Skip to content

GitLab

Commit 790e49e5 authored by stephenwu's avatar stephenwu
Browse files

Merge branch 'master' of https://github.com/tensorflow/models into run_superglue

parents 8ab018b0 5bb827c3
Hide whitespace changes
Inline Side-by-side
Showing with 790 additions and 146 deletions
official/vision/beta/modeling/backbones/resnet_deeplab_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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_deeplab models."""
# Import libraries
......
official/vision/beta/modeling/backbones/resnet_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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
......
official/vision/beta/modeling/backbones/revnet.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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
"""Contains definitions of RevNet."""
from typing import Any, Callable, Dict, Optional
......
official/vision/beta/modeling/backbones/revnet_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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 RevNet."""
# Import libraries
......
official/vision/beta/modeling/backbones/spinenet.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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
"""Contains definitions of SpineNet Networks."""
import math
......
official/vision/beta/modeling/backbones/spinenet_mobile.py 0 → 100644
View file @ 790e49e5
# Copyright 2021 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
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Contains definitions of Mobile SpineNet Networks."""
import math
# Import libraries
from absl import logging
import tensorflow as tf
from official.modeling import tf_utils
from official.vision.beta.modeling.backbones import factory
from official.vision.beta.modeling.layers import nn_blocks
from official.vision.beta.modeling.layers import nn_layers
from official.vision.beta.ops import spatial_transform_ops
layers = tf.keras.layers
FILTER_SIZE_MAP = {
0: 8,
1: 16,
2: 24,
3: 40,
4: 80,
5: 112,
6: 112,
7: 112,
}
# The fixed SpineNet architecture discovered by NAS.
# Each element represents a specification of a building block:
# (block_level, block_fn, (input_offset0, input_offset1), is_output).
SPINENET_BLOCK_SPECS = [
(2, 'mbconv', (0, 1), False),
(2, 'mbconv', (1, 2), False),
(4, 'mbconv', (1, 2), False),
(3, 'mbconv', (3, 4), False),
(4, 'mbconv', (3, 5), False),
(6, 'mbconv', (4, 6), False),
(4, 'mbconv', (4, 6), False),
(5, 'mbconv', (7, 8), False),
(7, 'mbconv', (7, 9), False),
(5, 'mbconv', (9, 10), False),
(5, 'mbconv', (9, 11), False),
(4, 'mbconv', (6, 11), True),
(3, 'mbconv', (5, 11), True),
(5, 'mbconv', (8, 13), True),
(7, 'mbconv', (6, 15), True),
(6, 'mbconv', (13, 15), True),
]
SCALING_MAP = {
'49': {
'endpoints_num_filters': 48,
'filter_size_scale': 1.0,
'block_repeats': 1,
},
'49S': {
'endpoints_num_filters': 40,
'filter_size_scale': 0.65,
'block_repeats': 1,
},
'49XS': {
'endpoints_num_filters': 24,
'filter_size_scale': 0.6,
'block_repeats': 1,
},
}
class BlockSpec(object):
"""A container class that specifies the block configuration for SpineNet."""
def __init__(self, level, block_fn, input_offsets, is_output):
self.level = level
self.block_fn = block_fn
self.input_offsets = input_offsets
self.is_output = is_output
def build_block_specs(block_specs=None):
"""Builds the list of BlockSpec objects for SpineNet."""
if not block_specs:
block_specs = SPINENET_BLOCK_SPECS
logging.info('Building SpineNet block specs: %s', block_specs)
return [BlockSpec(*b) for b in block_specs]
@tf.keras.utils.register_keras_serializable(package='Vision')
class SpineNetMobile(tf.keras.Model):
"""Creates a Mobile SpineNet family model.
This implements:
[1] Xianzhi Du, Tsung-Yi Lin, Pengchong Jin, Golnaz Ghiasi, Mingxing Tan,
Yin Cui, Quoc V. Le, Xiaodan Song.
SpineNet: Learning Scale-Permuted Backbone for Recognition and Localization.
(https://arxiv.org/abs/1912.05027).
