Skip to content

GitLab

".github/git@developer.sourcefind.cn:Fzc7075/nunchaku.git" did not exist on "102a0f7d422b9101cf5b0f7caffe1bda7ec09150"
Commit 8a9a607c authored by Yeqing Li's avatar Yeqing Li Committed by A. Unique TensorFlower
Browse files

Internal change 

PiperOrigin-RevId: 425419954
parent b7bb52f0
Hide whitespace changes
Inline Side-by-side
Showing with 6966 additions and 0 deletions
official/vision/modeling/heads/instance_heads.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains definitions of instance prediction heads."""
from typing import List, Union, Optional
# Import libraries
import tensorflow as tf
from official.modeling import tf_utils
@tf.keras.utils.register_keras_serializable(package='Vision')
class DetectionHead(tf.keras.layers.Layer):
"""Creates a detection head."""
def __init__(
self,
num_classes: int,
num_convs: int = 0,
num_filters: int = 256,
use_separable_conv: bool = False,
num_fcs: int = 2,
fc_dims: int = 1024,
class_agnostic_bbox_pred: bool = False,
activation: str = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
bias_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
**kwargs):
"""Initializes a detection head.
Args:
num_classes: An `int` for the number of classes.
num_convs: An `int` number that represents the number of the intermediate
convolution layers before the FC layers.
num_filters: An `int` number that represents the number of filters of the
intermediate convolution layers.
use_separable_conv: A `bool` that indicates whether the separable
convolution layers is used.
num_fcs: An `int` number that represents the number of FC layers before
the predictions.
fc_dims: An `int` number that represents the number of dimension of the FC
layers.
class_agnostic_bbox_pred: `bool`, indicating whether bboxes should be
predicted for every class or not.
activation: A `str` that indicates which activation is used, e.g. 'relu',
'swish', etc.
use_sync_bn: A `bool` that indicates whether to use synchronized batch
normalization across different replicas.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
**kwargs: Additional keyword arguments to be passed.
"""
super(DetectionHead, self).__init__(**kwargs)
self._config_dict = {
'num_classes': num_classes,
'num_convs': num_convs,
'num_filters': num_filters,
'use_separable_conv': use_separable_conv,
'num_fcs': num_fcs,
'fc_dims': fc_dims,
'class_agnostic_bbox_pred': class_agnostic_bbox_pred,
'activation': activation,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
}
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation = tf_utils.get_activation(activation)
def build(self, input_shape: Union[tf.TensorShape, List[tf.TensorShape]]):
"""Creates the variables of the head."""
conv_op = (tf.keras.layers.SeparableConv2D
if self._config_dict['use_separable_conv']
else tf.keras.layers.Conv2D)
conv_kwargs = {
'filters': self._config_dict['num_filters'],
'kernel_size': 3,
'padding': 'same',
}
if self._config_dict['use_separable_conv']:
conv_kwargs.update({
'depthwise_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'pointwise_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'depthwise_regularizer': self._config_dict['kernel_regularizer'],
'pointwise_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
else:
conv_kwargs.update({
'kernel_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'kernel_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
bn_op = (tf.keras.layers.experimental.SyncBatchNormalization
if self._config_dict['use_sync_bn']
else tf.keras.layers.BatchNormalization)
bn_kwargs = {
'axis': self._bn_axis,
'momentum': self._config_dict['norm_momentum'],
'epsilon': self._config_dict['norm_epsilon'],
}
self._convs = []
self._conv_norms = []
for i in range(self._config_dict['num_convs']):
conv_name = 'detection-conv_{}'.format(i)
self._convs.append(conv_op(name=conv_name, **conv_kwargs))
bn_name = 'detection-conv-bn_{}'.format(i)
self._conv_norms.append(bn_op(name=bn_name, **bn_kwargs))
self._fcs = []
self._fc_norms = []
for i in range(self._config_dict['num_fcs']):
fc_name = 'detection-fc_{}'.format(i)
self._fcs.append(
tf.keras.layers.Dense(
units=self._config_dict['fc_dims'],
kernel_initializer=tf.keras.initializers.VarianceScaling(
scale=1 / 3.0, mode='fan_out', distribution='uniform'),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'],
name=fc_name))
bn_name = 'detection-fc-bn_{}'.format(i)
self._fc_norms.append(bn_op(name=bn_name, **bn_kwargs))
self._classifier = tf.keras.layers.Dense(
units=self._config_dict['num_classes'],
kernel_initializer=tf.keras.initializers.RandomNormal(stddev=0.01),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'],
name='detection-scores')
num_box_outputs = (4 if self._config_dict['class_agnostic_bbox_pred'] else
self._config_dict['num_classes'] * 4)
self._box_regressor = tf.keras.layers.Dense(
units=num_box_outputs,
kernel_initializer=tf.keras.initializers.RandomNormal(stddev=0.001),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'],
name='detection-boxes')
super(DetectionHead, self).build(input_shape)
def call(self, inputs: tf.Tensor, training: bool = None):
"""Forward pass of box and class branches for the Mask-RCNN model.
Args:
inputs: A `tf.Tensor` of the shape [batch_size, num_instances, roi_height,
roi_width, roi_channels], representing the ROI features.
training: a `bool` indicating whether it is in `training` mode.
Returns:
class_outputs: A `tf.Tensor` of the shape
[batch_size, num_rois, num_classes], representing the class predictions.
box_outputs: A `tf.Tensor` of the shape
[batch_size, num_rois, num_classes * 4], representing the box
predictions.
"""
roi_features = inputs
_, num_rois, height, width, filters = roi_features.get_shape().as_list()
x = tf.reshape(roi_features, [-1, height, width, filters])
for conv, bn in zip(self._convs, self._conv_norms):
x = conv(x)
x = bn(x)
x = self._activation(x)
_, _, _, filters = x.get_shape().as_list()
x = tf.reshape(x, [-1, num_rois, height * width * filters])
for fc, bn in zip(self._fcs, self._fc_norms):
x = fc(x)
x = bn(x)
x = self._activation(x)
classes = self._classifier(x)
boxes = self._box_regressor(x)
return classes, boxes
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config):
return cls(**config)
@tf.keras.utils.register_keras_serializable(package='Vision')
class MaskHead(tf.keras.layers.Layer):
"""Creates a mask head."""
def __init__(
self,
num_classes: int,
upsample_factor: int = 2,
num_convs: int = 4,
num_filters: int = 256,
use_separable_conv: bool = False,
activation: str = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
bias_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
class_agnostic: bool = False,
**kwargs):
"""Initializes a mask head.
Args:
num_classes: An `int` of the number of classes.
upsample_factor: An `int` that indicates the upsample factor to generate
the final predicted masks. It should be >= 1.
num_convs: An `int` number that represents the number of the intermediate
convolution layers before the mask prediction layers.
num_filters: An `int` number that represents the number of filters of the
intermediate convolution layers.
use_separable_conv: A `bool` that indicates whether the separable
convolution layers is used.
activation: A `str` that indicates which activation is used, e.g. 'relu',
'swish', etc.
use_sync_bn: A `bool` that indicates whether to use synchronized batch
normalization across different replicas.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
class_agnostic: A `bool`. If set, we use a single channel mask head that
is shared between all classes.
**kwargs: Additional keyword arguments to be passed.
"""
super(MaskHead, self).__init__(**kwargs)
self._config_dict = {
'num_classes': num_classes,
'upsample_factor': upsample_factor,
'num_convs': num_convs,
'num_filters': num_filters,
'use_separable_conv': use_separable_conv,
'activation': activation,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
'class_agnostic': class_agnostic
}
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation = tf_utils.get_activation(activation)
def build(self, input_shape: Union[tf.TensorShape, List[tf.TensorShape]]):
"""Creates the variables of the head."""
conv_op = (tf.keras.layers.SeparableConv2D
if self._config_dict['use_separable_conv']
else tf.keras.layers.Conv2D)
conv_kwargs = {
'filters': self._config_dict['num_filters'],
'kernel_size': 3,
'padding': 'same',
}
if self._config_dict['use_separable_conv']:
conv_kwargs.update({
'depthwise_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'pointwise_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'depthwise_regularizer': self._config_dict['kernel_regularizer'],
'pointwise_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
else:
conv_kwargs.update({
'kernel_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'kernel_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
bn_op = (tf.keras.layers.experimental.SyncBatchNormalization
if self._config_dict['use_sync_bn']
else tf.keras.layers.BatchNormalization)
bn_kwargs = {
'axis': self._bn_axis,
'momentum': self._config_dict['norm_momentum'],
'epsilon': self._config_dict['norm_epsilon'],
}
self._convs = []
self._conv_norms = []
for i in range(self._config_dict['num_convs']):
conv_name = 'mask-conv_{}'.format(i)
self._convs.append(conv_op(name=conv_name, **conv_kwargs))
bn_name = 'mask-conv-bn_{}'.format(i)
self._conv_norms.append(bn_op(name=bn_name, **bn_kwargs))
self._deconv = tf.keras.layers.Conv2DTranspose(
filters=self._config_dict['num_filters'],
kernel_size=self._config_dict['upsample_factor'],
strides=self._config_dict['upsample_factor'],
padding='valid',
kernel_initializer=tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'],
name='mask-upsampling')
self._deconv_bn = bn_op(name='mask-deconv-bn', **bn_kwargs)
if self._config_dict['class_agnostic']:
num_filters = 1
else:
num_filters = self._config_dict['num_classes']
conv_kwargs = {
'filters': num_filters,
'kernel_size': 1,
'padding': 'valid',
}
if self._config_dict['use_separable_conv']:
conv_kwargs.update({
'depthwise_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'pointwise_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'depthwise_regularizer': self._config_dict['kernel_regularizer'],
'pointwise_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
else:
conv_kwargs.update({
'kernel_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'kernel_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
self._mask_regressor = conv_op(name='mask-logits', **conv_kwargs)
super(MaskHead, self).build(input_shape)
def call(self, inputs: List[tf.Tensor], training: bool = None):
"""Forward pass of mask branch for the Mask-RCNN model.
Args:
inputs: A `list` of two tensors where
inputs[0]: A `tf.Tensor` of shape [batch_size, num_instances,
roi_height, roi_width, roi_channels], representing the ROI features.
inputs[1]: A `tf.Tensor` of shape [batch_size, num_instances],
representing the classes of the ROIs.
training: A `bool` indicating whether it is in `training` mode.
Returns:
mask_outputs: A `tf.Tensor` of shape
[batch_size, num_instances, roi_height * upsample_factor,
roi_width * upsample_factor], representing the mask predictions.
"""
roi_features, roi_classes = inputs
batch_size, num_rois, height, width, filters = (
roi_features.get_shape().as_list())
if batch_size is None:
batch_size = tf.shape(roi_features)[0]
x = tf.reshape(roi_features, [-1, height, width, filters])
for conv, bn in zip(self._convs, self._conv_norms):
x = conv(x)
x = bn(x)
x = self._activation(x)
x = self._deconv(x)
x = self._deconv_bn(x)
x = self._activation(x)
logits = self._mask_regressor(x)
mask_height = height * self._config_dict['upsample_factor']
mask_width = width * self._config_dict['upsample_factor']
if self._config_dict['class_agnostic']:
logits = tf.reshape(logits, [-1, num_rois, mask_height, mask_width, 1])
else:
logits = tf.reshape(
logits,
[-1, num_rois, mask_height, mask_width,
self._config_dict['num_classes']])
batch_indices = tf.tile(
tf.expand_dims(tf.range(batch_size), axis=1), [1, num_rois])
mask_indices = tf.tile(
tf.expand_dims(tf.range(num_rois), axis=0), [batch_size, 1])
if self._config_dict['class_agnostic']:
class_gather_indices = tf.zeros_like(roi_classes, dtype=tf.int32)
else:
class_gather_indices = tf.cast(roi_classes, dtype=tf.int32)
gather_indices = tf.stack(
[batch_indices, mask_indices, class_gather_indices],
axis=2)
mask_outputs = tf.gather_nd(
tf.transpose(logits, [0, 1, 4, 2, 3]), gather_indices)
return mask_outputs
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config):
return cls(**config)
official/vision/modeling/heads/instance_heads_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
# Lint as: python3
"""Tests for instance_heads.py."""
# Import libraries
from absl.testing import parameterized
import numpy as np
import tensorflow as tf
from official.vision.modeling.heads import instance_heads
class DetectionHeadTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(0, 0, False, False),
(0, 1, False, False),
(1, 0, False, False),
(1, 1, False, False),
)
def test_forward(self, num_convs, num_fcs, use_separable_conv, use_sync_bn):
detection_head = instance_heads.DetectionHead(
num_classes=3,
num_convs=num_convs,
num_filters=16,
use_separable_conv=use_separable_conv,
num_fcs=num_fcs,
fc_dims=4,
activation='relu',
use_sync_bn=use_sync_bn,
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_regularizer=None,
bias_regularizer=None,
)
roi_features = np.random.rand(2, 10, 128, 128, 16)
scores, boxes = detection_head(roi_features)
self.assertAllEqual(scores.numpy().shape, [2, 10, 3])
self.assertAllEqual(boxes.numpy().shape, [2, 10, 12])
def test_serialize_deserialize(self):
detection_head = instance_heads.DetectionHead(
num_classes=91,
num_convs=0,
num_filters=256,
use_separable_conv=False,
num_fcs=2,
fc_dims=1024,
activation='relu',
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_regularizer=None,
bias_regularizer=None,
)
config = detection_head.get_config()
new_detection_head = instance_heads.DetectionHead.from_config(config)
self.assertAllEqual(
detection_head.get_config(), new_detection_head.get_config())
class MaskHeadTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(1, 1, False),
(1, 2, False),
(2, 1, False),
(2, 2, False),
)
def test_forward(self, upsample_factor, num_convs, use_sync_bn):
mask_head = instance_heads.MaskHead(
num_classes=3,
upsample_factor=upsample_factor,
num_convs=num_convs,
num_filters=16,
use_separable_conv=False,
activation='relu',
use_sync_bn=use_sync_bn,
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_regularizer=None,
bias_regularizer=None,
)
roi_features = np.random.rand(2, 10, 14, 14, 16)
roi_classes = np.zeros((2, 10))
masks = mask_head([roi_features, roi_classes])
self.assertAllEqual(
masks.numpy().shape,
[2, 10, 14 * upsample_factor, 14 * upsample_factor])
def test_serialize_deserialize(self):
mask_head = instance_heads.MaskHead(
num_classes=3,
upsample_factor=2,
num_convs=1,
num_filters=256,
use_separable_conv=False,
activation='relu',
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_regularizer=None,
bias_regularizer=None,
)
config = mask_head.get_config()
new_mask_head = instance_heads.MaskHead.from_config(config)
self.assertAllEqual(
mask_head.get_config(), new_mask_head.get_config())
def test_forward_class_agnostic(self):
mask_head = instance_heads.MaskHead(
num_classes=3,
class_agnostic=True
)
roi_features = np.random.rand(2, 10, 14, 14, 16)
roi_classes = np.zeros((2, 10))
masks = mask_head([roi_features, roi_classes])
self.assertAllEqual(masks.numpy().shape, [2, 10, 28, 28])
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/heads/segmentation_heads.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains definitions of segmentation heads."""
from typing import List, Union, Optional, Mapping, Tuple, Any
import tensorflow as tf
from official.modeling import tf_utils
from official.vision.modeling.layers import nn_layers
from official.vision.ops import spatial_transform_ops
class MaskScoring(tf.keras.Model):
"""Creates a mask scoring layer.
This implements mask scoring layer from the paper:
Zhaojin Huang, Lichao Huang, Yongchao Gong, Chang Huang, Xinggang Wang.
Mask Scoring R-CNN.
(https://arxiv.org/pdf/1903.00241.pdf)
"""
def __init__(
self,
num_classes: int,
fc_input_size: List[int],
num_convs: int = 3,
num_filters: int = 256,
fc_dims: int = 1024,
num_fcs: int = 2,
activation: str = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
bias_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
**kwargs):
"""Initializes mask scoring layer.
Args:
num_classes: An `int` for number of classes.
fc_input_size: A List of `int` for the input size of the
fully connected layers.
num_convs: An`int` for number of conv layers.
num_filters: An `int` for the number of filters for conv layers.
fc_dims: An `int` number of filters for each fully connected layers.
num_fcs: An `int` for number of fully connected layers.
activation: A `str` name of the activation function.
use_sync_bn: A bool, whether or not to use sync batch normalization.
norm_momentum: A float for the momentum in BatchNorm. Defaults to 0.99.
norm_epsilon: A float for the epsilon value in BatchNorm. Defaults to
0.001.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
**kwargs: Additional keyword arguments to be passed.
"""
super(MaskScoring, self).__init__(**kwargs)
self._config_dict = {
'num_classes': num_classes,
'num_convs': num_convs,
'num_filters': num_filters,
'fc_input_size': fc_input_size,
'fc_dims': fc_dims,
'num_fcs': num_fcs,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
'activation': activation,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
}
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation = tf_utils.get_activation(activation)
def build(self, input_shape: Union[tf.TensorShape, List[tf.TensorShape]]):
"""Creates the variables of the mask scoring head."""
conv_op = tf.keras.layers.Conv2D
conv_kwargs = {
'filters': self._config_dict['num_filters'],
'kernel_size': 3,
'padding': 'same',
}
conv_kwargs.update({
'kernel_initializer': tf.keras.initializers.VarianceScaling(
scale=2, mode='fan_out', distribution='untruncated_normal'),
'bias_initializer': tf.zeros_initializer(),
'kernel_regularizer': self._config_dict['kernel_regularizer'],
'bias_regularizer': self._config_dict['bias_regularizer'],
})
bn_op = (tf.keras.layers.experimental.SyncBatchNormalization
if self._config_dict['use_sync_bn']
else tf.keras.layers.BatchNormalization)
bn_kwargs = {
'axis': self._bn_axis,
'momentum': self._config_dict['norm_momentum'],
'epsilon': self._config_dict['norm_epsilon'],
}
self._convs = []
self._conv_norms = []
for i in range(self._config_dict['num_convs']):
conv_name = 'mask-scoring_{}'.format(i)
self._convs.append(conv_op(name=conv_name, **conv_kwargs))
bn_name = 'mask-scoring-bn_{}'.format(i)
self._conv_norms.append(bn_op(name=bn_name, **bn_kwargs))
self._fcs = []
self._fc_norms = []
for i in range(self._config_dict['num_fcs']):
fc_name = 'mask-scoring-fc_{}'.format(i)
self._fcs.append(
tf.keras.layers.Dense(
units=self._config_dict['fc_dims'],
kernel_initializer=tf.keras.initializers.VarianceScaling(
scale=1 / 3.0, mode='fan_out', distribution='uniform'),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'],
name=fc_name))
bn_name = 'mask-scoring-fc-bn_{}'.format(i)
self._fc_norms.append(bn_op(name=bn_name, **bn_kwargs))
self._classifier = tf.keras.layers.Dense(
units=self._config_dict['num_classes'],
kernel_initializer=tf.keras.initializers.RandomNormal(stddev=0.01),
bias_initializer=tf.zeros_initializer(),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'],
name='iou-scores')
super(MaskScoring, self).build(input_shape)
def call(self, inputs: tf.Tensor, training: bool = None):
"""Forward pass mask scoring head.
Args:
inputs: A `tf.Tensor` of the shape [batch_size, width, size, num_classes],
representing the segmentation logits.
training: a `bool` indicating whether it is in `training` mode.
Returns:
mask_scores: A `tf.Tensor` of predicted mask scores
[batch_size, num_classes].
"""
x = tf.stop_gradient(inputs)
for conv, bn in zip(self._convs, self._conv_norms):
x = conv(x)
x = bn(x)
x = self._activation(x)
# Casts feat to float32 so the resize op can be run on TPU.
x = tf.cast(x, tf.float32)
x = tf.image.resize(x, size=self._config_dict['fc_input_size'],
method=tf.image.ResizeMethod.BILINEAR)
# Casts it back to be compatible with the rest opetations.
x = tf.cast(x, inputs.dtype)
_, h, w, filters = x.get_shape().as_list()
x = tf.reshape(x, [-1, h * w * filters])
for fc, bn in zip(self._fcs, self._fc_norms):
x = fc(x)
x = bn(x)
x = self._activation(x)
ious = self._classifier(x)
return ious
def get_config(self) -> Mapping[str, Any]:
return self._config_dict
@classmethod
def from_config(cls, config, custom_objects=None):
return cls(**config)
@tf.keras.utils.register_keras_serializable(package='Vision')
class SegmentationHead(tf.keras.layers.Layer):
"""Creates a segmentation head."""
def __init__(
self,
num_classes: int,
level: Union[int, str],
num_convs: int = 2,
num_filters: int = 256,
use_depthwise_convolution: bool = False,
prediction_kernel_size: int = 1,
upsample_factor: int = 1,
feature_fusion: Optional[str] = None,
decoder_min_level: Optional[int] = None,
decoder_max_level: Optional[int] = None,
low_level: int = 2,
low_level_num_filters: int = 48,
num_decoder_filters: int = 256,
activation: str = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
bias_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
**kwargs):
"""Initializes a segmentation head.
Args:
num_classes: An `int` number of mask classification categories. The number
of classes does not include background class.
level: An `int` or `str`, level to use to build segmentation head.
num_convs: An `int` number of stacked convolution before the last
prediction layer.
num_filters: An `int` number to specify the number of filters used.
Default is 256.
use_depthwise_convolution: A bool to specify if use depthwise separable
convolutions.
prediction_kernel_size: An `int` number to specify the kernel size of the
prediction layer.
upsample_factor: An `int` number to specify the upsampling factor to
generate finer mask. Default 1 means no upsampling is applied.
feature_fusion: One of `deeplabv3plus`, `pyramid_fusion`,
`panoptic_fpn_fusion`, or None. If `deeplabv3plus`, features from
decoder_features[level] will be fused with low level feature maps from
backbone. If `pyramid_fusion`, multiscale features will be resized and
fused at the target level.
decoder_min_level: An `int` of minimum level from decoder to use in
feature fusion. It is only used when feature_fusion is set to
`panoptic_fpn_fusion`.
decoder_max_level: An `int` of maximum level from decoder to use in
feature fusion. It is only used when feature_fusion is set to
`panoptic_fpn_fusion`.
low_level: An `int` of backbone level to be used for feature fusion. It is
used when feature_fusion is set to `deeplabv3plus`.
low_level_num_filters: An `int` of reduced number of filters for the low
level features before fusing it with higher level features. It is only
used when feature_fusion is set to `deeplabv3plus`.
num_decoder_filters: An `int` of number of filters in the decoder outputs.
It is only used when feature_fusion is set to `panoptic_fpn_fusion`.
activation: A `str` that indicates which activation is used, e.g. 'relu',
'swish', etc.
use_sync_bn: A `bool` that indicates whether to use synchronized batch
normalization across different replicas.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
**kwargs: Additional keyword arguments to be passed.
"""
super(SegmentationHead, self).__init__(**kwargs)
self._config_dict = {
'num_classes': num_classes,
'level': level,
'num_convs': num_convs,
'num_filters': num_filters,
'use_depthwise_convolution': use_depthwise_convolution,
'prediction_kernel_size': prediction_kernel_size,
'upsample_factor': upsample_factor,
'feature_fusion': feature_fusion,
'decoder_min_level': decoder_min_level,
'decoder_max_level': decoder_max_level,
'low_level': low_level,
'low_level_num_filters': low_level_num_filters,
'num_decoder_filters': num_decoder_filters,
'activation': activation,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer
}
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation = tf_utils.get_activation(activation)
def build(self, input_shape: Union[tf.TensorShape, List[tf.TensorShape]]):
"""Creates the variables of the segmentation head."""
use_depthwise_convolution = self._config_dict['use_depthwise_convolution']
random_initializer = tf.keras.initializers.RandomNormal(stddev=0.01)
conv_op = tf.keras.layers.Conv2D
conv_kwargs = {
'kernel_size': 3 if not use_depthwise_convolution else 1,
'padding': 'same',
'use_bias': False,
'kernel_initializer': random_initializer,
'kernel_regularizer': self._config_dict['kernel_regularizer'],
}
bn_op = (tf.keras.layers.experimental.SyncBatchNormalization
if self._config_dict['use_sync_bn']
else tf.keras.layers.BatchNormalization)
bn_kwargs = {
'axis': self._bn_axis,
'momentum': self._config_dict['norm_momentum'],
'epsilon': self._config_dict['norm_epsilon'],
}
if self._config_dict['feature_fusion'] == 'deeplabv3plus':
# Deeplabv3+ feature fusion layers.
self._dlv3p_conv = conv_op(
kernel_size=1,
padding='same',
use_bias=False,
kernel_initializer=tf.keras.initializers.RandomNormal(stddev=0.01),
kernel_regularizer=self._config_dict['kernel_regularizer'],
name='segmentation_head_deeplabv3p_fusion_conv',
filters=self._config_dict['low_level_num_filters'])
self._dlv3p_norm = bn_op(
name='segmentation_head_deeplabv3p_fusion_norm', **bn_kwargs)
elif self._config_dict['feature_fusion'] == 'panoptic_fpn_fusion':
self._panoptic_fpn_fusion = nn_layers.PanopticFPNFusion(
min_level=self._config_dict['decoder_min_level'],
max_level=self._config_dict['decoder_max_level'],
target_level=self._config_dict['level'],
num_filters=self._config_dict['num_filters'],
num_fpn_filters=self._config_dict['num_decoder_filters'],
activation=self._config_dict['activation'],
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'])
# Segmentation head layers.
self._convs = []
self._norms = []
for i in range(self._config_dict['num_convs']):
if use_depthwise_convolution:
self._convs.append(
tf.keras.layers.DepthwiseConv2D(
name='segmentation_head_depthwise_conv_{}'.format(i),
kernel_size=3,
padding='same',
use_bias=False,
depthwise_initializer=random_initializer,
depthwise_regularizer=self._config_dict['kernel_regularizer'],
depth_multiplier=1))
norm_name = 'segmentation_head_depthwise_norm_{}'.format(i)
self._norms.append(bn_op(name=norm_name, **bn_kwargs))
conv_name = 'segmentation_head_conv_{}'.format(i)
self._convs.append(
conv_op(
name=conv_name,
filters=self._config_dict['num_filters'],
**conv_kwargs))
norm_name = 'segmentation_head_norm_{}'.format(i)
self._norms.append(bn_op(name=norm_name, **bn_kwargs))
self._classifier = conv_op(
name='segmentation_output',
filters=self._config_dict['num_classes'],
kernel_size=self._config_dict['prediction_kernel_size'],
padding='same',
bias_initializer=tf.zeros_initializer(),
kernel_initializer=tf.keras.initializers.RandomNormal(stddev=0.01),
kernel_regularizer=self._config_dict['kernel_regularizer'],
bias_regularizer=self._config_dict['bias_regularizer'])
super().build(input_shape)
def call(self, inputs: Tuple[Union[tf.Tensor, Mapping[str, tf.Tensor]],
Union[tf.Tensor, Mapping[str, tf.Tensor]]]):
"""Forward pass of the segmentation head.
