Internal change

PiperOrigin-RevId: 302937425

official/vision/image_classification/README.md

 # Image Classification
 
-This folder contains the TF 2.0 model examples for image classification:
+This folder contains TF 2.0 model examples for image classification:
 
-* [ResNet](#resnet)
 * [MNIST](#mnist)
+* [Classifier Trainer](#classifier-trainer), a framework that uses the Keras
+compile/fit methods for image classification models, including:
+  * ResNet
+  * EfficientNet[^1]
 
+[^1]: Currently a work in progress. We cannot match "AutoAugment (AA)" in [the original version](https://github.com/tensorflow/tpu/tree/master/models/official/efficientnet).
 For more information about other types of models, please refer to this
 [README file](../../README.md).
 
-## ResNet
-
-Similar to the [estimator implementation](../../r1/resnet), the Keras
-implementation has code for the ImageNet dataset. The ImageNet
-version uses a ResNet50 model implemented in
-[`resnet_model.py`](./resnet/resnet_model.py).
-
+## Before you begin
 Please make sure that you have the latest version of TensorFlow
 installed and
 [add the models folder to your Python path](/official/#running-the-models).
 
-### Pretrained Models
-
-* [ResNet50 Checkpoints](https://storage.googleapis.com/cloud-tpu-checkpoints/resnet/resnet50.tar.gz)
-
-* ResNet50 TFHub: [feature vector](https://tfhub.dev/tensorflow/resnet_50/feature_vector/1)
-and [classification](https://tfhub.dev/tensorflow/resnet_50/classification/1)
-
-### ImageNet Training
+### ImageNet preparation
 
 Download the ImageNet dataset and convert it to TFRecord format.
 The following [script](https://github.com/tensorflow/tpu/blob/master/tools/datasets/imagenet_to_gcs.py)
 and [README](https://github.com/tensorflow/tpu/tree/master/tools/datasets#imagenet_to_gcspy)
 provide a few options.
 
-Once your dataset is ready, you can begin training the model as follows:
-
-```bash
-python resnet/resnet_imagenet_main.py
-```
-
-Again, if you did not download the data to the default directory, specify the
-location with the `--data_dir` flag:
-
-```bash
-python resnet/resnet_imagenet_main.py --data_dir=/path/to/imagenet
-```
-
-There are more flag options you can specify. Here are some examples:
-
-- `--use_synthetic_data`: when set to true, synthetic data, rather than real
-data, are used;
-- `--batch_size`: the batch size used for the model;
-- `--model_dir`: the directory to save the model checkpoint;
-- `--train_epochs`: number of epoches to run for training the model;
-- `--train_steps`: number of steps to run for training the model. We now only
-support a number that is smaller than the number of batches in an epoch.
-- `--skip_eval`: when set to true, evaluation as well as validation during
-training is skipped
-
-For example, this is a typical command line to run with ImageNet data with
-batch size 128 per GPU:
-
-```bash
-python -m resnet/resnet_imagenet_main.py \
-    --model_dir=/tmp/model_dir/something \
-    --num_gpus=2 \
-    --batch_size=128 \
-    --train_epochs=90 \
-    --train_steps=10 \
-    --use_synthetic_data=false
-```
-
-See [`common.py`](common.py) for full list of options.
-
-### Using multiple GPUs
-
-You can train these models on multiple GPUs using `tf.distribute.Strategy` API.
-You can read more about them in this
-[guide](https://www.tensorflow.org/guide/distribute_strategy).
-
-In this example, we have made it easier to use is with just a command line flag
-`--num_gpus`. By default this flag is 1 if TensorFlow is compiled with CUDA,
-and 0 otherwise.
-
-- --num_gpus=0: Uses tf.distribute.OneDeviceStrategy with CPU as the device.
-- --num_gpus=1: Uses tf.distribute.OneDeviceStrategy with GPU as the device.
-- --num_gpus=2+: Uses tf.distribute.MirroredStrategy to run synchronous
-distributed training across the GPUs.
-
-If you wish to run without `tf.distribute.Strategy`, you can do so by setting
-`--distribution_strategy=off`.
-
-### Running on multiple GPU hosts
-
-You can also train these models on multiple hosts, each with GPUs, using
-`tf.distribute.Strategy`.
-
-The easiest way to run multi-host benchmarks is to set the
-[`TF_CONFIG`](https://www.tensorflow.org/guide/distributed_training#TF_CONFIG)
-appropriately at each host.  e.g., to run using `MultiWorkerMirroredStrategy` on
-2 hosts, the `cluster` in `TF_CONFIG` should have 2 `host:port` entries, and
-host `i` should have the `task` in `TF_CONFIG` set to `{"type": "worker",
-"index": i}`.  `MultiWorkerMirroredStrategy` will automatically use all the
-available GPUs at each host.
-
 ### Running on Cloud TPUs
 
-Note: This model will **not** work with TPUs on Colab.
+Note: These models will **not** work with TPUs on Colab.
 