[2] Xianzhi Du, Tsung-Yi Lin, Pengchong Jin, Yin Cui, Mingxing Tan,
Quoc Le, Xiaodan Song.
Efficient Scale-Permuted Backbone with Learned Resource Distribution.
(https://arxiv.org/abs/2010.11426).
"""
def __init__(self,
input_specs=tf.keras.layers.InputSpec(shape=[None, 512, 512, 3]),
min_level=3,
max_level=7,
block_specs=build_block_specs(),
endpoints_num_filters=48,
se_ratio=0.2,
block_repeats=1,
filter_size_scale=1.0,
expand_ratio=6,
init_stochastic_depth_rate=0.0,
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 Mobile SpineNet model.
Args:
input_specs: A `tf.keras.layers.InputSpec` of the input tensor.
min_level: An `int` of min level for output mutiscale features.
max_level: An `int` of max level for output mutiscale features.
block_specs: The block specifications for the SpineNet model discovered by
NAS.
endpoints_num_filters: An `int` of feature dimension for the output
endpoints.
se_ratio: A `float` of Squeeze-and-Excitation ratio.
block_repeats: An `int` of number of blocks contained in the layer.
filter_size_scale: A `float` of multiplier for the filters (number of
channels) for all convolution ops. The value must be greater than zero.
Typical usage will be to set this value in (0, 1) to reduce the number
of parameters or computation cost of the model.
expand_ratio: An `integer` of expansion ratios for inverted bottleneck
blocks.
init_stochastic_depth_rate: A `float` of initial stochastic depth rate.
kernel_initializer: A str for kernel initializer of 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: If True, use synchronized batch normalization.
norm_momentum: A `float` of normalization momentum for the moving average.
norm_epsilon: A small `float` added to variance to avoid dividing by zero.
**kwargs: Additional keyword arguments to be passed.
"""
self._input_specs = input_specs
self._min_level = min_level
self._max_level = max_level
self._block_specs = block_specs
self._endpoints_num_filters = endpoints_num_filters
self._se_ratio = se_ratio
self._block_repeats = block_repeats
self._filter_size_scale = filter_size_scale
self._expand_ratio = expand_ratio
self._init_stochastic_depth_rate = init_stochastic_depth_rate
self._kernel_initializer = kernel_initializer
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
self._activation = activation
self._use_sync_bn = use_sync_bn
self._norm_momentum = norm_momentum
self._norm_epsilon = norm_epsilon
if activation == 'relu':
self._activation_fn = tf.nn.relu
elif activation == 'swish':
self._activation_fn = tf.nn.swish
else:
raise ValueError('Activation {} not implemented.'.format(activation))
self._num_init_blocks = 2
if use_sync_bn:
self._norm = layers.experimental.SyncBatchNormalization
else:
self._norm = layers.BatchNormalization
if tf.keras.backend.image_data_format() == 'channels_last':
self._bn_axis = -1
else:
self._bn_axis = 1
# Build SpineNet.
inputs = tf.keras.Input(shape=input_specs.shape[1:])
net = self._build_stem(inputs=inputs)
net = self._build_scale_permuted_network(
net=net, input_width=input_specs.shape[2])
endpoints = self._build_endpoints(net=net)
self._output_specs = {l: endpoints[l].get_shape() for l in endpoints}
super().__init__(inputs=inputs, outputs=endpoints)
def _block_group(self,
inputs,
in_filters,
out_filters,
strides,
expand_ratio=6,
block_repeats=1,
se_ratio=0.2,
stochastic_depth_drop_rate=None,
name='block_group'):
"""Creates one group of blocks for the SpineNet model."""
x = nn_blocks.InvertedBottleneckBlock(
in_filters=in_filters,
out_filters=out_filters,
strides=strides,
se_ratio=se_ratio,
expand_ratio=expand_ratio,
stochastic_depth_drop_rate=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)(
inputs)
for _ in range(1, block_repeats):
x = nn_blocks.InvertedBottleneckBlock(
in_filters=in_filters,
out_filters=out_filters,
strides=1,
se_ratio=se_ratio,
expand_ratio=expand_ratio,
stochastic_depth_drop_rate=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)(
inputs)
return tf.identity(x, name=name)
def _build_stem(self, inputs):
"""Builds SpineNet stem."""