It supports both a tuple of 2 tensors or 2 dictionaries. The first is
backbone endpoints, and the second is decoder endpoints. When inputs are
tensors, they are from a single level of feature maps. When inputs are
dictionaries, they contain multiple levels of feature maps, where the key
is the index of feature map.
Args:
inputs: A tuple of 2 feature map tensors of shape
[batch, height_l, width_l, channels] or 2 dictionaries of tensors:
- key: A `str` of the level of the multilevel features.
- values: A `tf.Tensor` of the feature map tensors, whose shape is
[batch, height_l, width_l, channels].
The first is backbone endpoints, and the second is decoder endpoints.
Returns:
segmentation prediction mask: A `tf.Tensor` of the segmentation mask
scores predicted from input features.
"""
backbone_output = inputs[0]
decoder_output = inputs[1]
if self._config_dict['feature_fusion'] == 'deeplabv3plus':
# deeplabv3+ feature fusion
x = decoder_output[str(self._config_dict['level'])] if isinstance(
decoder_output, dict) else decoder_output
y = backbone_output[str(self._config_dict['low_level'])] if isinstance(
backbone_output, dict) else backbone_output
y = self._dlv3p_norm(self._dlv3p_conv(y))
y = self._activation(y)
x = tf.image.resize(
x, tf.shape(y)[1:3], method=tf.image.ResizeMethod.BILINEAR)
x = tf.cast(x, dtype=y.dtype)
x = tf.concat([x, y], axis=self._bn_axis)
elif self._config_dict['feature_fusion'] == 'pyramid_fusion':
if not isinstance(decoder_output, dict):
raise ValueError('Only support dictionary decoder_output.')
x = nn_layers.pyramid_feature_fusion(decoder_output,
self._config_dict['level'])
elif self._config_dict['feature_fusion'] == 'panoptic_fpn_fusion':
x = self._panoptic_fpn_fusion(decoder_output)
else:
x = decoder_output[str(self._config_dict['level'])] if isinstance(
decoder_output, dict) else decoder_output
for conv, norm in zip(self._convs, self._norms):
x = conv(x)
x = norm(x)
x = self._activation(x)
if self._config_dict['upsample_factor'] > 1:
x = spatial_transform_ops.nearest_upsampling(
x, scale=self._config_dict['upsample_factor'])
return self._classifier(x)
def get_config(self):
base_config = super().get_config()
return dict(list(base_config.items()) + list(self._config_dict.items()))
@classmethod
def from_config(cls, config):
return cls(**config)
official/vision/modeling/heads/segmentation_heads_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
# Lint as: python3
"""Tests for segmentation_heads.py."""
# Import libraries
from absl.testing import parameterized
import numpy as np
import tensorflow as tf
from official.vision.modeling.heads import segmentation_heads
class SegmentationHeadTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(2, 'pyramid_fusion', None, None),
(3, 'pyramid_fusion', None, None),
(2, 'panoptic_fpn_fusion', 2, 5),
(2, 'panoptic_fpn_fusion', 2, 6),
(3, 'panoptic_fpn_fusion', 3, 5),
(3, 'panoptic_fpn_fusion', 3, 6))
def test_forward(self, level, feature_fusion,
decoder_min_level, decoder_max_level):
backbone_features = {
'3': np.random.rand(2, 128, 128, 16),
'4': np.random.rand(2, 64, 64, 16),
'5': np.random.rand(2, 32, 32, 16),
}
decoder_features = {
'3': np.random.rand(2, 128, 128, 64),
'4': np.random.rand(2, 64, 64, 64),
'5': np.random.rand(2, 32, 32, 64),
'6': np.random.rand(2, 16, 16, 64),
}
if feature_fusion == 'panoptic_fpn_fusion':
backbone_features['2'] = np.random.rand(2, 256, 256, 16)
decoder_features['2'] = np.random.rand(2, 256, 256, 64)
head = segmentation_heads.SegmentationHead(
num_classes=10,
level=level,
feature_fusion=feature_fusion,
decoder_min_level=decoder_min_level,
decoder_max_level=decoder_max_level,
num_decoder_filters=64)
logits = head((backbone_features, decoder_features))
if level in decoder_features:
self.assertAllEqual(logits.numpy().shape, [
2, decoder_features[str(level)].shape[1],
decoder_features[str(level)].shape[2], 10
])
def test_serialize_deserialize(self):
head = segmentation_heads.SegmentationHead(num_classes=10, level=3)
config = head.get_config()
new_head = segmentation_heads.SegmentationHead.from_config(config)
self.assertAllEqual(head.get_config(), new_head.get_config())
class MaskScoringHeadTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(1, 1, 64, [4, 4]),
(2, 1, 64, [4, 4]),
(3, 1, 64, [4, 4]),
(1, 2, 32, [8, 8]),
(2, 2, 32, [8, 8]),
(3, 2, 32, [8, 8]),)
def test_forward(self, num_convs, num_fcs,
num_filters, fc_input_size):
features = np.random.rand(2, 64, 64, 16)
head = segmentation_heads.MaskScoring(
num_classes=2,
num_convs=num_convs,
num_filters=num_filters,
fc_dims=128,
fc_input_size=fc_input_size)
scores = head(features)
self.assertAllEqual(scores.numpy().shape, [2, 2])
def test_serialize_deserialize(self):
head = segmentation_heads.MaskScoring(
num_classes=2, fc_input_size=[4, 4], fc_dims=128)
config = head.get_config()
new_head = segmentation_heads.MaskScoring.from_config(config)
self.assertAllEqual(head.get_config(), new_head.get_config())
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/__init__.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
# Lint as: python3
"""Layers package definition."""
from official.vision.modeling.layers.box_sampler import BoxSampler
from official.vision.modeling.layers.detection_generator import DetectionGenerator
from official.vision.modeling.layers.detection_generator import MultilevelDetectionGenerator
from official.vision.modeling.layers.mask_sampler import MaskSampler
from official.vision.modeling.layers.nn_blocks import BottleneckBlock
from official.vision.modeling.layers.nn_blocks import BottleneckResidualInner
from official.vision.modeling.layers.nn_blocks import DepthwiseSeparableConvBlock
from official.vision.modeling.layers.nn_blocks import InvertedBottleneckBlock
from official.vision.modeling.layers.nn_blocks import ResidualBlock
from official.vision.modeling.layers.nn_blocks import ResidualInner
from official.vision.modeling.layers.nn_blocks import ReversibleLayer
from official.vision.modeling.layers.nn_blocks_3d import BottleneckBlock3D
from official.vision.modeling.layers.nn_blocks_3d import SelfGating
from official.vision.modeling.layers.nn_layers import CausalConvMixin
from official.vision.modeling.layers.nn_layers import Conv2D
from official.vision.modeling.layers.nn_layers import Conv3D
from official.vision.modeling.layers.nn_layers import DepthwiseConv2D
from official.vision.modeling.layers.nn_layers import GlobalAveragePool3D
from official.vision.modeling.layers.nn_layers import PositionalEncoding
from official.vision.modeling.layers.nn_layers import Scale
from official.vision.modeling.layers.nn_layers import SpatialAveragePool3D
from official.vision.modeling.layers.nn_layers import SqueezeExcitation
from official.vision.modeling.layers.nn_layers import StochasticDepth
from official.vision.modeling.layers.nn_layers import TemporalSoftmaxPool
from official.vision.modeling.layers.roi_aligner import MultilevelROIAligner
from official.vision.modeling.layers.roi_generator import MultilevelROIGenerator
from official.vision.modeling.layers.roi_sampler import ROISampler
official/vision/modeling/layers/box_sampler.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains definitions of box sampler."""
# Import libraries
import tensorflow as tf
from official.vision.ops import sampling_ops
@tf.keras.utils.register_keras_serializable(package='Vision')
class BoxSampler(tf.keras.layers.Layer):
"""Creates a BoxSampler to sample positive and negative boxes."""
def __init__(self,
num_samples: int = 512,
foreground_fraction: float = 0.25,
**kwargs):
"""Initializes a box sampler.
Args:
num_samples: An `int` of the number of sampled boxes per image.
foreground_fraction: A `float` in [0, 1], what percentage of boxes should
be sampled from the positive examples.
**kwargs: Additional keyword arguments passed to Layer.
"""
self._config_dict = {
'num_samples': num_samples,
'foreground_fraction': foreground_fraction,
}
super(BoxSampler, self).__init__(**kwargs)
def call(self, positive_matches: tf.Tensor, negative_matches: tf.Tensor,
ignored_matches: tf.Tensor):
"""Samples and selects positive and negative instances.
Args:
positive_matches: A `bool` tensor of shape of [batch, N] where N is the
number of instances. For each element, `True` means the instance
corresponds to a positive example.
negative_matches: A `bool` tensor of shape of [batch, N] where N is the
number of instances. For each element, `True` means the instance
corresponds to a negative example.
ignored_matches: A `bool` tensor of shape of [batch, N] where N is the
number of instances. For each element, `True` means the instance should
be ignored.
Returns:
A `tf.tensor` of shape of [batch_size, K], storing the indices of the
sampled examples, where K is `num_samples`.
"""
sample_candidates = tf.logical_and(
tf.logical_or(positive_matches, negative_matches),
tf.logical_not(ignored_matches))
sampler = sampling_ops.BalancedPositiveNegativeSampler(
positive_fraction=self._config_dict['foreground_fraction'],
is_static=True)
batch_size = sample_candidates.shape[0]
sampled_indicators = []
for i in range(batch_size):
sampled_indicator = sampler.subsample(
sample_candidates[i],
self._config_dict['num_samples'],
positive_matches[i])
sampled_indicators.append(sampled_indicator)
sampled_indicators = tf.stack(sampled_indicators)
_, selected_indices = tf.nn.top_k(
tf.cast(sampled_indicators, dtype=tf.int32),
k=self._config_dict['num_samples'],
sorted=True)
return selected_indices
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config):
return cls(**config)
official/vision/modeling/layers/box_sampler_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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 roi_sampler.py."""
# Import libraries
import numpy as np
import tensorflow as tf
from official.vision.modeling.layers import box_sampler
class BoxSamplerTest(tf.test.TestCase):
def test_box_sampler(self):
positive_matches = np.array(
[[True, False, False, False, True, True, False],
[False, False, False, False, False, True, True]])
negative_matches = np.array(
[[False, True, True, True, False, False, False],
[True, True, True, True, False, False, False]])
ignored_matches = np.array(
[[False, False, False, False, False, False, True],
[False, False, False, False, True, False, False]])
sampler = box_sampler.BoxSampler(num_samples=2, foreground_fraction=0.5)
# Runs on TPU.
strategy = tf.distribute.TPUStrategy()
with strategy.scope():
selected_indices_tpu = sampler(
positive_matches, negative_matches, ignored_matches)
self.assertEqual(2, tf.shape(selected_indices_tpu)[1])
# Runs on CPU.
selected_indices_cpu = sampler(
positive_matches, negative_matches, ignored_matches)
self.assertEqual(2, tf.shape(selected_indices_cpu)[1])
def test_serialize_deserialize(self):
kwargs = dict(
num_samples=512,
foreground_fraction=0.25,
)
sampler = box_sampler.BoxSampler(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(sampler.get_config(), expected_config)
new_sampler = box_sampler.BoxSampler.from_config(
sampler.get_config())
self.assertAllEqual(sampler.get_config(), new_sampler.get_config())
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/deeplab.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Layers for DeepLabV3."""
import tensorflow as tf
class SpatialPyramidPooling(tf.keras.layers.Layer):
"""Implements the Atrous Spatial Pyramid Pooling.
References:
[Rethinking Atrous Convolution for Semantic Image Segmentation](
https://arxiv.org/pdf/1706.05587.pdf)
[Encoder-Decoder with Atrous Separable Convolution for Semantic Image
Segmentation](https://arxiv.org/pdf/1802.02611.pdf)
"""
def __init__(
self,
output_channels,
dilation_rates,
pool_kernel_size=None,
use_sync_bn=False,
batchnorm_momentum=0.99,
batchnorm_epsilon=0.001,
activation='relu',
dropout=0.5,
kernel_initializer='glorot_uniform',
kernel_regularizer=None,
interpolation='bilinear',
use_depthwise_convolution=False,
**kwargs):
"""Initializes `SpatialPyramidPooling`.
Args:
output_channels: Number of channels produced by SpatialPyramidPooling.
dilation_rates: A list of integers for parallel dilated conv.
pool_kernel_size: A list of integers or None. If None, global average
pooling is applied, otherwise an average pooling of pool_kernel_size
is applied.
use_sync_bn: A bool, whether or not to use sync batch normalization.
batchnorm_momentum: A float for the momentum in BatchNorm. Defaults to
0.99.
batchnorm_epsilon: A float for the epsilon value in BatchNorm. Defaults to
0.001.
activation: A `str` for type of activation to be used. Defaults to 'relu'.
dropout: A float for the dropout rate before output. Defaults to 0.5.
kernel_initializer: Kernel initializer for conv layers. Defaults to
`glorot_uniform`.
kernel_regularizer: Kernel regularizer for conv layers. Defaults to None.
interpolation: The interpolation method for upsampling. Defaults to
`bilinear`.
use_depthwise_convolution: Allows spatial pooling to be separable
depthwise convolusions. [Encoder-Decoder with Atrous Separable
Convolution for Semantic Image Segmentation](
https://arxiv.org/pdf/1802.02611.pdf)
**kwargs: Other keyword arguments for the layer.
"""
super(SpatialPyramidPooling, self).__init__(**kwargs)
self.output_channels = output_channels
self.dilation_rates = dilation_rates
self.use_sync_bn = use_sync_bn
self.batchnorm_momentum = batchnorm_momentum
self.batchnorm_epsilon = batchnorm_epsilon
self.activation = activation
self.dropout = dropout
self.kernel_initializer = tf.keras.initializers.get(kernel_initializer)
self.kernel_regularizer = tf.keras.regularizers.get(kernel_regularizer)
self.interpolation = interpolation
self.input_spec = tf.keras.layers.InputSpec(ndim=4)
self.pool_kernel_size = pool_kernel_size
self.use_depthwise_convolution = use_depthwise_convolution
def build(self, input_shape):
height = input_shape[1]
width = input_shape[2]
channels = input_shape[3]
self.aspp_layers = []
if self.use_sync_bn:
bn_op = tf.keras.layers.experimental.SyncBatchNormalization
else:
bn_op = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
bn_axis = -1
else:
bn_axis = 1
conv_sequential = tf.keras.Sequential([
tf.keras.layers.Conv2D(
filters=self.output_channels, kernel_size=(1, 1),
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.kernel_regularizer,
use_bias=False),
bn_op(
axis=bn_axis,
momentum=self.batchnorm_momentum,
epsilon=self.batchnorm_epsilon),
tf.keras.layers.Activation(self.activation)
])
self.aspp_layers.append(conv_sequential)
for dilation_rate in self.dilation_rates:
leading_layers = []
kernel_size = (3, 3)
if self.use_depthwise_convolution:
leading_layers += [
tf.keras.layers.DepthwiseConv2D(
depth_multiplier=1, kernel_size=kernel_size,
padding='same', depthwise_regularizer=self.kernel_regularizer,
depthwise_initializer=self.kernel_initializer,
dilation_rate=dilation_rate, use_bias=False)
]
kernel_size = (1, 1)
conv_sequential = tf.keras.Sequential(leading_layers + [
tf.keras.layers.Conv2D(
filters=self.output_channels, kernel_size=kernel_size,
padding='same', kernel_regularizer=self.kernel_regularizer,
kernel_initializer=self.kernel_initializer,
dilation_rate=dilation_rate, use_bias=False),
bn_op(axis=bn_axis, momentum=self.batchnorm_momentum,
epsilon=self.batchnorm_epsilon),
tf.keras.layers.Activation(self.activation)])
self.aspp_layers.append(conv_sequential)
if self.pool_kernel_size is None:
pool_sequential = tf.keras.Sequential([
tf.keras.layers.GlobalAveragePooling2D(),
tf.keras.layers.Reshape((1, 1, channels))])
else:
pool_sequential = tf.keras.Sequential([
tf.keras.layers.AveragePooling2D(self.pool_kernel_size)])
pool_sequential.add(
tf.keras.Sequential([
tf.keras.layers.Conv2D(
filters=self.output_channels,
kernel_size=(1, 1),
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.kernel_regularizer,
use_bias=False),
bn_op(
axis=bn_axis,
momentum=self.batchnorm_momentum,
epsilon=self.batchnorm_epsilon),
tf.keras.layers.Activation(self.activation),
tf.keras.layers.experimental.preprocessing.Resizing(
height,
width,
interpolation=self.interpolation,
dtype=tf.float32)
]))
self.aspp_layers.append(pool_sequential)
self.projection = tf.keras.Sequential([
tf.keras.layers.Conv2D(
filters=self.output_channels, kernel_size=(1, 1),
kernel_initializer=self.kernel_initializer,
kernel_regularizer=self.kernel_regularizer,
use_bias=False),
bn_op(
axis=bn_axis,
momentum=self.batchnorm_momentum,
epsilon=self.batchnorm_epsilon),
tf.keras.layers.Activation(self.activation),
tf.keras.layers.Dropout(rate=self.dropout)])
def call(self, inputs, training=None):
if training is None:
training = tf.keras.backend.learning_phase()
result = []
for layer in self.aspp_layers:
result.append(tf.cast(layer(inputs, training=training), inputs.dtype))
result = tf.concat(result, axis=-1)
result = self.projection(result, training=training)
return result
def get_config(self):
config = {
'output_channels': self.output_channels,
'dilation_rates': self.dilation_rates,
'pool_kernel_size': self.pool_kernel_size,
'use_sync_bn': self.use_sync_bn,
'batchnorm_momentum': self.batchnorm_momentum,
'batchnorm_epsilon': self.batchnorm_epsilon,
'activation': self.activation,
'dropout': self.dropout,
'kernel_initializer': tf.keras.initializers.serialize(
self.kernel_initializer),
'kernel_regularizer': tf.keras.regularizers.serialize(
self.kernel_regularizer),
'interpolation': self.interpolation,
}
base_config = super(SpatialPyramidPooling, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
official/vision/modeling/layers/deeplab_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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 ASPP."""
import tensorflow as tf
from tensorflow.python.keras import keras_parameterized
from official.vision.modeling.layers import deeplab
@keras_parameterized.run_all_keras_modes
class DeeplabTest(keras_parameterized.TestCase):
@keras_parameterized.parameterized.parameters(
(None,),
([32, 32],),
)
def test_aspp(self, pool_kernel_size):
inputs = tf.keras.Input(shape=(64, 64, 128), dtype=tf.float32)
layer = deeplab.SpatialPyramidPooling(output_channels=256,
dilation_rates=[6, 12, 18],
pool_kernel_size=None)
output = layer(inputs)
self.assertAllEqual([None, 64, 64, 256], output.shape)
def test_aspp_invalid_shape(self):
inputs = tf.keras.Input(shape=(64, 64), dtype=tf.float32)
layer = deeplab.SpatialPyramidPooling(output_channels=256,
dilation_rates=[6, 12, 18])
with self.assertRaises(ValueError):
_ = layer(inputs)
def test_config_with_custom_name(self):
layer = deeplab.SpatialPyramidPooling(256, [5], name='aspp')
config = layer.get_config()
layer_1 = deeplab.SpatialPyramidPooling.from_config(config)
self.assertEqual(layer_1.name, layer.name)
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/detection_generator.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains definitions of generators to generate the final detections."""
import contextlib
from typing import List, Optional, Mapping
# Import libraries
import tensorflow as tf
from official.vision.ops import box_ops
from official.vision.ops import nms
from official.vision.ops import preprocess_ops
def _generate_detections_v1(boxes: tf.Tensor,
scores: tf.Tensor,
attributes: Optional[Mapping[str,
tf.Tensor]] = None,
pre_nms_top_k: int = 5000,
pre_nms_score_threshold: float = 0.05,
nms_iou_threshold: float = 0.5,
max_num_detections: int = 100,
soft_nms_sigma: Optional[float] = None):
"""Generates the final detections given the model outputs.
The implementation unrolls the batch dimension and process images one by one.
It required the batch dimension to be statically known and it is TPU
compatible.
Args:
boxes: A `tf.Tensor` with shape `[batch_size, N, num_classes, 4]` or
`[batch_size, N, 1, 4]` for box predictions on all feature levels. The
N is the number of total anchors on all levels.
scores: A `tf.Tensor` with shape `[batch_size, N, num_classes]`, which
stacks class probability on all feature levels. The N is the number of
total anchors on all levels. The num_classes is the number of classes
predicted by the model. Note that the class_outputs here is the raw score.
attributes: None or a dict of (attribute_name, attributes) pairs. Each
attributes is a `tf.Tensor` with shape
`[batch_size, N, num_classes, attribute_size]` or
`[batch_size, N, 1, attribute_size]` for attribute predictions on all
feature levels. The N is the number of total anchors on all levels. Can
be None if no attribute learning is required.
pre_nms_top_k: An `int` number of top candidate detections per class before
NMS.
pre_nms_score_threshold: A `float` representing the threshold for deciding
when to remove boxes based on score.
nms_iou_threshold: A `float` representing the threshold for deciding whether
boxes overlap too much with respect to IOU.
max_num_detections: A scalar representing maximum number of boxes retained
over all classes.
soft_nms_sigma: A `float` representing the sigma parameter for Soft NMS.
When soft_nms_sigma=0.0 (which is default), we fall back to standard NMS.
Returns:
nms_boxes: A `float` type `tf.Tensor` of shape
`[batch_size, max_num_detections, 4]` representing top detected boxes in
`[y1, x1, y2, x2]`.
nms_scores: A `float` type `tf.Tensor` of shape
`[batch_size, max_num_detections]` representing sorted confidence scores
for detected boxes. The values are between `[0, 1]`.
nms_classes: An `int` type `tf.Tensor` of shape
`[batch_size, max_num_detections]` representing classes for detected
boxes.
valid_detections: An `int` type `tf.Tensor` of shape `[batch_size]` only the
top `valid_detections` boxes are valid detections.
nms_attributes: None or a dict of (attribute_name, attributes). Each
attribute is a `float` type `tf.Tensor` of shape
`[batch_size, max_num_detections, attribute_size]` representing attribute
predictions for detected boxes. Can be an empty dict if no attribute
learning is required.
"""
with tf.name_scope('generate_detections'):
batch_size = scores.get_shape().as_list()[0]
nmsed_boxes = []
nmsed_classes = []
nmsed_scores = []
valid_detections = []
if attributes:
nmsed_attributes = {att_name: [] for att_name in attributes.keys()}
else:
nmsed_attributes = {}
for i in range(batch_size):
(nmsed_boxes_i, nmsed_scores_i, nmsed_classes_i, valid_detections_i,
nmsed_att_i) = _generate_detections_per_image(
boxes[i],
scores[i],
attributes={
att_name: att[i] for att_name, att in attributes.items()
} if attributes else {},
pre_nms_top_k=pre_nms_top_k,
pre_nms_score_threshold=pre_nms_score_threshold,
nms_iou_threshold=nms_iou_threshold,
max_num_detections=max_num_detections,
soft_nms_sigma=soft_nms_sigma)
nmsed_boxes.append(nmsed_boxes_i)
nmsed_scores.append(nmsed_scores_i)
nmsed_classes.append(nmsed_classes_i)
valid_detections.append(valid_detections_i)
if attributes:
for att_name in attributes.keys():
nmsed_attributes[att_name].append(nmsed_att_i[att_name])
nmsed_boxes = tf.stack(nmsed_boxes, axis=0)
nmsed_scores = tf.stack(nmsed_scores, axis=0)
nmsed_classes = tf.stack(nmsed_classes, axis=0)
valid_detections = tf.stack(valid_detections, axis=0)
if attributes:
for att_name in attributes.keys():
nmsed_attributes[att_name] = tf.stack(nmsed_attributes[att_name], axis=0)
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections, nmsed_attributes
def _generate_detections_per_image(
boxes: tf.Tensor,
scores: tf.Tensor,
attributes: Optional[Mapping[str, tf.Tensor]] = None,
pre_nms_top_k: int = 5000,
pre_nms_score_threshold: float = 0.05,
nms_iou_threshold: float = 0.5,
max_num_detections: int = 100,
soft_nms_sigma: Optional[float] = None):
"""Generates the final detections per image given the model outputs.
Args:
boxes: A `tf.Tensor` with shape `[N, num_classes, 4]` or `[N, 1, 4]`, which
box predictions on all feature levels. The N is the number of total
anchors on all levels.
scores: A `tf.Tensor` with shape `[N, num_classes]`, which stacks class
probability on all feature levels. The N is the number of total anchors on
all levels. The num_classes is the number of classes predicted by the
model. Note that the class_outputs here is the raw score.
attributes: If not None, a dict of `tf.Tensor`. Each value is in shape
`[N, num_classes, attribute_size]` or `[N, 1, attribute_size]` of
attribute predictions on all feature levels. The N is the number of total
anchors on all levels.
pre_nms_top_k: An `int` number of top candidate detections per class before
NMS.
pre_nms_score_threshold: A `float` representing the threshold for deciding
when to remove boxes based on score.
nms_iou_threshold: A `float` representing the threshold for deciding whether
boxes overlap too much with respect to IOU.
max_num_detections: A `scalar` representing maximum number of boxes retained
over all classes.
soft_nms_sigma: A `float` representing the sigma parameter for Soft NMS.