-You can train the ResNet CTL model on Cloud TPUs using
+You can train image classification models on Cloud TPUs using
 `tf.distribute.TPUStrategy`. If you are not familiar with Cloud TPUs, it is
 strongly recommended that you go through the
 [quickstart](https://cloud.google.com/tpu/docs/quickstart) to learn how to
 create a TPU and GCE VM.
 
-To run ResNet model on a TPU, you must set `--distribution_strategy=tpu` and
-`--tpu=$TPU_NAME`, where `$TPU_NAME` the name of your TPU in the Cloud Console.
-From a GCE VM, you can run the following command to train ResNet for one epoch
-on a v2-8 or v3-8 TPU:
+## MNIST
+
+To download the data and run the MNIST sample model locally for the first time,
+run one of the following command:
 
 ```bash
-python resnet/resnet_ctl_imagenet_main.py \
-  --tpu=$TPU_NAME \
+python3 mnist_main.py \
   --model_dir=$MODEL_DIR \
   --data_dir=$DATA_DIR \
-  --batch_size=1024 \
-  --steps_per_loop=500 \
-  --train_epochs=1 \
-  --use_synthetic_data=false \
-  --dtype=fp32 \
-  --enable_eager=true \
-  --enable_tensorboard=true \
-  --distribution_strategy=tpu \
-  --log_steps=50 \
-  --single_l2_loss_op=true \
-  --use_tf_function=true
+  --train_epochs=10 \
+  --distribution_strategy=one_device \
+  --num_gpus=$NUM_GPUS \
+  --download
 ```
 
-To train the ResNet to convergence, run it for 90 epochs:
+To train the model on a Cloud TPU, run the following command:
 
 ```bash
-python resnet/resnet_ctl_imagenet_main.py \
+python3 mnist_main.py \
   --tpu=$TPU_NAME \
   --model_dir=$MODEL_DIR \
   --data_dir=$DATA_DIR \
-  --batch_size=1024 \
-  --steps_per_loop=500 \
-  --train_epochs=90 \
-  --use_synthetic_data=false \
-  --dtype=fp32 \
-  --enable_eager=true \
-  --enable_tensorboard=true \
+  --train_epochs=10 \
   --distribution_strategy=tpu \
-  --log_steps=50 \
-  --single_l2_loss_op=true \
-  --use_tf_function=true
+  --download
 ```
 
-Note: `$MODEL_DIR` and `$DATA_DIR` must be GCS paths.
+Note: the `--download` flag is only required the first time you run the model.
 
 
-## MNIST
+## Classifier Trainer
+The classifier trainer is a unified framework for running image classification
+models using Keras's compile/fit methods. Experiments should be provided in the
+form of YAML files, some examples are included within the configs/examples
+folder. Please see [configs/examples](./configs/examples) for more example
+configurations.
 
-To download the data and run the MNIST sample model locally for the first time,
-run one of the following command:
+The provided configuration files use a per replica batch size and is scaled
+by the number of devices. For instance, if `batch size` = 64, then for 1 GPU
+the global batch size would be 64 * 1 = 64. For 8 GPUs, the global batch size
+would be 64 * 8 = 512. Similarly, for a v3-8 TPU, the global batch size would
+be 64 * 8 = 512, and for a v3-32, the global batch size is 64 * 32 = 2048.
 