x = layers.Conv2D(
filters=int(FILTER_SIZE_MAP[0] * self._filter_size_scale),
kernel_size=3,
strides=2,
use_bias=False,
padding='same',
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
inputs)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
x = tf_utils.get_activation(self._activation_fn)(x)
net = []
stem_strides = [1, 2]
# Build the initial level 2 blocks.
for i in range(self._num_init_blocks):
x = self._block_group(
inputs=x,
in_filters=int(FILTER_SIZE_MAP[i] * self._filter_size_scale),
out_filters=int(FILTER_SIZE_MAP[i + 1] * self._filter_size_scale),
expand_ratio=self._expand_ratio,
strides=stem_strides[i],
se_ratio=self._se_ratio,
block_repeats=self._block_repeats,
name='stem_block_{}'.format(i + 1))
net.append(x)
return net
def _build_scale_permuted_network(self,
net,
input_width,
weighted_fusion=False):
"""Builds scale-permuted network."""
net_sizes = [
int(math.ceil(input_width / 2)),
int(math.ceil(input_width / 2**2))
]
num_outgoing_connections = [0] * len(net)
endpoints = {}
for i, block_spec in enumerate(self._block_specs):
# Find out specs for the target block.
target_width = int(math.ceil(input_width / 2**block_spec.level))
target_num_filters = int(FILTER_SIZE_MAP[block_spec.level] *
self._filter_size_scale)
# Resample then merge input0 and input1.
parents = []
input0 = block_spec.input_offsets[0]
input1 = block_spec.input_offsets[1]
x0 = self._resample_with_sepconv(
inputs=net[input0],
input_width=net_sizes[input0],
target_width=target_width,
target_num_filters=target_num_filters)
parents.append(x0)
num_outgoing_connections[input0] += 1
x1 = self._resample_with_sepconv(
inputs=net[input1],
input_width=net_sizes[input1],
target_width=target_width,
target_num_filters=target_num_filters)
parents.append(x1)
num_outgoing_connections[input1] += 1
# Merge 0 outdegree blocks to the output block.
if block_spec.is_output:
for j, (j_feat,
j_connections) in enumerate(zip(net, num_outgoing_connections)):
if j_connections == 0 and (j_feat.shape[2] == target_width and
j_feat.shape[3] == x0.shape[3]):
parents.append(j_feat)
num_outgoing_connections[j] += 1
# pylint: disable=g-direct-tensorflow-import
if weighted_fusion:
dtype = parents[0].dtype
parent_weights = [
tf.nn.relu(tf.cast(tf.Variable(1.0, name='block{}_fusion{}'.format(
i, j)), dtype=dtype)) for j in range(len(parents))]
weights_sum = tf.add_n(parent_weights)
parents = [
parents[i] * parent_weights[i] / (weights_sum + 0.0001)
for i in range(len(parents))
]
# Fuse all parent nodes then build a new block.
x = tf_utils.get_activation(self._activation_fn)(tf.add_n(parents))
x = self._block_group(
inputs=x,
in_filters=target_num_filters,
out_filters=target_num_filters,
strides=1,
se_ratio=self._se_ratio,
expand_ratio=self._expand_ratio,
block_repeats=self._block_repeats,
stochastic_depth_drop_rate=nn_layers.get_stochastic_depth_rate(
self._init_stochastic_depth_rate, i + 1, len(self._block_specs)),
name='scale_permuted_block_{}'.format(i + 1))
net.append(x)
net_sizes.append(target_width)
num_outgoing_connections.append(0)
# Save output feats.
if block_spec.is_output:
if block_spec.level in endpoints:
raise ValueError('Duplicate feats found for output level {}.'.format(
block_spec.level))
if (block_spec.level < self._min_level or
block_spec.level > self._max_level):
raise ValueError('Output level is out of range [{}, {}]'.format(
self._min_level, self._max_level))
endpoints[str(block_spec.level)] = x
return endpoints
def _build_endpoints(self, net):
"""Matches filter size for endpoints before sharing conv layers."""