When soft_nms_sigma=0.0, we fall back to standard NMS.
If set to None, `tf.image.non_max_suppression_padded` is called instead.
Returns:
nms_boxes: A `float` tf.Tensor of shape `[max_num_detections, 4]`
representing top detected boxes in `[y1, x1, y2, x2]`.
nms_scores: A `float` tf.Tensor of shape `[max_num_detections]` representing
sorted confidence scores for detected boxes. The values are between [0,
1].
nms_classes: An `int` tf.Tensor of shape `[max_num_detections]` representing
classes for detected boxes.
valid_detections: An `int` tf.Tensor of shape [1] only the top
`valid_detections` boxes are valid detections.
nms_attributes: None or a dict. Each value is a `float` tf.Tensor of shape
`[max_num_detections, attribute_size]` representing attribute predictions
for detected boxes. Can be an empty dict if `attributes` is None.
"""
nmsed_boxes = []
nmsed_scores = []
nmsed_classes = []
num_classes_for_box = boxes.get_shape().as_list()[1]
num_classes = scores.get_shape().as_list()[1]
if attributes:
nmsed_attributes = {att_name: [] for att_name in attributes.keys()}
else:
nmsed_attributes = {}
for i in range(num_classes):
boxes_i = boxes[:, min(num_classes_for_box - 1, i)]
scores_i = scores[:, i]
# Obtains pre_nms_top_k before running NMS.
scores_i, indices = tf.nn.top_k(
scores_i, k=tf.minimum(tf.shape(scores_i)[-1], pre_nms_top_k))
boxes_i = tf.gather(boxes_i, indices)
if soft_nms_sigma is not None:
(nmsed_indices_i,
nmsed_scores_i) = tf.image.non_max_suppression_with_scores(
tf.cast(boxes_i, tf.float32),
tf.cast(scores_i, tf.float32),
max_num_detections,
iou_threshold=nms_iou_threshold,
score_threshold=pre_nms_score_threshold,
soft_nms_sigma=soft_nms_sigma,
name='nms_detections_' + str(i))
nmsed_boxes_i = tf.gather(boxes_i, nmsed_indices_i)
nmsed_boxes_i = preprocess_ops.clip_or_pad_to_fixed_size(
nmsed_boxes_i, max_num_detections, 0.0)
nmsed_scores_i = preprocess_ops.clip_or_pad_to_fixed_size(
nmsed_scores_i, max_num_detections, -1.0)
else:
(nmsed_indices_i,
nmsed_num_valid_i) = tf.image.non_max_suppression_padded(
tf.cast(boxes_i, tf.float32),
tf.cast(scores_i, tf.float32),
max_num_detections,
iou_threshold=nms_iou_threshold,
score_threshold=pre_nms_score_threshold,
pad_to_max_output_size=True,
name='nms_detections_' + str(i))
nmsed_boxes_i = tf.gather(boxes_i, nmsed_indices_i)
nmsed_scores_i = tf.gather(scores_i, nmsed_indices_i)
# Sets scores of invalid boxes to -1.
nmsed_scores_i = tf.where(
tf.less(tf.range(max_num_detections), [nmsed_num_valid_i]),
nmsed_scores_i, -tf.ones_like(nmsed_scores_i))
nmsed_classes_i = tf.fill([max_num_detections], i)
nmsed_boxes.append(nmsed_boxes_i)
nmsed_scores.append(nmsed_scores_i)
nmsed_classes.append(nmsed_classes_i)
if attributes:
for att_name, att in attributes.items():
num_classes_for_attr = att.get_shape().as_list()[1]
att_i = att[:, min(num_classes_for_attr - 1, i)]
att_i = tf.gather(att_i, indices)
nmsed_att_i = tf.gather(att_i, nmsed_indices_i)
nmsed_att_i = preprocess_ops.clip_or_pad_to_fixed_size(
nmsed_att_i, max_num_detections, 0.0)
nmsed_attributes[att_name].append(nmsed_att_i)
# Concats results from all classes and sort them.
nmsed_boxes = tf.concat(nmsed_boxes, axis=0)
nmsed_scores = tf.concat(nmsed_scores, axis=0)
nmsed_classes = tf.concat(nmsed_classes, axis=0)
nmsed_scores, indices = tf.nn.top_k(
nmsed_scores, k=max_num_detections, sorted=True)
nmsed_boxes = tf.gather(nmsed_boxes, indices)
nmsed_classes = tf.gather(nmsed_classes, indices)
valid_detections = tf.reduce_sum(
tf.cast(tf.greater(nmsed_scores, -1), tf.int32))
if attributes:
for att_name in attributes.keys():
nmsed_attributes[att_name] = tf.concat(nmsed_attributes[att_name], axis=0)
nmsed_attributes[att_name] = tf.gather(nmsed_attributes[att_name],
indices)
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections, nmsed_attributes
def _select_top_k_scores(scores_in: tf.Tensor, pre_nms_num_detections: int):
"""Selects top_k scores and indices for each class.
Args:
scores_in: A `tf.Tensor` with shape `[batch_size, N, num_classes]`, which
stacks class logit outputs on all feature levels. The N is the number of
total anchors on all levels. The num_classes is the number of classes
predicted by the model.
pre_nms_num_detections: Number of candidates before NMS.
Returns:
scores and indices: A `tf.Tensor` with shape
`[batch_size, pre_nms_num_detections, num_classes]`.
"""
batch_size, num_anchors, num_class = scores_in.get_shape().as_list()
if batch_size is None:
batch_size = tf.shape(scores_in)[0]
scores_trans = tf.transpose(scores_in, perm=[0, 2, 1])
scores_trans = tf.reshape(scores_trans, [-1, num_anchors])
top_k_scores, top_k_indices = tf.nn.top_k(
scores_trans, k=pre_nms_num_detections, sorted=True)
top_k_scores = tf.reshape(top_k_scores,
[batch_size, num_class, pre_nms_num_detections])
top_k_indices = tf.reshape(top_k_indices,
[batch_size, num_class, pre_nms_num_detections])
return tf.transpose(top_k_scores,
[0, 2, 1]), tf.transpose(top_k_indices, [0, 2, 1])
def _generate_detections_v2(boxes: tf.Tensor,
scores: tf.Tensor,
pre_nms_top_k: int = 5000,
pre_nms_score_threshold: float = 0.05,
nms_iou_threshold: float = 0.5,
max_num_detections: int = 100):
"""Generates the final detections given the model outputs.
This implementation unrolls classes dimension while using the tf.while_loop
to implement the batched NMS, so that it can be parallelized at the batch
dimension. It should give better performance comparing to v1 implementation.
It is TPU compatible.
Args:
boxes: A `tf.Tensor` with shape `[batch_size, N, num_classes, 4]` or
`[batch_size, N, 1, 4]`, which box predictions on all feature levels. The
N is the number of total anchors on all levels.
scores: A `tf.Tensor` with shape `[batch_size, N, num_classes]`, which
stacks class probability on all feature levels. The N is the number of
total anchors on all levels. The num_classes is the number of classes
predicted by the model. Note that the class_outputs here is the raw score.
pre_nms_top_k: An `int` number of top candidate detections per class before
NMS.
pre_nms_score_threshold: A `float` representing the threshold for deciding
when to remove boxes based on score.
nms_iou_threshold: A `float` representing the threshold for deciding whether
boxes overlap too much with respect to IOU.
max_num_detections: A `scalar` representing maximum number of boxes retained
over all classes.
Returns:
nms_boxes: A `float` tf.Tensor of shape [batch_size, max_num_detections, 4]
representing top detected boxes in [y1, x1, y2, x2].
nms_scores: A `float` tf.Tensor of shape [batch_size, max_num_detections]
representing sorted confidence scores for detected boxes. The values are
between [0, 1].
nms_classes: An `int` tf.Tensor of shape [batch_size, max_num_detections]
representing classes for detected boxes.
valid_detections: An `int` tf.Tensor of shape [batch_size] only the top
`valid_detections` boxes are valid detections.
"""
with tf.name_scope('generate_detections'):
nmsed_boxes = []
nmsed_classes = []
nmsed_scores = []
valid_detections = []
batch_size, _, num_classes_for_box, _ = boxes.get_shape().as_list()
if batch_size is None:
batch_size = tf.shape(boxes)[0]
_, total_anchors, num_classes = scores.get_shape().as_list()
# Selects top pre_nms_num scores and indices before NMS.
scores, indices = _select_top_k_scores(
scores, min(total_anchors, pre_nms_top_k))
for i in range(num_classes):
boxes_i = boxes[:, :, min(num_classes_for_box - 1, i), :]
scores_i = scores[:, :, i]
# Obtains pre_nms_top_k before running NMS.
boxes_i = tf.gather(boxes_i, indices[:, :, i], batch_dims=1, axis=1)
# Filter out scores.
boxes_i, scores_i = box_ops.filter_boxes_by_scores(
boxes_i, scores_i, min_score_threshold=pre_nms_score_threshold)
(nmsed_scores_i, nmsed_boxes_i) = nms.sorted_non_max_suppression_padded(
tf.cast(scores_i, tf.float32),
tf.cast(boxes_i, tf.float32),
max_num_detections,
iou_threshold=nms_iou_threshold)
nmsed_classes_i = tf.fill([batch_size, max_num_detections], i)
nmsed_boxes.append(nmsed_boxes_i)
nmsed_scores.append(nmsed_scores_i)
nmsed_classes.append(nmsed_classes_i)
nmsed_boxes = tf.concat(nmsed_boxes, axis=1)
nmsed_scores = tf.concat(nmsed_scores, axis=1)
nmsed_classes = tf.concat(nmsed_classes, axis=1)
nmsed_scores, indices = tf.nn.top_k(
nmsed_scores, k=max_num_detections, sorted=True)
nmsed_boxes = tf.gather(nmsed_boxes, indices, batch_dims=1, axis=1)
nmsed_classes = tf.gather(nmsed_classes, indices, batch_dims=1)
valid_detections = tf.reduce_sum(
input_tensor=tf.cast(tf.greater(nmsed_scores, 0.0), tf.int32), axis=1)
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections
def _generate_detections_batched(boxes: tf.Tensor, scores: tf.Tensor,
pre_nms_score_threshold: float,
nms_iou_threshold: float,
max_num_detections: int):
"""Generates detected boxes with scores and classes for one-stage detector.
The function takes output of multi-level ConvNets and anchor boxes and
generates detected boxes. Note that this used batched nms, which is not
supported on TPU currently.
Args:
boxes: A `tf.Tensor` with shape `[batch_size, N, num_classes, 4]` or
`[batch_size, N, 1, 4]`, which box predictions on all feature levels. The
N is the number of total anchors on all levels.
scores: A `tf.Tensor` with shape `[batch_size, N, num_classes]`, which
stacks class probability on all feature levels. The N is the number of
total anchors on all levels. The num_classes is the number of classes
predicted by the model. Note that the class_outputs here is the raw score.
pre_nms_score_threshold: A `float` representing the threshold for deciding
when to remove boxes based on score.
nms_iou_threshold: A `float` representing the threshold for deciding whether
boxes overlap too much with respect to IOU.
max_num_detections: A `scalar` representing maximum number of boxes retained
over all classes.
Returns:
nms_boxes: A `float` tf.Tensor of shape [batch_size, max_num_detections, 4]
representing top detected boxes in [y1, x1, y2, x2].
nms_scores: A `float` tf.Tensor of shape [batch_size, max_num_detections]
representing sorted confidence scores for detected boxes. The values are
between [0, 1].
nms_classes: An `int` tf.Tensor of shape [batch_size, max_num_detections]
representing classes for detected boxes.
valid_detections: An `int` tf.Tensor of shape [batch_size] only the top
`valid_detections` boxes are valid detections.
"""
with tf.name_scope('generate_detections'):
nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections = (
tf.image.combined_non_max_suppression(
boxes,
scores,
max_output_size_per_class=max_num_detections,
max_total_size=max_num_detections,
iou_threshold=nms_iou_threshold,
score_threshold=pre_nms_score_threshold,
pad_per_class=False,
clip_boxes=False))
nmsed_classes = tf.cast(nmsed_classes, tf.int32)
return nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections
@tf.keras.utils.register_keras_serializable(package='Vision')
class DetectionGenerator(tf.keras.layers.Layer):
"""Generates the final detected boxes with scores and classes."""
def __init__(self,
apply_nms: bool = True,
pre_nms_top_k: int = 5000,
pre_nms_score_threshold: float = 0.05,
nms_iou_threshold: float = 0.5,
max_num_detections: int = 100,
nms_version: str = 'v2',
use_cpu_nms: bool = False,
soft_nms_sigma: Optional[float] = None,
**kwargs):
"""Initializes a detection generator.
Args:
apply_nms: A `bool` of whether or not apply non maximum suppression.
If False, the decoded boxes and their scores are returned.
pre_nms_top_k: An `int` of the number of top scores proposals to be kept
before applying NMS.
pre_nms_score_threshold: A `float` of the score threshold to apply before
applying NMS. Proposals whose scores are below this threshold are
thrown away.
nms_iou_threshold: A `float` in [0, 1], the NMS IoU threshold.
max_num_detections: An `int` of the final number of total detections to
generate.
nms_version: A string of `batched`, `v1` or `v2` specifies NMS version.
use_cpu_nms: A `bool` of whether or not enforce NMS to run on CPU.
soft_nms_sigma: A `float` representing the sigma parameter for Soft NMS.
When soft_nms_sigma=0.0, we fall back to standard NMS.
**kwargs: Additional keyword arguments passed to Layer.
"""
self._config_dict = {
'apply_nms': apply_nms,
'pre_nms_top_k': pre_nms_top_k,
'pre_nms_score_threshold': pre_nms_score_threshold,
'nms_iou_threshold': nms_iou_threshold,
'max_num_detections': max_num_detections,
'nms_version': nms_version,
'use_cpu_nms': use_cpu_nms,
'soft_nms_sigma': soft_nms_sigma,
}
super(DetectionGenerator, self).__init__(**kwargs)
def __call__(self,
raw_boxes: tf.Tensor,
raw_scores: tf.Tensor,
anchor_boxes: tf.Tensor,
image_shape: tf.Tensor,
regression_weights: Optional[List[float]] = None,
bbox_per_class: bool = True):
"""Generates final detections.
Args:
raw_boxes: A `tf.Tensor` of shape of `[batch_size, K, num_classes * 4]`
representing the class-specific box coordinates relative to anchors.
raw_scores: A `tf.Tensor` of shape of `[batch_size, K, num_classes]`
representing the class logits before applying score activiation.
anchor_boxes: A `tf.Tensor` of shape of `[batch_size, K, 4]` representing
the corresponding anchor boxes w.r.t `box_outputs`.
image_shape: A `tf.Tensor` of shape of `[batch_size, 2]` storing the image
height and width w.r.t. the scaled image, i.e. the same image space as
`box_outputs` and `anchor_boxes`.
regression_weights: A list of four float numbers to scale coordinates.
bbox_per_class: A `bool`. If True, perform per-class box regression.
Returns:
If `apply_nms` = True, the return is a dictionary with keys:
`detection_boxes`: A `float` tf.Tensor of shape
[batch, max_num_detections, 4] representing top detected boxes in
[y1, x1, y2, x2].
`detection_scores`: A `float` `tf.Tensor` of shape
[batch, max_num_detections] representing sorted confidence scores for
detected boxes. The values are between [0, 1].
`detection_classes`: An `int` tf.Tensor of shape
[batch, max_num_detections] representing classes for detected boxes.
`num_detections`: An `int` tf.Tensor of shape [batch] only the first
`num_detections` boxes are valid detections
If `apply_nms` = False, the return is a dictionary with keys:
`decoded_boxes`: A `float` tf.Tensor of shape [batch, num_raw_boxes, 4]
representing all the decoded boxes.
`decoded_box_scores`: A `float` tf.Tensor of shape
[batch, num_raw_boxes] representing socres of all the decoded boxes.
"""
box_scores = tf.nn.softmax(raw_scores, axis=-1)
# Removes the background class.
box_scores_shape = tf.shape(box_scores)
box_scores_shape_list = box_scores.get_shape().as_list()
batch_size = box_scores_shape[0]
num_locations = box_scores_shape_list[1]
num_classes = box_scores_shape_list[-1]
box_scores = tf.slice(box_scores, [0, 0, 1], [-1, -1, -1])
if bbox_per_class:
num_detections = num_locations * (num_classes - 1)
raw_boxes = tf.reshape(raw_boxes,
[batch_size, num_locations, num_classes, 4])
raw_boxes = tf.slice(raw_boxes, [0, 0, 1, 0], [-1, -1, -1, -1])
anchor_boxes = tf.tile(
tf.expand_dims(anchor_boxes, axis=2), [1, 1, num_classes - 1, 1])
raw_boxes = tf.reshape(raw_boxes, [batch_size, num_detections, 4])
anchor_boxes = tf.reshape(anchor_boxes, [batch_size, num_detections, 4])
# Box decoding.
decoded_boxes = box_ops.decode_boxes(
raw_boxes, anchor_boxes, weights=regression_weights)
# Box clipping
decoded_boxes = box_ops.clip_boxes(
decoded_boxes, tf.expand_dims(image_shape, axis=1))
if bbox_per_class:
decoded_boxes = tf.reshape(
decoded_boxes, [batch_size, num_locations, num_classes - 1, 4])
else:
decoded_boxes = tf.expand_dims(decoded_boxes, axis=2)
if not self._config_dict['apply_nms']:
return {
'decoded_boxes': decoded_boxes,
'decoded_box_scores': box_scores,
}
# Optionally force the NMS be run on CPU.
if self._config_dict['use_cpu_nms']:
nms_context = tf.device('cpu:0')
else:
nms_context = contextlib.nullcontext()
with nms_context:
if self._config_dict['nms_version'] == 'batched':
(nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
_generate_detections_batched(
decoded_boxes, box_scores,
self._config_dict['pre_nms_score_threshold'],
self._config_dict['nms_iou_threshold'],
self._config_dict['max_num_detections']))
elif self._config_dict['nms_version'] == 'v1':
(nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections, _) = (
_generate_detections_v1(
decoded_boxes,
box_scores,
pre_nms_top_k=self._config_dict['pre_nms_top_k'],
pre_nms_score_threshold=self
._config_dict['pre_nms_score_threshold'],
nms_iou_threshold=self._config_dict['nms_iou_threshold'],
max_num_detections=self._config_dict['max_num_detections'],
soft_nms_sigma=self._config_dict['soft_nms_sigma']))
elif self._config_dict['nms_version'] == 'v2':
(nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
_generate_detections_v2(
decoded_boxes,
box_scores,
pre_nms_top_k=self._config_dict['pre_nms_top_k'],
pre_nms_score_threshold=self
._config_dict['pre_nms_score_threshold'],
nms_iou_threshold=self._config_dict['nms_iou_threshold'],
max_num_detections=self._config_dict['max_num_detections']))
else:
raise ValueError('NMS version {} not supported.'.format(
self._config_dict['nms_version']))
# Adds 1 to offset the background class which has index 0.
nmsed_classes += 1
return {
'num_detections': valid_detections,
'detection_boxes': nmsed_boxes,
'detection_classes': nmsed_classes,
'detection_scores': nmsed_scores,
}
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config):
return cls(**config)
@tf.keras.utils.register_keras_serializable(package='Vision')
class MultilevelDetectionGenerator(tf.keras.layers.Layer):
"""Generates detected boxes with scores and classes for one-stage detector."""
def __init__(self,
apply_nms: bool = True,
pre_nms_top_k: int = 5000,
pre_nms_score_threshold: float = 0.05,
nms_iou_threshold: float = 0.5,
max_num_detections: int = 100,
nms_version: str = 'v1',
use_cpu_nms: bool = False,
soft_nms_sigma: Optional[float] = None,
**kwargs):
"""Initializes a multi-level detection generator.
Args:
apply_nms: A `bool` of whether or not apply non maximum suppression. If
False, the decoded boxes and their scores are returned.
pre_nms_top_k: An `int` of the number of top scores proposals to be kept
before applying NMS.
pre_nms_score_threshold: A `float` of the score threshold to apply before
applying NMS. Proposals whose scores are below this threshold are thrown
away.
nms_iou_threshold: A `float` in [0, 1], the NMS IoU threshold.
max_num_detections: An `int` of the final number of total detections to
generate.
nms_version: A string of `batched`, `v1` or `v2` specifies NMS version
use_cpu_nms: A `bool` of whether or not enforce NMS to run on CPU.
soft_nms_sigma: A `float` representing the sigma parameter for Soft NMS.
When soft_nms_sigma=0.0, we fall back to standard NMS.
**kwargs: Additional keyword arguments passed to Layer.
"""
self._config_dict = {
'apply_nms': apply_nms,
'pre_nms_top_k': pre_nms_top_k,
'pre_nms_score_threshold': pre_nms_score_threshold,
'nms_iou_threshold': nms_iou_threshold,
'max_num_detections': max_num_detections,
'nms_version': nms_version,
'use_cpu_nms': use_cpu_nms,
'soft_nms_sigma': soft_nms_sigma,
}
super(MultilevelDetectionGenerator, self).__init__(**kwargs)
def _decode_multilevel_outputs(
self,
raw_boxes: Mapping[str, tf.Tensor],
raw_scores: Mapping[str, tf.Tensor],
anchor_boxes: tf.Tensor,
image_shape: tf.Tensor,
raw_attributes: Optional[Mapping[str, tf.Tensor]] = None):
"""Collects dict of multilevel boxes, scores, attributes into lists."""
boxes = []
scores = []
if raw_attributes:
attributes = {att_name: [] for att_name in raw_attributes.keys()}
else:
attributes = {}
levels = list(raw_boxes.keys())
min_level = int(min(levels))
max_level = int(max(levels))
for i in range(min_level, max_level + 1):
raw_boxes_i = raw_boxes[str(i)]
raw_scores_i = raw_scores[str(i)]
batch_size = tf.shape(raw_boxes_i)[0]
(_, feature_h_i, feature_w_i,
num_anchors_per_locations_times_4) = raw_boxes_i.get_shape().as_list()
num_locations = feature_h_i * feature_w_i
num_anchors_per_locations = num_anchors_per_locations_times_4 // 4
num_classes = raw_scores_i.get_shape().as_list(
)[-1] // num_anchors_per_locations
# Applies score transformation and remove the implicit background class.
scores_i = tf.sigmoid(
tf.reshape(raw_scores_i, [
batch_size, num_locations * num_anchors_per_locations, num_classes
]))
scores_i = tf.slice(scores_i, [0, 0, 1], [-1, -1, -1])
# Box decoding.
# The anchor boxes are shared for all data in a batch.
# One stage detector only supports class agnostic box regression.
anchor_boxes_i = tf.reshape(
anchor_boxes[str(i)],
[batch_size, num_locations * num_anchors_per_locations, 4])
raw_boxes_i = tf.reshape(
raw_boxes_i,
[batch_size, num_locations * num_anchors_per_locations, 4])
boxes_i = box_ops.decode_boxes(raw_boxes_i, anchor_boxes_i)
# Box clipping.
boxes_i = box_ops.clip_boxes(
boxes_i, tf.expand_dims(image_shape, axis=1))
boxes.append(boxes_i)
scores.append(scores_i)
if raw_attributes:
for att_name, raw_att in raw_attributes.items():
attribute_size = raw_att[str(
i)].get_shape().as_list()[-1] // num_anchors_per_locations
att_i = tf.reshape(raw_att[str(i)], [
batch_size, num_locations * num_anchors_per_locations,
attribute_size
])
attributes[att_name].append(att_i)
boxes = tf.concat(boxes, axis=1)
boxes = tf.expand_dims(boxes, axis=2)
scores = tf.concat(scores, axis=1)
if raw_attributes:
for att_name in raw_attributes.keys():
attributes[att_name] = tf.concat(attributes[att_name], axis=1)
attributes[att_name] = tf.expand_dims(attributes[att_name], axis=2)
return boxes, scores, attributes
def __call__(self,
raw_boxes: Mapping[str, tf.Tensor],
raw_scores: Mapping[str, tf.Tensor],
anchor_boxes: tf.Tensor,
image_shape: tf.Tensor,
raw_attributes: Optional[Mapping[str, tf.Tensor]] = None):
"""Generates final detections.
Args:
raw_boxes: A `dict` with keys representing FPN levels and values
representing box tenors of shape `[batch, feature_h, feature_w,
num_anchors * 4]`.
raw_scores: A `dict` with keys representing FPN levels and values
representing logit tensors of shape `[batch, feature_h, feature_w,
num_anchors]`.
anchor_boxes: A `tf.Tensor` of shape of [batch_size, K, 4] representing
the corresponding anchor boxes w.r.t `box_outputs`.
image_shape: A `tf.Tensor` of shape of [batch_size, 2] storing the image
height and width w.r.t. the scaled image, i.e. the same image space as
`box_outputs` and `anchor_boxes`.
raw_attributes: If not None, a `dict` of (attribute_name,
attribute_prediction) pairs. `attribute_prediction` is a dict that
contains keys representing FPN levels and values representing tenors of
shape `[batch, feature_h, feature_w, num_anchors * attribute_size]`.
Returns:
If `apply_nms` = True, the return is a dictionary with keys:
`detection_boxes`: A `float` tf.Tensor of shape
[batch, max_num_detections, 4] representing top detected boxes in
[y1, x1, y2, x2].
`detection_scores`: A `float` tf.Tensor of shape
[batch, max_num_detections] representing sorted confidence scores for
detected boxes. The values are between [0, 1].
`detection_classes`: An `int` tf.Tensor of shape
[batch, max_num_detections] representing classes for detected boxes.
`num_detections`: An `int` tf.Tensor of shape [batch] only the first
`num_detections` boxes are valid detections
`detection_attributes`: A dict. Values of the dict is a `float`
tf.Tensor of shape [batch, max_num_detections, attribute_size]
representing attribute predictions for detected boxes.
If `apply_nms` = False, the return is a dictionary with keys:
`decoded_boxes`: A `float` tf.Tensor of shape [batch, num_raw_boxes, 4]
representing all the decoded boxes.
