv conv = Conv2dResBlock( out_channels=ch, kernel_size=self.kernel_size, ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) conv = hk.Conv2D(output_channels=self.out_channels, kernel_shape=1) conv = hk.remat(conv) if self.remat else conv out = conv(emb) return out def __call__( # type: ignore[no-untyped-def] self, image: jnp.ndarray, **kwargs, # noqa: ARG002 ) -> jnp.ndarray: """Forward pass. Args: image: (batch, h, w, d, in_channels). kwargs: unused arguments. Returns: Predictions (batch, h, w, d, out_channels). Raises: ValueError: if input shape does not match. """ if image.shape[-4:] != (*self.in_shape, self.in_channels): raise ValueError( f"Input shape {image.shape[-3:]} does not match" f" configs {(*self.in_shape, self.in_channels)}" ) # (batch, h, w, d, in_channels) -> (batch, d, h, w, in_channels) image = jnp.transpose(image, (0, 3, 1, 2, 4)) # (batch, d, h, w, in_channels) -> (batch*d, h, w, in_channels) image = jnp.reshape(image, (-1, *self.in_shape[:2], self.in_channels)) embeddings = self.encoder(image=image) out = self.decoder(embeddings=embeddings) # (batch*d, h, w, out_channels) -> (batch, d, h, w, out_channels) out = jnp.reshape( out, (-1, self.in_shape[2], *self.in_shape[:2], self.out_channels) ) # (batch, d, h, w, out_channels) -> (batch, h, w, d, out_channels) out = jnp.transpose(out, (0, 2, 3, 1, 4)) return out
7,387
27.635659
85
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/model/unet_3d.py
"""UNet for segmentation.""" import dataclasses from typing import Callable, List, Tuple import haiku as hk import jax from jax import numpy as jnp from imgx.model.basic import instance_norm @dataclasses.dataclass class Conv3dNormAct(hk.Module): """Block with conv3d-norm-act.""" out_channels: int kernel_size: int activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.gelu def __call__( self, x: jnp.ndarray, ) -> jnp.ndarray: """Forward pass. Args: x: tensor to be up-sampled. Returns: Tensor. """ x = hk.Conv3D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x) return x @dataclasses.dataclass class Conv3dResBlock(hk.Module): """Block with two conv3d-norm-act layers and residual link.""" out_channels: int kernel_size: int activation: Callable[[jnp.ndarray], jnp.ndarray] = jax.nn.gelu def __call__( self, x: jnp.ndarray, ) -> jnp.ndarray: """Forward pass. Args: x: tensor to be up-sampled. Returns: Tensor. """ res = x x = hk.Conv3D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x) x = hk.Conv3D( output_channels=self.out_channels, kernel_shape=self.kernel_size, with_bias=False, )(x) x = instance_norm(x) x = self.activation(x + res) return x @dataclasses.dataclass class Conv3dPool(hk.Module): """Patch merging, a down-sample layer.""" out_channels: int scale_factor: int def __call__( self, x: jnp.ndarray, ) -> jnp.ndarray: """Forward. Args: x: shape (batch, h, w, d, in_channels). Returns: Down-sampled array. """ x = hk.Conv3D( output_channels=self.out_channels, kernel_shape=self.scale_factor, stride=self.scale_factor, with_bias=False, )(x) x = instance_norm(x) return x @dataclasses.dataclass class Unet3d(hk.Module): """3D UNet. https://github.com/Project-MONAI/MONAI/blob/dev/monai/networks/nets/basic_unet.py """ in_shape: Tuple[int, int, int] # spatial shape in_channels: int # input channels out_channels: int num_channels: Tuple[int, ...] # channel at each depth, including the bottom kernel_size: int = 3 scale_factor: int = 2 # spatial down-sampling/up-sampling remat: bool = False # remat reduces memory cost at cost of compute speed def encoder( self, image: jnp.ndarray, ) -> List[jnp.ndarray]: """Encoder the image. Args: image: image tensor of shape (batch, H, W, D, C). Returns: List of embeddings from each layer. """ conv = hk.Sequential( [ Conv3dNormAct( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ), Conv3dResBlock( out_channels=self.num_channels[0], kernel_size=self.kernel_size, ), ] ) conv = hk.remat(conv) if self.remat else conv emb = conv(image) embeddings = [emb] for ch in self.num_channels: conv = hk.Sequential( [ Conv3dPool(out_channels=ch, scale_factor=self.scale_factor), Conv3dResBlock( out_channels=ch, kernel_size=self.kernel_size, ), ] ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) embeddings.append(emb) return embeddings def decoder( self, embeddings: List[jnp.ndarray], ) -> jnp.ndarray: """Decode the embedding and perform prediction. Args: embeddings: list of embeddings from each layer. Starting with the first layer. Returns: Unnormalized logits. """ if len(embeddings) != len(self.num_channels) + 1: raise ValueError("UNet decoder input length does not match") emb = embeddings[-1] # calculate up-sampled channel # [32, 64, 128, 256] -> [32, 32, 64, 128] channels = self.num_channels[:1] + self.num_channels[:-1] for i, ch in enumerate(channels[::-1]): # skipped.shape <= up-scaled shape # as padding may be added when down-sampling skipped = embeddings[-i - 2] skipped_shape = skipped.shape[-4:-1] # deconv and pad to make emb of same shape as skipped conv = hk.Conv3DTranspose( output_channels=ch, kernel_shape=self.scale_factor, stride=self.scale_factor, ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) emb = emb[ ..., : skipped_shape[0], : skipped_shape[1], : skipped_shape[2], :, ] # add skipped emb += skipped # conv conv = Conv3dResBlock( out_channels=ch, kernel_size=self.kernel_size, ) conv = hk.remat(conv) if self.remat else conv emb = conv(emb) conv = hk.Conv3D(output_channels=self.out_channels, kernel_shape=1) conv = hk.remat(conv) if self.remat else conv out = conv(emb) return out def __call__( # type: ignore[no-untyped-def] self, image: jnp.ndarray, **kwargs, # noqa: ARG002 ) -> jnp.ndarray: """Forward pass. Args: image: (batch, h, w, d, in_channels). kwargs: unused arguments. Returns: Predictions (batch, h, w, d, out_channels). Raises: ValueError: if input shape does not match. """ if image.shape[-4:] != (*self.in_shape, self.in_channels): raise ValueError( f"Input shape {image.shape[-3:]} does not match" f" configs {(*self.in_shape, self.in_channels)}" ) embeddings = self.encoder(image=image) out = self.decoder(embeddings=embeddings) return out
6,811
26.691057
85
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/model/basic.py
"""Basic functions and modules.""" import haiku as hk from jax import numpy as jnp def layer_norm(x: jnp.ndarray) -> jnp.ndarray: """Applies a unique LayerNorm at the last axis. Args: x: input Returns: Normalised input. """ return hk.LayerNorm(axis=-1, create_scale=True, create_offset=True)(x) def instance_norm(x: jnp.ndarray) -> jnp.ndarray: """Applies a unique InstanceNorm. Args: x: input Returns: Normalised input. """ return hk.InstanceNorm(create_scale=True, create_offset=True)(x) def dropout(x: jnp.ndarray, dropout_rate: float) -> jnp.ndarray: """Applies dropout only if necessary. This function is necessary to avoid defining random key for testing. Otherwise, calling `hk.dropout` will result in the following error: You must pass a non-None PRNGKey to init and/or apply if you make use of random numbers. Args: x: input dropout_rate: rate of dropout Returns: Dropout applied input. """ if dropout_rate == 0.0: # noqa: PLR2004 return x return hk.dropout(hk.next_rng_key(), dropout_rate, x) def sinusoidal_positional_embedding( x: jnp.ndarray, dim: int, max_period: int = 10000 ) -> jnp.ndarray: """Create sinusoidal timestep embeddings. Half defined by sin, half by cos. For position x, the embeddings are (for i = 0,...,half_dim-1) sin(x / (max_period ** (i/half_dim))) cos(x / (max_period ** (i/half_dim))) Args: x: (batch, ), with non-negative values. dim: embedding dimension, assume to be evenly divided by two. max_period: controls the minimum frequency of the embeddings. Returns: Embedding of size (batch, dim). """ half_dim = dim // 2 # (half_dim,) freq = jnp.arange(0, half_dim, dtype=jnp.float32) freq = jnp.exp(-jnp.log(max_period) * freq / half_dim) # (batch, half_dim) args = x[:, None] * freq[None, :] # (batch, dim) return jnp.concatenate([jnp.cos(args), jnp.sin(args)], axis=-1)
2,088
26.486842
74
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/exp/train_state.py
"""Training state and checkpoints.""" import pickle from pathlib import Path from typing import Optional, Tuple import chex import haiku as hk import jax import jax.numpy as jnp import jmp import numpy as np import optax from imgx.device import broadcast_to_local_devices, get_first_replica_values CHECKPOINT_ATTRS = [ "params", "network_state", "opt_state", "ema_network_state", "ema_params", ] @chex.dataclass class TrainState: """Dataclass to keep track of state of training. The state of training is structured as a chex.dataclass, which enables instances of this class to be passed into jax transformations like tree_map and pmap. The stored values are broadcast across devices. """ params: hk.Params network_state: hk.State opt_state: optax.OptState loss_scale: jmp.LossScale global_step: jnp.array rng: jax.random.PRNGKey ema_params: Optional[hk.Params] = None ema_network_state: Optional[hk.State] = None def save_array_tree(ckpt_dir: Path, state: chex.ArrayTree) -> None: """Save the state with arrays and tree saved separately. Args: ckpt_dir: directory to save. state: state to save, including params, optimizer, etc. """ ckpt_dir.mkdir(parents=True, exist_ok=True) with open(ckpt_dir / "arrays.npy", "wb") as f: for x in jax.tree_util.tree_leaves(state): np.save(f, x, allow_pickle=False) tree_struct = jax.tree_map(lambda _: 0, state) with open(ckpt_dir / "tree.pkl", "wb") as f: pickle.dump(tree_struct, f) def restore_array_tree(ckpt_dir: Path) -> chex.ArrayTree: """Restore the state from saved files. Args: ckpt_dir: directory to load. Returns: Restored state, including params, optimizer, etc. """ with open(ckpt_dir / "tree.pkl", "rb") as f: tree_struct = pickle.load(f) leaves, treedef = jax.tree_util.tree_flatten(tree_struct) with open(ckpt_dir / "arrays.npy", "rb") as f: flat_state = [np.load(f) for _ in leaves] return jax.tree_util.tree_unflatten(treedef, flat_state) def save_ckpt(train_state: TrainState, ckpt_dir: Path) -> None: """Save the state with arrays and tree saved separately. Args: train_state: checkpoint to save. ckpt_dir: directory to save. """ train_state = jax.tree_map(get_first_replica_values, train_state) state_dict = dict(train_state) # type: ignore[call-overload] # loss_scale needs to be stored differently loss_scale = state_dict["loss_scale"] del state_dict["loss_scale"] loss_scale_type = loss_scale.