Migrate official/vision/image_classification to official/legacy/image_classification

PiperOrigin-RevId: 410878028

official/legacy/image_classification/efficientnet/efficientnet_config.py

+# 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
+"""Configuration definitions for EfficientNet losses, learning rates, and optimizers."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import dataclasses
+from official.legacy.image_classification.configs import base_configs
+from official.modeling.hyperparams import base_config
+
+
+@dataclasses.dataclass
+class EfficientNetModelConfig(base_configs.ModelConfig):
+  """Configuration for the EfficientNet model.
+
+  This configuration will default to settings used for training efficientnet-b0
+  on a v3-8 TPU on ImageNet.
+
+  Attributes:
+    name: The name of the model. Defaults to 'EfficientNet'.
+    num_classes: The number of classes in the model.
+    model_params: A dictionary that represents the parameters of the
+      EfficientNet model. These will be passed in to the "from_name" function.
+    loss: The configuration for loss. Defaults to a categorical cross entropy
+      implementation.
+    optimizer: The configuration for optimizations. Defaults to an RMSProp
+      configuration.
+    learning_rate: The configuration for learning rate. Defaults to an
+      exponential configuration.
+  """
+  name: str = 'EfficientNet'
+  num_classes: int = 1000
+  model_params: base_config.Config = dataclasses.field(
+      default_factory=lambda: {
+          'model_name': 'efficientnet-b0',
+          'model_weights_path': '',
+          'weights_format': 'saved_model',
+          'overrides': {
+              'batch_norm': 'default',
+              'rescale_input': True,
+              'num_classes': 1000,
+              'activation': 'swish',
+              'dtype': 'float32',
+          }
+      })
+  loss: base_configs.LossConfig = base_configs.LossConfig(
+      name='categorical_crossentropy', label_smoothing=0.1)
+  optimizer: base_configs.OptimizerConfig = base_configs.OptimizerConfig(
+      name='rmsprop',
+      decay=0.9,
+      epsilon=0.001,
+      momentum=0.9,
+      moving_average_decay=None)
+  learning_rate: base_configs.LearningRateConfig = base_configs.LearningRateConfig(  # pylint: disable=line-too-long
+      name='exponential',
+      initial_lr=0.008,
+      decay_epochs=2.4,
+      decay_rate=0.97,
+      warmup_epochs=5,
+      scale_by_batch_size=1. / 128.,
+      staircase=True)