`decoded_box_scores`: A `float` tf.Tensor of shape
[batch, num_raw_boxes] representing socres of all the decoded boxes.
`decoded_box_attributes`: A dict. Values in the dict is a
`float` tf.Tensor of shape [batch, num_raw_boxes, attribute_size]
representing attribute predictions of all the decoded boxes.
"""
boxes, scores, attributes = self._decode_multilevel_outputs(
raw_boxes, raw_scores, anchor_boxes, image_shape, raw_attributes)
if not self._config_dict['apply_nms']:
return {
'decoded_boxes': boxes,
'decoded_box_scores': scores,
'decoded_box_attributes': attributes,
}
# Optionally force the NMS to run on CPU.
if self._config_dict['use_cpu_nms']:
nms_context = tf.device('cpu:0')
else:
nms_context = contextlib.nullcontext()
with nms_context:
if raw_attributes and (self._config_dict['nms_version'] != 'v1'):
raise ValueError(
'Attribute learning is only supported for NMSv1 but NMS {} is used.'
.format(self._config_dict['nms_version']))
if self._config_dict['nms_version'] == 'batched':
(nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
_generate_detections_batched(
boxes, scores, self._config_dict['pre_nms_score_threshold'],
self._config_dict['nms_iou_threshold'],
self._config_dict['max_num_detections']))
# Set `nmsed_attributes` to None for batched NMS.
nmsed_attributes = {}
elif self._config_dict['nms_version'] == 'v1':
(nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections,
nmsed_attributes) = (
_generate_detections_v1(
boxes,
scores,
attributes=attributes if raw_attributes else None,
pre_nms_top_k=self._config_dict['pre_nms_top_k'],
pre_nms_score_threshold=self
._config_dict['pre_nms_score_threshold'],
nms_iou_threshold=self._config_dict['nms_iou_threshold'],
max_num_detections=self._config_dict['max_num_detections'],
soft_nms_sigma=self._config_dict['soft_nms_sigma']))
elif self._config_dict['nms_version'] == 'v2':
(nmsed_boxes, nmsed_scores, nmsed_classes, valid_detections) = (
_generate_detections_v2(
boxes,
scores,
pre_nms_top_k=self._config_dict['pre_nms_top_k'],
pre_nms_score_threshold=self
._config_dict['pre_nms_score_threshold'],
nms_iou_threshold=self._config_dict['nms_iou_threshold'],
max_num_detections=self._config_dict['max_num_detections']))
# Set `nmsed_attributes` to None for v2.
nmsed_attributes = {}
else:
raise ValueError('NMS version {} not supported.'.format(
self._config_dict['nms_version']))
# Adds 1 to offset the background class which has index 0.
nmsed_classes += 1
return {
'num_detections': valid_detections,
'detection_boxes': nmsed_boxes,
'detection_classes': nmsed_classes,
'detection_scores': nmsed_scores,
'detection_attributes': nmsed_attributes,
}
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config):
return cls(**config)
official/vision/modeling/layers/detection_generator_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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 detection_generator.py."""
# Import libraries
from absl.testing import parameterized
import numpy as np
import tensorflow as tf
from official.vision.modeling.layers import detection_generator
from official.vision.ops import anchor
class SelectTopKScoresTest(tf.test.TestCase):
def testSelectTopKScores(self):
pre_nms_num_boxes = 2
scores_data = [[[0.2, 0.2], [0.1, 0.9], [0.5, 0.1], [0.3, 0.5]]]
scores_in = tf.constant(scores_data, dtype=tf.float32)
top_k_scores, top_k_indices = detection_generator._select_top_k_scores(
scores_in, pre_nms_num_detections=pre_nms_num_boxes)
expected_top_k_scores = np.array([[[0.5, 0.9], [0.3, 0.5]]],
dtype=np.float32)
expected_top_k_indices = [[[2, 1], [3, 3]]]
self.assertAllEqual(top_k_scores.numpy(), expected_top_k_scores)
self.assertAllEqual(top_k_indices.numpy(), expected_top_k_indices)
class DetectionGeneratorTest(
parameterized.TestCase, tf.test.TestCase):
@parameterized.product(
nms_version=['batched', 'v1', 'v2'],
use_cpu_nms=[True, False],
soft_nms_sigma=[None, 0.1])
def testDetectionsOutputShape(self, nms_version, use_cpu_nms, soft_nms_sigma):
max_num_detections = 10
num_classes = 4
pre_nms_top_k = 5000
pre_nms_score_threshold = 0.01
batch_size = 1
kwargs = {
'apply_nms': True,
'pre_nms_top_k': pre_nms_top_k,
'pre_nms_score_threshold': pre_nms_score_threshold,
'nms_iou_threshold': 0.5,
'max_num_detections': max_num_detections,
'nms_version': nms_version,
'use_cpu_nms': use_cpu_nms,
'soft_nms_sigma': soft_nms_sigma,
}
generator = detection_generator.DetectionGenerator(**kwargs)
cls_outputs_all = (
np.random.rand(84, num_classes) - 0.5) * 3 # random 84x3 outputs.
box_outputs_all = np.random.rand(84, 4 * num_classes) # random 84 boxes.
anchor_boxes_all = np.random.rand(84, 4) # random 84 boxes.
class_outputs = tf.reshape(
tf.convert_to_tensor(cls_outputs_all, dtype=tf.float32),
[1, 84, num_classes])
box_outputs = tf.reshape(
tf.convert_to_tensor(box_outputs_all, dtype=tf.float32),
[1, 84, 4 * num_classes])
anchor_boxes = tf.reshape(
tf.convert_to_tensor(anchor_boxes_all, dtype=tf.float32),
[1, 84, 4])
image_info = tf.constant(
[[[1000, 1000], [100, 100], [0.1, 0.1], [0, 0]]],
dtype=tf.float32)
results = generator(
box_outputs, class_outputs, anchor_boxes, image_info[:, 1, :])
boxes = results['detection_boxes']
classes = results['detection_classes']
scores = results['detection_scores']
valid_detections = results['num_detections']
self.assertEqual(boxes.numpy().shape, (batch_size, max_num_detections, 4))
self.assertEqual(scores.numpy().shape, (batch_size, max_num_detections,))
self.assertEqual(classes.numpy().shape, (batch_size, max_num_detections,))
self.assertEqual(valid_detections.numpy().shape, (batch_size,))
def test_serialize_deserialize(self):
kwargs = {
'apply_nms': True,
'pre_nms_top_k': 1000,
'pre_nms_score_threshold': 0.1,
'nms_iou_threshold': 0.5,
'max_num_detections': 10,
'nms_version': 'v2',
'use_cpu_nms': False,
'soft_nms_sigma': None,
}
generator = detection_generator.DetectionGenerator(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(generator.get_config(), expected_config)
new_generator = (
detection_generator.DetectionGenerator.from_config(
generator.get_config()))
self.assertAllEqual(generator.get_config(), new_generator.get_config())
class MultilevelDetectionGeneratorTest(
parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
('batched', False, True, None),
('batched', False, False, None),
('v2', False, True, None),
('v2', False, False, None),
('v1', True, True, 0.0),
('v1', True, False, 0.1),
('v1', True, False, None),
)
def testDetectionsOutputShape(self, nms_version, has_att_heads, use_cpu_nms,
soft_nms_sigma):
min_level = 4
max_level = 6
num_scales = 2
max_num_detections = 10
aspect_ratios = [1.0, 2.0]
anchor_scale = 2.0
output_size = [64, 64]
num_classes = 4
pre_nms_top_k = 5000
pre_nms_score_threshold = 0.01
batch_size = 1
kwargs = {
'apply_nms': True,
'pre_nms_top_k': pre_nms_top_k,
'pre_nms_score_threshold': pre_nms_score_threshold,
'nms_iou_threshold': 0.5,
'max_num_detections': max_num_detections,
'nms_version': nms_version,
'use_cpu_nms': use_cpu_nms,
'soft_nms_sigma': soft_nms_sigma,
}
input_anchor = anchor.build_anchor_generator(min_level, max_level,
num_scales, aspect_ratios,
anchor_scale)
anchor_boxes = input_anchor(output_size)
cls_outputs_all = (
np.random.rand(84, num_classes) - 0.5) * 3 # random 84x3 outputs.
box_outputs_all = np.random.rand(84, 4) # random 84 boxes.
class_outputs = {
'4':
tf.reshape(
tf.convert_to_tensor(cls_outputs_all[0:64], dtype=tf.float32),
[1, 8, 8, num_classes]),
'5':
tf.reshape(
tf.convert_to_tensor(cls_outputs_all[64:80], dtype=tf.float32),
[1, 4, 4, num_classes]),
'6':
tf.reshape(
tf.convert_to_tensor(cls_outputs_all[80:84], dtype=tf.float32),
[1, 2, 2, num_classes]),
}
box_outputs = {
'4': tf.reshape(tf.convert_to_tensor(
box_outputs_all[0:64], dtype=tf.float32), [1, 8, 8, 4]),
'5': tf.reshape(tf.convert_to_tensor(
box_outputs_all[64:80], dtype=tf.float32), [1, 4, 4, 4]),
'6': tf.reshape(tf.convert_to_tensor(
box_outputs_all[80:84], dtype=tf.float32), [1, 2, 2, 4]),
}
if has_att_heads:
att_outputs_all = np.random.rand(84, 1) # random attributes.
att_outputs = {
'depth': {
'4':
tf.reshape(
tf.convert_to_tensor(
att_outputs_all[0:64], dtype=tf.float32),
[1, 8, 8, 1]),
'5':
tf.reshape(
tf.convert_to_tensor(
att_outputs_all[64:80], dtype=tf.float32),
[1, 4, 4, 1]),
'6':
tf.reshape(
tf.convert_to_tensor(
att_outputs_all[80:84], dtype=tf.float32),
[1, 2, 2, 1]),
}
}
else:
att_outputs = None
image_info = tf.constant([[[1000, 1000], [100, 100], [0.1, 0.1], [0, 0]]],
dtype=tf.float32)
generator = detection_generator.MultilevelDetectionGenerator(**kwargs)
results = generator(box_outputs, class_outputs, anchor_boxes,
image_info[:, 1, :], att_outputs)
boxes = results['detection_boxes']
classes = results['detection_classes']
scores = results['detection_scores']
valid_detections = results['num_detections']
self.assertEqual(boxes.numpy().shape, (batch_size, max_num_detections, 4))
self.assertEqual(scores.numpy().shape, (batch_size, max_num_detections,))
self.assertEqual(classes.numpy().shape, (batch_size, max_num_detections,))
self.assertEqual(valid_detections.numpy().shape, (batch_size,))
if has_att_heads:
for att in results['detection_attributes'].values():
self.assertEqual(att.numpy().shape, (batch_size, max_num_detections, 1))
def test_serialize_deserialize(self):
kwargs = {
'apply_nms': True,
'pre_nms_top_k': 1000,
'pre_nms_score_threshold': 0.1,
'nms_iou_threshold': 0.5,
'max_num_detections': 10,
'nms_version': 'v2',
'use_cpu_nms': False,
'soft_nms_sigma': None,
}
generator = detection_generator.MultilevelDetectionGenerator(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(generator.get_config(), expected_config)
new_generator = (
detection_generator.MultilevelDetectionGenerator.from_config(
generator.get_config()))
self.assertAllEqual(generator.get_config(), new_generator.get_config())
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/mask_sampler.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains definitions of mask sampler."""
# Import libraries
import tensorflow as tf
from official.vision.ops import spatial_transform_ops
def _sample_and_crop_foreground_masks(candidate_rois: tf.Tensor,
candidate_gt_boxes: tf.Tensor,
candidate_gt_classes: tf.Tensor,
candidate_gt_indices: tf.Tensor,
gt_masks: tf.Tensor,
num_sampled_masks: int = 128,
mask_target_size: int = 28):
"""Samples and creates cropped foreground masks for training.
Args:
candidate_rois: A `tf.Tensor` of shape of [batch_size, N, 4], where N is the
number of candidate RoIs to be considered for mask sampling. It includes
both positive and negative RoIs. The `num_mask_samples_per_image` positive
RoIs will be sampled to create mask training targets.
candidate_gt_boxes: A `tf.Tensor` of shape of [batch_size, N, 4], storing
the corresponding groundtruth boxes to the `candidate_rois`.
candidate_gt_classes: A `tf.Tensor` of shape of [batch_size, N], storing the
corresponding groundtruth classes to the `candidate_rois`. 0 in the tensor
corresponds to the background class, i.e. negative RoIs.
candidate_gt_indices: A `tf.Tensor` of shape [batch_size, N], storing the
corresponding groundtruth instance indices to the `candidate_gt_boxes`,
i.e. gt_boxes[candidate_gt_indices[:, i]] = candidate_gt_boxes[:, i] and
gt_boxes which is of shape [batch_size, MAX_INSTANCES, 4], M >= N, is
the superset of candidate_gt_boxes.
gt_masks: A `tf.Tensor` of [batch_size, MAX_INSTANCES, mask_height,
mask_width] containing all the groundtruth masks which sample masks are
drawn from.
num_sampled_masks: An `int` that specifies the number of masks to sample.
mask_target_size: An `int` that specifies the final cropped mask size after
sampling. The output masks are resized w.r.t the sampled RoIs.
Returns:
foreground_rois: A `tf.Tensor` of shape of [batch_size, K, 4] storing the
RoI that corresponds to the sampled foreground masks, where
K = num_mask_samples_per_image.
foreground_classes: A `tf.Tensor` of shape of [batch_size, K] storing the
classes corresponding to the sampled foreground masks.
cropoped_foreground_masks: A `tf.Tensor` of shape of
[batch_size, K, mask_target_size, mask_target_size] storing the cropped
foreground masks used for training.
"""
_, fg_instance_indices = tf.nn.top_k(
tf.cast(tf.greater(candidate_gt_classes, 0), dtype=tf.int32),
k=num_sampled_masks)
fg_instance_indices_shape = tf.shape(fg_instance_indices)
batch_indices = (
tf.expand_dims(tf.range(fg_instance_indices_shape[0]), axis=-1) *
tf.ones([1, fg_instance_indices_shape[-1]], dtype=tf.int32))
gather_nd_instance_indices = tf.stack(
[batch_indices, fg_instance_indices], axis=-1)
foreground_rois = tf.gather_nd(
candidate_rois, gather_nd_instance_indices)
foreground_boxes = tf.gather_nd(
candidate_gt_boxes, gather_nd_instance_indices)
foreground_classes = tf.gather_nd(
candidate_gt_classes, gather_nd_instance_indices)
foreground_gt_indices = tf.gather_nd(
candidate_gt_indices, gather_nd_instance_indices)
foreground_gt_indices = tf.where(
tf.equal(foreground_gt_indices, -1),
tf.zeros_like(foreground_gt_indices),
foreground_gt_indices)
foreground_gt_indices_shape = tf.shape(foreground_gt_indices)
batch_indices = (
tf.expand_dims(tf.range(foreground_gt_indices_shape[0]), axis=-1) *
tf.ones([1, foreground_gt_indices_shape[-1]], dtype=tf.int32))
gather_nd_gt_indices = tf.stack(
[batch_indices, foreground_gt_indices], axis=-1)
foreground_masks = tf.gather_nd(gt_masks, gather_nd_gt_indices)
cropped_foreground_masks = spatial_transform_ops.crop_mask_in_target_box(
foreground_masks, foreground_boxes, foreground_rois, mask_target_size,
sample_offset=0.5)
return foreground_rois, foreground_classes, cropped_foreground_masks
@tf.keras.utils.register_keras_serializable(package='Vision')
class MaskSampler(tf.keras.layers.Layer):
"""Samples and creates mask training targets."""
def __init__(self, mask_target_size: int, num_sampled_masks: int, **kwargs):
self._config_dict = {
'mask_target_size': mask_target_size,
'num_sampled_masks': num_sampled_masks,
}
super(MaskSampler, self).__init__(**kwargs)
def call(self, candidate_rois: tf.Tensor, candidate_gt_boxes: tf.Tensor,
candidate_gt_classes: tf.Tensor, candidate_gt_indices: tf.Tensor,
gt_masks: tf.Tensor):
"""Samples and creates mask targets for training.
Args:
candidate_rois: A `tf.Tensor` of shape of [batch_size, N, 4], where N is
the number of candidate RoIs to be considered for mask sampling. It
includes both positive and negative RoIs. The
`num_mask_samples_per_image` positive RoIs will be sampled to create
mask training targets.
candidate_gt_boxes: A `tf.Tensor` of shape of [batch_size, N, 4], storing
the corresponding groundtruth boxes to the `candidate_rois`.
candidate_gt_classes: A `tf.Tensor` of shape of [batch_size, N], storing
the corresponding groundtruth classes to the `candidate_rois`. 0 in the
tensor corresponds to the background class, i.e. negative RoIs.
candidate_gt_indices: A `tf.Tensor` of shape [batch_size, N], storing the
corresponding groundtruth instance indices to the `candidate_gt_boxes`,
i.e. gt_boxes[candidate_gt_indices[:, i]] = candidate_gt_boxes[:, i],
where gt_boxes which is of shape [batch_size, MAX_INSTANCES, 4], M >=
N, is the superset of candidate_gt_boxes.
gt_masks: A `tf.Tensor` of [batch_size, MAX_INSTANCES, mask_height,
mask_width] containing all the groundtruth masks which sample masks are
drawn from. after sampling. The output masks are resized w.r.t the
sampled RoIs.
Returns:
foreground_rois: A `tf.Tensor` of shape of [batch_size, K, 4] storing the
RoI that corresponds to the sampled foreground masks, where
K = num_mask_samples_per_image.
foreground_classes: A `tf.Tensor` of shape of [batch_size, K] storing the
classes corresponding to the sampled foreground masks.
cropoped_foreground_masks: A `tf.Tensor` of shape of
[batch_size, K, mask_target_size, mask_target_size] storing the
cropped foreground masks used for training.
"""
foreground_rois, foreground_classes, cropped_foreground_masks = (
_sample_and_crop_foreground_masks(
candidate_rois,
candidate_gt_boxes,
candidate_gt_classes,
candidate_gt_indices,
gt_masks,
self._config_dict['num_sampled_masks'],
self._config_dict['mask_target_size']))
return foreground_rois, foreground_classes, cropped_foreground_masks
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config):
return cls(**config)
official/vision/modeling/layers/mask_sampler_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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 mask_sampler.py."""
# Import libraries
import numpy as np
import tensorflow as tf
from official.vision.modeling.layers import mask_sampler
class SampleAndCropForegroundMasksTest(tf.test.TestCase):
def test_sample_and_crop_foreground_masks(self):
candidate_rois_np = np.array(
[[[0, 0, 0.5, 0.5], [0.5, 0.5, 1, 1],
[2, 2, 4, 4], [1, 1, 5, 5]]])
candidate_rois = tf.constant(candidate_rois_np, dtype=tf.float32)
candidate_gt_boxes_np = np.array(
[[[0, 0, 0.6, 0.6], [0, 0, 0, 0],
[1, 1, 3, 3], [1, 1, 3, 3]]])
candidate_gt_boxes = tf.constant(candidate_gt_boxes_np, dtype=tf.float32)
candidate_gt_classes_np = np.array([[4, 0, 0, 2]])
candidate_gt_classes = tf.constant(
candidate_gt_classes_np, dtype=tf.float32)
candidate_gt_indices_np = np.array([[10, -1, -1, 20]])
candidate_gt_indices = tf.constant(
candidate_gt_indices_np, dtype=tf.int32)
gt_masks_np = np.random.rand(1, 100, 32, 32)
gt_masks = tf.constant(gt_masks_np, dtype=tf.float32)
num_mask_samples_per_image = 2
mask_target_size = 28
# Runs on TPU.
strategy = tf.distribute.TPUStrategy()
with strategy.scope():
foreground_rois, foreground_classes, cropped_foreground_masks = (
mask_sampler._sample_and_crop_foreground_masks(
candidate_rois, candidate_gt_boxes, candidate_gt_classes,
candidate_gt_indices, gt_masks, num_mask_samples_per_image,
mask_target_size))
foreground_rois_tpu = foreground_rois.numpy()
foreground_classes_tpu = foreground_classes.numpy()
cropped_foreground_masks_tpu = cropped_foreground_masks.numpy()
foreground_rois, foreground_classes, cropped_foreground_masks = (
mask_sampler._sample_and_crop_foreground_masks(
candidate_rois, candidate_gt_boxes, candidate_gt_classes,
candidate_gt_indices, gt_masks, num_mask_samples_per_image,
mask_target_size))
foreground_rois_cpu = foreground_rois.numpy()
foreground_classes_cpu = foreground_classes.numpy()
cropped_foreground_masks_cpu = cropped_foreground_masks.numpy()
# consistency.
self.assertAllEqual(foreground_rois_tpu.shape, foreground_rois_cpu.shape)
self.assertAllEqual(
foreground_classes_tpu.shape, foreground_classes_cpu.shape)
self.assertAllEqual(
cropped_foreground_masks_tpu.shape, cropped_foreground_masks_cpu.shape)
# correctnesss.
self.assertAllEqual(foreground_rois_tpu.shape, [1, 2, 4])
self.assertAllEqual(foreground_classes_tpu.shape, [1, 2])
self.assertAllEqual(cropped_foreground_masks_tpu.shape, [1, 2, 28, 28])
class MaskSamplerTest(tf.test.TestCase):
def test_mask_sampler(self):
candidate_rois_np = np.array(
[[[0, 0, 0.5, 0.5], [0.5, 0.5, 1, 1],
[2, 2, 4, 4], [1, 1, 5, 5]]])
candidate_rois = tf.constant(candidate_rois_np, dtype=tf.float32)
candidate_gt_boxes_np = np.array(
[[[0, 0, 0.6, 0.6], [0, 0, 0, 0],
[1, 1, 3, 3], [1, 1, 3, 3]]])
candidate_gt_boxes = tf.constant(candidate_gt_boxes_np, dtype=tf.float32)
candidate_gt_classes_np = np.array([[4, 0, 0, 2]])
candidate_gt_classes = tf.constant(
candidate_gt_classes_np, dtype=tf.float32)
candidate_gt_indices_np = np.array([[10, -1, -1, 20]])
candidate_gt_indices = tf.constant(
candidate_gt_indices_np, dtype=tf.int32)
gt_masks_np = np.random.rand(1, 100, 32, 32)
gt_masks = tf.constant(gt_masks_np, dtype=tf.float32)
sampler = mask_sampler.MaskSampler(28, 2)
foreground_rois, foreground_classes, cropped_foreground_masks = sampler(
candidate_rois, candidate_gt_boxes, candidate_gt_classes,
candidate_gt_indices, gt_masks)
# correctnesss.
self.assertAllEqual(foreground_rois.numpy().shape, [1, 2, 4])
self.assertAllEqual(foreground_classes.numpy().shape, [1, 2])
self.assertAllEqual(cropped_foreground_masks.numpy().shape, [1, 2, 28, 28])
def test_serialize_deserialize(self):
kwargs = dict(
mask_target_size=7,
num_sampled_masks=10,
)
sampler = mask_sampler.MaskSampler(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(sampler.get_config(), expected_config)
new_sampler = mask_sampler.MaskSampler.from_config(
sampler.get_config())
self.assertAllEqual(sampler.get_config(), new_sampler.get_config())
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/nn_blocks.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains common building blocks for neural networks."""
from typing import Any, Callable, Dict, List, Optional, Tuple, Union, Text
# Import libraries
from absl import logging
import tensorflow as tf
from official.modeling import tf_utils
from official.vision.modeling.layers import nn_layers
def _pad_strides(strides: int, axis: int) -> Tuple[int, int, int, int]:
"""Converts int to len 4 strides (`tf.nn.avg_pool` uses length 4)."""
if axis == 1:
return (1, 1, strides, strides)
else:
return (1, strides, strides, 1)
def _maybe_downsample(x: tf.Tensor, out_filter: int, strides: int,
axis: int) -> tf.Tensor:
"""Downsamples feature map and 0-pads tensor if in_filter != out_filter."""
data_format = 'NCHW' if axis == 1 else 'NHWC'
strides = _pad_strides(strides, axis=axis)
x = tf.nn.avg_pool(x, strides, strides, 'VALID', data_format=data_format)
in_filter = x.shape[axis]
if in_filter < out_filter:
# Pad on channel dimension with 0s: half on top half on bottom.
pad_size = [(out_filter - in_filter) // 2, (out_filter - in_filter) // 2]
if axis == 1:
x = tf.pad(x, [[0, 0], pad_size, [0, 0], [0, 0]])
else:
x = tf.pad(x, [[0, 0], [0, 0], [0, 0], pad_size])
return x + 0.
@tf.keras.utils.register_keras_serializable(package='Vision')
class ResidualBlock(tf.keras.layers.Layer):
"""A residual block."""
def __init__(self,
filters,
strides,
use_projection=False,
se_ratio=None,
resnetd_shortcut=False,
stochastic_depth_drop_rate=None,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
activation='relu',
use_explicit_padding: bool = False,
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
bn_trainable=True,
**kwargs):
"""Initializes a residual block with BN after convolutions.
Args:
filters: An `int` number of filters for the first two convolutions. Note
that the third and final convolution will use 4 times as many filters.
strides: An `int` block stride. If greater than 1, this block will
ultimately downsample the input.
use_projection: A `bool` for whether this block should use a projection
shortcut (versus the default identity shortcut). This is usually `True`
for the first block of a block group, which may change the number of
filters and the resolution.
se_ratio: A `float` or None. Ratio of the Squeeze-and-Excitation layer.
resnetd_shortcut: A `bool` if True, apply the resnetd style modification
to the shortcut connection. Not implemented in residual blocks.
stochastic_depth_drop_rate: A `float` or None. if not None, drop rate for
the stochastic depth layer.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2d.
Default to None.
activation: A `str` name of the activation function.
use_explicit_padding: Use 'VALID' padding for convolutions, but prepad
inputs so that the output dimensions are the same as if 'SAME' padding
were used.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
bn_trainable: A `bool` that indicates whether batch norm layers should be
trainable. Default to True.