endpoints = {}
for level in range(self._min_level, self._max_level + 1):
x = layers.Conv2D(
filters=self._endpoints_num_filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
net[str(level)])
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
x = tf_utils.get_activation(self._activation_fn)(x)
endpoints[str(level)] = x
return endpoints
def _resample_with_sepconv(self, inputs, input_width, target_width,
target_num_filters):
"""Matches resolution and feature dimension."""
x = inputs
# Spatial resampling.
if input_width > target_width:
while input_width > target_width:
x = layers.DepthwiseConv2D(
kernel_size=3,
strides=2,
padding='SAME',
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
x)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
x = tf_utils.get_activation(self._activation_fn)(x)
input_width /= 2
elif input_width < target_width:
scale = target_width // input_width
x = spatial_transform_ops.nearest_upsampling(x, scale=scale)
# Last 1x1 conv to match filter size.
x = layers.Conv2D(
filters=target_num_filters,
kernel_size=1,
strides=1,
use_bias=False,
kernel_initializer=self._kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
x)
x = self._norm(
axis=self._bn_axis,
momentum=self._norm_momentum,
epsilon=self._norm_epsilon)(
x)
return x
def get_config(self):
config_dict = {
'min_level': self._min_level,
'max_level': self._max_level,
'endpoints_num_filters': self._endpoints_num_filters,
'se_ratio': self._se_ratio,
'expand_ratio': self._expand_ratio,
'block_repeats': self._block_repeats,
'filter_size_scale': self._filter_size_scale,
'init_stochastic_depth_rate': self._init_stochastic_depth_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
}
return config_dict
@classmethod
def from_config(cls, config, custom_objects=None):
return cls(**config)
@property
def output_specs(self):
"""A dict of {level: TensorShape} pairs for the model output."""
return self._output_specs
@factory.register_backbone_builder('spinenet_mobile')
def build_spinenet_mobile(
input_specs: tf.keras.layers.InputSpec,
model_config,
l2_regularizer: tf.keras.regularizers.Regularizer = None) -> tf.keras.Model:
"""Builds Mobile SpineNet backbone from a config."""
backbone_type = model_config.backbone.type
backbone_cfg = model_config.backbone.get()
norm_activation_config = model_config.norm_activation
assert backbone_type == 'spinenet_mobile', (f'Inconsistent backbone type '
f'{backbone_type}')
model_id = backbone_cfg.model_id
if model_id not in SCALING_MAP:
raise ValueError(
'Mobile SpineNet-{} is not a valid architecture.'.format(model_id))
scaling_params = SCALING_MAP[model_id]
return SpineNetMobile(
input_specs=input_specs,
min_level=model_config.min_level,
max_level=model_config.max_level,
endpoints_num_filters=scaling_params['endpoints_num_filters'],
block_repeats=scaling_params['block_repeats'],
filter_size_scale=scaling_params['filter_size_scale'],
se_ratio=backbone_cfg.se_ratio,
expand_ratio=backbone_cfg.expand_ratio,
init_stochastic_depth_rate=backbone_cfg.stochastic_depth_drop_rate,
kernel_regularizer=l2_regularizer,
activation=norm_activation_config.activation,
use_sync_bn=norm_activation_config.use_sync_bn,
norm_momentum=norm_activation_config.norm_momentum,
norm_epsilon=norm_activation_config.norm_epsilon)
official/vision/beta/modeling/backbones/spinenet_mobile_test.py 0 → 100644
View file @ 790e49e5
# Copyright 2021 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
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Tests for SpineNet."""