official/legacy/image_classification/efficientnet/efficientnet_model.py

+# 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
+"""Contains definitions for EfficientNet model.
+
+[1] Mingxing Tan, Quoc V. Le
+  EfficientNet: Rethinking Model Scaling for Convolutional Neural Networks.
+  ICML'19, https://arxiv.org/abs/1905.11946
+"""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+import dataclasses
+import math
+from typing import Any, Dict, Optional, Text, Tuple
+
+from absl import logging
+import tensorflow as tf
+from official.legacy.image_classification import preprocessing
+from official.legacy.image_classification.efficientnet import common_modules
+from official.modeling import tf_utils
+from official.modeling.hyperparams import base_config
+
+
+@dataclasses.dataclass
+class BlockConfig(base_config.Config):
+  """Config for a single MB Conv Block."""
+  input_filters: int = 0
+  output_filters: int = 0
+  kernel_size: int = 3
+  num_repeat: int = 1
+  expand_ratio: int = 1
+  strides: Tuple[int, int] = (1, 1)
+  se_ratio: Optional[float] = None
+  id_skip: bool = True
+  fused_conv: bool = False
+  conv_type: str = 'depthwise'
+
+
+@dataclasses.dataclass
+class ModelConfig(base_config.Config):
+  """Default Config for Efficientnet-B0."""
+  width_coefficient: float = 1.0
+  depth_coefficient: float = 1.0
+  resolution: int = 224
+  dropout_rate: float = 0.2
+  blocks: Tuple[BlockConfig, ...] = (
+      # (input_filters, output_filters, kernel_size, num_repeat,
+      #  expand_ratio, strides, se_ratio)
+      # pylint: disable=bad-whitespace
+      BlockConfig.from_args(32, 16, 3, 1, 1, (1, 1), 0.25),
+      BlockConfig.from_args(16, 24, 3, 2, 6, (2, 2), 0.25),
+      BlockConfig.from_args(24, 40, 5, 2, 6, (2, 2), 0.25),
+      BlockConfig.from_args(40, 80, 3, 3, 6, (2, 2), 0.25),
+      BlockConfig.from_args(80, 112, 5, 3, 6, (1, 1), 0.25),
+      BlockConfig.from_args(112, 192, 5, 4, 6, (2, 2), 0.25),
+      BlockConfig.from_args(192, 320, 3, 1, 6, (1, 1), 0.25),
+      # pylint: enable=bad-whitespace
+  )
+  stem_base_filters: int = 32
+  top_base_filters: int = 1280
+  activation: str = 'simple_swish'
+  batch_norm: str = 'default'
+  bn_momentum: float = 0.99
+  bn_epsilon: float = 1e-3
+  # While the original implementation used a weight decay of 1e-5,
+  # tf.nn.l2_loss divides it by 2, so we halve this to compensate in Keras
+  weight_decay: float = 5e-6
+  drop_connect_rate: float = 0.2
+  depth_divisor: int = 8
+  min_depth: Optional[int] = None
+  use_se: bool = True
+  input_channels: int = 3
+  num_classes: int = 1000
+  model_name: str = 'efficientnet'
+  rescale_input: bool = True
+  data_format: str = 'channels_last'
+  dtype: str = 'float32'
+
+
+MODEL_CONFIGS = {
+    # (width, depth, resolution, dropout)
+    'efficientnet-b0': ModelConfig.from_args(1.0, 1.0, 224, 0.2),
+    'efficientnet-b1': ModelConfig.from_args(1.0, 1.1, 240, 0.2),
+    'efficientnet-b2': ModelConfig.from_args(1.1, 1.2, 260, 0.3),
+    'efficientnet-b3': ModelConfig.from_args(1.2, 1.4, 300, 0.3),
+    'efficientnet-b4': ModelConfig.from_args(1.4, 1.8, 380, 0.4),
+    'efficientnet-b5': ModelConfig.from_args(1.6, 2.2, 456, 0.4),
+    'efficientnet-b6': ModelConfig.from_args(1.8, 2.6, 528, 0.5),
+    'efficientnet-b7': ModelConfig.from_args(2.0, 3.1, 600, 0.5),
+    'efficientnet-b8': ModelConfig.from_args(2.2, 3.6, 672, 0.5),
+    'efficientnet-l2': ModelConfig.from_args(4.3, 5.3, 800, 0.5),
+}
+
+CONV_KERNEL_INITIALIZER = {
+    'class_name': 'VarianceScaling',
+    'config': {
+        'scale': 2.0,
+        'mode': 'fan_out',
+        # Note: this is a truncated normal distribution
+        'distribution': 'normal'
+    }
+}
+
+DENSE_KERNEL_INITIALIZER = {
+    'class_name': 'VarianceScaling',
+    'config': {
+        'scale': 1 / 3.0,
+        'mode': 'fan_out',
+        'distribution': 'uniform'
+    }
+}
+
+
+def round_filters(filters: int, config: ModelConfig) -> int:
+  """Round number of filters based on width coefficient."""
+  width_coefficient = config.width_coefficient
+  min_depth = config.min_depth
+  divisor = config.depth_divisor
+  orig_filters = filters
+
+  if not width_coefficient:
+    return filters
+
+  filters *= width_coefficient
+  min_depth = min_depth or divisor
+  new_filters = max(min_depth, int(filters + divisor / 2) // divisor * divisor)
+  # Make sure that round down does not go down by more than 10%.
+  if new_filters < 0.9 * filters:
+    new_filters += divisor
+  logging.info('round_filter input=%s output=%s', orig_filters, new_filters)
+  return int(new_filters)
+
+
+def round_repeats(repeats: int, depth_coefficient: float) -> int:
+  """Round number of repeats based on depth coefficient."""
+  return int(math.ceil(depth_coefficient * repeats))
+
+
+def conv2d_block(inputs: tf.Tensor,
+                 conv_filters: Optional[int],
+                 config: ModelConfig,
+                 kernel_size: Any = (1, 1),
+                 strides: Any = (1, 1),
+                 use_batch_norm: bool = True,
+                 use_bias: bool = False,
+                 activation: Optional[Any] = None,
+                 depthwise: bool = False,
+                 name: Optional[Text] = None):
+  """A conv2d followed by batch norm and an activation."""
+  batch_norm = common_modules.get_batch_norm(config.batch_norm)
+  bn_momentum = config.bn_momentum
+  bn_epsilon = config.bn_epsilon
+  data_format = tf.keras.backend.image_data_format()