**kwargs: Additional keyword arguments to be passed.
"""
super(ResidualBlock, self).__init__(**kwargs)
self._filters = filters
self._strides = strides
self._use_projection = use_projection
self._se_ratio = se_ratio
self._resnetd_shortcut = resnetd_shortcut
self._use_explicit_padding = use_explicit_padding
self._use_sync_bn = use_sync_bn
self._activation = activation
self._stochastic_depth_drop_rate = stochastic_depth_drop_rate
self._kernel_initializer = kernel_initializer
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation_fn = tf_utils.get_activation(activation)
self._bn_trainable = bn_trainable
def build(self, input_shape):
if self._use_projection:
self._shortcut = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=1,
strides=self._strides,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
conv1_padding = 'same'
# explicit padding here is added for centernet
if self._use_explicit_padding:
self._pad = tf.keras.layers.ZeroPadding2D(padding=(1, 1))
conv1_padding = 'valid'
self._conv1 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=3,
strides=self._strides,
padding=conv1_padding,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._conv2 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=3,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters,
out_filters=self._filters,
se_ratio=self._se_ratio,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
super(ResidualBlock, self).build(input_shape)
def get_config(self):
config = {
'filters': self._filters,
'strides': self._strides,
'use_projection': self._use_projection,
'se_ratio': self._se_ratio,
'resnetd_shortcut': self._resnetd_shortcut,
'stochastic_depth_drop_rate': self._stochastic_depth_drop_rate,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'bias_regularizer': self._bias_regularizer,
'activation': self._activation,
'use_explicit_padding': self._use_explicit_padding,
'use_sync_bn': self._use_sync_bn,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon,
'bn_trainable': self._bn_trainable
}
base_config = super(ResidualBlock, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, training=None):
shortcut = inputs
if self._use_projection:
shortcut = self._shortcut(shortcut)
shortcut = self._norm0(shortcut)
if self._use_explicit_padding:
inputs = self._pad(inputs)
x = self._conv1(inputs)
x = self._norm1(x)
x = self._activation_fn(x)
x = self._conv2(x)
x = self._norm2(x)
if self._squeeze_excitation:
x = self._squeeze_excitation(x)
if self._stochastic_depth:
x = self._stochastic_depth(x, training=training)
return self._activation_fn(x + shortcut)
@tf.keras.utils.register_keras_serializable(package='Vision')
class BottleneckBlock(tf.keras.layers.Layer):
"""A standard bottleneck block."""
def __init__(self,
filters,
strides,
dilation_rate=1,
use_projection=False,
se_ratio=None,
resnetd_shortcut=False,
stochastic_depth_drop_rate=None,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
activation='relu',
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
bn_trainable=True,
**kwargs):
"""Initializes a standard bottleneck block with BN after convolutions.
Args:
filters: An `int` number of filters for the first two convolutions. Note
that the third and final convolution will use 4 times as many filters.
strides: An `int` block stride. If greater than 1, this block will
ultimately downsample the input.
dilation_rate: An `int` dilation_rate of convolutions. Default to 1.
use_projection: A `bool` for whether this block should use a projection
shortcut (versus the default identity shortcut). This is usually `True`
for the first block of a block group, which may change the number of
filters and the resolution.
se_ratio: A `float` or None. Ratio of the Squeeze-and-Excitation layer.
resnetd_shortcut: A `bool`. If True, apply the resnetd style modification
to the shortcut connection.
stochastic_depth_drop_rate: A `float` or None. If not None, drop rate for
the stochastic depth layer.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2d.
Default to None.
activation: A `str` name of the activation function.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
bn_trainable: A `bool` that indicates whether batch norm layers should be
trainable. Default to True.
**kwargs: Additional keyword arguments to be passed.
"""
super(BottleneckBlock, self).__init__(**kwargs)
self._filters = filters
self._strides = strides
self._dilation_rate = dilation_rate
self._use_projection = use_projection
self._se_ratio = se_ratio
self._resnetd_shortcut = resnetd_shortcut
self._use_sync_bn = use_sync_bn
self._activation = activation
self._stochastic_depth_drop_rate = stochastic_depth_drop_rate
self._kernel_initializer = kernel_initializer
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._bn_trainable = bn_trainable
def build(self, input_shape):
if self._use_projection:
if self._resnetd_shortcut:
self._shortcut0 = tf.keras.layers.AveragePooling2D(
pool_size=2, strides=self._strides, padding='same')
self._shortcut1 = tf.keras.layers.Conv2D(
filters=self._filters * 4,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._shortcut = tf.keras.layers.Conv2D(
filters=self._filters * 4,
kernel_size=1,
strides=self._strides,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._conv1 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._activation1 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
self._conv2 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=3,
strides=self._strides,
dilation_rate=self._dilation_rate,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._activation2 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
self._conv3 = tf.keras.layers.Conv2D(
filters=self._filters * 4,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm3 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon,
trainable=self._bn_trainable)
self._activation3 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters * 4,
out_filters=self._filters * 4,
se_ratio=self._se_ratio,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
self._add = tf.keras.layers.Add()
super(BottleneckBlock, self).build(input_shape)
def get_config(self):
config = {
'filters': self._filters,
'strides': self._strides,
'dilation_rate': self._dilation_rate,
'use_projection': self._use_projection,
'se_ratio': self._se_ratio,
'resnetd_shortcut': self._resnetd_shortcut,
'stochastic_depth_drop_rate': self._stochastic_depth_drop_rate,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'bias_regularizer': self._bias_regularizer,
'activation': self._activation,
'use_sync_bn': self._use_sync_bn,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon,
'bn_trainable': self._bn_trainable
}
base_config = super(BottleneckBlock, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, training=None):
shortcut = inputs
if self._use_projection:
if self._resnetd_shortcut:
shortcut = self._shortcut0(shortcut)
shortcut = self._shortcut1(shortcut)
else:
shortcut = self._shortcut(shortcut)
shortcut = self._norm0(shortcut)
x = self._conv1(inputs)
x = self._norm1(x)
x = self._activation1(x)
x = self._conv2(x)
x = self._norm2(x)
x = self._activation2(x)
x = self._conv3(x)
x = self._norm3(x)
if self._squeeze_excitation:
x = self._squeeze_excitation(x)
if self._stochastic_depth:
x = self._stochastic_depth(x, training=training)
x = self._add([x, shortcut])
return self._activation3(x)
@tf.keras.utils.register_keras_serializable(package='Vision')
class InvertedBottleneckBlock(tf.keras.layers.Layer):
"""An inverted bottleneck block."""
def __init__(self,
in_filters,
out_filters,
expand_ratio,
strides,
kernel_size=3,
se_ratio=None,
stochastic_depth_drop_rate=None,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
activation='relu',
se_inner_activation='relu',
se_gating_activation='sigmoid',
se_round_down_protect=True,
expand_se_in_filters=False,
depthwise_activation=None,
use_sync_bn=False,
dilation_rate=1,
divisible_by=1,
regularize_depthwise=False,
use_depthwise=True,
use_residual=True,
norm_momentum=0.99,
norm_epsilon=0.001,
output_intermediate_endpoints=False,
**kwargs):
"""Initializes an inverted bottleneck block with BN after convolutions.
Args:
in_filters: An `int` number of filters of the input tensor.
out_filters: An `int` number of filters of the output tensor.
expand_ratio: An `int` of expand_ratio for an inverted bottleneck block.
strides: An `int` block stride. If greater than 1, this block will
ultimately downsample the input.
kernel_size: An `int` kernel_size of the depthwise conv layer.
se_ratio: A `float` or None. If not None, se ratio for the squeeze and
excitation layer.
stochastic_depth_drop_rate: A `float` or None. if not None, drop rate for
the stochastic depth layer.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2d.
Default to None.
activation: A `str` name of the activation function.
se_inner_activation: A `str` name of squeeze-excitation inner activation.
se_gating_activation: A `str` name of squeeze-excitation gating
activation.
se_round_down_protect: A `bool` of whether round down more than 10%
will be allowed in SE layer.
expand_se_in_filters: A `bool` of whether or not to expand in_filter in
squeeze and excitation layer.
depthwise_activation: A `str` name of the activation function for
depthwise only.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
dilation_rate: An `int` that specifies the dilation rate to use for.
divisible_by: An `int` that ensures all inner dimensions are divisible by
this number.
dilated convolution: An `int` to specify the same value for all spatial
dimensions.
regularize_depthwise: A `bool` of whether or not apply regularization on
depthwise.
use_depthwise: A `bool` of whether to uses fused convolutions instead of
depthwise.
use_residual: A `bool` of whether to include residual connection between
input and output.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
output_intermediate_endpoints: A `bool` of whether or not output the
intermediate endpoints.
**kwargs: Additional keyword arguments to be passed.
"""
super(InvertedBottleneckBlock, self).__init__(**kwargs)
self._in_filters = in_filters
self._out_filters = out_filters
self._expand_ratio = expand_ratio
self._strides = strides
self._kernel_size = kernel_size
self._se_ratio = se_ratio
self._divisible_by = divisible_by
self._stochastic_depth_drop_rate = stochastic_depth_drop_rate
self._dilation_rate = dilation_rate
self._use_sync_bn = use_sync_bn
self._regularize_depthwise = regularize_depthwise
self._use_depthwise = use_depthwise
self._use_residual = use_residual
self._activation = activation
self._se_inner_activation = se_inner_activation
self._se_gating_activation = se_gating_activation
self._depthwise_activation = depthwise_activation
self._se_round_down_protect = se_round_down_protect
self._kernel_initializer = kernel_initializer
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
self._expand_se_in_filters = expand_se_in_filters
self._output_intermediate_endpoints = output_intermediate_endpoints
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
if not depthwise_activation:
self._depthwise_activation = activation
if regularize_depthwise:
self._depthsize_regularizer = kernel_regularizer
else:
self._depthsize_regularizer = None
def build(self, input_shape):
expand_filters = self._in_filters
if self._expand_ratio > 1:
# First 1x1 conv for channel expansion.
expand_filters = nn_layers.make_divisible(
self._in_filters * self._expand_ratio, self._divisible_by)
expand_kernel = 1 if self._use_depthwise else self._kernel_size
expand_stride = 1 if self._use_depthwise else self._strides
self._conv0 = tf.keras.layers.Conv2D(
filters=expand_filters,
kernel_size=expand_kernel,
strides=expand_stride,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._activation_layer = tf_utils.get_activation(
self._activation, use_keras_layer=True)
if self._use_depthwise:
# Depthwise conv.
self._conv1 = tf.keras.layers.DepthwiseConv2D(
kernel_size=(self._kernel_size, self._kernel_size),
strides=self._strides,
padding='same',
depth_multiplier=1,
dilation_rate=self._dilation_rate,
use_bias=False,
depthwise_initializer=self._kernel_initializer,
depthwise_regularizer=self._depthsize_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._depthwise_activation_layer = tf_utils.get_activation(
self._depthwise_activation, use_keras_layer=True)
# Squeeze and excitation.
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
logging.info('Use Squeeze and excitation.')
in_filters = self._in_filters
if self._expand_se_in_filters:
in_filters = expand_filters
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=in_filters,
out_filters=expand_filters,
se_ratio=self._se_ratio,
divisible_by=self._divisible_by,
round_down_protect=self._se_round_down_protect,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer,
activation=self._se_inner_activation,
gating_activation=self._se_gating_activation)
else:
self._squeeze_excitation = None
# Last 1x1 conv.
self._conv2 = tf.keras.layers.Conv2D(
filters=self._out_filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
self._add = tf.keras.layers.Add()
super(InvertedBottleneckBlock, self).build(input_shape)
def get_config(self):
config = {
'in_filters': self._in_filters,
'out_filters': self._out_filters,
'expand_ratio': self._expand_ratio,
'strides': self._strides,
'kernel_size': self._kernel_size,
'se_ratio': self._se_ratio,
'divisible_by': self._divisible_by,
'stochastic_depth_drop_rate': self._stochastic_depth_drop_rate,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'bias_regularizer': self._bias_regularizer,
'activation': self._activation,
'se_inner_activation': self._se_inner_activation,
'se_gating_activation': self._se_gating_activation,
'se_round_down_protect': self._se_round_down_protect,
'expand_se_in_filters': self._expand_se_in_filters,
'depthwise_activation': self._depthwise_activation,
'dilation_rate': self._dilation_rate,
'use_sync_bn': self._use_sync_bn,
'regularize_depthwise': self._regularize_depthwise,
'use_depthwise': self._use_depthwise,
'use_residual': self._use_residual,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon
}
base_config = super(InvertedBottleneckBlock, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, training=None):
endpoints = {}
shortcut = inputs
if self._expand_ratio > 1:
x = self._conv0(inputs)
x = self._norm0(x)
x = self._activation_layer(x)
else:
x = inputs
if self._use_depthwise:
x = self._conv1(x)
x = self._norm1(x)
x = self._depthwise_activation_layer(x)
if self._output_intermediate_endpoints:
endpoints['depthwise'] = x
if self._squeeze_excitation:
x = self._squeeze_excitation(x)
x = self._conv2(x)
x = self._norm2(x)
if (self._use_residual and self._in_filters == self._out_filters and
self._strides == 1):
if self._stochastic_depth:
x = self._stochastic_depth(x, training=training)
x = self._add([x, shortcut])
if self._output_intermediate_endpoints:
return x, endpoints
return x
@tf.keras.utils.register_keras_serializable(package='Vision')
class ResidualInner(tf.keras.layers.Layer):
"""Creates a single inner block of a residual.
This corresponds to `F`/`G` functions in the RevNet paper:
Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse.
The Reversible Residual Network: Backpropagation Without Storing Activations.
(https://arxiv.org/pdf/1707.04585.pdf)
"""
def __init__(
self,
filters: int,
strides: int,
kernel_initializer: Union[str, Callable[
..., tf.keras.initializers.Initializer]] = 'VarianceScaling',
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
activation: Union[str, Callable[..., tf.Tensor]] = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
batch_norm_first: bool = True,
**kwargs):
"""Initializes a ResidualInner.
Args:
filters: An `int` of output filter size.
strides: An `int` of stride size for convolution for the residual block.
kernel_initializer: A `str` or `tf.keras.initializers.Initializer`
instance for convolutional layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` for Conv2D.
activation: A `str` or `callable` instance of the activation function.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
batch_norm_first: A `bool` of whether to apply activation and batch norm
before conv.
**kwargs: Additional keyword arguments to be passed.
"""
super(ResidualInner, self).__init__(**kwargs)
self.strides = strides
self.filters = filters
self._kernel_initializer = tf.keras.initializers.get(kernel_initializer)
self._kernel_regularizer = kernel_regularizer
self._activation = tf.keras.activations.get(activation)
self._use_sync_bn = use_sync_bn
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._batch_norm_first = batch_norm_first
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation_fn = tf_utils.get_activation(activation)
def build(self, input_shape: tf.TensorShape):
if self._batch_norm_first:
self._batch_norm_0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv2d_1 = tf.keras.layers.Conv2D(
filters=self.filters,
kernel_size=3,
strides=self.strides,
use_bias=False,
padding='same',
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer)
self._batch_norm_1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv2d_2 = tf.keras.layers.Conv2D(
filters=self.filters,
kernel_size=3,
strides=1,
use_bias=False,
padding='same',
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer)
super(ResidualInner, self).build(input_shape)
def get_config(self) -> Dict[str, Any]:
config = {
'filters': self.filters,
'strides': self.strides,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'activation': self._activation,
'use_sync_bn': self._use_sync_bn,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon,
'batch_norm_first': self._batch_norm_first,
}
base_config = super(ResidualInner, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self,
inputs: tf.Tensor,
training: Optional[bool] = None) -> tf.Tensor:
x = inputs
if self._batch_norm_first:
x = self._batch_norm_0(x, training=training)
x = self._activation_fn(x)
x = self._conv2d_1(x)
x = self._batch_norm_1(x, training=training)
x = self._activation_fn(x)
x = self._conv2d_2(x)
return x
@tf.keras.utils.register_keras_serializable(package='Vision')
class BottleneckResidualInner(tf.keras.layers.Layer):
"""Creates a single inner block of a bottleneck.
This corresponds to `F`/`G` functions in the RevNet paper:
Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse.
The Reversible Residual Network: Backpropagation Without Storing Activations.
(https://arxiv.org/pdf/1707.04585.pdf)
"""
def __init__(
self,
filters: int,
strides: int,
kernel_initializer: Union[str, Callable[
..., tf.keras.initializers.Initializer]] = 'VarianceScaling',
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
activation: Union[str, Callable[..., tf.Tensor]] = 'relu',
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
batch_norm_first: bool = True,
**kwargs):
"""Initializes a BottleneckResidualInner.
Args:
filters: An `int` number of filters for first 2 convolutions. Last Last,
and thus the number of output channels from the bottlneck block is
`4*filters`
strides: An `int` of stride size for convolution for the residual block.
kernel_initializer: A `str` or `tf.keras.initializers.Initializer`
instance for convolutional layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` for Conv2D.
activation: A `str` or `callable` instance of the activation function.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
batch_norm_first: A `bool` of whether to apply activation and batch norm
before conv.
**kwargs: Additional keyword arguments to be passed.
"""
super(BottleneckResidualInner, self).__init__(**kwargs)
self.strides = strides
self.filters = filters
self._kernel_initializer = tf.keras.initializers.get(kernel_initializer)
self._kernel_regularizer = kernel_regularizer
self._activation = tf.keras.activations.get(activation)
self._use_sync_bn = use_sync_bn
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._batch_norm_first = batch_norm_first
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation_fn = tf_utils.get_activation(activation)
def build(self, input_shape: tf.TensorShape):
if self._batch_norm_first:
self._batch_norm_0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv2d_1 = tf.keras.layers.Conv2D(
filters=self.filters,
kernel_size=1,
strides=self.strides,
use_bias=False,
padding='same',
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer)
self._batch_norm_1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv2d_2 = tf.keras.layers.Conv2D(
filters=self.filters,
kernel_size=3,
strides=1,
use_bias=False,
padding='same',
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer)
self._batch_norm_2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv2d_3 = tf.keras.layers.Conv2D(
filters=self.filters * 4,
kernel_size=1,
strides=1,
use_bias=False,
padding='same',
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer)
super(BottleneckResidualInner, self).build(input_shape)
def get_config(self) -> Dict[str, Any]:
config = {
'filters': self.filters,
'strides': self.strides,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'activation': self._activation,
'use_sync_bn': self._use_sync_bn,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon,
'batch_norm_first': self._batch_norm_first,
}
base_config = super(BottleneckResidualInner, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self,
inputs: tf.Tensor,
training: Optional[bool] = None) -> tf.Tensor:
x = inputs
if self._batch_norm_first:
x = self._batch_norm_0(x, training=training)
x = self._activation_fn(x)
x = self._conv2d_1(x)
x = self._batch_norm_1(x, training=training)
x = self._activation_fn(x)
x = self._conv2d_2(x)
x = self._batch_norm_2(x, training=training)
x = self._activation_fn(x)
x = self._conv2d_3(x)
return x
@tf.keras.utils.register_keras_serializable(package='Vision')
class ReversibleLayer(tf.keras.layers.Layer):
"""Creates a reversible layer.
Computes y1 = x1 + f(x2), y2 = x2 + g(y1), where f and g can be arbitrary
layers that are stateless, which in this case are `ResidualInner` layers.
"""
def __init__(self,
f: tf.keras.layers.Layer,
g: tf.keras.layers.Layer,
manual_grads: bool = True,
**kwargs):
"""Initializes a ReversibleLayer.
Args:
f: A `tf.keras.layers.Layer` instance of `f` inner block referred to in
paper. Each reversible layer consists of two inner functions. For
example, in RevNet the reversible residual consists of two f/g inner
(bottleneck) residual functions. Where the input to the reversible layer
is x, the input gets partitioned in the channel dimension and the
forward pass follows (eq8): x = [x1; x2], z1 = x1 + f(x2), y2 = x2 +
g(z1), y1 = stop_gradient(z1).
g: A `tf.keras.layers.Layer` instance of `g` inner block referred to in
paper. Detailed explanation same as above as `f` arg.
manual_grads: A `bool` [Testing Only] of whether to manually take
gradients as in Algorithm 1 or defer to autograd.
**kwargs: Additional keyword arguments to be passed.
"""
super(ReversibleLayer, self).__init__(**kwargs)
self._f = f
self._g = g
self._manual_grads = manual_grads
if tf.keras.backend.image_data_format() == 'channels_last':
self._axis = -1
else:
self._axis = 1
def get_config(self) -> Dict[str, Any]:
config = {
'f': self._f,
'g': self._g,
'manual_grads': self._manual_grads,
}
base_config = super(ReversibleLayer, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def _ckpt_non_trainable_vars(self):
self._f_non_trainable_vars = [
v.read_value() for v in self._f.non_trainable_variables
]
self._g_non_trainable_vars = [
v.read_value() for v in self._g.non_trainable_variables
]
def _load_ckpt_non_trainable_vars(self):
for v, v_chkpt in zip(self._f.non_trainable_variables,
self._f_non_trainable_vars):
v.assign(v_chkpt)
for v, v_chkpt in zip(self._g.non_trainable_variables,
self._g_non_trainable_vars):
v.assign(v_chkpt)
def call(self,
inputs: tf.Tensor,
training: Optional[bool] = None) -> tf.Tensor:
@tf.custom_gradient
def reversible(
x: tf.Tensor
) -> Tuple[tf.Tensor, Callable[[Any], Tuple[List[tf.Tensor],
List[tf.Tensor]]]]:
"""Implements Algorithm 1 in the RevNet paper.
Aidan N. Gomez, Mengye Ren, Raquel Urtasun, Roger B. Grosse.
The Reversible Residual Network: Backpropagation Without Storing
Activations.
(https://arxiv.org/pdf/1707.04585.pdf)
Args:
x: An input `tf.Tensor.
Returns:
y: The output [y1; y2] in Algorithm 1.
grad_fn: A callable function that computes the gradients.
"""
with tf.GradientTape() as fwdtape:
fwdtape.watch(x)
x1, x2 = tf.split(x, num_or_size_splits=2, axis=self._axis)
f_x2 = self._f(x2, training=training)
x1_down = _maybe_downsample(x1, f_x2.shape[self._axis], self._f.strides,
self._axis)
z1 = f_x2 + x1_down
g_z1 = self._g(z1, training=training)
x2_down = _maybe_downsample(x2, g_z1.shape[self._axis], self._f.strides,
self._axis)
y2 = x2_down + g_z1
# Equation 8: https://arxiv.org/pdf/1707.04585.pdf
# Decouple y1 and z1 so that their derivatives are different.
y1 = tf.identity(z1)
y = tf.concat([y1, y2], axis=self._axis)
irreversible = ((self._f.strides != 1 or self._g.strides != 1) or
(y.shape[self._axis] != inputs.shape[self._axis]))
# Checkpointing moving mean/variance for batch normalization layers
# as they shouldn't be updated during the custom gradient pass of f/g.
self._ckpt_non_trainable_vars()
def grad_fn(
dy: tf.Tensor,
variables: Optional[List[tf.Variable]] = None,
) -> Tuple[List[tf.Tensor], List[tf.Tensor]]:
"""Given dy calculate (dy/dx)|_{x_{input}} using f/g."""
if irreversible or not self._manual_grads:
grads_combined = fwdtape.gradient(
y, [x] + variables, output_gradients=dy)
dx = grads_combined[0]
grad_vars = grads_combined[1:]
else:
y1_nograd = tf.stop_gradient(y1)
y2_nograd = tf.stop_gradient(y2)
dy1, dy2 = tf.split(dy, num_or_size_splits=2, axis=self._axis)
# Index mapping from self.f/g.trainable_variables to grad_fn
# input `variables` kwarg so that we can reorder dwf + dwg
# variable gradient list to match `variables` order.
f_var_refs = [v.ref() for v in self._f.trainable_variables]
g_var_refs = [v.ref() for v in self._g.trainable_variables]
fg_var_refs = f_var_refs + g_var_refs
self_to_var_index = [fg_var_refs.index(v.ref()) for v in variables]
# Algorithm 1 in paper (line # documented in-line)
z1 = y1_nograd # line 2
with tf.GradientTape() as gtape:
gtape.watch(z1)
g_z1 = self._g(z1, training=training)
x2 = y2_nograd - g_z1 # line 3
with tf.GradientTape() as ftape:
ftape.watch(x2)
f_x2 = self._f(x2, training=training)
x1 = z1 - f_x2 # pylint: disable=unused-variable # line 4
# Compute gradients
g_grads_combined = gtape.gradient(
g_z1, [z1] + self._g.trainable_variables, output_gradients=dy2)
dz1 = dy1 + g_grads_combined[0] # line 5
dwg = g_grads_combined[1:] # line 9
f_grads_combined = ftape.gradient(
f_x2, [x2] + self._f.trainable_variables, output_gradients=dz1)
dx2 = dy2 + f_grads_combined[0] # line 6
dwf = f_grads_combined[1:] # line 8
dx1 = dz1 # line 7
# Pack the input and variable gradients.
dx = tf.concat([dx1, dx2], axis=self._axis)
grad_vars = dwf + dwg
# Reorder gradients (trainable_variables to variables kwarg order)
grad_vars = [grad_vars[i] for i in self_to_var_index]
# Restore batch normalization moving mean/variance for correctness.
self._load_ckpt_non_trainable_vars()
return dx, grad_vars # grad_fn end
return y, grad_fn # reversible end
activations = reversible(inputs)
return activations
@tf.keras.utils.register_keras_serializable(package='Vision')
class DepthwiseSeparableConvBlock(tf.keras.layers.Layer):
"""Creates an depthwise separable convolution block with batch normalization."""
def __init__(
self,
filters: int,
kernel_size: int = 3,
strides: int = 1,
regularize_depthwise=False,
activation: Text = 'relu6',
kernel_initializer: Text = 'VarianceScaling',
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
dilation_rate: int = 1,
use_sync_bn: bool = False,
norm_momentum: float = 0.99,
norm_epsilon: float = 0.001,
**kwargs):
"""Initializes a convolution block with batch normalization.