# Import libraries
from absl.testing import parameterized
import tensorflow as tf
from official.vision.beta.modeling.backbones import spinenet_mobile
class SpineNetMobileTest(parameterized.TestCase, tf.test.TestCase):
@parameterized.parameters(
(128, 0.6, 1, 0.0, 24),
(128, 0.65, 1, 0.2, 40),
(256, 1.0, 1, 0.2, 48),
)
def test_network_creation(self, input_size, filter_size_scale, block_repeats,
se_ratio, endpoints_num_filters):
"""Test creation of SpineNet models."""
min_level = 3
max_level = 7
tf.keras.backend.set_image_data_format('channels_last')
input_specs = tf.keras.layers.InputSpec(
shape=[None, input_size, input_size, 3])
model = spinenet_mobile.SpineNetMobile(
input_specs=input_specs,
min_level=min_level,
max_level=max_level,
endpoints_num_filters=endpoints_num_filters,
resample_alpha=se_ratio,
block_repeats=block_repeats,
filter_size_scale=filter_size_scale,
init_stochastic_depth_rate=0.2,
)
inputs = tf.keras.Input(shape=(input_size, input_size, 3), batch_size=1)
endpoints = model(inputs)
for l in range(min_level, max_level + 1):
self.assertIn(str(l), endpoints.keys())
self.assertAllEqual(
[1, input_size / 2**l, input_size / 2**l, endpoints_num_filters],
endpoints[str(l)].shape.as_list())
def test_serialize_deserialize(self):
# Create a network object that sets all of its config options.
kwargs = dict(
min_level=3,
max_level=7,
endpoints_num_filters=256,
se_ratio=0.2,
expand_ratio=6,
block_repeats=1,
filter_size_scale=1.0,
init_stochastic_depth_rate=0.2,
use_sync_bn=False,
activation='relu',
norm_momentum=0.99,
norm_epsilon=0.001,
kernel_initializer='VarianceScaling',
kernel_regularizer=None,
bias_regularizer=None,
)
network = spinenet_mobile.SpineNetMobile(**kwargs)
expected_config = dict(kwargs)
self.assertEqual(network.get_config(), expected_config)
# Create another network object from the first object's config.
new_network = spinenet_mobile.SpineNetMobile.from_config(
network.get_config())
# Validate that the config can be forced to JSON.
_ = new_network.to_json()
# If the serialization was successful, the new config should match the old.
self.assertAllEqual(network.get_config(), new_network.get_config())
if __name__ == '__main__':
tf.test.main()
official/vision/beta/modeling/backbones/spinenet_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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 SpineNet."""
# Import libraries
from absl.testing import parameterized
......
official/vision/beta/modeling/classification_model.py
View file @ 790e49e5
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -11,7 +11,7 @@
# 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.
# ==============================================================================
"""Build classification models."""
# Import libraries
......@@ -59,28 +59,10 @@ class ClassificationModel(tf.keras.Model):
skip_logits_layer: `bool`, whether to skip the prediction layer.
**kwargs: keyword arguments to be passed.
"""
self._self_setattr_tracking = False
self._config_dict = {
'backbone': backbone,
'num_classes': num_classes,
'input_specs': input_specs,
'dropout_rate': dropout_rate,
'kernel_initializer': kernel_initializer,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
'add_head_batch_norm': add_head_batch_norm,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
}
self._input_specs = input_specs
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
self._backbone = backbone
if use_sync_bn:
self._norm = tf.keras.layers.experimental.SyncBatchNormalization
norm = tf.keras.layers.experimental.SyncBatchNormalization
else:
self._norm = tf.keras.layers.BatchNormalization
norm = tf.keras.layers.BatchNormalization
axis = -1 if tf.keras.backend.image_data_format() == 'channels_last' else 1
inputs = tf.keras.Input(shape=input_specs.shape[1:])
......@@ -88,18 +70,37 @@ class ClassificationModel(tf.keras.Model):
x = endpoints[max(endpoints.keys())]
if add_head_batch_norm:
x = self._norm(axis=axis, momentum=norm_momentum, epsilon=norm_epsilon)(x)
x = norm(axis=axis, momentum=norm_momentum, epsilon=norm_epsilon)(x)
x = tf.keras.layers.GlobalAveragePooling2D()(x)
if not skip_logits_layer:
x = tf.keras.layers.Dropout(dropout_rate)(x)
x = tf.keras.layers.Dense(
num_classes, kernel_initializer=kernel_initializer,
kernel_regularizer=self._kernel_regularizer,
bias_regularizer=self._bias_regularizer)(
num_classes,
kernel_initializer=kernel_initializer,
kernel_regularizer=kernel_regularizer,
bias_regularizer=bias_regularizer)(
x)
super(ClassificationModel, self).__init__(
inputs=inputs, outputs=x, **kwargs)
self._config_dict = {
'backbone': backbone,
'num_classes': num_classes,
'input_specs': input_specs,
'dropout_rate': dropout_rate,
'kernel_initializer': kernel_initializer,
'kernel_regularizer': kernel_regularizer,
'bias_regularizer': bias_regularizer,
'add_head_batch_norm': add_head_batch_norm,
'use_sync_bn': use_sync_bn,
'norm_momentum': norm_momentum,
'norm_epsilon': norm_epsilon,
}
self._input_specs = input_specs
self._kernel_regularizer = kernel_regularizer
self._bias_regularizer = bias_regularizer
self._backbone = backbone
self._norm = norm
@property
def checkpoint_items(self):
......