+  weight_decay = config.weight_decay
+
+  name = name or ''
+
+  # Collect args based on what kind of conv2d block is desired
+  init_kwargs = {
+      'kernel_size': kernel_size,
+      'strides': strides,
+      'use_bias': use_bias,
+      'padding': 'same',
+      'name': name + '_conv2d',
+      'kernel_regularizer': tf.keras.regularizers.l2(weight_decay),
+      'bias_regularizer': tf.keras.regularizers.l2(weight_decay),
+  }
+
+  if depthwise:
+    conv2d = tf.keras.layers.DepthwiseConv2D
+    init_kwargs.update({'depthwise_initializer': CONV_KERNEL_INITIALIZER})
+  else:
+    conv2d = tf.keras.layers.Conv2D
+    init_kwargs.update({
+        'filters': conv_filters,
+        'kernel_initializer': CONV_KERNEL_INITIALIZER
+    })
+
+  x = conv2d(**init_kwargs)(inputs)
+
+  if use_batch_norm:
+    bn_axis = 1 if data_format == 'channels_first' else -1
+    x = batch_norm(
+        axis=bn_axis,
+        momentum=bn_momentum,
+        epsilon=bn_epsilon,
+        name=name + '_bn')(
+            x)
+
+  if activation is not None:
+    x = tf.keras.layers.Activation(activation, name=name + '_activation')(x)
+  return x
+
+
+def mb_conv_block(inputs: tf.Tensor,
+                  block: BlockConfig,
+                  config: ModelConfig,
+                  prefix: Optional[Text] = None):
+  """Mobile Inverted Residual Bottleneck.
+
+  Args:
+    inputs: the Keras input to the block
+    block: BlockConfig, arguments to create a Block
+    config: ModelConfig, a set of model parameters
+    prefix: prefix for naming all layers
+
+  Returns:
+    the output of the block
+  """
+  use_se = config.use_se
+  activation = tf_utils.get_activation(config.activation)
+  drop_connect_rate = config.drop_connect_rate
+  data_format = tf.keras.backend.image_data_format()
+  use_depthwise = block.conv_type != 'no_depthwise'
+  prefix = prefix or ''
+
+  filters = block.input_filters * block.expand_ratio
+
+  x = inputs
+
+  if block.fused_conv:
+    # If we use fused mbconv, skip expansion and use regular conv.
+    x = conv2d_block(
+        x,
+        filters,
+        config,
+        kernel_size=block.kernel_size,
+        strides=block.strides,
+        activation=activation,
+        name=prefix + 'fused')
+  else:
+    if block.expand_ratio != 1:
+      # Expansion phase
+      kernel_size = (1, 1) if use_depthwise else (3, 3)
+      x = conv2d_block(
+          x,
+          filters,
+          config,
+          kernel_size=kernel_size,
+          activation=activation,
+          name=prefix + 'expand')
+
+    # Depthwise Convolution
+    if use_depthwise:
+      x = conv2d_block(
+          x,
+          conv_filters=None,
+          config=config,
+          kernel_size=block.kernel_size,
+          strides=block.strides,
+          activation=activation,
+          depthwise=True,
+          name=prefix + 'depthwise')
+
+  # Squeeze and Excitation phase
+  if use_se:
+    assert block.se_ratio is not None
+    assert 0 < block.se_ratio <= 1
+    num_reduced_filters = max(1, int(block.input_filters * block.se_ratio))
+
+    if data_format == 'channels_first':
+      se_shape = (filters, 1, 1)
+    else:
+      se_shape = (1, 1, filters)
+
+    se = tf.keras.layers.GlobalAveragePooling2D(name=prefix + 'se_squeeze')(x)
+    se = tf.keras.layers.Reshape(se_shape, name=prefix + 'se_reshape')(se)
+
+    se = conv2d_block(
+        se,
+        num_reduced_filters,
+        config,
+        use_bias=True,
+        use_batch_norm=False,
+        activation=activation,
+        name=prefix + 'se_reduce')
+    se = conv2d_block(
+        se,
+        filters,
+        config,
+        use_bias=True,
+        use_batch_norm=False,
+        activation='sigmoid',
+        name=prefix + 'se_expand')
+    x = tf.keras.layers.multiply([x, se], name=prefix + 'se_excite')
+
+  # Output phase
+  x = conv2d_block(
+      x, block.output_filters, config, activation=None, name=prefix + 'project')
+
+  # Add identity so that quantization-aware training can insert quantization
+  # ops correctly.
+  x = tf.keras.layers.Activation(
+      tf_utils.get_activation('identity'), name=prefix + 'id')(
+          x)
+
+  if (block.id_skip and all(s == 1 for s in block.strides) and
+      block.input_filters == block.output_filters):
+    if drop_connect_rate and drop_connect_rate > 0:
+      # Apply dropconnect
+      # The only difference between dropout and dropconnect in TF is scaling by
+      # drop_connect_rate during training. See:
+      # https://github.com/keras-team/keras/pull/9898#issuecomment-380577612
+      x = tf.keras.layers.Dropout(
+          drop_connect_rate, noise_shape=(None, 1, 1, 1), name=prefix + 'drop')(
+              x)
+
+    x = tf.keras.layers.add([x, inputs], name=prefix + 'add')
+
+  return x
+
+
+def efficientnet(image_input: tf.keras.layers.Input, config: ModelConfig):  # pytype: disable=invalid-annotation  # typed-keras
+  """Creates an EfficientNet graph given the model parameters.
+
+  This function is wrapped by the `EfficientNet` class to make a tf.keras.Model.
+
+  Args:
+    image_input: the input batch of images
+    config: the model config
+
+  Returns:
+    the output of efficientnet
+  """
+  depth_coefficient = config.depth_coefficient
+  blocks = config.blocks
+  stem_base_filters = config.stem_base_filters
+  top_base_filters = config.top_base_filters
+  activation = tf_utils.get_activation(config.activation)