Args:
filters: An `int` number of filters for the first two convolutions. Note
that the third and final convolution will use 4 times as many filters.
kernel_size: An `int` that specifies the height and width of the 2D
convolution window.
strides: An `int` of block stride. If greater than 1, this block will
ultimately downsample the input.
regularize_depthwise: A `bool`. If Ture, apply regularization on
depthwise.
activation: A `str` name of the activation function.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
dilation_rate: An `int` or tuple/list of 2 `int`, specifying the dilation
rate to use for dilated convolution. Can be a single integer to specify
the same value for all spatial dimensions.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
**kwargs: Additional keyword arguments to be passed.
"""
super(DepthwiseSeparableConvBlock, self).__init__(**kwargs)
self._filters = filters
self._kernel_size = kernel_size
self._strides = strides
self._activation = activation
self._regularize_depthwise = regularize_depthwise
self._kernel_initializer = kernel_initializer
self._kernel_regularizer = kernel_regularizer
self._dilation_rate = dilation_rate
self._use_sync_bn = use_sync_bn
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation_fn = tf_utils.get_activation(activation)
if regularize_depthwise:
self._depthsize_regularizer = kernel_regularizer
else:
self._depthsize_regularizer = None
def get_config(self):
config = {
'filters': self._filters,
'strides': self._strides,
'regularize_depthwise': self._regularize_depthwise,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'activation': self._activation,
'use_sync_bn': self._use_sync_bn,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon
}
base_config = super(DepthwiseSeparableConvBlock, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def build(self, input_shape):
self._dwconv0 = tf.keras.layers.DepthwiseConv2D(
kernel_size=self._kernel_size,
strides=self._strides,
padding='same',
depth_multiplier=1,
dilation_rate=self._dilation_rate,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._depthsize_regularizer,
use_bias=False)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._conv1 = tf.keras.layers.Conv2D(
filters=self._filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
super(DepthwiseSeparableConvBlock, self).build(input_shape)
def call(self, inputs, training=None):
x = self._dwconv0(inputs)
x = self._norm0(x)
x = self._activation_fn(x)
x = self._conv1(x)
x = self._norm1(x)
return self._activation_fn(x)
@tf.keras.utils.register_keras_serializable(package='Vision')
class TuckerConvBlock(tf.keras.layers.Layer):
"""An Tucker block (generalized bottleneck)."""
def __init__(self,
in_filters,
out_filters,
input_compression_ratio,
output_compression_ratio,
strides,
kernel_size=3,
stochastic_depth_drop_rate=None,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
activation='relu',
use_sync_bn=False,
divisible_by=1,
use_residual=True,
norm_momentum=0.99,
norm_epsilon=0.001,
**kwargs):
"""Initializes an inverted bottleneck block with BN after convolutions.
Args:
in_filters: An `int` number of filters of the input tensor.
out_filters: An `int` number of filters of the output tensor.
input_compression_ratio: An `float` of compression ratio for
input filters.
output_compression_ratio: An `float` of compression ratio for
output filters.
strides: An `int` block stride. If greater than 1, this block will
ultimately downsample the input.
kernel_size: An `int` kernel_size of the depthwise conv layer.
stochastic_depth_drop_rate: A `float` or None. if not None, drop rate for
the stochastic depth layer.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2d.
Default to None.
activation: A `str` name of the activation function.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
divisible_by: An `int` that ensures all inner dimensions are divisible by
this number.
use_residual: A `bool` of whether to include residual connection between
input and output.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
**kwargs: Additional keyword arguments to be passed.
"""
super(TuckerConvBlock, self).__init__(**kwargs)
self._in_filters = in_filters
self._out_filters = out_filters
self._input_compression_ratio = input_compression_ratio
self._output_compression_ratio = output_compression_ratio
self._strides = strides
self._kernel_size = kernel_size
self._divisible_by = divisible_by
self._stochastic_depth_drop_rate = stochastic_depth_drop_rate
self._use_sync_bn = use_sync_bn
self._use_residual = use_residual
self._activation = activation
self._kernel_initializer = kernel_initializer
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
def build(self, input_shape):
input_compressed_filters = nn_layers.make_divisible(
value=self._in_filters * self._input_compression_ratio,
divisor=self._divisible_by,
round_down_protect=False)
self._conv0 = tf.keras.layers.Conv2D(
filters=input_compressed_filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._activation_layer0 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
output_compressed_filters = nn_layers.make_divisible(
value=self._out_filters * self._output_compression_ratio,
divisor=self._divisible_by,
round_down_protect=False)
self._conv1 = tf.keras.layers.Conv2D(
filters=output_compressed_filters,
kernel_size=self._kernel_size,
strides=self._strides,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._activation_layer1 = tf_utils.get_activation(
self._activation, use_keras_layer=True)
# Last 1x1 conv.
self._conv2 = tf.keras.layers.Conv2D(
filters=self._out_filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
self._add = tf.keras.layers.Add()
super(TuckerConvBlock, self).build(input_shape)
def get_config(self):
config = {
'in_filters': self._in_filters,
'out_filters': self._out_filters,
'input_compression_ratio': self._input_compression_ratio,
'output_compression_ratio': self._output_compression_ratio,
'strides': self._strides,
'kernel_size': self._kernel_size,
'divisible_by': self._divisible_by,
'stochastic_depth_drop_rate': self._stochastic_depth_drop_rate,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'bias_regularizer': self._bias_regularizer,
'activation': self._activation,
'use_sync_bn': self._use_sync_bn,
'use_residual': self._use_residual,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon
}
base_config = super(TuckerConvBlock, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, training=None):
shortcut = inputs
x = self._conv0(inputs)
x = self._norm0(x)
x = self._activation_layer0(x)
x = self._conv1(x)
x = self._norm1(x)
x = self._activation_layer1(x)
x = self._conv2(x)
x = self._norm2(x)
if (self._use_residual and
self._in_filters == self._out_filters and
self._strides == 1):
if self._stochastic_depth:
x = self._stochastic_depth(x, training=training)
x = self._add([x, shortcut])
return x
official/vision/modeling/layers/nn_blocks_3d.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains common building blocks for 3D networks."""
# Import libraries
import tensorflow as tf
from official.modeling import tf_utils
from official.vision.modeling.layers import nn_layers
@tf.keras.utils.register_keras_serializable(package='Vision')
class SelfGating(tf.keras.layers.Layer):
"""Feature gating as used in S3D-G.
This implements the S3D-G network from:
Saining Xie, Chen Sun, Jonathan Huang, Zhuowen Tu, Kevin Murphy.
Rethinking Spatiotemporal Feature Learning: Speed-Accuracy Trade-offs in Video
Classification.
(https://arxiv.org/pdf/1712.04851.pdf)
"""
def __init__(self, filters, **kwargs):
"""Initializes a self-gating layer.
Args:
filters: An `int` number of filters for the convolutional layer.
**kwargs: Additional keyword arguments to be passed.
"""
super(SelfGating, self).__init__(**kwargs)
self._filters = filters
def build(self, input_shape):
self._spatial_temporal_average = tf.keras.layers.GlobalAveragePooling3D()
# No BN and activation after conv.
self._transformer_w = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=[1, 1, 1],
use_bias=True,
kernel_initializer=tf.keras.initializers.TruncatedNormal(
mean=0.0, stddev=0.01))
super(SelfGating, self).build(input_shape)
def call(self, inputs):
x = self._spatial_temporal_average(inputs)
x = tf.expand_dims(x, 1)
x = tf.expand_dims(x, 2)
x = tf.expand_dims(x, 3)
x = self._transformer_w(x)
x = tf.nn.sigmoid(x)
return tf.math.multiply(x, inputs)
@tf.keras.utils.register_keras_serializable(package='Vision')
class BottleneckBlock3D(tf.keras.layers.Layer):
"""Creates a 3D bottleneck block."""
def __init__(self,
filters,
temporal_kernel_size,
temporal_strides,
spatial_strides,
stochastic_depth_drop_rate=0.0,
se_ratio=None,
use_self_gating=False,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
activation='relu',
use_sync_bn=False,
norm_momentum=0.99,
norm_epsilon=0.001,
**kwargs):
"""Initializes a 3D bottleneck block with BN after convolutions.
Args:
filters: An `int` number of filters for the first two convolutions. Note
that the third and final convolution will use 4 times as many filters.
temporal_kernel_size: An `int` of kernel size for the temporal
convolutional layer.
temporal_strides: An `int` of ftemporal stride for the temporal
convolutional layer.
spatial_strides: An `int` of spatial stride for the spatial convolutional
layer.
stochastic_depth_drop_rate: A `float` or None. If not None, drop rate for
the stochastic depth layer.
se_ratio: A `float` or None. Ratio of the Squeeze-and-Excitation layer.
use_self_gating: A `bool` of whether to apply self-gating module or not.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2d.
Default to None.
activation: A `str` name of the activation function.
use_sync_bn: A `bool`. If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A `float` added to variance to avoid dividing by zero.
**kwargs: Additional keyword arguments to be passed.
"""
super(BottleneckBlock3D, self).__init__(**kwargs)
self._filters = filters
self._temporal_kernel_size = temporal_kernel_size
self._spatial_strides = spatial_strides
self._temporal_strides = temporal_strides
self._stochastic_depth_drop_rate = stochastic_depth_drop_rate
self._use_self_gating = use_self_gating
self._se_ratio = se_ratio
self._use_sync_bn = use_sync_bn
self._activation = activation
self._kernel_initializer = kernel_initializer
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
self._activation_fn = tf_utils.get_activation(activation)
def build(self, input_shape):
self._shortcut_maxpool = tf.keras.layers.MaxPool3D(
pool_size=[1, 1, 1],
strides=[
self._temporal_strides, self._spatial_strides, self._spatial_strides
])
self._shortcut_conv = tf.keras.layers.Conv3D(
filters=4 * self._filters,
kernel_size=1,
strides=[
self._temporal_strides, self._spatial_strides, self._spatial_strides
],
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm0 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._temporal_conv = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=[self._temporal_kernel_size, 1, 1],
strides=[self._temporal_strides, 1, 1],
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm1 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._spatial_conv = tf.keras.layers.Conv3D(
filters=self._filters,
kernel_size=[1, 3, 3],
strides=[1, self._spatial_strides, self._spatial_strides],
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm2 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
self._expand_conv = tf.keras.layers.Conv3D(
filters=4 * self._filters,
kernel_size=[1, 1, 1],
strides=[1, 1, 1],
padding='same',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._norm3 = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)
if self._se_ratio and self._se_ratio > 0 and self._se_ratio <= 1:
self._squeeze_excitation = nn_layers.SqueezeExcitation(
in_filters=self._filters * 4,
out_filters=self._filters * 4,
se_ratio=self._se_ratio,
use_3d_input=True,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
else:
self._squeeze_excitation = None
if self._stochastic_depth_drop_rate:
self._stochastic_depth = nn_layers.StochasticDepth(
self._stochastic_depth_drop_rate)
else:
self._stochastic_depth = None
if self._use_self_gating:
self._self_gating = SelfGating(filters=4 * self._filters)
else:
self._self_gating = None
super(BottleneckBlock3D, self).build(input_shape)
def get_config(self):
config = {
'filters': self._filters,
'temporal_kernel_size': self._temporal_kernel_size,
'temporal_strides': self._temporal_strides,
'spatial_strides': self._spatial_strides,
'use_self_gating': self._use_self_gating,
'se_ratio': self._se_ratio,
'stochastic_depth_drop_rate': self._stochastic_depth_drop_rate,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'bias_regularizer': self._bias_regularizer,
'activation': self._activation,
'use_sync_bn': self._use_sync_bn,
'norm_momentum': self._norm_momentum,
'norm_epsilon': self._norm_epsilon
}
base_config = super(BottleneckBlock3D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, training=None):
in_filters = inputs.shape.as_list()[-1]
if in_filters == 4 * self._filters:
if self._temporal_strides == 1 and self._spatial_strides == 1:
shortcut = inputs
else:
shortcut = self._shortcut_maxpool(inputs)
else:
shortcut = self._shortcut_conv(inputs)
shortcut = self._norm0(shortcut)
x = self._temporal_conv(inputs)
x = self._norm1(x)
x = self._activation_fn(x)
x = self._spatial_conv(x)
x = self._norm2(x)
x = self._activation_fn(x)
x = self._expand_conv(x)
x = self._norm3(x)
# Apply self-gating, SE, stochastic depth.
if self._self_gating:
x = self._self_gating(x)
if self._squeeze_excitation:
x = self._squeeze_excitation(x)
if self._stochastic_depth:
x = self._stochastic_depth(x, training=training)
# Apply activation before additional modules.
x = self._activation_fn(x + shortcut)
return x
official/vision/modeling/layers/nn_blocks_3d_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
# Lint as: python3
"""Tests for resnet."""
# Import libraries
from absl.testing import parameterized
import tensorflow as tf
from official.vision.modeling.layers import nn_blocks_3d
class NNBlocksTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(nn_blocks_3d.BottleneckBlock3D, 1, 1, 2, True, 0.2, 0.1),
(nn_blocks_3d.BottleneckBlock3D, 3, 2, 1, False, 0.0, 0.0),
)
def test_bottleneck_block_creation(self, block_fn, temporal_kernel_size,
temporal_strides, spatial_strides,
use_self_gating, se_ratio,
stochastic_depth):
temporal_size = 16
spatial_size = 128
filters = 256
inputs = tf.keras.Input(
shape=(temporal_size, spatial_size, spatial_size, filters * 4),
batch_size=1)
block = block_fn(
filters=filters,
temporal_kernel_size=temporal_kernel_size,
temporal_strides=temporal_strides,
spatial_strides=spatial_strides,
use_self_gating=use_self_gating,
se_ratio=se_ratio,
stochastic_depth_drop_rate=stochastic_depth)
features = block(inputs)
self.assertAllEqual([
1, temporal_size // temporal_strides, spatial_size // spatial_strides,
spatial_size // spatial_strides, filters * 4
], features.shape.as_list())
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/nn_blocks_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
# Lint as: python3
"""Tests for nn_blocks."""
from typing import Any, Iterable, Tuple
# Import libraries
from absl.testing import parameterized
import tensorflow as tf
from tensorflow.python.distribute import combinations
from tensorflow.python.distribute import strategy_combinations
from official.vision.modeling.layers import nn_blocks
def distribution_strategy_combinations() -> Iterable[Tuple[Any, ...]]:
"""Returns the combinations of end-to-end tests to run."""
return combinations.combine(
distribution=[
strategy_combinations.default_strategy,
strategy_combinations.cloud_tpu_strategy,
strategy_combinations.one_device_strategy_gpu,
],)
class NNBlocksTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(nn_blocks.ResidualBlock, 1, False, 0.0, None),
(nn_blocks.ResidualBlock, 2, True, 0.2, 0.25),
)
def test_residual_block_creation(self, block_fn, strides, use_projection,
stochastic_depth_drop_rate, se_ratio):
input_size = 128
filter_size = 256
inputs = tf.keras.Input(
shape=(input_size, input_size, filter_size), batch_size=1)
block = block_fn(
filter_size,
strides,
use_projection=use_projection,
se_ratio=se_ratio,
stochastic_depth_drop_rate=stochastic_depth_drop_rate,
)
features = block(inputs)
self.assertAllEqual(
[1, input_size // strides, input_size // strides, filter_size],
features.shape.as_list())
@parameterized.parameters(
(nn_blocks.BottleneckBlock, 1, False, 0.0, None),
(nn_blocks.BottleneckBlock, 2, True, 0.2, 0.25),
)
def test_bottleneck_block_creation(self, block_fn, strides, use_projection,
stochastic_depth_drop_rate, se_ratio):
input_size = 128
filter_size = 256
inputs = tf.keras.Input(
shape=(input_size, input_size, filter_size * 4), batch_size=1)
block = block_fn(
filter_size,
strides,
use_projection=use_projection,
se_ratio=se_ratio,
stochastic_depth_drop_rate=stochastic_depth_drop_rate)
features = block(inputs)
self.assertAllEqual(
[1, input_size // strides, input_size // strides, filter_size * 4],
features.shape.as_list())
@parameterized.parameters(
(nn_blocks.InvertedBottleneckBlock, 1, 1, None, None),
(nn_blocks.InvertedBottleneckBlock, 6, 1, None, None),
(nn_blocks.InvertedBottleneckBlock, 1, 2, None, None),
(nn_blocks.InvertedBottleneckBlock, 1, 1, 0.2, None),
(nn_blocks.InvertedBottleneckBlock, 1, 1, None, 0.2),
)
def test_invertedbottleneck_block_creation(self, block_fn, expand_ratio,
strides, se_ratio,
stochastic_depth_drop_rate):
input_size = 128
in_filters = 24
out_filters = 40
inputs = tf.keras.Input(
shape=(input_size, input_size, in_filters), batch_size=1)
block = block_fn(
in_filters=in_filters,
out_filters=out_filters,
expand_ratio=expand_ratio,
strides=strides,
se_ratio=se_ratio,
stochastic_depth_drop_rate=stochastic_depth_drop_rate)
features = block(inputs)
self.assertAllEqual(
[1, input_size // strides, input_size // strides, out_filters],
features.shape.as_list())
@parameterized.parameters(
(nn_blocks.TuckerConvBlock, 1, 0.25, 0.25),
(nn_blocks.TuckerConvBlock, 2, 0.25, 0.25),
)
def test_tucker_conv_block(
self, block_fn, strides,
input_compression_ratio, output_compression_ratio):
input_size = 128
in_filters = 24
out_filters = 24
inputs = tf.keras.Input(
shape=(input_size, input_size, in_filters), batch_size=1)
block = block_fn(
in_filters=in_filters,
out_filters=out_filters,
input_compression_ratio=input_compression_ratio,
output_compression_ratio=output_compression_ratio,
strides=strides)
features = block(inputs)
self.assertAllEqual(
[1, input_size // strides, input_size // strides, out_filters],
features.shape.as_list())
class ResidualInnerTest(parameterized.TestCase, tf.test.TestCase):
@combinations.generate(distribution_strategy_combinations())
def test_shape(self, distribution):
bsz, h, w, c = 8, 32, 32, 32
filters = 64
strides = 2
input_tensor = tf.random.uniform(shape=[bsz, h, w, c])
with distribution.scope():
test_layer = nn_blocks.ResidualInner(filters, strides)
output = test_layer(input_tensor)
expected_output_shape = [bsz, h // strides, w // strides, filters]
self.assertEqual(expected_output_shape, output.shape.as_list())
class BottleneckResidualInnerTest(parameterized.TestCase, tf.test.TestCase):
@combinations.generate(distribution_strategy_combinations())
def test_shape(self, distribution):
bsz, h, w, c = 8, 32, 32, 32
filters = 64
strides = 2
input_tensor = tf.random.uniform(shape=[bsz, h, w, c])
with distribution.scope():
test_layer = nn_blocks.BottleneckResidualInner(filters, strides)
output = test_layer(input_tensor)
expected_output_shape = [bsz, h // strides, w // strides, filters * 4]
self.assertEqual(expected_output_shape, output.shape.as_list())
class DepthwiseSeparableConvBlockTest(parameterized.TestCase, tf.test.TestCase):
@combinations.generate(distribution_strategy_combinations())
def test_shape(self, distribution):
batch_size, height, width, num_channels = 8, 32, 32, 32
num_filters = 64
strides = 2
input_tensor = tf.random.normal(
shape=[batch_size, height, width, num_channels])
with distribution.scope():
block = nn_blocks.DepthwiseSeparableConvBlock(
num_filters, strides=strides)
config_dict = block.get_config()
recreate_block = nn_blocks.DepthwiseSeparableConvBlock(**config_dict)
output_tensor = block(input_tensor)
expected_output_shape = [
batch_size, height // strides, width // strides, num_filters
]
self.assertEqual(output_tensor.shape.as_list(), expected_output_shape)
output_tensor = recreate_block(input_tensor)
self.assertEqual(output_tensor.shape.as_list(), expected_output_shape)
class ReversibleLayerTest(parameterized.TestCase, tf.test.TestCase):
@combinations.generate(distribution_strategy_combinations())
def test_downsampling_non_reversible_step(self, distribution):
bsz, h, w, c = 8, 32, 32, 32
filters = 64
strides = 2
input_tensor = tf.random.uniform(shape=[bsz, h, w, c])
with distribution.scope():
f = nn_blocks.ResidualInner(
filters=filters // 2, strides=strides, batch_norm_first=True)
g = nn_blocks.ResidualInner(
filters=filters // 2, strides=1, batch_norm_first=True)
test_layer = nn_blocks.ReversibleLayer(f, g)
test_layer.build(input_tensor.shape)
optimizer = tf.keras.optimizers.SGD(learning_rate=0.01)
@tf.function
def step_fn():
with tf.GradientTape() as tape:
output = test_layer(input_tensor, training=True)
grads = tape.gradient(output, test_layer.trainable_variables)
# Test applying gradients with optimizer works
optimizer.apply_gradients(zip(grads, test_layer.trainable_variables))
return output
replica_output = distribution.run(step_fn)
outputs = distribution.experimental_local_results(replica_output)
# Assert forward pass shape
expected_output_shape = [bsz, h // strides, w // strides, filters]
for output in outputs:
self.assertEqual(expected_output_shape, output.shape.as_list())
@combinations.generate(distribution_strategy_combinations())
def test_reversible_step(self, distribution):
# Reversible layers satisfy: (a) strides = 1 (b) in_filter = out_filter
bsz, h, w, c = 8, 32, 32, 32
filters = c
strides = 1
input_tensor = tf.random.uniform(shape=[bsz, h, w, c])
with distribution.scope():
f = nn_blocks.ResidualInner(
filters=filters // 2, strides=strides, batch_norm_first=False)
g = nn_blocks.ResidualInner(
filters=filters // 2, strides=1, batch_norm_first=False)
test_layer = nn_blocks.ReversibleLayer(f, g)
test_layer(input_tensor, training=False) # init weights
optimizer = tf.keras.optimizers.SGD(learning_rate=0.01)
@tf.function
def step_fn():
with tf.GradientTape() as tape:
output = test_layer(input_tensor, training=True)
grads = tape.gradient(output, test_layer.trainable_variables)
# Test applying gradients with optimizer works
optimizer.apply_gradients(zip(grads, test_layer.trainable_variables))
return output
@tf.function
def fwd():
test_layer(input_tensor)
distribution.run(fwd) # Initialize variables
prev_variables = tf.identity_n(test_layer.trainable_variables)
replica_output = distribution.run(step_fn)
outputs = distribution.experimental_local_results(replica_output)
# Assert variables values have changed values
for v0, v1 in zip(prev_variables, test_layer.trainable_variables):
self.assertNotAllEqual(v0, v1)
# Assert forward pass shape
expected_output_shape = [bsz, h // strides, w // strides, filters]
for output in outputs:
self.assertEqual(expected_output_shape, output.shape.as_list())
@combinations.generate(distribution_strategy_combinations())
def test_manual_gradients_correctness(self, distribution):
bsz, h, w, c = 8, 32, 32, 32
filters = c
strides = 1
input_tensor = tf.random.uniform(shape=[bsz, h, w, c * 4]) # bottleneck
with distribution.scope():
f_manual = nn_blocks.BottleneckResidualInner(
filters=filters // 2, strides=strides, batch_norm_first=False)
g_manual = nn_blocks.BottleneckResidualInner(
filters=filters // 2, strides=1, batch_norm_first=False)
manual_grad_layer = nn_blocks.ReversibleLayer(f_manual, g_manual)
manual_grad_layer(input_tensor, training=False) # init weights
f_auto = nn_blocks.BottleneckResidualInner(
filters=filters // 2, strides=strides, batch_norm_first=False)
g_auto = nn_blocks.BottleneckResidualInner(
filters=filters // 2, strides=1, batch_norm_first=False)
auto_grad_layer = nn_blocks.ReversibleLayer(
f_auto, g_auto, manual_grads=False)
auto_grad_layer(input_tensor) # init weights
# Clone all weights (tf.keras.layers.Layer has no .clone())
auto_grad_layer._f.set_weights(manual_grad_layer._f.get_weights())
auto_grad_layer._g.set_weights(manual_grad_layer._g.get_weights())
@tf.function
def manual_fn():
with tf.GradientTape() as tape:
output = manual_grad_layer(input_tensor, training=True)
grads = tape.gradient(output, manual_grad_layer.trainable_variables)
return grads
@tf.function
def auto_fn():
with tf.GradientTape() as tape:
output = auto_grad_layer(input_tensor, training=True)
grads = tape.gradient(output, auto_grad_layer.trainable_variables)
return grads
manual_grads = distribution.run(manual_fn)
auto_grads = distribution.run(auto_fn)
# Assert gradients calculated manually are close to that from autograd
for manual_grad, auto_grad in zip(manual_grads, auto_grads):
self.assertAllClose(
distribution.experimental_local_results(manual_grad),
distribution.experimental_local_results(auto_grad),
atol=5e-3,
rtol=5e-3)
# Verify that BN moving mean and variance is correct.
for manual_var, auto_var in zip(manual_grad_layer.non_trainable_variables,
auto_grad_layer.non_trainable_variables):
self.assertAllClose(manual_var, auto_var)
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/nn_layers.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains common building blocks for neural networks."""
from typing import Any, Callable, Dict, List, Mapping, Optional, Tuple, Union
from absl import logging
import tensorflow as tf
import tensorflow_addons as tfa
from official.modeling import tf_utils
from official.vision.ops import spatial_transform_ops
# Type annotations.
States = Dict[str, tf.Tensor]
Activation = Union[str, Callable]
def make_divisible(value: float,
divisor: int,
min_value: Optional[float] = None,
round_down_protect: bool = True,
) -> int:
"""This is to ensure that all layers have channels that are divisible by 8.
Args:
value: A `float` of original value.
divisor: An `int` of the divisor that need to be checked upon.
min_value: A `float` of minimum value threshold.
round_down_protect: A `bool` indicating whether round down more than 10%
will be allowed.
Returns:
The adjusted value in `int` that is divisible against divisor.