official/vision/beta/modeling/classification_model_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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 classification network."""
# Import libraries
......
official/vision/beta/modeling/decoders/__init__.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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
"""Decoders package definition."""
from official.vision.beta.modeling.decoders.aspp import ASPP
......
official/vision/beta/modeling/decoders/aspp.py
View file @ 790e49e5
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -11,8 +11,8 @@
# 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.
# ==============================================================================
"""ASPP decoder."""
"""Contains definitions of Atrous Spatial Pyramid Pooling (ASPP) decoder."""
# Import libraries
import tensorflow as tf
......@@ -22,7 +22,7 @@ from official.vision import keras_cv
@tf.keras.utils.register_keras_serializable(package='Vision')
class ASPP(tf.keras.layers.Layer):
"""ASPP."""
"""Creates an Atrous Spatial Pyramid Pooling (ASPP) layer."""
def __init__(self,
level,
......@@ -38,26 +38,28 @@ class ASPP(tf.keras.layers.Layer):
kernel_regularizer=None,
interpolation='bilinear',
**kwargs):
"""ASPP initialization function.
"""Initializes an Atrous Spatial Pyramid Pooling (ASPP) layer.
Args:
level: `int` level to apply ASPP.
dilation_rates: `list` of dilation rates.
num_filters: `int` number of output filters in ASPP.
pool_kernel_size: `list` of [height, width] of pooling kernel size or
level: An `int` level to apply ASPP.
dilation_rates: A `list` of dilation rates.
num_filters: An `int` number of output filters in ASPP.
pool_kernel_size: A `list` of [height, width] of pooling kernel size or
None. Pooling size is with respect to original image size, it will be
scaled down by 2**level. If None, global average pooling is used.
use_sync_bn: if True, use synchronized batch normalization.
norm_momentum: `float` normalization omentum for the moving average.
norm_epsilon: `float` small float added to variance to avoid dividing by
zero.
activation: `str` activation to be used in ASPP.
dropout_rate: `float` rate for dropout regularization.
kernel_initializer: kernel_initializer for convolutional layers.
kernel_regularizer: tf.keras.regularizers.Regularizer object for Conv2D.
interpolation: interpolation method, one of bilinear, nearest, bicubic,
area, lanczos3, lanczos5, gaussian, or mitchellcubic.
**kwargs: keyword arguments to be passed.
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.
activation: A `str` activation to be used in ASPP.
dropout_rate: A `float` rate for dropout regularization.
kernel_initializer: A `str` name of kernel_initializer for convolutional
layers.
kernel_regularizer: A `tf.keras.regularizers.Regularizer` object for
Conv2D. Default is None.
interpolation: A `str` of interpolation method. It should be one of
`bilinear`, `nearest`, `bicubic`, `area`, `lanczos3`, `lanczos5`,
`gaussian`, or `mitchellcubic`.
**kwargs: Additional keyword arguments to be passed.
"""
super(ASPP, self).__init__(**kwargs)
self._config_dict = {
......@@ -96,20 +98,22 @@ class ASPP(tf.keras.layers.Layer):
interpolation=self._config_dict['interpolation'])
def call(self, inputs):
"""ASPP call method.
"""Calls the Atrous Spatial Pyramid Pooling (ASPP) layer on an input.