+  dropout_rate = config.dropout_rate
+  drop_connect_rate = config.drop_connect_rate
+  num_classes = config.num_classes
+  input_channels = config.input_channels
+  rescale_input = config.rescale_input
+  data_format = tf.keras.backend.image_data_format()
+  dtype = config.dtype
+  weight_decay = config.weight_decay
+
+  x = image_input
+  if data_format == 'channels_first':
+    # Happens on GPU/TPU if available.
+    x = tf.keras.layers.Permute((3, 1, 2))(x)
+  if rescale_input:
+    x = preprocessing.normalize_images(
+        x, num_channels=input_channels, dtype=dtype, data_format=data_format)
+
+  # Build stem
+  x = conv2d_block(
+      x,
+      round_filters(stem_base_filters, config),
+      config,
+      kernel_size=[3, 3],
+      strides=[2, 2],
+      activation=activation,
+      name='stem')
+
+  # Build blocks
+  num_blocks_total = sum(
+      round_repeats(block.num_repeat, depth_coefficient) for block in blocks)
+  block_num = 0
+
+  for stack_idx, block in enumerate(blocks):
+    assert block.num_repeat > 0
+    # Update block input and output filters based on depth multiplier
+    block = block.replace(
+        input_filters=round_filters(block.input_filters, config),
+        output_filters=round_filters(block.output_filters, config),
+        num_repeat=round_repeats(block.num_repeat, depth_coefficient))
+
+    # The first block needs to take care of stride and filter size increase
+    drop_rate = drop_connect_rate * float(block_num) / num_blocks_total
+    config = config.replace(drop_connect_rate=drop_rate)
+    block_prefix = 'stack_{}/block_0/'.format(stack_idx)
+    x = mb_conv_block(x, block, config, block_prefix)
+    block_num += 1
+    if block.num_repeat > 1:
+      block = block.replace(input_filters=block.output_filters, strides=[1, 1])
+
+      for block_idx in range(block.num_repeat - 1):
+        drop_rate = drop_connect_rate * float(block_num) / num_blocks_total
+        config = config.replace(drop_connect_rate=drop_rate)
+        block_prefix = 'stack_{}/block_{}/'.format(stack_idx, block_idx + 1)
+        x = mb_conv_block(x, block, config, prefix=block_prefix)
+        block_num += 1
+
+  # Build top
+  x = conv2d_block(
+      x,
+      round_filters(top_base_filters, config),
+      config,
+      activation=activation,
+      name='top')
+
+  # Build classifier
+  x = tf.keras.layers.GlobalAveragePooling2D(name='top_pool')(x)
+  if dropout_rate and dropout_rate > 0:
+    x = tf.keras.layers.Dropout(dropout_rate, name='top_dropout')(x)
+  x = tf.keras.layers.Dense(
+      num_classes,
+      kernel_initializer=DENSE_KERNEL_INITIALIZER,
+      kernel_regularizer=tf.keras.regularizers.l2(weight_decay),
+      bias_regularizer=tf.keras.regularizers.l2(weight_decay),
+      name='logits')(
+          x)
+  x = tf.keras.layers.Activation('softmax', name='probs')(x)
+
+  return x
+
+
+class EfficientNet(tf.keras.Model):
+  """Wrapper class for an EfficientNet Keras model.
+
+  Contains helper methods to build, manage, and save metadata about the model.
+  """
+
+  def __init__(self,
+               config: Optional[ModelConfig] = None,
+               overrides: Optional[Dict[Text, Any]] = None):
+    """Create an EfficientNet model.
+
+    Args:
+      config: (optional) the main model parameters to create the model
+      overrides: (optional) a dict containing keys that can override config
+    """
+    overrides = overrides or {}
+    config = config or ModelConfig()
+
+    self.config = config.replace(**overrides)
+
+    input_channels = self.config.input_channels
+    model_name = self.config.model_name
+    input_shape = (None, None, input_channels)  # Should handle any size image
+    image_input = tf.keras.layers.Input(shape=input_shape)
+
+    output = efficientnet(image_input, self.config)
+
+    # Cast to float32 in case we have a different model dtype
+    output = tf.cast(output, tf.float32)
+
+    logging.info('Building model %s with params %s', model_name, self.config)
+
+    super(EfficientNet, self).__init__(
+        inputs=image_input, outputs=output, name=model_name)
+
+  @classmethod
+  def from_name(cls,
+                model_name: Text,
+                model_weights_path: Optional[Text] = None,
+                weights_format: Text = 'saved_model',
+                overrides: Optional[Dict[Text, Any]] = None):
+    """Construct an EfficientNet model from a predefined model name.
+
+    E.g., `EfficientNet.from_name('efficientnet-b0')`.
+
+    Args:
+      model_name: the predefined model name
+      model_weights_path: the path to the weights (h5 file or saved model dir)
+      weights_format: the model weights format. One of 'saved_model', 'h5', or
+        'checkpoint'.
+      overrides: (optional) a dict containing keys that can override config
+
+    Returns:
+      A constructed EfficientNet instance.
+    """
+    model_configs = dict(MODEL_CONFIGS)
+    overrides = dict(overrides) if overrides else {}
+
+    # One can define their own custom models if necessary
+    model_configs.update(overrides.pop('model_config', {}))
+
+    if model_name not in model_configs:
+      raise ValueError('Unknown model name {}'.format(model_name))
+
+    config = model_configs[model_name]
+
+    model = cls(config=config, overrides=overrides)
+
+    if model_weights_path:
+      common_modules.load_weights(
+          model, model_weights_path, weights_format=weights_format)
+
+    return model