"""
if min_value is None:
min_value = divisor
new_value = max(min_value, int(value + divisor / 2) // divisor * divisor)
# Make sure that round down does not go down by more than 10%.
if round_down_protect and new_value < 0.9 * value:
new_value += divisor
return int(new_value)
def round_filters(filters: int,
multiplier: float,
divisor: int = 8,
min_depth: Optional[int] = None,
round_down_protect: bool = True,
skip: bool = False) -> int:
"""Rounds number of filters based on width multiplier."""
orig_f = filters
if skip or not multiplier:
return filters
new_filters = make_divisible(value=filters * multiplier,
divisor=divisor,
min_value=min_depth,
round_down_protect=round_down_protect)
logging.info('round_filter input=%s output=%s', orig_f, new_filters)
return int(new_filters)
def get_padding_for_kernel_size(kernel_size):
"""Compute padding size given kernel size."""
if kernel_size == 7:
return (3, 3)
elif kernel_size == 3:
return (1, 1)
else:
raise ValueError('Padding for kernel size {} not known.'.format(
kernel_size))
@tf.keras.utils.register_keras_serializable(package='Vision')
class SqueezeExcitation(tf.keras.layers.Layer):
"""Creates a squeeze and excitation layer."""
def __init__(self,
in_filters,
out_filters,
se_ratio,
divisible_by=1,
use_3d_input=False,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
activation='relu',
gating_activation='sigmoid',
round_down_protect=True,
**kwargs):
"""Initializes a squeeze and excitation layer.
Args:
in_filters: An `int` number of filters of the input tensor.
out_filters: An `int` number of filters of the output tensor.
se_ratio: A `float` or None. If not None, se ratio for the squeeze and
excitation layer.
divisible_by: An `int` that ensures all inner dimensions are divisible by
this number.
use_3d_input: A `bool` of whether input is 2D or 3D image.
kernel_initializer: A `str` of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default to None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2d.
Default to None.
activation: A `str` name of the activation function.
gating_activation: A `str` name of the activation function for final
gating function.
round_down_protect: A `bool` of whether round down more than 10% will be
allowed.
**kwargs: Additional keyword arguments to be passed.
"""
super(SqueezeExcitation, self).__init__(**kwargs)
self._in_filters = in_filters
self._out_filters = out_filters
self._se_ratio = se_ratio
self._divisible_by = divisible_by
self._round_down_protect = round_down_protect
self._use_3d_input = use_3d_input
self._activation = activation
self._gating_activation = gating_activation
self._kernel_initializer = kernel_initializer
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
if tf.keras.backend.image_data_format() == 'channels_last':
if not use_3d_input:
self._spatial_axis = [1, 2]
else:
self._spatial_axis = [1, 2, 3]
else:
if not use_3d_input:
self._spatial_axis = [2, 3]
else:
self._spatial_axis = [2, 3, 4]
self._activation_fn = tf_utils.get_activation(activation)
self._gating_activation_fn = tf_utils.get_activation(gating_activation)
def build(self, input_shape):
num_reduced_filters = make_divisible(
max(1, int(self._in_filters * self._se_ratio)),
divisor=self._divisible_by,
round_down_protect=self._round_down_protect)
self._se_reduce = tf.keras.layers.Conv2D(
filters=num_reduced_filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=True,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
self._se_expand = tf.keras.layers.Conv2D(
filters=self._out_filters,
kernel_size=1,
strides=1,
padding='same',
use_bias=True,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)
super(SqueezeExcitation, self).build(input_shape)
def get_config(self):
config = {
'in_filters': self._in_filters,
'out_filters': self._out_filters,
'se_ratio': self._se_ratio,
'divisible_by': self._divisible_by,
'use_3d_input': self._use_3d_input,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'bias_regularizer': self._bias_regularizer,
'activation': self._activation,
'gating_activation': self._gating_activation,
'round_down_protect': self._round_down_protect,
}
base_config = super(SqueezeExcitation, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs):
x = tf.reduce_mean(inputs, self._spatial_axis, keepdims=True)
x = self._activation_fn(self._se_reduce(x))
x = self._gating_activation_fn(self._se_expand(x))
return x * inputs
def get_stochastic_depth_rate(init_rate, i, n):
"""Get drop connect rate for the ith block.
Args:
init_rate: A `float` of initial drop rate.
i: An `int` of order of the current block.
n: An `int` total number of blocks.
Returns:
Drop rate of the ith block.
"""
if init_rate is not None:
if init_rate < 0 or init_rate > 1:
raise ValueError('Initial drop rate must be within 0 and 1.')
rate = init_rate * float(i) / n
else:
rate = None
return rate
@tf.keras.utils.register_keras_serializable(package='Vision')
class StochasticDepth(tf.keras.layers.Layer):
"""Creates a stochastic depth layer."""
def __init__(self, stochastic_depth_drop_rate, **kwargs):
"""Initializes a stochastic depth layer.
Args:
stochastic_depth_drop_rate: A `float` of drop rate.
**kwargs: Additional keyword arguments to be passed.
Returns:
A output `tf.Tensor` of which should have the same shape as input.
"""
super(StochasticDepth, self).__init__(**kwargs)
self._drop_rate = stochastic_depth_drop_rate
def get_config(self):
config = {'drop_rate': self._drop_rate}
base_config = super(StochasticDepth, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs, training=None):
if training is None:
training = tf.keras.backend.learning_phase()
if not training or self._drop_rate is None or self._drop_rate == 0:
return inputs
keep_prob = 1.0 - self._drop_rate
batch_size = tf.shape(inputs)[0]
random_tensor = keep_prob
random_tensor += tf.random.uniform(
[batch_size] + [1] * (inputs.shape.rank - 1), dtype=inputs.dtype)
binary_tensor = tf.floor(random_tensor)
output = tf.math.divide(inputs, keep_prob) * binary_tensor
return output
@tf.keras.utils.register_keras_serializable(package='Vision')
def pyramid_feature_fusion(inputs, target_level):
"""Fuses all feature maps in the feature pyramid at the target level.
Args:
inputs: A dictionary containing the feature pyramid. The size of the input
tensor needs to be fixed.
target_level: An `int` of the target feature level for feature fusion.
Returns:
A `float` `tf.Tensor` of shape [batch_size, feature_height, feature_width,
feature_channel].
"""
# Convert keys to int.
pyramid_feats = {int(k): v for k, v in inputs.items()}
min_level = min(pyramid_feats.keys())
max_level = max(pyramid_feats.keys())
resampled_feats = []
for l in range(min_level, max_level + 1):
if l == target_level:
resampled_feats.append(pyramid_feats[l])
else:
feat = pyramid_feats[l]
target_size = list(feat.shape[1:3])
target_size[0] *= 2**(l - target_level)
target_size[1] *= 2**(l - target_level)
# Casts feat to float32 so the resize op can be run on TPU.
feat = tf.cast(feat, tf.float32)
feat = tf.image.resize(
feat, size=target_size, method=tf.image.ResizeMethod.BILINEAR)
# Casts it back to be compatible with the rest opetations.
feat = tf.cast(feat, pyramid_feats[l].dtype)
resampled_feats.append(feat)
return tf.math.add_n(resampled_feats)
class PanopticFPNFusion(tf.keras.Model):
"""Creates a Panoptic FPN feature Fusion layer.
This implements feature fusion for semantic segmentation head from the paper:
Alexander Kirillov, Ross Girshick, Kaiming He and Piotr Dollar.
Panoptic Feature Pyramid Networks.
(https://arxiv.org/pdf/1901.02446.pdf)
"""
def __init__(
self,
min_level: int = 2,
max_level: int = 5,
target_level: int = 2,
num_filters: int = 128,
num_fpn_filters: int = 256,
activation: str = 'relu',
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
bias_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
**kwargs):
"""Initializes panoptic FPN feature fusion layer.
Args:
min_level: An `int` of minimum level to use in feature fusion.
max_level: An `int` of maximum level to use in feature fusion.
target_level: An `int` of the target feature level for feature fusion.
num_filters: An `int` number of filters in conv2d layers.
num_fpn_filters: An `int` number of filters in the FPN outputs
activation: A `str` name of the activation function.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
bias_regularizer: A `tf.keras.regularizers.Regularizer` object for Conv2D.
**kwargs: Additional keyword arguments to be passed.
Returns:
A `float` `tf.Tensor` of shape [batch_size, feature_height, feature_width,
feature_channel].
"""
if target_level > max_level:
raise ValueError('target_level should be less than max_level')
self._config_dict = {
'min_level': min_level,
'max_level': max_level,
'target_level': target_level,
'num_filters': num_filters,
'num_fpn_filters': num_fpn_filters,
'activation': activation,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
}
norm = tfa.layers.GroupNormalization
conv2d = tf.keras.layers.Conv2D
activation_fn = tf_utils.get_activation(activation)
if tf.keras.backend.image_data_format() == 'channels_last':
norm_axis = -1
else:
norm_axis = 1
inputs = self._build_inputs(num_fpn_filters, min_level, max_level)
upscaled_features = []
for level in range(min_level, max_level + 1):
num_conv_layers = max(1, level - target_level)
x = inputs[str(level)]
for i in range(num_conv_layers):
x = conv2d(
filters=num_filters,
kernel_size=3,
padding='same',
kernel_initializer=tf.keras.initializers.VarianceScaling(),
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer)(x)
x = norm(groups=32, axis=norm_axis)(x)
x = activation_fn(x)
if level != target_level:
x = spatial_transform_ops.nearest_upsampling(x, scale=2)
upscaled_features.append(x)
fused_features = tf.math.add_n(upscaled_features)
self._output_specs = {str(target_level): fused_features.get_shape()}
super(PanopticFPNFusion, self).__init__(
inputs=inputs, outputs=fused_features, **kwargs)
def _build_inputs(self, num_filters: int,
min_level: int, max_level: int):
inputs = {}
for level in range(min_level, max_level + 1):
inputs[str(level)] = tf.keras.Input(shape=[None, None, num_filters])
return inputs
def get_config(self) -> Mapping[str, Any]:
return self._config_dict
@classmethod
def from_config(cls, config, custom_objects=None):
return cls(**config)
@property
def output_specs(self) -> Mapping[str, tf.TensorShape]:
"""A dict of {level: TensorShape} pairs for the model output."""
return self._output_specs
@tf.keras.utils.register_keras_serializable(package='Vision')
class Scale(tf.keras.layers.Layer):
"""Scales the input by a trainable scalar weight.
This is useful for applying ReZero to layers, which improves convergence
speed. This implements the paper:
ReZero is All You Need: Fast Convergence at Large Depth.
(https://arxiv.org/pdf/2003.04887.pdf).
"""
def __init__(
self,
initializer: tf.keras.initializers.Initializer = 'ones',
regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
**kwargs):
"""Initializes a scale layer.
Args:
initializer: A `str` of initializer for the scalar weight.
regularizer: A `tf.keras.regularizers.Regularizer` for the scalar weight.
**kwargs: Additional keyword arguments to be passed to this layer.
Returns:
An `tf.Tensor` of which should have the same shape as input.
"""
super(Scale, self).__init__(**kwargs)
self._initializer = initializer
self._regularizer = regularizer
self._scale = self.add_weight(
name='scale',
shape=[],
dtype=self.dtype,
initializer=self._initializer,
regularizer=self._regularizer,
trainable=True)
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'initializer': self._initializer,
'regularizer': self._regularizer,
}
base_config = super(Scale, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs):
"""Calls the layer with the given inputs."""
scale = tf.cast(self._scale, inputs.dtype)
return scale * inputs
@tf.keras.utils.register_keras_serializable(package='Vision')
class TemporalSoftmaxPool(tf.keras.layers.Layer):
"""Creates a network layer corresponding to temporal softmax pooling.
This is useful for multi-class logits (used in e.g., Charades). Modified from
AssembleNet Charades evaluation from:
Michael S. Ryoo, AJ Piergiovanni, Mingxing Tan, Anelia Angelova.
AssembleNet: Searching for Multi-Stream Neural Connectivity in Video
Architectures.
(https://arxiv.org/pdf/1905.13209.pdf).
"""
def call(self, inputs):
"""Calls the layer with the given inputs."""
assert inputs.shape.rank in (3, 4, 5)
frames = tf.shape(inputs)[1]
pre_logits = inputs / tf.sqrt(tf.cast(frames, inputs.dtype))
activations = tf.nn.softmax(pre_logits, axis=1)
outputs = inputs * activations
return outputs
@tf.keras.utils.register_keras_serializable(package='Vision')
class PositionalEncoding(tf.keras.layers.Layer):
"""Creates a network layer that adds a sinusoidal positional encoding.
Positional encoding is incremented across frames, and is added to the input.
The positional encoding is first weighted at 0 so that the network can choose
to ignore it. This implements:
Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszkoreit, Llion Jones,
Aidan N. Gomez, Lukasz Kaiser, Illia Polosukhin.
Attention Is All You Need.
(https://arxiv.org/pdf/1706.03762.pdf).
"""
def __init__(self,
initializer: tf.keras.initializers.Initializer = 'zeros',
cache_encoding: bool = False,
state_prefix: Optional[str] = None,
**kwargs):
"""Initializes positional encoding.
Args:
initializer: A `str` of initializer for weighting the positional encoding.
cache_encoding: A `bool`. If True, cache the positional encoding tensor
after calling build. Otherwise, rebuild the tensor for every call.
Setting this to False can be useful when we want to input a variable
number of frames, so the positional encoding tensor can change shape.
state_prefix: a prefix string to identify states.
**kwargs: Additional keyword arguments to be passed to this layer.
Returns:
A `tf.Tensor` of which should have the same shape as input.
"""
super(PositionalEncoding, self).__init__(**kwargs)
self._initializer = initializer
self._cache_encoding = cache_encoding
self._pos_encoding = None
self._rezero = Scale(initializer=initializer, name='rezero')
state_prefix = state_prefix if state_prefix is not None else ''
self._state_prefix = state_prefix
self._frame_count_name = f'{state_prefix}_pos_enc_frame_count'
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'initializer': self._initializer,
'cache_encoding': self._cache_encoding,
'state_prefix': self._state_prefix,
}
base_config = super(PositionalEncoding, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def _positional_encoding(self,
num_positions: Union[int, tf.Tensor],
hidden_size: Union[int, tf.Tensor],
start_position: Union[int, tf.Tensor] = 0,
dtype: str = 'float32') -> tf.Tensor:
"""Creates a sequence of sinusoidal positional encoding vectors.
Args:
num_positions: the total number of positions (frames).
hidden_size: the number of channels used for the hidden vectors.
start_position: the start position.
dtype: the dtype of the output tensor.
Returns:
The positional encoding tensor with shape [num_positions, hidden_size].
"""
if isinstance(start_position, tf.Tensor) and start_position.shape.rank == 1:
start_position = start_position[0]
# Calling `tf.range` with `dtype=tf.bfloat16` results in an error,
# so we cast afterward.
positions = tf.range(start_position, start_position + num_positions)
positions = tf.cast(positions, dtype)[:, tf.newaxis]
idx = tf.range(hidden_size)[tf.newaxis, :]
power = tf.cast(2 * (idx // 2), dtype)
power /= tf.cast(hidden_size, dtype)
angles = 1. / tf.math.pow(10_000., power)
radians = positions * angles
sin = tf.math.sin(radians[:, 0::2])
cos = tf.math.cos(radians[:, 1::2])
pos_encoding = tf.concat([sin, cos], axis=-1)
return pos_encoding
def _get_pos_encoding(self,
input_shape: tf.Tensor,
frame_count: int = 0) -> tf.Tensor:
"""Calculates the positional encoding from the input shape.
Args:
input_shape: the shape of the input.
frame_count: a count of frames that indicates the index of the first
frame.
Returns:
The positional encoding tensor with shape [num_positions, hidden_size].
"""
frames = input_shape[1]
channels = input_shape[-1]
pos_encoding = self._positional_encoding(
frames, channels, start_position=frame_count, dtype=self.dtype)
pos_encoding = tf.reshape(pos_encoding, [1, frames, 1, 1, channels])
return pos_encoding
def build(self, input_shape):
"""Builds the layer with the given input shape.
Args:
input_shape: The input shape.
Raises:
ValueError: If using 'channels_first' data format.
"""
if tf.keras.backend.image_data_format() == 'channels_first':
raise ValueError('"channels_first" mode is unsupported.')
if self._cache_encoding:
self._pos_encoding = self._get_pos_encoding(input_shape)
super(PositionalEncoding, self).build(input_shape)
def call(
self,
inputs: tf.Tensor,
states: Optional[States] = None,
output_states: bool = True,
) -> Union[tf.Tensor, Tuple[tf.Tensor, States]]:
"""Calls the layer with the given inputs.
Args:
inputs: An input `tf.Tensor`.
states: A `dict` of states such that, if any of the keys match for this
layer, will overwrite the contents of the buffer(s). Expected keys
include `state_prefix + '_pos_enc_frame_count'`.
output_states: A `bool`. If True, returns the output tensor and output
states. Returns just the output tensor otherwise.
Returns:
An output `tf.Tensor` (and optionally the states if `output_states=True`).
Raises:
ValueError: If using 'channels_first' data format.
"""
states = dict(states) if states is not None else {}
# Keep a count of frames encountered across input iterations in
# num_frames to be able to accurately update the positional encoding.
num_frames = tf.shape(inputs)[1]
frame_count = tf.cast(states.get(self._frame_count_name, [0]), tf.int32)
states[self._frame_count_name] = frame_count + num_frames
if self._cache_encoding:
pos_encoding = self._pos_encoding
else:
pos_encoding = self._get_pos_encoding(
tf.shape(inputs), frame_count=frame_count)
pos_encoding = tf.cast(pos_encoding, inputs.dtype)
pos_encoding = self._rezero(pos_encoding)
outputs = inputs + pos_encoding
return (outputs, states) if output_states else outputs
@tf.keras.utils.register_keras_serializable(package='Vision')
class GlobalAveragePool3D(tf.keras.layers.Layer):
"""Creates a global average pooling layer with causal mode.
Implements causal mode, which runs a cumulative sum (with `tf.cumsum`) across
frames in the time dimension, allowing the use of a stream buffer. Sums any
valid input state with the current input to allow state to accumulate over
several iterations.
"""
def __init__(self,
keepdims: bool = False,
causal: bool = False,
state_prefix: Optional[str] = None,
**kwargs):
"""Initializes a global average pool layer.
Args:
keepdims: A `bool`. If True, keep the averaged dimensions.
causal: A `bool` of whether to run in causal mode with a cumulative sum
across frames.
state_prefix: a prefix string to identify states.
**kwargs: Additional keyword arguments to be passed to this layer.
Returns:
An output `tf.Tensor`.
"""
super(GlobalAveragePool3D, self).__init__(**kwargs)
self._keepdims = keepdims
self._causal = causal
state_prefix = state_prefix if state_prefix is not None else ''
self._state_prefix = state_prefix
self._state_name = f'{state_prefix}_pool_buffer'
self._frame_count_name = f'{state_prefix}_pool_frame_count'
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'keepdims': self._keepdims,
'causal': self._causal,
'state_prefix': self._state_prefix,
}
base_config = super(GlobalAveragePool3D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self,
inputs: tf.Tensor,
states: Optional[States] = None,
output_states: bool = True
) -> Union[tf.Tensor, Tuple[tf.Tensor, States]]:
"""Calls the layer with the given inputs.
Args:
inputs: An input `tf.Tensor`.
states: A `dict` of states such that, if any of the keys match for this
layer, will overwrite the contents of the buffer(s).
Expected keys include `state_prefix + '__pool_buffer'` and
`state_prefix + '__pool_frame_count'`.
output_states: A `bool`. If True, returns the output tensor and output
states. Returns just the output tensor otherwise.
Returns:
An output `tf.Tensor` (and optionally the states if `output_states=True`).
If `causal=True`, the output tensor will have shape
`[batch_size, num_frames, 1, 1, channels]` if `keepdims=True`. We keep
the frame dimension in this case to simulate a cumulative global average
as if we are inputting one frame at a time. If `causal=False`, the output
is equivalent to `tf.keras.layers.GlobalAveragePooling3D` with shape
`[batch_size, 1, 1, 1, channels]` if `keepdims=True` (plus the optional
buffer stored in `states`).
Raises:
ValueError: If using 'channels_first' data format.
"""
states = dict(states) if states is not None else {}
if tf.keras.backend.image_data_format() == 'channels_first':
raise ValueError('"channels_first" mode is unsupported.')
# Shape: [batch_size, 1, 1, 1, channels]
buffer = states.get(self._state_name, None)
if buffer is None:
buffer = tf.zeros_like(inputs[:, :1, :1, :1], dtype=inputs.dtype)
states[self._state_name] = buffer
# Keep a count of frames encountered across input iterations in
# num_frames to be able to accurately take a cumulative average across
# all frames when running in streaming mode
num_frames = tf.shape(inputs)[1]
frame_count = states.get(self._frame_count_name, tf.constant([0]))
frame_count = tf.cast(frame_count, tf.int32)
states[self._frame_count_name] = frame_count + num_frames
if self._causal:
# Take a mean of spatial dimensions to make computation more efficient.
x = tf.reduce_mean(inputs, axis=[2, 3], keepdims=True)
x = tf.cumsum(x, axis=1)
x = x + buffer
# The last frame will be the value of the next state
# Shape: [batch_size, 1, 1, 1, channels]
states[self._state_name] = x[:, -1:]
# In causal mode, the divisor increments by 1 for every frame to
# calculate cumulative averages instead of one global average
mean_divisors = tf.range(num_frames) + frame_count + 1
mean_divisors = tf.reshape(mean_divisors, [1, num_frames, 1, 1, 1])
mean_divisors = tf.cast(mean_divisors, x.dtype)
# Shape: [batch_size, num_frames, 1, 1, channels]
x = x / mean_divisors
else:
# In non-causal mode, we (optionally) sum across frames to take a
# cumulative average across input iterations rather than individual
# frames. If no buffer state is passed, this essentially becomes
# regular global average pooling.
# Shape: [batch_size, 1, 1, 1, channels]
x = tf.reduce_sum(inputs, axis=(1, 2, 3), keepdims=True)
x = x / tf.cast(tf.shape(inputs)[2] * tf.shape(inputs)[3], x.dtype)
x = x + buffer
# Shape: [batch_size, 1, 1, 1, channels]
states[self._state_name] = x
x = x / tf.cast(frame_count + num_frames, x.dtype)
if not self._keepdims:
x = tf.squeeze(x, axis=(1, 2, 3))
return (x, states) if output_states else x
@tf.keras.utils.register_keras_serializable(package='Vision')
class SpatialAveragePool3D(tf.keras.layers.Layer):
"""Creates a global average pooling layer pooling across spatial dimentions."""
def __init__(self, keepdims: bool = False, **kwargs):
"""Initializes a global average pool layer.
Args:
keepdims: A `bool`. If True, keep the averaged dimensions.
**kwargs: Additional keyword arguments to be passed to this layer.
Returns:
An output `tf.Tensor`.
"""
super(SpatialAveragePool3D, self).__init__(**kwargs)
self._keepdims = keepdims
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'keepdims': self._keepdims,
}
base_config = super(SpatialAveragePool3D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def build(self, input_shape):
"""Builds the layer with the given input shape."""
if tf.keras.backend.image_data_format() == 'channels_first':
raise ValueError('"channels_first" mode is unsupported.')
super(SpatialAveragePool3D, self).build(input_shape)
def call(self, inputs):
"""Calls the layer with the given inputs."""
if inputs.shape.rank != 5:
raise ValueError(
'Input should have rank {}, got {}'.format(5, inputs.shape.rank))
return tf.reduce_mean(inputs, axis=(2, 3), keepdims=self._keepdims)
class CausalConvMixin:
"""Mixin class to implement CausalConv for `tf.keras.layers.Conv` layers."""
@property
def use_buffered_input(self) -> bool:
return self._use_buffered_input
@use_buffered_input.setter
def use_buffered_input(self, variable: bool):
self._use_buffered_input = variable
def _compute_buffered_causal_padding(self,
inputs: tf.Tensor,
use_buffered_input: bool = False,
time_axis: int = 1,
) -> List[List[int]]:
"""Calculates padding for 'causal' option for conv layers.
Args:
inputs: An optional input `tf.Tensor` to be padded.
use_buffered_input: A `bool`. If True, use 'valid' padding along the time
dimension. This should be set when applying the stream buffer.
time_axis: An `int` of the axis of the time dimension.
Returns:
A list of paddings for `tf.pad`.
"""
input_shape = tf.shape(inputs)[1:-1]
if tf.keras.backend.image_data_format() == 'channels_first':
raise ValueError('"channels_first" mode is unsupported.')
kernel_size_effective = [
(self.kernel_size[i] +
(self.kernel_size[i] - 1) * (self.dilation_rate[i] - 1))
for i in range(self.rank)
]
pad_total = [kernel_size_effective[0] - 1]
for i in range(1, self.rank):
overlap = (input_shape[i] - 1) % self.strides[i] + 1
pad_total.append(tf.maximum(kernel_size_effective[i] - overlap, 0))
pad_beg = [pad_total[i] // 2 for i in range(self.rank)]
pad_end = [pad_total[i] - pad_beg[i] for i in range(self.rank)]
padding = [[pad_beg[i], pad_end[i]] for i in range(self.rank)]
padding = [[0, 0]] + padding + [[0, 0]]
if use_buffered_input:
padding[time_axis] = [0, 0]
else:
padding[time_axis] = [padding[time_axis][0] + padding[time_axis][1], 0]
return padding
def _causal_validate_init(self):
"""Validates the Conv layer initial configuration."""
# Overriding this method is meant to circumvent unnecessary errors when
# using causal padding.
if (self.filters is not None
and self.filters % self.groups != 0):
raise ValueError(
'The number of filters must be evenly divisible by the number of '
'groups. Received: groups={}, filters={}'.format(
self.groups, self.filters))
if not all(self.kernel_size):
raise ValueError('The argument `kernel_size` cannot contain 0(s). '
'Received: %s' % (self.kernel_size,))
def _buffered_spatial_output_shape(self, spatial_output_shape: List[int]):
"""Computes the spatial output shape from the input shape."""
# When buffer padding, use 'valid' padding across time. The output shape
# across time should be the input shape minus any padding, assuming
# the stride across time is 1.
if self._use_buffered_input and spatial_output_shape[0] is not None:
padding = self._compute_buffered_causal_padding(
tf.zeros([1] + spatial_output_shape + [1]), use_buffered_input=False)
spatial_output_shape[0] -= sum(padding[1])
return spatial_output_shape
@tf.keras.utils.register_keras_serializable(package='Vision')
class Conv2D(tf.keras.layers.Conv2D, CausalConvMixin):
"""Conv2D layer supporting CausalConv.