The output of ASPP will be a dict of level, Tensor even if only one
The output of ASPP will be a dict of {`level`, `tf.Tensor`} even if only one
level is present. Hence, this will be compatible with the rest of the
segmentation model interfaces..
segmentation model interfaces.
Args:
inputs: A dict of tensors
- key: `str`, the level of the multilevel feature maps.
- values: `Tensor`, [batch, height_l, width_l, filter_size].
inputs: A `dict` of `tf.Tensor` where
- key: A `str` of the level of the multilevel feature maps.
- values: A `tf.Tensor` of shape [batch, height_l, width_l,
filter_size].
Returns:
A dict of tensors
- key: `str`, the level of the multilevel feature maps.
- values: `Tensor`, output of ASPP module.
A `dict` of `tf.Tensor` where
- key: A `str` of the level of the multilevel feature maps.
- values: A `tf.Tensor` of output of ASPP module.
"""
outputs = {}
level = str(self._config_dict['level'])
......
official/vision/beta/modeling/decoders/aspp_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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 aspp."""
# Import libraries
......
official/vision/beta/modeling/decoders/factory.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,8 +11,10 @@
# 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.
# ==============================================================================
"""factory method."""
# Lint as: python3
"""Contains the factory method to create decoders."""
# Import libraries
import tensorflow as tf
......@@ -26,13 +27,14 @@ def build_decoder(input_specs,
"""Builds decoder from a config.
Args:
input_specs: `dict` input specifications. A dictionary consists of
input_specs: A `dict` of input specifications. A dictionary consists of
{level: TensorShape} from a backbone.
model_config: A OneOfConfig. Model config.
l2_regularizer: tf.keras.regularizers.Regularizer instance. Default to None.
l2_regularizer: A `tf.keras.regularizers.Regularizer` instance. Default to
None.
Returns:
tf.keras.Model instance of the decoder.
A `tf.keras.Model` instance of the decoder.
"""
decoder_type = model_config.decoder.type
decoder_cfg = model_config.decoder.get()
......
official/vision/beta/modeling/decoders/fpn.py
View file @ 790e49e5
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -11,14 +11,8 @@
# 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.
# ==============================================================================
"""Feature Pyramid Networks.
Feature Pyramid Networks were proposed in:
[1] Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan,
, and Serge Belongie
Feature Pyramid Networks for Object Detection. CVPR 2017.
"""
"""Contains the definitions of Feature Pyramid Networks (FPN)."""
# Import libraries
import tensorflow as tf
......@@ -29,7 +23,14 @@ from official.vision.beta.ops import spatial_transform_ops
@tf.keras.utils.register_keras_serializable(package='Vision')
class FPN(tf.keras.Model):
"""Feature pyramid network."""
"""Creates a Feature Pyramid Network (FPN).
This implemets the paper:
Tsung-Yi Lin, Piotr Dollar, Ross Girshick, Kaiming He, Bharath Hariharan, and
Serge Belongie.
Feature Pyramid Networks for Object Detection.
(https://arxiv.org/pdf/1612.03144)
"""
def __init__(self,
input_specs,
......@@ -45,25 +46,26 @@ class FPN(tf.keras.Model):
kernel_regularizer=None,
bias_regularizer=None,
**kwargs):
"""FPN initialization function.
"""Initializes a Feature Pyramid Network (FPN).
Args:
input_specs: `dict` input specifications. A dictionary consists of
input_specs: A `dict` of input specifications. A dictionary consists of
{level: TensorShape} from a backbone.
min_level: `int` minimum level in FPN output feature maps.
max_level: `int` maximum level in FPN output feature maps.
num_filters: `int` number of filters in FPN layers.
use_separable_conv: `bool`, if True use separable convolution for
min_level: An `int` of minimum level in FPN output feature maps.
max_level: An `int` of maximum level in FPN output feature maps.
num_filters: An `int` number of filters in FPN layers.
use_separable_conv: A `bool`. If True use separable convolution for
convolution in FPN layers.
activation: `str` name of the activation function.
use_sync_bn: if True, use synchronized batch normalization.
norm_momentum: `float` normalization omentum for the moving average.
norm_epsilon: `float` small float added to variance to avoid dividing by
zero.
kernel_initializer: kernel_initializer for convolutional layers.
kernel_regularizer: tf.keras.regularizers.Regularizer object for Conv2D.
bias_regularizer: tf.keras.regularizers.Regularizer object for Conv2d.