official/legacy/image_classification/efficientnet/tfhub_export.py

+# 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.
+
+"""A script to export TF-Hub SavedModel."""
+
+from __future__ import absolute_import
+from __future__ import division
+# from __future__ import google_type_annotations
+from __future__ import print_function
+
+import os
+
+from absl import app
+from absl import flags
+
+import tensorflow as tf
+
+from official.legacy.image_classification.efficientnet import efficientnet_model
+
+FLAGS = flags.FLAGS
+
+flags.DEFINE_string("model_name", None, "EfficientNet model name.")
+flags.DEFINE_string("model_path", None, "File path to TF model checkpoint.")
+flags.DEFINE_string("export_path", None,
+                    "TF-Hub SavedModel destination path to export.")
+
+
+def export_tfhub(model_path, hub_destination, model_name):
+  """Restores a tf.keras.Model and saves for TF-Hub."""
+  model_configs = dict(efficientnet_model.MODEL_CONFIGS)
+  config = model_configs[model_name]
+
+  image_input = tf.keras.layers.Input(
+      shape=(None, None, 3), name="image_input", dtype=tf.float32)
+  x = image_input * 255.0
+  ouputs = efficientnet_model.efficientnet(x, config)
+  hub_model = tf.keras.Model(image_input, ouputs)
+  ckpt = tf.train.Checkpoint(model=hub_model)
+  ckpt.restore(model_path).assert_existing_objects_matched()
+  hub_model.save(
+      os.path.join(hub_destination, "classification"), include_optimizer=False)
+
+  feature_vector_output = hub_model.get_layer(name="top_pool").get_output_at(0)
+  hub_model2 = tf.keras.Model(image_input, feature_vector_output)
+  hub_model2.save(
+      os.path.join(hub_destination, "feature-vector"), include_optimizer=False)
+
+
+def main(argv):
+  if len(argv) > 1:
+    raise app.UsageError("Too many command-line arguments.")
+
+  export_tfhub(FLAGS.model_path, FLAGS.export_path, FLAGS.model_name)
+
+
+if __name__ == "__main__":
+  app.run(main)

official/legacy/image_classification/learning_rate.py

+# 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
+"""Learning rate utilities for vision tasks."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from typing import Any, Mapping, Optional
+
+import numpy as np
+import tensorflow as tf
+
+BASE_LEARNING_RATE = 0.1
+
+
+class WarmupDecaySchedule(tf.keras.optimizers.schedules.LearningRateSchedule):
+  """A wrapper for LearningRateSchedule that includes warmup steps."""
+
+  def __init__(self,
+               lr_schedule: tf.keras.optimizers.schedules.LearningRateSchedule,
+               warmup_steps: int,
+               warmup_lr: Optional[float] = None):
+    """Add warmup decay to a learning rate schedule.
+
+    Args:
+      lr_schedule: base learning rate scheduler
+      warmup_steps: number of warmup steps
+      warmup_lr: an optional field for the final warmup learning rate. This
+        should be provided if the base `lr_schedule` does not contain this
+        field.
+    """
+    super(WarmupDecaySchedule, self).__init__()
+    self._lr_schedule = lr_schedule
+    self._warmup_steps = warmup_steps
+    self._warmup_lr = warmup_lr
+
+  def __call__(self, step: int):
+    lr = self._lr_schedule(step)
+    if self._warmup_steps:
+      if self._warmup_lr is not None:
+        initial_learning_rate = tf.convert_to_tensor(
+            self._warmup_lr, name="initial_learning_rate")
+      else:
+        initial_learning_rate = tf.convert_to_tensor(
+            self._lr_schedule.initial_learning_rate,
+            name="initial_learning_rate")
+      dtype = initial_learning_rate.dtype
+      global_step_recomp = tf.cast(step, dtype)
+      warmup_steps = tf.cast(self._warmup_steps, dtype)
+      warmup_lr = initial_learning_rate * global_step_recomp / warmup_steps
+      lr = tf.cond(global_step_recomp < warmup_steps, lambda: warmup_lr,
+                   lambda: lr)
+    return lr
+
+  def get_config(self) -> Mapping[str, Any]:
+    config = self._lr_schedule.get_config()
+    config.update({
+        "warmup_steps": self._warmup_steps,
+        "warmup_lr": self._warmup_lr,
+    })
+    return config
+
+
+class CosineDecayWithWarmup(tf.keras.optimizers.schedules.LearningRateSchedule):
+  """Class to generate learning rate tensor."""
+
+  def __init__(self, batch_size: int, total_steps: int, warmup_steps: int):
+    """Creates the consine learning rate tensor with linear warmup.
+
+    Args:
+      batch_size: The training batch size used in the experiment.
+      total_steps: Total training steps.
+      warmup_steps: Steps for the warm up period.
+    """
+    super(CosineDecayWithWarmup, self).__init__()
+    base_lr_batch_size = 256
+    self._total_steps = total_steps
+    self._init_learning_rate = BASE_LEARNING_RATE * batch_size / base_lr_batch_size
+    self._warmup_steps = warmup_steps
+
+  def __call__(self, global_step: int):
+    global_step = tf.cast(global_step, dtype=tf.float32)
+    warmup_steps = self._warmup_steps
+    init_lr = self._init_learning_rate
+    total_steps = self._total_steps
+
+    linear_warmup = global_step / warmup_steps * init_lr
+
+    cosine_learning_rate = init_lr * (tf.cos(np.pi *
+                                             (global_step - warmup_steps) /
+                                             (total_steps - warmup_steps)) +
+                                      1.0) / 2.0
+
+    learning_rate = tf.where(global_step < warmup_steps, linear_warmup,
+                             cosine_learning_rate)
+    return learning_rate
+
+  def get_config(self):
+    return {
+        "total_steps": self._total_steps,
+        "warmup_learning_rate": self._warmup_learning_rate,
+        "warmup_steps": self._warmup_steps,
+        "init_learning_rate": self._init_learning_rate,
+    }