Supports `padding='causal'` option (like in `tf.keras.layers.Conv1D`),
which applies causal padding to the temporal dimension, and same padding in
the spatial dimensions.
"""
def __init__(self, *args, use_buffered_input=False, **kwargs):
"""Initializes conv2d.
Args:
*args: Arguments to be passed.
use_buffered_input: A `bool`. If True, the input is expected to be padded
beforehand. In effect, calling this layer will use 'valid' padding on
the temporal dimension to simulate 'causal' padding.
**kwargs: Additional keyword arguments to be passed.
Returns:
An output `tf.Tensor` of the Conv2D operation.
"""
super(Conv2D, self).__init__(*args, **kwargs)
self._use_buffered_input = use_buffered_input
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'use_buffered_input': self._use_buffered_input,
}
base_config = super(Conv2D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def _compute_causal_padding(self, inputs):
"""Computes causal padding dimensions for the given inputs."""
return self._compute_buffered_causal_padding(
inputs, use_buffered_input=self._use_buffered_input)
def _validate_init(self):
"""Validates the Conv layer initial configuration."""
self._causal_validate_init()
def _spatial_output_shape(self, spatial_input_shape: List[int]):
"""Computes the spatial output shape from the input shape."""
shape = super(Conv2D, self)._spatial_output_shape(spatial_input_shape)
return self._buffered_spatial_output_shape(shape)
@tf.keras.utils.register_keras_serializable(package='Vision')
class DepthwiseConv2D(tf.keras.layers.DepthwiseConv2D, CausalConvMixin):
"""DepthwiseConv2D layer supporting CausalConv.
Supports `padding='causal'` option (like in `tf.keras.layers.Conv1D`),
which applies causal padding to the temporal dimension, and same padding in
the spatial dimensions.
"""
def __init__(self, *args, use_buffered_input=False, **kwargs):
"""Initializes depthwise conv2d.
Args:
*args: Arguments to be passed.
use_buffered_input: A `bool`. If True, the input is expected to be padded
beforehand. In effect, calling this layer will use 'valid' padding on
the temporal dimension to simulate 'causal' padding.
**kwargs: Additional keyword arguments to be passed.
Returns:
An output `tf.Tensor` of the DepthwiseConv2D operation.
"""
super(DepthwiseConv2D, self).__init__(*args, **kwargs)
self._use_buffered_input = use_buffered_input
# Causal padding is unsupported by default for DepthwiseConv2D,
# so we resort to valid padding internally. However, we handle
# causal padding as a special case with `self._is_causal`, which is
# defined by the super class.
if self.padding == 'causal':
self.padding = 'valid'
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'use_buffered_input': self._use_buffered_input,
}
base_config = super(DepthwiseConv2D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs):
"""Calls the layer with the given inputs."""
if self._is_causal:
inputs = tf.pad(inputs, self._compute_causal_padding(inputs))
return super(DepthwiseConv2D, self).call(inputs)
def _compute_causal_padding(self, inputs):
"""Computes causal padding dimensions for the given inputs."""
return self._compute_buffered_causal_padding(
inputs, use_buffered_input=self._use_buffered_input)
def _validate_init(self):
"""Validates the Conv layer initial configuration."""
self._causal_validate_init()
def _spatial_output_shape(self, spatial_input_shape: List[int]):
"""Computes the spatial output shape from the input shape."""
shape = super(DepthwiseConv2D, self)._spatial_output_shape(
spatial_input_shape)
return self._buffered_spatial_output_shape(shape)
@tf.keras.utils.register_keras_serializable(package='Vision')
class Conv3D(tf.keras.layers.Conv3D, CausalConvMixin):
"""Conv3D layer supporting CausalConv.
Supports `padding='causal'` option (like in `tf.keras.layers.Conv1D`),
which applies causal padding to the temporal dimension, and same padding in
the spatial dimensions.
"""
def __init__(self, *args, use_buffered_input=False, **kwargs):
"""Initializes conv3d.
Args:
*args: Arguments to be passed.
use_buffered_input: A `bool`. If True, the input is expected to be padded
beforehand. In effect, calling this layer will use 'valid' padding on
the temporal dimension to simulate 'causal' padding.
**kwargs: Additional keyword arguments to be passed.
Returns:
An output `tf.Tensor` of the Conv3D operation.
"""
super(Conv3D, self).__init__(*args, **kwargs)
self._use_buffered_input = use_buffered_input
def get_config(self):
"""Returns a dictionary containing the config used for initialization."""
config = {
'use_buffered_input': self._use_buffered_input,
}
base_config = super(Conv3D, self).get_config()
return dict(list(base_config.items()) + list(config.items()))
def call(self, inputs):
"""Call the layer with the given inputs."""
# Note: tf.nn.conv3d with depthwise kernels on CPU is currently only
# supported when compiling with TF graph (XLA) using tf.function, so it
# is compiled by default here (b/186463870).
conv_fn = tf.function(super(Conv3D, self).call, jit_compile=True)
return conv_fn(inputs)
def _compute_causal_padding(self, inputs):
"""Computes causal padding dimensions for the given inputs."""
return self._compute_buffered_causal_padding(
inputs, use_buffered_input=self._use_buffered_input)
def _validate_init(self):
"""Validates the Conv layer initial configuration."""
self._causal_validate_init()
def _spatial_output_shape(self, spatial_input_shape: List[int]):
"""Computes the spatial output shape from the input shape."""
shape = super(Conv3D, self)._spatial_output_shape(spatial_input_shape)
return self._buffered_spatial_output_shape(shape)
@tf.keras.utils.register_keras_serializable(package='Vision')
class SpatialPyramidPooling(tf.keras.layers.Layer):
"""Implements the Atrous Spatial Pyramid Pooling.
References:
[Rethinking Atrous Convolution for Semantic Image Segmentation](
https://arxiv.org/pdf/1706.05587.pdf)
[Encoder-Decoder with Atrous Separable Convolution for Semantic Image
Segmentation](https://arxiv.org/pdf/1802.02611.pdf)
"""
def __init__(
self,
output_channels: int,
dilation_rates: List[int],
pool_kernel_size: Optional[List[int]] = None,
use_sync_bn: bool = False,
batchnorm_momentum: float = 0.99,
batchnorm_epsilon: float = 0.001,
activation: str = 'relu',
dropout: float = 0.5,
kernel_initializer: str = 'GlorotUniform',
kernel_regularizer: Optional[tf.keras.regularizers.Regularizer] = None,
interpolation: str = 'bilinear',
use_depthwise_convolution: bool = False,
**kwargs):
"""Initializes `SpatialPyramidPooling`.
Args:
output_channels: Number of channels produced by SpatialPyramidPooling.
dilation_rates: A list of integers for parallel dilated conv.
pool_kernel_size: A list of integers or None. If None, global average
pooling is applied, otherwise an average pooling of pool_kernel_size is
applied.
use_sync_bn: A bool, whether or not to use sync batch normalization.
batchnorm_momentum: A float for the momentum in BatchNorm. Defaults to
0.99.
batchnorm_epsilon: A float for the epsilon value in BatchNorm. Defaults to
0.001.
activation: A `str` for type of activation to be used. Defaults to 'relu'.
dropout: A float for the dropout rate before output. Defaults to 0.5.
kernel_initializer: Kernel initializer for conv layers. Defaults to
`glorot_uniform`.
kernel_regularizer: Kernel regularizer for conv layers. Defaults to None.
interpolation: The interpolation method for upsampling. Defaults to
`bilinear`.
use_depthwise_convolution: Allows spatial pooling to be separable
depthwise convolusions. [Encoder-Decoder with Atrous Separable
Convolution for Semantic Image Segmentation](
https://arxiv.org/pdf/1802.02611.pdf)
**kwargs: Other keyword arguments for the layer.
"""
super().__init__(**kwargs)
self._output_channels = output_channels
self._dilation_rates = dilation_rates
self._use_sync_bn = use_sync_bn
self._batchnorm_momentum = batchnorm_momentum
self._batchnorm_epsilon = batchnorm_epsilon
self._activation = activation
self._dropout = dropout
self._kernel_initializer = kernel_initializer
self._kernel_regularizer = kernel_regularizer
self._interpolation = interpolation
self._pool_kernel_size = pool_kernel_size
self._use_depthwise_convolution = use_depthwise_convolution
self._activation_fn = tf_utils.get_activation(activation)
if self._use_sync_bn:
self._bn_op = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._bn_op = tf.keras.layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
def build(self, input_shape):
height = input_shape[1]
width = input_shape[2]
channels = input_shape[3]
self.aspp_layers = []
conv1 = tf.keras.layers.Conv2D(
filters=self._output_channels,
kernel_size=(1, 1),
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
use_bias=False)
norm1 = self._bn_op(
axis=self._bn_axis,
momentum=self._batchnorm_momentum,
epsilon=self._batchnorm_epsilon)
self.aspp_layers.append([conv1, norm1])
for dilation_rate in self._dilation_rates:
leading_layers = []
kernel_size = (3, 3)
if self._use_depthwise_convolution:
leading_layers += [
tf.keras.layers.DepthwiseConv2D(
depth_multiplier=1,
kernel_size=kernel_size,
padding='same',
depthwise_regularizer=self._kernel_regularizer,
depthwise_initializer=self._kernel_initializer,
dilation_rate=dilation_rate,
use_bias=False)
]
kernel_size = (1, 1)
conv_dilation = leading_layers + [
tf.keras.layers.Conv2D(
filters=self._output_channels,
kernel_size=kernel_size,
padding='same',
kernel_regularizer=self._kernel_regularizer,
kernel_initializer=self._kernel_initializer,
dilation_rate=dilation_rate,
use_bias=False)
]
norm_dilation = self._bn_op(
axis=self._bn_axis,
momentum=self._batchnorm_momentum,
epsilon=self._batchnorm_epsilon)
self.aspp_layers.append(conv_dilation + [norm_dilation])
if self._pool_kernel_size is None:
pooling = [
tf.keras.layers.GlobalAveragePooling2D(),
tf.keras.layers.Reshape((1, 1, channels))
]
else:
pooling = [tf.keras.layers.AveragePooling2D(self._pool_kernel_size)]
conv2 = tf.keras.layers.Conv2D(
filters=self._output_channels,
kernel_size=(1, 1),
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
use_bias=False)
norm2 = self._bn_op(
axis=self._bn_axis,
momentum=self._batchnorm_momentum,
epsilon=self._batchnorm_epsilon)
self.aspp_layers.append(pooling + [conv2, norm2])
self._resizing_layer = tf.keras.layers.Resizing(
height, width, interpolation=self._interpolation, dtype=tf.float32)
self._projection = [
tf.keras.layers.Conv2D(
filters=self._output_channels,
kernel_size=(1, 1),
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
use_bias=False),
self._bn_op(
axis=self._bn_axis,
momentum=self._batchnorm_momentum,
epsilon=self._batchnorm_epsilon)
]
self._dropout_layer = tf.keras.layers.Dropout(rate=self._dropout)
self._concat_layer = tf.keras.layers.Concatenate(axis=-1)
def call(self,
inputs: tf.Tensor,
training: Optional[bool] = None) -> tf.Tensor:
if training is None:
training = tf.keras.backend.learning_phase()
result = []
for i, layers in enumerate(self.aspp_layers):
x = inputs
for layer in layers:
# Apply layers sequentially.
x = layer(x, training=training)
x = self._activation_fn(x)
# Apply resize layer to the end of the last set of layers.
if i == len(self.aspp_layers) - 1:
x = self._resizing_layer(x)
result.append(tf.cast(x, inputs.dtype))
x = self._concat_layer(result)
for layer in self._projection:
x = layer(x, training=training)
x = self._activation_fn(x)
return self._dropout_layer(x)
def get_config(self):
config = {
'output_channels': self._output_channels,
'dilation_rates': self._dilation_rates,
'pool_kernel_size': self._pool_kernel_size,
'use_sync_bn': self._use_sync_bn,
'batchnorm_momentum': self._batchnorm_momentum,
'batchnorm_epsilon': self._batchnorm_epsilon,
'activation': self._activation,
'dropout': self._dropout,
'kernel_initializer': self._kernel_initializer,
'kernel_regularizer': self._kernel_regularizer,
'interpolation': self._interpolation,
}
base_config = super().get_config()
return dict(list(base_config.items()) + list(config.items()))
official/vision/modeling/layers/nn_layers_test.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
# Lint as: python3
"""Tests for nn_layers."""
# Import libraries
from absl.testing import parameterized
import tensorflow as tf
from official.vision.modeling.layers import nn_layers
class NNLayersTest(parameterized.TestCase, tf.test.TestCase):
def test_scale(self):
scale = nn_layers.Scale(initializer=tf.keras.initializers.constant(10.))
output = scale(3.)
self.assertAllEqual(output, 30.)
def test_temporal_softmax_pool(self):
inputs = tf.range(4, dtype=tf.float32) + 1.
inputs = tf.reshape(inputs, [1, 4, 1, 1, 1])
layer = nn_layers.TemporalSoftmaxPool()
output = layer(inputs)
self.assertAllClose(
output,
[[[[[0.10153633]]],
[[[0.33481020]]],
[[[0.82801306]]],
[[[1.82021690]]]]])
def test_positional_encoding(self):
pos_encoding = nn_layers.PositionalEncoding(
initializer='ones', cache_encoding=False)
pos_encoding_cached = nn_layers.PositionalEncoding(
initializer='ones', cache_encoding=True)
inputs = tf.ones([1, 4, 1, 1, 3])
outputs, _ = pos_encoding(inputs)
outputs_cached, _ = pos_encoding_cached(inputs)
expected = tf.constant(
[[[[[1.0000000, 1.0000000, 2.0000000]]],
[[[1.8414710, 1.0021545, 1.5403023]]],
[[[1.9092975, 1.0043088, 0.5838531]]],
[[[1.1411200, 1.0064633, 0.0100075]]]]])
self.assertEqual(outputs.shape, expected.shape)
self.assertAllClose(outputs, expected)
self.assertEqual(outputs.shape, outputs_cached.shape)
self.assertAllClose(outputs, outputs_cached)
inputs = tf.ones([1, 5, 1, 1, 3])
_ = pos_encoding(inputs)
def test_positional_encoding_bfloat16(self):
pos_encoding = nn_layers.PositionalEncoding(initializer='ones')
inputs = tf.ones([1, 4, 1, 1, 3], dtype=tf.bfloat16)
outputs, _ = pos_encoding(inputs)
expected = tf.constant(
[[[[[1.0000000, 1.0000000, 2.0000000]]],
[[[1.8414710, 1.0021545, 1.5403023]]],
[[[1.9092975, 1.0043088, 0.5838531]]],
[[[1.1411200, 1.0064633, 0.0100075]]]]])
self.assertEqual(outputs.shape, expected.shape)
self.assertAllClose(outputs, expected)
def test_global_average_pool_basic(self):
pool = nn_layers.GlobalAveragePool3D(keepdims=True)
inputs = tf.ones([1, 2, 3, 4, 1])
outputs = pool(inputs, output_states=False)
expected = tf.ones([1, 1, 1, 1, 1])
self.assertEqual(outputs.shape, expected.shape)
self.assertAllEqual(outputs, expected)
def test_positional_encoding_stream(self):
pos_encoding = nn_layers.PositionalEncoding(
initializer='ones', cache_encoding=False)
inputs = tf.range(4, dtype=tf.float32) + 1.
inputs = tf.reshape(inputs, [1, 4, 1, 1, 1])
inputs = tf.tile(inputs, [1, 1, 1, 1, 3])
expected, _ = pos_encoding(inputs)
for num_splits in [1, 2, 4]:
frames = tf.split(inputs, num_splits, axis=1)
states = {}
predicted = []
for frame in frames:
output, states = pos_encoding(frame, states=states)
predicted.append(output)
predicted = tf.concat(predicted, axis=1)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
self.assertAllClose(predicted, [[[[[1.0000000, 1.0000000, 2.0000000]]],
[[[2.8414710, 2.0021544, 2.5403023]]],
[[[3.9092975, 3.0043090, 2.5838532]]],
[[[4.1411200, 4.0064630, 3.0100074]]]]])
def test_global_average_pool_keras(self):
pool = nn_layers.GlobalAveragePool3D(keepdims=False)
keras_pool = tf.keras.layers.GlobalAveragePooling3D()
inputs = 10 * tf.random.normal([1, 2, 3, 4, 1])
outputs = pool(inputs, output_states=False)
keras_output = keras_pool(inputs)
self.assertAllEqual(outputs.shape, keras_output.shape)
self.assertAllClose(outputs, keras_output)
def test_stream_global_average_pool(self):
gap = nn_layers.GlobalAveragePool3D(keepdims=True, causal=False)
inputs = tf.range(4, dtype=tf.float32) + 1.
inputs = tf.reshape(inputs, [1, 4, 1, 1, 1])
inputs = tf.tile(inputs, [1, 1, 2, 2, 3])
expected, _ = gap(inputs)
for num_splits in [1, 2, 4]:
frames = tf.split(inputs, num_splits, axis=1)
states = {}
predicted = None
for frame in frames:
predicted, states = gap(frame, states=states)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
self.assertAllClose(
predicted,
[[[[[2.5, 2.5, 2.5]]]]])
def test_causal_stream_global_average_pool(self):
gap = nn_layers.GlobalAveragePool3D(keepdims=True, causal=True)
inputs = tf.range(4, dtype=tf.float32) + 1.
inputs = tf.reshape(inputs, [1, 4, 1, 1, 1])
inputs = tf.tile(inputs, [1, 1, 2, 2, 3])
expected, _ = gap(inputs)
for num_splits in [1, 2, 4]:
frames = tf.split(inputs, num_splits, axis=1)
states = {}
predicted = []
for frame in frames:
x, states = gap(frame, states=states)
predicted.append(x)
predicted = tf.concat(predicted, axis=1)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
self.assertAllClose(
predicted,
[[[[[1.0, 1.0, 1.0]]],
[[[1.5, 1.5, 1.5]]],
[[[2.0, 2.0, 2.0]]],
[[[2.5, 2.5, 2.5]]]]])
def test_spatial_average_pool(self):
pool = nn_layers.SpatialAveragePool3D(keepdims=True)
inputs = tf.range(64, dtype=tf.float32) + 1.
inputs = tf.reshape(inputs, [1, 4, 4, 4, 1])
output = pool(inputs)
self.assertEqual(output.shape, [1, 4, 1, 1, 1])
self.assertAllClose(
output,
[[[[[8.50]]],
[[[24.5]]],
[[[40.5]]],
[[[56.5]]]]])
def test_conv2d_causal(self):
conv2d = nn_layers.Conv2D(
filters=3,
kernel_size=(3, 3),
strides=(1, 2),
padding='causal',
use_buffered_input=True,
kernel_initializer='ones',
use_bias=False,
)
inputs = tf.ones([1, 4, 2, 3])
paddings = [[0, 0], [2, 0], [0, 0], [0, 0]]
padded_inputs = tf.pad(inputs, paddings)
predicted = conv2d(padded_inputs)
expected = tf.constant(
[[[[6.0, 6.0, 6.0]],
[[12., 12., 12.]],
[[18., 18., 18.]],
[[18., 18., 18.]]]])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
conv2d.use_buffered_input = False
predicted = conv2d(inputs)
self.assertFalse(conv2d.use_buffered_input)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
def test_depthwise_conv2d_causal(self):
conv2d = nn_layers.DepthwiseConv2D(
kernel_size=(3, 3),
strides=(1, 1),
padding='causal',
use_buffered_input=True,
depthwise_initializer='ones',
use_bias=False,
)
inputs = tf.ones([1, 2, 2, 3])
paddings = [[0, 0], [2, 0], [0, 0], [0, 0]]
padded_inputs = tf.pad(inputs, paddings)
predicted = conv2d(padded_inputs)
expected = tf.constant(
[[[[2., 2., 2.],
[2., 2., 2.]],
[[4., 4., 4.],
[4., 4., 4.]]]])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
conv2d.use_buffered_input = False
predicted = conv2d(inputs)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
def test_conv3d_causal(self):
conv3d = nn_layers.Conv3D(
filters=3,
kernel_size=(3, 3, 3),
strides=(1, 2, 2),
padding='causal',
use_buffered_input=True,
kernel_initializer='ones',
use_bias=False,
)
inputs = tf.ones([1, 2, 4, 4, 3])
paddings = [[0, 0], [2, 0], [0, 0], [0, 0], [0, 0]]
padded_inputs = tf.pad(inputs, paddings)
predicted = conv3d(padded_inputs)
expected = tf.constant(
[[[[[27., 27., 27.],
[18., 18., 18.]],
[[18., 18., 18.],
[12., 12., 12.]]],
[[[54., 54., 54.],
[36., 36., 36.]],
[[36., 36., 36.],
[24., 24., 24.]]]]])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
conv3d.use_buffered_input = False
predicted = conv3d(inputs)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
def test_depthwise_conv3d_causal(self):
conv3d = nn_layers.Conv3D(
filters=3,
kernel_size=(3, 3, 3),
strides=(1, 2, 2),
padding='causal',
use_buffered_input=True,
kernel_initializer='ones',
use_bias=False,
groups=3,
)
inputs = tf.ones([1, 2, 4, 4, 3])
paddings = [[0, 0], [2, 0], [0, 0], [0, 0], [0, 0]]
padded_inputs = tf.pad(inputs, paddings)
predicted = conv3d(padded_inputs)
expected = tf.constant(
[[[[[9.0, 9.0, 9.0],
[6.0, 6.0, 6.0]],
[[6.0, 6.0, 6.0],
[4.0, 4.0, 4.0]]],
[[[18.0, 18.0, 18.0],
[12., 12., 12.]],
[[12., 12., 12.],
[8., 8., 8.]]]]])
output_shape = conv3d._spatial_output_shape([4, 4, 4])
self.assertAllClose(output_shape, [2, 2, 2])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
conv3d.use_buffered_input = False
predicted = conv3d(inputs)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
def test_conv3d_causal_padding_2d(self):
"""Test to ensure causal padding works like standard padding."""
conv3d = nn_layers.Conv3D(
filters=1,
kernel_size=(1, 3, 3),
strides=(1, 2, 2),
padding='causal',
use_buffered_input=False,
kernel_initializer='ones',
use_bias=False,
)
keras_conv3d = tf.keras.layers.Conv3D(
filters=1,
kernel_size=(1, 3, 3),
strides=(1, 2, 2),
padding='same',
kernel_initializer='ones',
use_bias=False,
)
inputs = tf.ones([1, 1, 4, 4, 1])
predicted = conv3d(inputs)
expected = keras_conv3d(inputs)
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
self.assertAllClose(predicted,
[[[[[9.],
[6.]],
[[6.],
[4.]]]]])
def test_conv3d_causal_padding_1d(self):
"""Test to ensure causal padding works like standard padding."""
conv3d = nn_layers.Conv3D(
filters=1,
kernel_size=(3, 1, 1),
strides=(2, 1, 1),
padding='causal',
use_buffered_input=False,
kernel_initializer='ones',
use_bias=False,
)
keras_conv1d = tf.keras.layers.Conv1D(
filters=1,
kernel_size=3,
strides=2,
padding='causal',
kernel_initializer='ones',
use_bias=False,
)
inputs = tf.ones([1, 4, 1, 1, 1])
predicted = conv3d(inputs)
expected = keras_conv1d(tf.squeeze(inputs, axis=[2, 3]))
expected = tf.reshape(expected, [1, 2, 1, 1, 1])
self.assertEqual(predicted.shape, expected.shape)
self.assertAllClose(predicted, expected)
self.assertAllClose(predicted,
[[[[[1.]]],
[[[3.]]]]])
@parameterized.parameters(
(None, []),
(None, [6, 12, 18]),
([32, 32], [6, 12, 18]),
)
def test_aspp(self, pool_kernel_size, dilation_rates):
inputs = tf.keras.Input(shape=(64, 64, 128), dtype=tf.float32)
layer = nn_layers.SpatialPyramidPooling(
output_channels=256,
dilation_rates=dilation_rates,
pool_kernel_size=pool_kernel_size)
output = layer(inputs)
self.assertAllEqual([None, 64, 64, 256], output.shape)
if __name__ == '__main__':
tf.test.main()
official/vision/modeling/layers/roi_aligner.py 0 → 100644
View file @ 8a9a607c
# Copyright 2022 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.
"""Contains definitions of ROI aligner."""
from typing import Mapping
import tensorflow as tf
from official.vision.ops import spatial_transform_ops
@tf.keras.utils.register_keras_serializable(package='Vision')
class MultilevelROIAligner(tf.keras.layers.Layer):
"""Performs ROIAlign for the second stage processing."""
def __init__(self, crop_size: int = 7, sample_offset: float = 0.5, **kwargs):
"""Initializes a ROI aligner.
Args:
crop_size: An `int` of the output size of the cropped features.
sample_offset: A `float` in [0, 1] of the subpixel sample offset.
**kwargs: Additional keyword arguments passed to Layer.
"""
self._config_dict = {
'crop_size': crop_size,
'sample_offset': sample_offset,
}
super(MultilevelROIAligner, self).__init__(**kwargs)
def call(self,
features: Mapping[str, tf.Tensor],
boxes: tf.Tensor,
training: bool = None):
"""Generates ROIs.
Args:
features: A dictionary with key as pyramid level and value as features.
The features are in shape of
[batch_size, height_l, width_l, num_filters].
boxes: A 3-D `tf.Tensor` of shape [batch_size, num_boxes, 4]. Each row
represents a box with [y1, x1, y2, x2] in un-normalized coordinates.
from grid point.
training: A `bool` of whether it is in training mode.
Returns:
A 5-D `tf.Tensor` representing feature crop of shape
[batch_size, num_boxes, crop_size, crop_size, num_filters].
"""
roi_features = spatial_transform_ops.multilevel_crop_and_resize(
features,
boxes,
output_size=self._config_dict['crop_size'],
sample_offset=self._config_dict['sample_offset'])
return roi_features
def get_config(self):
return self._config_dict
@classmethod
def from_config(cls, config, custom_objects=None):
return cls(**config)
Markdown is supported
0% or .
You are about to add 0 people to the discussion. Proceed with caution.
Finish editing this message first!
Please register or to comment