__class__.__name__ state_dict["loss_scale_type"] = loss_scale_type if isinstance(loss_scale, jmp.StaticLossScale): state_dict["loss_scale"] = loss_scale.loss_scale elif isinstance(loss_scale, jmp.DynamicLossScale): state_dict["loss_scale"] = loss_scale.loss_scale state_dict["loss_scale_counter"] = loss_scale.counter state_dict["loss_scale_period"] = loss_scale.period state_dict["loss_scale_factor"] = loss_scale.factor save_array_tree(ckpt_dir=ckpt_dir, state=state_dict) def restore_ckpt(ckpt_dir: Path) -> TrainState: """Restore the state from saved files. Args: ckpt_dir: directory to load. Returns: train_state: checkpoint to save. global_step: number of batch consumed. """ state_dict = restore_array_tree(ckpt_dir) # loss_scale needs to be loaded differently loss_scale_type = state_dict["loss_scale_type"] del state_dict["loss_scale_type"] if loss_scale_type == "NoOpLossScale": loss_scale = jmp.NoOpLossScale() elif loss_scale_type == "StaticLossScale": loss_scale = state_dict["loss_scale"] del state_dict["loss_scale"] loss_scale = jmp.StaticLossScale(loss_scale) elif loss_scale_type == "DynamicLossScale": loss_scale = state_dict["loss_scale"] counter = state_dict["loss_scale_counter"] # factor and period are ints not arrays period = int(state_dict["loss_scale_period"]) factor = int(state_dict["loss_scale_factor"]) del state_dict["loss_scale"] del state_dict["loss_scale_counter"] del state_dict["loss_scale_period"] del state_dict["loss_scale_factor"] loss_scale = jmp.DynamicLossScale( loss_scale=loss_scale, counter=counter, period=period, factor=factor ) else: raise ValueError(f"Unknown loss_scale type {loss_scale_type}.") # TODO should consider shards state_dict = jax.tree_map(broadcast_to_local_devices, state_dict) train_state = TrainState( # type: ignore[call-arg] params=state_dict["params"], network_state=state_dict["network_state"], opt_state=state_dict["opt_state"], loss_scale=loss_scale, global_step=state_dict["global_step"], rng=state_dict["rng"], ema_params=state_dict.get("ema_params", None), ema_network_state=state_dict.get("ema_network_state", None), ) return train_state def get_eval_params_and_state( train_state: TrainState, ) -> Tuple[hk.Params, hk.State]: """Get the parameters and state for evaluation. Args: train_state: Train State. Returns: params, state. """ if train_state.ema_params is not None: params = train_state.ema_params state = train_state.ema_network_state else: params = train_state.params state = train_state.network_state return get_first_replica_values(params), get_first_replica_values(state) def get_eval_params_and_state_from_ckpt( ckpt_dir: Path, use_ema: bool, ) -> Tuple[hk.Params, hk.State]: """Get the parameters and state for evaluation from checkpoint. Args: ckpt_dir: directory to load. use_ema: use EMA or not. Returns: Broadcast params, state. """ state_dict = restore_array_tree(ckpt_dir) if use_ema: params = state_dict["ema_params"] state = state_dict["ema_network_state"] else: params = state_dict["params"] state = state_dict["network_state"] # make sure arrays are initialised in CPU with jax.default_device(jax.devices("cpu")[0]): params = jax.tree_map(jnp.asarray, params) state = jax.tree_map(jnp.asarray, state) # broadcast to other devices params = broadcast_to_local_devices(params) state = broadcast_to_local_devices(state) return params, state
6,560
30.695652
80
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/exp/optim.py
"""Module for optimization.""" import logging from typing import Tuple import jax import jax.numpy as jnp import optax from omegaconf import DictConfig def ema_update( ema_value: jnp.ndarray, current_value: jnp.ndarray, decay: float, step: jnp.ndarray, ) -> jnp.ndarray: """Implements exponential moving average (EMA) with TF1-style decay warmup. Args: ema_value: previous value. current_value: current value. decay: decay ratio. step: number of steps so far. Returns: updated value. """ decay = jnp.minimum(decay, (1.0 + step) / (10.0 + step)) return ema_value * decay + current_value * (1 - decay) def get_lr_schedule(config: DictConfig) -> optax.Schedule: """Get learning rate scheduler. Args: config: entire configuration. Returns: Scheduler """ return optax.warmup_cosine_decay_schedule(**config.optimizer.lr_schedule) def get_every_k_schedule(config: DictConfig) -> int: """Get k for gradient accumulations. Args: config: entire configuration. Returns: k, where gradients are accumulated every k step. """ num_devices_per_replica = config.training.num_devices_per_replica batch_size_per_replica = config.training.batch_size_per_replica num_replicas = jax.local_device_count() // num_devices_per_replica batch_size_per_step = batch_size_per_replica * num_replicas if config.training.batch_size < batch_size_per_step: raise