official/legacy/image_classification/learning_rate_test.py

+# 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.
+
+"""Tests for learning_rate."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import tensorflow as tf
+
+from official.legacy.image_classification import learning_rate
+
+
+class LearningRateTests(tf.test.TestCase):
+
+  def test_warmup_decay(self):
+    """Basic computational test for warmup decay."""
+    initial_lr = 0.01
+    decay_steps = 100
+    decay_rate = 0.01
+    warmup_steps = 10
+
+    base_lr = tf.keras.optimizers.schedules.ExponentialDecay(
+        initial_learning_rate=initial_lr,
+        decay_steps=decay_steps,
+        decay_rate=decay_rate)
+    lr = learning_rate.WarmupDecaySchedule(
+        lr_schedule=base_lr, warmup_steps=warmup_steps)
+
+    for step in range(warmup_steps - 1):
+      config = lr.get_config()
+      self.assertEqual(config['warmup_steps'], warmup_steps)
+      self.assertAllClose(
+          self.evaluate(lr(step)), step / warmup_steps * initial_lr)
+
+  def test_cosine_decay_with_warmup(self):
+    """Basic computational test for cosine decay with warmup."""
+    expected_lrs = [0.0, 0.1, 0.05, 0.0]
+
+    lr = learning_rate.CosineDecayWithWarmup(
+        batch_size=256, total_steps=3, warmup_steps=1)
+
+    for step in [0, 1, 2, 3]:
+      self.assertAllClose(lr(step), expected_lrs[step])
+
+
+if __name__ == '__main__':
+  tf.test.main()