**kwargs: keyword arguments to be passed.
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.
kernel_initializer: A `str` name of kernel_initializer for convolutional
layers.
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.
"""
self._config_dict = {
'input_specs': input_specs,
......
official/vision/beta/modeling/decoders/fpn_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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 FPN."""
# Import libraries
......
official/vision/beta/modeling/decoders/nasfpn.py
View file @ 790e49e5
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -11,13 +11,8 @@
# 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.
# ==============================================================================
"""NAS-FPN.
Golnaz Ghiasi, Tsung-Yi Lin, Ruoming Pang, Quoc V. Le.
NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection.
https://arxiv.org/abs/1904.07392. CVPR 2019.
"""
"""Contains definitions of NAS-FPN."""
# Import libraries
from absl import logging
......@@ -60,7 +55,13 @@ def build_block_specs(block_specs=None):
@tf.keras.utils.register_keras_serializable(package='Vision')
class NASFPN(tf.keras.Model):
"""NAS-FPN."""
"""Creates a NAS-FPN model.
This implements the paper:
Golnaz Ghiasi, Tsung-Yi Lin, Ruoming Pang, Quoc V. Le.
NAS-FPN: Learning Scalable Feature Pyramid Architecture for Object Detection.
(https://arxiv.org/abs/1904.07392)
"""
def __init__(self,
input_specs,
......@@ -78,29 +79,30 @@ class NASFPN(tf.keras.Model):
kernel_regularizer=None,
bias_regularizer=None,
**kwargs):
"""FPN initialization function.
"""Initializes a NAS-FPN model.
Args:
input_specs: `dict` input specifications. A dictionary consists of
input_specs: A `dict` of input specifications. A dictionary consists of
{level: TensorShape} from a backbone.
min_level: `int` minimum level in FPN output feature maps.
max_level: `int` maximum level in FPN output feature maps.
min_level: An `int` of minimum level in FPN output feature maps.
max_level: An `int` of maximum level in FPN output feature maps.
block_specs: a list of BlockSpec objects that specifies the NAS-FPN
network topology. By default, the previously discovered architecture is
used.
num_filters: `int` number of filters in FPN layers.
num_filters: An `int` number of filters in FPN layers.
num_repeats: number of repeats for feature pyramid network.
use_separable_conv: `bool`, if True use separable convolution for
use_separable_conv: A `bool`. If True use separable convolution for
convolution in FPN layers.
activation: `str` name of the activation function.
use_sync_bn: if True, use synchronized batch normalization.
norm_momentum: `float` normalization omentum for the moving average.
norm_epsilon: `float` small float added to variance to avoid dividing by
zero.
kernel_initializer: kernel_initializer for convolutional layers.
kernel_regularizer: tf.keras.regularizers.Regularizer object for Conv2D.
bias_regularizer: tf.keras.regularizers.Regularizer object for Conv2d.
**kwargs: keyword arguments to be passed.
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.
kernel_initializer: A `str` name of kernel_initializer for convolutional
layers.
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.
"""
self._config_dict = {
'input_specs': input_specs,
......
official/vision/beta/modeling/decoders/nasfpn_test.py
View file @ 790e49e5
# Lint as: python3
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -12,7 +11,8 @@
# 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 NAS-FPN."""
# Import libraries
......
official/vision/beta/modeling/factory.py
View file @ 790e49e5
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -11,7 +11,7 @@
# 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.
# ==============================================================================
"""Factory methods to build models."""
# Import libraries
......
official/vision/beta/modeling/factory_3d.py
View file @ 790e49e5
# Copyright 2020 The TensorFlow Authors. All Rights Reserved.
# Copyright 2021 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.
......@@ -11,7 +11,7 @@
# 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.
# ==============================================================================
"""Factory methods to build models."""
# Import libraries
......
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