ValueError( f"Batch size {config.training.batch_size} is too small. " f"batch_size_per_replica * num_replicas = " f"{batch_size_per_replica} * {num_replicas} = " f"{batch_size_per_step}." ) if config.training.batch_size % batch_size_per_step != 0: raise ValueError( "Batch size cannot be evenly divided by batch size per step." ) every_k_schedule = config.training.batch_size // batch_size_per_step if every_k_schedule > 1: logging.info( f"Using gradient accumulation. " f"Each model duplicate is stored across {num_devices_per_replica} " f"shard{'s' if num_devices_per_replica > 1 else ''}. " f"Each step has {batch_size_per_step} samples. " f"Gradients are averaged every {every_k_schedule} steps. " f"Effective batch size is {config.training.batch_size}." ) return every_k_schedule def init_optimizer( config: DictConfig, ) -> Tuple[optax.GradientTransformation, int]: """Initialize optimizer. Args: config: entire configuration. Returns: optimizer and every_k_schedule. """ lr_schedule = get_lr_schedule(config) optimizer = optax.chain( optax.clip_by_global_norm(config.optimizer.grad_norm), getattr(optax, config.optimizer.name)( learning_rate=lr_schedule, **config.optimizer.kwargs ), ) # accumulate gradient when needed every_k_schedule = get_every_k_schedule(config) if every_k_schedule == 1: # no need to accumulate gradient return optimizer, every_k_schedule optimizer = optax.MultiSteps(optimizer, every_k_schedule=every_k_schedule) return optimizer, every_k_schedule
3,308
30.216981
79
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/exp/experiment.py
"""Module for launching experiments.""" import logging from functools import partial from pathlib import Path from typing import Callable, Dict, Mapping, Optional, Tuple, Union import chex import haiku as hk import jax import jax.numpy as jnp import jmp import optax import tensorflow as tf from omegaconf import DictConfig from imgx import IMAGE, REPLICA_AXIS, TEST_SPLIT, VALID_SPLIT from imgx.datasets import IMAGE_SHAPE_MAP from imgx.datasets.augmentation import build_aug_fn_from_config from imgx.datasets.iterator import get_image_tfds_dataset from imgx.datasets.util import py_prefetch from imgx.device import ( bind_rng_to_host_or_device, broadcast_to_local_devices, get_first_replica_values, is_tpu, ) from imgx.exp.eval import build_batch_eval_fn, build_dataset_eval_fn from imgx.exp.loss import build_loss_fn from imgx.exp.mixed_precision import get_mixed_precision_policy, select_tree from imgx.exp.optim import ema_update, get_lr_schedule, init_optimizer from imgx.exp.train_state import TrainState def init_train_state( batch: chex.ArrayTree, rng: jax.random.PRNGKey, loss_init: Callable, config: DictConfig, ) -> TrainState: """Initialize train_state. Args: batch: a batch example. rng: random key. loss_init: init function of loss. config: entire configuration. """ config_mp = config.training.mixed_precision config_ema = config.training.ema # init network rng, train_rng = jax.random.split(rng) params, network_state = loss_init(rng, batch) ema_params = params if config_ema.use else None ema_network_state = network_state if config_ema.use else None # count params on one device params_count = sum(x.size for x in jax.tree_util.tree_leaves(params)) logging.info(f"The model has {params_count:,} parameters.") # init optimizer state optimizer, _ = init_optimizer(config=config) opt_state = optimizer.init(params) # init loss_scale # it is necessary to use NoOpLossScale even not intended to use mp # otherwise, some unknown default policy may be used # resulted in non-converging losses and nans loss_scale = jmp.NoOpLossScale() if config_mp.use and (not is_tpu()): # no need to scale on TPU # https://cloud.google.com/tpu/docs/bfloat16 scale = jmp.half_dtype()(2**15) loss_scale = jmp.DynamicLossScale(scale) global_step = jnp.array(0, dtype=jnp.int32) return TrainState( # type: ignore[call-arg] params=params, network_state=network_state, opt_state=opt_state, loss_scale=loss_scale, global_step=global_step, rng=train_rng, ema_params=ema_params, ema_network_state=ema_network_state, ) def update_parameters( train_state: TrainState, batch: Mapping[str, chex.ArrayTree], loss_apply: Callable, config: DictConfig, ) -> Tuple[TrainState, chex.ArrayTree]: """Updates parameters. Mixed precision references: - https://github.com/deepmind/jmp - https://github.com/deepmind/dm-haiku/blob/main/examples/imagenet Args: train_state: training state. batch: training data. loss_apply: apply of loss function. config: entire configuration. Returns: train_state: training state. scalars: metric dict. """ def loss_fn( params: hk.Params, network_state: hk.State, loss_scale: jmp.LossScale, rng_key: jax.random.PRNGKey, batch_data: chex.ArrayTree, ) -> Tuple[chex.ArrayTree, Tuple[chex.ArrayTree, hk.State]]: """Regroup loss output. Args: params: network parameters. network_state: network state. loss_scale: scale loss for mixed precision. rng_key: random key. batch_data: data of a batch. Returns: - loss - (metric dict, model state) """ (loss, batch_scalars), network_state = loss_apply( params, network_state, rng_key, batch_data ) return loss_scale.scale(loss), (batch_scalars, network_state) config_mp = config.training.mixed_precision config_ema = config.training.ema aug_fn = build_aug_fn_from_config(config) # get random key for the step rng, step_rng = jax.random.split(train_state.rng) aug_rng, step_rng = jax.random.split(step_rng) aug_rng = bind_rng_to_host_or_device( aug_rng, bind_to="device", axis_name=REPLICA_AXIS ) step_rng = bind_rng_to_host_or_device( step_rng, bind_to="device", axis_name=REPLICA_AXIS ) # data augmentation batch = aug_fn(aug_rng, batch) # gradient calculation grad_loss_fn = jax.grad(loss_fn, has_aux=True) grads, (scalars, updated_network_state) = grad_loss_fn( train_state.params, train_state.network_state, train_state.loss_scale, step_rng, batch, ) scalars["grad_norm_before_pmean"] = optax.global_norm(grads) scalars["params_norm"] = optax.global_norm(train_state.params) # grads are in "param_dtype" (likely float32) # cast them back to compute dtype such that # we do the all-reduce below in the compute precision # which is typically lower than the param precision policy = get_mixed_precision_policy(config_mp.use) grads = policy.cast_to_compute(grads) grads = train_state.loss_scale.unscale(grads) # take the mean across all replicas to keep params in sync grads = jax.lax.pmean(grads, axis_name=REPLICA_AXIS) # compute our optimizer update in the same precision as params grads = policy.cast_to_param(grads) # update parameters optimizer, every_k_schedule = init_optimizer(config=config) updates, updated_opt_state = optimizer.update( grads, train_state.opt_state, train_state.params ) updated_params = optax.apply_updates(train_state.params, updates) scalars["lr"] = get_lr_schedule(config)( train_state.global_step // every_k_schedule ) scalars["grad_norm"] = optax.global_norm(grads) scalars["grad_update_norm"] = optax.global_norm(updates) grads_finite = jmp.all_finite(grads) updated_loss_scale = train_state.loss_scale.adjust(grads_finite) # mixed precision or not, skip non-finite gradients (updated_params, updated_network_state, updated_opt_state) = select_tree( grads_finite, (updated_params, updated_network_state, updated_opt_state), ( train_state.params, train_state.network_state, train_state.opt_state, ), ) scalars["loss_scale"] = updated_loss_scale.loss_scale # average metrics across replicas min_scalars = {} max_scalars = {} mean_scalars = {} for k in scalars: if k.startswith("min_"): min_scalars[k] = scalars[k] elif k.startswith("max_"): max_scalars[k] = scalars[k] else: mean_scalars[k] = scalars[k] min_scalars = jax.lax.pmin(min_scalars, axis_name=REPLICA_AXIS) max_scalars = jax.lax.pmax(max_scalars, axis_name=REPLICA_AXIS) mean_scalars = jax.lax.pmean(mean_scalars, axis_name=REPLICA_AXIS) scalars = { **min_scalars, **max_scalars, **mean_scalars, } # update train_state train_state = train_state.replace( params=updated_params, network_state=updated_network_state, opt_state=updated_opt_state, loss_scale=updated_loss_scale, rng=rng, ) if train_state.ema_params is not None: ema = partial( ema_update, decay=config_ema.decay, step=train_state.global_step, ) ema_params = jax.tree_map(ema, train_state.ema_params, updated_params) ema_network_state = jax.tree_map( ema, train_state.ema_network_state, updated_network_state ) train_state = train_state.replace( ema_params=ema_params, ema_network_state=ema_network_state, ) train_state = train_state.replace( global_step=train_state.global_step + 1, ) return train_state, scalars def batch_eval( batch: Mapping[str, chex.ArrayTree], config: DictConfig, ) -> Tuple[chex.ArrayTree, chex.ArrayTree]: """Calculate prediction and metrics given a batch, without loss. Args: batch: input batch data. config: entire configuration. Returns: - metrics. - prediction. """ eval_fn = build_batch_eval_fn( config=config, ) return eval_fn(batch) class Experiment: """Experiment for supervised training.""" def __init__(self, config: DictConfig) -> None: """Initializes experiment. Args: config: experiment config. """ # Do not use accelerators in data pipeline. tf.config.experimental.set_visible_devices([], device_type="GPU") tf.config.experimental.set_visible_devices([], device_type="TPU") # save args self.config = config # init data loaders and networks self.dataset = get_image_tfds_dataset( dataset_name=self.config.data.name, config=self.config, ) self.train_iter = py_prefetch(lambda: self.dataset.train_iter) self.valid_iter = py_prefetch(lambda: self.dataset.valid_iter) self.test_iter = py_prefetch(lambda: self.dataset.test_iter) def train_init(self) -> TrainState: """Initialize data loader, loss, networks for training. Returns: initialized training state. """ # init loss loss = hk.transform_with_state(build_loss_fn(config=self.config)) # the batch is for multi-devices # (num_models, ...) # num_models is not the same as num_devices_per_replica batch = next(self.train_iter) batch = get_first_replica_values(batch) # check image size data_config = self.config["data"] image_shape = IMAGE_SHAPE_MAP[data_config["name"]] chex.assert_equal(batch[IMAGE].shape[1:4], image_shape) aug_fn = build_aug_fn_from_config(self.config) aug_rng = jax.random.PRNGKey(self.config["seed"]) batch = aug_fn(aug_rng, batch) # init train state on cpu first rng = jax.random.PRNGKey(self.config.seed) train_state = jax.jit( partial(init_train_state, loss_init=loss.init, config=self.config), backend="cpu", )( batch=batch, rng=rng, ) # then broadcast train_state to devices train_state = broadcast_to_local_devices(train_state) # define pmap-ed update func self.update_params_pmap = jax.pmap( partial( update_parameters, loss_apply=loss.apply, config=self.config, ), axis_name=REPLICA_AXIS, donate_argnums=(0,), ) return train_state def train_step( self, train_state: TrainState, ) -> Union[TrainState, chex.ArrayTree]: """Training step. Args: train_state: training state. Returns: - updated train_state. - metric dict. """ batch = next(self.train_iter) train_state, scalars = self.update_params_pmap( train_state, batch, ) scalars = get_first_replica_values(scalars) scalars = jax.tree_map(lambda x: x.item(), scalars) # tensor to values return train_state, scalars def eval_init(self) -> None: """Initialize data loader, loss, networks for validation.""" evaluate = hk.transform_with_state( partial(batch_eval, config=self.config) ) self.evaluate_pmap = jax.pmap( evaluate.apply, axis_name=REPLICA_AXIS, ) self.eval_dataset = build_dataset_eval_fn(self.config) def eval_step( self, split: str, params: hk.Params, state: hk.State, rng: jax.random.PRNGKey, out_dir: Optional[Path], save_predictions: bool, ) -> Dict: """Validation step on entire validation data set. Args: split: data split. params: network parameters. state: network state. rng: random key. out_dir: output directory to save metrics and predictions, if None, no files will be saved. save_predictions: if True, save predicted masks. Returns: metric dict. Raises: ValueError: if split is not supported. """ if split not in [VALID_SPLIT, TEST_SPLIT]: raise ValueError( "Evaluation can only be performed on valid and test splits." ) if split == VALID_SPLIT: batch_iterator = self.valid_iter num_steps = self.dataset.num_valid_steps else: batch_iterator = self.test_iter num_steps = self.dataset.num_test_steps if out_dir is not None: out_dir.mkdir(parents=True, exist_ok=True) return self.eval_dataset( evaluate_pmap=self.evaluate_pmap, params=params, state=state, rng=rng, batch_iterator=batch_iterator, num_steps=num_steps, out_dir=out_dir, save_predictions=save_predictions, )
13,642
30.801865
79
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/exp/loss.py
"""Module for building models and losses.""" from typing import Callable, Dict, Tuple import haiku as hk import jax import jax.numpy as jnp from omegaconf import DictConfig from imgx import IMAGE, LABEL from imgx.datasets import NUM_CLASSES_MAP from imgx.diffusion.gaussian_diffusion import ( DiffusionModelOutputType, DiffusionModelVarianceType, GaussianDiffusion, ) from imgx.exp.mixed_precision import set_mixed_precision_policy from imgx.exp.model import build_diffusion_model, build_vision_model from imgx.loss import mean_cross_entropy, mean_focal_loss from imgx.loss.dice import ( dice_loss, mean_with_background, mean_without_background, ) from imgx.metric import class_proportion def segmentation_loss_with_aux( logits: jnp.ndarray, mask_true: jnp.ndarray, loss_config: DictConfig, ) -> Tuple[jnp.ndarray, Dict[str, jnp.ndarray]]: """Calculate segmentation loss with auxiliary losses and return metrics. Args: logits: unnormalised logits of shape (batch, ..., num_classes). mask_true: one hot label of shape (batch, ..., num_classes). loss_config: have weights of diff losses. Returns: - calculated loss. - metrics. """ scalars = {} # Dice # (batch, num_classes) dice_loss_batch_cls = dice_loss( logits=logits, mask_true=mask_true, ) # (1,) dice_loss_scalar = jax.lax.cond( loss_config["dice_include_background"], mean_with_background, mean_without_background, dice_loss_batch_cls, ) scalars["mean_dice_loss"] = dice_loss_scalar # metrics for i in range(dice_loss_batch_cls.shape[-1]): scalars[f"mean_dice_loss_class_{i}"] = jnp.nanmean( dice_loss_batch_cls[:, i] ) scalars[f"min_dice_loss_class_{i}"] = jnp.nanmin( dice_loss_batch_cls[:, i] ) scalars[f"max_dice_loss_class_{i}"] = jnp.nanmax( dice_loss_batch_cls[:, i] ) scalars["mean_dice_loss"] = jnp.nanmean(dice_loss_batch_cls) scalars["min_dice_loss"] = jnp.nanmin(dice_loss_batch_cls) scalars["max_dice_loss"] = jnp.nanmax(dice_loss_batch_cls) # cross entropy ce_loss_scalar = mean_cross_entropy( logits=logits, mask_true=mask_true, ) scalars["mean_cross_entropy_loss"] = ce_loss_scalar # focal loss focal_loss_scalar = mean_focal_loss( logits=logits, mask_true=mask_true, ) scalars["mean_focal_loss"] = focal_loss_scalar # total loss loss_scalar = 0 if loss_config["dice"] > 0: loss_scalar += dice_loss_scalar * loss_config["dice"] if loss_config["cross_entropy"] > 0: loss_scalar += ce_loss_scalar * loss_config["cross_entropy"] if loss_config["focal"] > 0: loss_scalar += focal_loss_scalar * loss_config["focal"] # class proportion # (batch, num_classes) cls_prop = class_proportion(mask_true) for i in range(dice_loss_batch_cls.shape[-1]): scalars[f"mean_proportion_class_{i}"] = jnp.nanmean(cls_prop[:, i]) scalars[f"min_proportion_class_{i}"] = jnp.nanmin(cls_prop[:, i]) scalars[f"max_proportion_class_{i}"] = jnp.nanmax(cls_prop[:, i]) return loss_scalar, scalars def segmentation_loss( input_dict: Dict[str, jnp.ndarray], model: hk.Module, num_classes: int, loss_config: DictConfig, ) -> Tuple[jnp.ndarray, Dict[str, jnp.ndarray]]: """Calculate segmentation loss and return metrics. Args: input_dict: input data having image and label. model: network instance. num_classes: number of classes including background. loss_config: have weights of diff losses. Returns: - calculated loss. - metrics. """ # (batch, ..., 1) image = jnp.expand_dims(input_dict[IMAGE], axis=-1) # (batch, ..., num_classes) logits = model(image=image, is_train=True) # (batch, ..., num_classes) mask_true = jax.nn.one_hot( input_dict[LABEL], num_classes=num_classes, axis=-1 ) return segmentation_loss_with_aux( logits=logits, mask_true=mask_true, loss_config=loss_config, ) def diffusion_loss( # pylint:disable=R0915 input_dict: Dict[str, jnp.ndarray], num_classes: int, gd: GaussianDiffusion, loss_config: DictConfig, recycle: bool, ) -> Tuple[jnp.ndarray, Dict[str, jnp.ndarray]]: """Calculate diffusion loss and return metrics. In diffusion, the noise is defined on segmentation mask. That is, x_t is segmentation logits. Args: input_dict: input data having image, label, and time_step. image: (batch, ...) label: (batch, ..., num_classes) time_step: (batch, ) num_classes: number of classes including background. gd: model for sampling. loss_config: have weights of diff losses. recycle: recycle model prediction or not. Returns: - calculated loss. - metrics. """ scalars = {} # (batch, ..., 1) image = jnp.expand_dims(input_dict[IMAGE], axis=-1) # (batch, ..., num_classes) mask_true = jax.nn.one_hot( input_dict[LABEL], num_classes=num_classes, axis=-1, dtype=image.dtype, ) # noise are standard normal distribution x_start = mask_true * 2 - 1 # (batch, ) t = gd.sample_timestep(batch_size=image.shape[0]) if recycle: # (batch, ..., num_classes) noise_recyle = gd.noise_sample(shape=x_start.shape, dtype=x_start.dtype) t_recycle = jnp.minimum(t + 1, gd.num_timesteps - 1) x_t_recycle = gd.q_sample( x_start=x_start, noise=noise_recyle, t=t_recycle ) # (batch, ..., ch_input + num_classes) model_in_recycle = jnp.concatenate([image, x_t_recycle], axis=-1) # (batch, ..., num_classes) or (batch, ..., 2*num_classes) # model outputs are always logits model_out_recycle = gd.model(model_in_recycle, t_recycle, is_train=True) x_start_recycle, _, _ = gd.p_mean_variance( model_out=model_out_recycle, x_t=x_t_recycle, t=t_recycle, ) x_start = jax.lax.stop_gradient(x_start_recycle) # (batch, ..., num_classes) noise = gd.noise_sample(shape=x_start.shape, dtype=x_start.dtype) x_t = gd.q_sample(x_start=x_start, noise=noise, t=t) # (batch, ..., ch_input + num_classes) model_in = jnp.concatenate([image, x_t], axis=-1) # (batch, ..., num_classes) or (batch, ..., 2*num_classes) # model outputs are always logits model_out = gd.model(model_in, t, is_train=True) model_out_vlb = jax.lax.stop_gradient(model_out) if gd.model_var_type in [ DiffusionModelVarianceType.LEARNED, DiffusionModelVarianceType.LEARNED_RANGE, ]: # model_out (batch, ..., num_classes) model_out, log_variance = jnp.split( model_out, indices_or_sections=2, axis=-1 ) # apply a stop-gradient to the mean output for the vlb to prevent # this loss change mean prediction model_out_vlb = jax.lax.stop_gradient(model_out) # model_out (batch, ..., num_classes*2) model_out_vlb = jnp.concatenate([model_out_vlb, log_variance], axis=-1) vlb_scalar = gd.variational_lower_bound( model_out=model_out_vlb, x_start=x_start, x_t=x_t, t=t, ) vlb_scalar = jnp.nanmean(vlb_scalar) scalars["vlb_loss"] = vlb_scalar if gd.model_out_type == DiffusionModelOutputType.EPSILON: mse_loss_scalar = jnp.mean((model_out - noise) ** 2) scalars["mse_loss"] = mse_loss_scalar x_start = gd.predict_xstart_from_epsilon_xt( x_t=x_t, epsilon=model_out, t=t ) logits = gd.x_to_logits(x_start) seg_loss_scalar, seg_scalars = segmentation_loss_with_aux( logits=logits, mask_true=mask_true, loss_config=loss_config, ) scalars = {**scalars, **seg_scalars} loss_scalar = loss_config["mse"] * mse_loss_scalar + seg_loss_scalar elif gd.model_out_type == DiffusionModelOutputType.X_START: logits = model_out loss_scalar, seg_scalars = segmentation_loss_with_aux( logits=logits, mask_true=mask_true, loss_config=loss_config, ) scalars = {**scalars, **seg_scalars} else: raise ValueError( f"Unknown DiffusionModelOutputType {gd.model_out_type}." ) if gd.model_var_type in [ DiffusionModelVarianceType.LEARNED, DiffusionModelVarianceType.LEARNED_RANGE, ]: # TODO nan values may happen loss_scalar += vlb_scalar * gd.num_timesteps / gd.num_timesteps_beta scalars["total_loss"] = loss_scalar scalars["mean_t"] = jnp.mean(t) scalars["max_t"] = jnp.max(t) scalars["min_t"] = jnp.min(t) return loss_scalar, scalars def