official/legacy/image_classification/mnist_main.py

+# 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.
+
+"""Runs a simple model on the MNIST dataset."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+
+# Import libraries
+from absl import app
+from absl import flags
+from absl import logging
+import tensorflow as tf
+import tensorflow_datasets as tfds
+from official.common import distribute_utils
+from official.legacy.image_classification.resnet import common
+from official.utils.flags import core as flags_core
+from official.utils.misc import model_helpers
+
+FLAGS = flags.FLAGS
+
+
+def build_model():
+  """Constructs the ML model used to predict handwritten digits."""
+
+  image = tf.keras.layers.Input(shape=(28, 28, 1))
+
+  y = tf.keras.layers.Conv2D(filters=32,
+                             kernel_size=5,
+                             padding='same',
+                             activation='relu')(image)
+  y = tf.keras.layers.MaxPooling2D(pool_size=(2, 2),
+                                   strides=(2, 2),
+                                   padding='same')(y)
+  y = tf.keras.layers.Conv2D(filters=32,
+                             kernel_size=5,
+                             padding='same',
+                             activation='relu')(y)
+  y = tf.keras.layers.MaxPooling2D(pool_size=(2, 2),
+                                   strides=(2, 2),
+                                   padding='same')(y)
+  y = tf.keras.layers.Flatten()(y)
+  y = tf.keras.layers.Dense(1024, activation='relu')(y)
+  y = tf.keras.layers.Dropout(0.4)(y)
+
+  probs = tf.keras.layers.Dense(10, activation='softmax')(y)
+
+  model = tf.keras.models.Model(image, probs, name='mnist')
+
+  return model
+
+
+@tfds.decode.make_decoder(output_dtype=tf.float32)
+def decode_image(example, feature):
+  """Convert image to float32 and normalize from [0, 255] to [0.0, 1.0]."""
+  return tf.cast(feature.decode_example(example), dtype=tf.float32) / 255
+
+
+def run(flags_obj, datasets_override=None, strategy_override=None):
+  """Run MNIST model training and eval loop using native Keras APIs.
+
+  Args:
+    flags_obj: An object containing parsed flag values.
+    datasets_override: A pair of `tf.data.Dataset` objects to train the model,
+                       representing the train and test sets.
+    strategy_override: A `tf.distribute.Strategy` object to use for model.
+
+  Returns:
+    Dictionary of training and eval stats.
+  """
+  # Start TF profiler server.
+  tf.profiler.experimental.server.start(flags_obj.profiler_port)
+
+  strategy = strategy_override or distribute_utils.get_distribution_strategy(
+      distribution_strategy=flags_obj.distribution_strategy,
+      num_gpus=flags_obj.num_gpus,
+      tpu_address=flags_obj.tpu)
+
+  strategy_scope = distribute_utils.get_strategy_scope(strategy)
+
+  mnist = tfds.builder('mnist', data_dir=flags_obj.data_dir)
+  if flags_obj.download:
+    mnist.download_and_prepare()
+
+  mnist_train, mnist_test = datasets_override or mnist.as_dataset(
+      split=['train', 'test'],
+      decoders={'image': decode_image()},  # pylint: disable=no-value-for-parameter
+      as_supervised=True)
+  train_input_dataset = mnist_train.cache().repeat().shuffle(
+      buffer_size=50000).batch(flags_obj.batch_size)
+  eval_input_dataset = mnist_test.cache().repeat().batch(flags_obj.batch_size)
+
+  with strategy_scope:
+    lr_schedule = tf.keras.optimizers.schedules.ExponentialDecay(
+        0.05, decay_steps=100000, decay_rate=0.96)
+    optimizer = tf.keras.optimizers.SGD(learning_rate=lr_schedule)
+
+    model = build_model()
+    model.compile(
+        optimizer=optimizer,
+        loss='sparse_categorical_crossentropy',
+        metrics=['sparse_categorical_accuracy'])
+
+  num_train_examples = mnist.info.splits['train'].num_examples
+  train_steps = num_train_examples // flags_obj.batch_size
+  train_epochs = flags_obj.train_epochs
+
+  ckpt_full_path = os.path.join(flags_obj.model_dir, 'model.ckpt-{epoch:04d}')
+  callbacks = [
+      tf.keras.callbacks.ModelCheckpoint(
+          ckpt_full_path, save_weights_only=True),
+      tf.keras.callbacks.TensorBoard(log_dir=flags_obj.model_dir),
+  ]
+
+  num_eval_examples = mnist.info.splits['test'].num_examples
+  num_eval_steps = num_eval_examples // flags_obj.batch_size
+
+  history = model.fit(
+      train_input_dataset,
+      epochs=train_epochs,
+      steps_per_epoch=train_steps,
+      callbacks=callbacks,
+      validation_steps=num_eval_steps,
+      validation_data=eval_input_dataset,
+      validation_freq=flags_obj.epochs_between_evals)
+
+  export_path = os.path.join(flags_obj.model_dir, 'saved_model')
+  model.save(export_path, include_optimizer=False)
+
+  eval_output = model.evaluate(
+      eval_input_dataset, steps=num_eval_steps, verbose=2)
+
+  stats = common.build_stats(history, eval_output, callbacks)
+  return stats
+
+
+def define_mnist_flags():
+  """Define command line flags for MNIST model."""
+  flags_core.define_base(
+      clean=True,
+      num_gpu=True,
+      train_epochs=True,
+      epochs_between_evals=True,
+      distribution_strategy=True)
+  flags_core.define_device()
+  flags_core.define_distribution()
+  flags.DEFINE_bool('download', True,
+                    'Whether to download data to `--data_dir`.')
+  flags.DEFINE_integer('profiler_port', 9012,
+                       'Port to start profiler server on.')
+  FLAGS.set_default('batch_size', 1024)
+
+
+def main(_):
+  model_helpers.apply_clean(FLAGS)
+  stats = run(flags.FLAGS)
+  logging.info('Run stats:\n%s', stats)
+
+
+if __name__ == '__main__':
+  logging.set_verbosity(logging.INFO)
+  define_mnist_flags()
+  app.run(main)

official/legacy/image_classification/mnist_test.py

+# 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.
+
+"""Test the Keras MNIST model on GPU."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import functools
+
+from absl.testing import parameterized
+import tensorflow as tf
+
+from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import strategy_combinations
+from official.legacy.image_classification import mnist_main
+from official.utils.testing import integration
+
+
+mnist_main.define_mnist_flags()
+
+
+def eager_strategy_combinations():
+  return combinations.combine(
+      distribution=[
+          strategy_combinations.default_strategy,
+          strategy_combinations.cloud_tpu_strategy,
+          strategy_combinations.one_device_strategy_gpu,
+      ],)
+
+
+class KerasMnistTest(tf.test.TestCase, parameterized.TestCase):
+  """Unit tests for sample Keras MNIST model."""
+  _tempdir = None
+
+  @classmethod
+  def setUpClass(cls):  # pylint: disable=invalid-name
+    super(KerasMnistTest, cls).setUpClass()
+
+  def tearDown(self):
+    super(KerasMnistTest, self).tearDown()
+    tf.io.gfile.rmtree(self.get_temp_dir())
+
+  @combinations.generate(eager_strategy_combinations())
+  def test_end_to_end(self, distribution):
+    """Test Keras MNIST model with `strategy`."""
+
+    extra_flags = [
+        "-train_epochs",
+        "1",
+        # Let TFDS find the metadata folder automatically
+        "--data_dir="
+    ]
+
+    dummy_data = (
+        tf.ones(shape=(10, 28, 28, 1), dtype=tf.int32),
+        tf.range(10),
+    )
+    datasets = (
+        tf.data.Dataset.from_tensor_slices(dummy_data),
+        tf.data.Dataset.from_tensor_slices(dummy_data),
+    )
+
+    run = functools.partial(
+        mnist_main.run,
+        datasets_override=datasets,
+        strategy_override=distribution)
+
+    integration.run_synthetic(
+        main=run,
+        synth=False,
+        tmp_root=self.create_tempdir().full_path,
+        extra_flags=extra_flags)
+
+
+if __name__ == "__main__":
+  tf.test.main()

official/legacy/image_classification/optimizer_factory_test.py

+# 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.
+
+"""Tests for optimizer_factory."""
+
+from __future__ import absolute_import
+from __future__ import division
+# from __future__ import google_type_annotations
+from __future__ import print_function
+
+from absl.testing import parameterized
+
+import tensorflow as tf
+from official.legacy.image_classification import optimizer_factory
+from official.legacy.image_classification.configs import base_configs
+
+
+class OptimizerFactoryTest(tf.test.TestCase, parameterized.TestCase):
+
+  def build_toy_model(self) -> tf.keras.Model:
+    """Creates a toy `tf.Keras.Model`."""
+    model = tf.keras.Sequential()
+    model.add(tf.keras.layers.Dense(1, input_shape=(1,)))
+    return model
+
+  @parameterized.named_parameters(
+      ('sgd', 'sgd', 0., False), ('momentum', 'momentum', 0., False),
+      ('rmsprop', 'rmsprop', 0., False), ('adam', 'adam', 0., False),
+      ('adamw', 'adamw', 0., False),
+      ('momentum_lookahead', 'momentum', 0., True),
+      ('sgd_ema', 'sgd', 0.999, False),
+      ('momentum_ema', 'momentum', 0.999, False),
+      ('rmsprop_ema', 'rmsprop', 0.999, False))
+  def test_optimizer(self, optimizer_name, moving_average_decay, lookahead):
+    """Smoke test to be sure no syntax errors."""
+    model = self.build_toy_model()
+    params = {
+        'learning_rate': 0.001,
+        'rho': 0.09,
+        'momentum': 0.,
+        'epsilon': 1e-07,
+        'moving_average_decay': moving_average_decay,
+        'lookahead': lookahead,
+    }
+    optimizer = optimizer_factory.build_optimizer(
+        optimizer_name=optimizer_name,
+        base_learning_rate=params['learning_rate'],
+        params=params,
+        model=model)
+    self.assertTrue(issubclass(type(optimizer), tf.keras.optimizers.Optimizer))
+
+  def test_unknown_optimizer(self):
+    with self.assertRaises(ValueError):
+      optimizer_factory.build_optimizer(
+          optimizer_name='this_optimizer_does_not_exist',
+          base_learning_rate=None,
+          params=None)
+
+  def test_learning_rate_without_decay_or_warmups(self):
+    params = base_configs.LearningRateConfig(
+        name='exponential',
+        initial_lr=0.01,
+        decay_rate=0.01,
+        decay_epochs=None,
+        warmup_epochs=None,
+        scale_by_batch_size=0.01,
+        examples_per_epoch=1,
+        boundaries=[0],
+        multipliers=[0, 1])
+    batch_size = 1
+    train_steps = 1
+
+    lr = optimizer_factory.build_learning_rate(
+        params=params, batch_size=batch_size, train_steps=train_steps)
+    self.assertTrue(
+        issubclass(
+            type(lr), tf.keras.optimizers.schedules.LearningRateSchedule))
+
+  @parameterized.named_parameters(('exponential', 'exponential'),
+                                  ('cosine_with_warmup', 'cosine_with_warmup'))
+  def test_learning_rate_with_decay_and_warmup(self, lr_decay_type):
+    """Basic smoke test for syntax."""
+    params = base_configs.LearningRateConfig(
+        name=lr_decay_type,
+        initial_lr=0.01,
+        decay_rate=0.01,
+        decay_epochs=1,
+        warmup_epochs=1,
+        scale_by_batch_size=0.01,
+        examples_per_epoch=1,
+        boundaries=[0],
+        multipliers=[0, 1])
+    batch_size = 1
+    train_epochs = 1
+    train_steps = 1
+
+    lr = optimizer_factory.build_learning_rate(
+        params=params,
+        batch_size=batch_size,
+        train_epochs=train_epochs,
+        train_steps=train_steps)
+    self.assertTrue(
+        issubclass(
+            type(lr), tf.keras.optimizers.schedules.LearningRateSchedule))
+
+
+if __name__ == '__main__':
+  tf.test.main()