build_loss_fn( config: DictConfig, ) -> Callable[ [Dict[str, jnp.ndarray]], Tuple[jnp.ndarray, Dict[str, jnp.ndarray]] ]: """Build model from config. Args: config: entire config. Returns: Loss function. Raises: ValueError: if config is wrong or not supported. """ data_config = config.data task_config = config.task model_config = config.model loss_config = config.loss mp_config = config.training.mixed_precision set_mixed_precision_policy( use_mp=mp_config.use, model_name=model_config.name ) # number of classes including background num_classes = NUM_CLASSES_MAP[data_config["name"]] if task_config["name"] == "segmentation": def seg_loss_fn( input_dict: Dict[str, jnp.ndarray] ) -> Tuple[jnp.ndarray, Dict[str, jnp.ndarray]]: vision_model = build_vision_model( data_config=data_config, task_config=task_config, model_config=model_config, ) return segmentation_loss( input_dict=input_dict, model=vision_model, num_classes=num_classes, loss_config=loss_config, ) return seg_loss_fn if task_config["name"] == "diffusion": def diffusion_loss_fn( input_dict: Dict[str, jnp.ndarray] ) -> Tuple[jnp.ndarray, Dict[str, jnp.ndarray]]: vision_model = build_vision_model( data_config=data_config, task_config=task_config, model_config=model_config, ) diffusion_model = build_diffusion_model( model=vision_model, diffusion_config=task_config["diffusion"], ) recycle = task_config["diffusion"]["recycle"] return diffusion_loss( input_dict=input_dict, gd=diffusion_model, num_classes=num_classes, loss_config=loss_config, recycle=recycle, ) return diffusion_loss_fn raise ValueError(f"Unknown task {task_config['name']}.")
11,245
30.858357
80
py
ImgX-DiffSeg
ImgX-DiffSeg-main/imgx/exp/model.py
"""Module for building models.""" import haiku as hk from omegaconf import DictConfig from imgx.datasets import IMAGE_SHAPE_MAP, NUM_CLASSES_MAP from imgx.diffusion.gaussian_diffusion import ( DiffusionBetaSchedule, DiffusionModelOutputType, DiffusionModelVarianceType, DiffusionSpace, GaussianDiffusion, ) from imgx.model import Unet3d, Unet3dSlice, Unet3dSliceTime, Unet3dTime def build_vision_model( data_config: DictConfig, task_config: DictConfig, model_config: DictConfig, ) -> hk.Module: """Build model from config. Args: data_config: have in_shape and out_channels. task_config: have task name and configs. model_config: have model name attribute. Returns: Model. Raises: ValueError: if config is wrong or not supported. """ if model_config.name not in model_config: raise ValueError(f"Missing configuration for {model_config.name}.") dataset_name = data_config["name"] image_shape = IMAGE_SHAPE_MAP[dataset_name] num_classes = NUM_CLASSES_MAP[dataset_name] if task_config.name == "segmentation": # TODO use enum in_channels = 1 # forward will expand dimension out_channels = num_classes elif task_config.name == "diffusion": # diffusion model takes the image and a noised mask/logits as input in_channels = 1 + num_classes # diffusion model may output variance per class out_channels = num_classes model_var_type = DiffusionModelVarianceType[ task_config["diffusion"]["model_var_type"].upper() ] if model_var_type in [ DiffusionModelVarianceType.LEARNED, DiffusionModelVarianceType.LEARNED_RANGE, ]: out_channels *= 2 else: raise ValueError(f"Unknown task {task_config.name}.") total_config = { "remat": model_config.remat, "in_shape": image_shape, "in_channels": in_channels, "out_channels": out_channels, **model_config[model_config.name], } if model_config.name == "unet3d": return Unet3d(**total_config) if model_config.name == "unet3d_slice": return Unet3dSlice(**total_config) if model_config.name == "unet3d_time": num_timesteps = task_config["diffusion"]["num_timesteps"] return Unet3dTime(num_timesteps=num_timesteps, **total_config) if model_config.name == "unet3d_slice_time": num_timesteps = task_config["diffusion"]["num_timesteps"] return Unet3dSliceTime(num_timesteps=num_timesteps, **total_config) raise ValueError(f"Unknown model {model_config.name}.") def build_diffusion_model( model: hk.Module, diffusion_config: DictConfig, ) -> GaussianDiffusion: """Build diffusion model from config and vision model. Args: model: the model used in diffusion. diffusion_config: config for diffusion setting. Returns: A GaussianDiffusion model. """ num_timesteps = diffusion_config["num_timesteps"] num_timesteps_beta = diffusion_config["num_timesteps_beta"] beta_config = diffusion_config["beta"].copy() beta_config["beta_schedule"] = DiffusionBetaSchedule[ beta_config["beta_schedule"].upper() ] model_out_type = DiffusionModelOutputType[ diffusion_config["model_out_type"].upper() ] model_var_type = DiffusionModelVarianceType[ diffusion_config["model_var_type"].upper() ] x_space = DiffusionSpace[diffusion_config["x_space"].upper()] x_limit = diffusion_config["x_limit"] use_ddim = diffusion_config["use_ddim"] return GaussianDiffusion( model=model, num_timesteps=num_timesteps, num_timesteps_beta=num_timesteps_beta, model_out_type=model_out_type, model_var_type=model_var_type, x_space=x_space, x_limit=x_limit, use_ddim=use_ddim, **beta_config, )
3,969
32.644068
75
py