+### ResNet50
+
+#### On GPU:
 ```bash
-python mnist_main.py \
+python3 classifier_trainer.py \
+  --mode=train_and_eval \
+  --model_type=resnet \
+  --dataset=imagenet \
   --model_dir=$MODEL_DIR \
   --data_dir=$DATA_DIR \
-  --train_epochs=10 \
-  --distribution_strategy=one_device \
-  --num_gpus=$NUM_GPUS \
-  --download
+  --config_file=configs/examples/resnet/imagenet/gpu.yaml \
+  --params_override='runtime.num_gpus=$NUM_GPUS'
 ```
 
-To train the model on a Cloud TPU, run the following command:
+#### On TPU:
+```bash
+python3 classifier_trainer.py \
+  --mode=train_and_eval \
+  --model_type=resnet \
+  --dataset=imagenet \
+  --tpu=$TPU_NAME \
+  --model_dir=$MODEL_DIR \
+  --data_dir=$DATA_DIR \
+  --config_file=config/examples/resnet/imagenet/tpu.yaml
+```
+
+### EfficientNet
+**Note: EfficientNet development is a work in progress.**
+#### On GPU:
+```bash
+python3 classifier_trainer.py \
+  --mode=train_and_eval \
+  --model_type=efficientnet \
+  --dataset=imagenet \
+  --model_dir=$MODEL_DIR \
+  --data_dir=$DATA_DIR \
+  --config_file=configs/examples/efficientnet/imagenet/efficientnet-b0-gpu.yaml \
+  --params_override='runtime.num_gpus=$NUM_GPUS'
+```
 
+
+#### On TPU:
 ```bash
-python mnist_main.py \
+python3 classifier_trainer.py \
+  --mode=train_and_eval \
+  --model_type=efficientnet \
+  --dataset=imagenet \
   --tpu=$TPU_NAME \
   --model_dir=$MODEL_DIR \
   --data_dir=$DATA_DIR \
-  --train_epochs=10 \
-  --distribution_strategy=tpu \
-  --download
+  --config_file=config/examples/efficientnet/imagenet/efficientnet-b0-tpu.yaml
 ```
 
-Note: the `--download` flag is only required the first time you run the model.
+Note that the number of GPU devices can be overridden in the command line using
+`--params_overrides`. The TPU does not need this override as the device is fixed
+by providing the TPU address or name with the `--tpu` flag.
+

official/vision/image_classification/augment_test.py

+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for autoaugment."""
+
+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.compat.v2 as tf
+
+from official.vision.image_classification import augment
+
+
+def get_dtype_test_cases():
+  return [
+      ('uint8', tf.uint8),
+      ('int32', tf.int32),
+      ('float16', tf.float16),
+      ('float32', tf.float32),
+  ]
+
+
+@parameterized.named_parameters(get_dtype_test_cases())
+class TransformsTest(parameterized.TestCase, tf.test.TestCase):
+  """Basic tests for fundamental transformations."""
+
+  def test_to_from_4d(self, dtype):
+    for shape in [(10, 10), (10, 10, 10), (10, 10, 10, 10)]:
+      original_ndims = len(shape)
+      image = tf.zeros(shape, dtype=dtype)
+      image_4d = augment.to_4d(image)
+      self.assertEqual(4, tf.rank(image_4d))
+      self.assertAllEqual(image, augment.from_4d(image_4d, original_ndims))
+
+  def test_transform(self, dtype):
+    image = tf.constant([[1, 2], [3, 4]], dtype=dtype)
+    self.assertAllEqual(augment.transform(image, transforms=[1]*8),
+                        [[4, 4], [4, 4]])
+
+  def test_translate(self, dtype):
+    image = tf.constant(
+        [[1, 0, 1, 0], [0, 1, 0, 1], [1, 0, 1, 0], [0, 1, 0, 1]],
+        dtype=dtype)
+    translations = [-1, -1]
+    translated = augment.translate(image=image,
+                                   translations=translations)
+    expected = [[1, 0, 1, 0], [0, 1, 0, 0], [1, 0, 1, 0], [0, 0, 0, 0]]
+    self.assertAllEqual(translated, expected)
+
+  def test_translate_shapes(self, dtype):
+    translation = [0, 0]
+    for shape in [(3, 3), (5, 5), (224, 224, 3)]:
+      image = tf.zeros(shape, dtype=dtype)
+      self.assertAllEqual(image, augment.translate(image, translation))
+
+  def test_translate_invalid_translation(self, dtype):
+    image = tf.zeros((1, 1), dtype=dtype)
+    invalid_translation = [[[1, 1]]]
+    with self.assertRaisesRegex(TypeError, 'rank 1 or 2'):
+      _ = augment.translate(image, invalid_translation)
+
+  def test_rotate(self, dtype):
+    image = tf.reshape(tf.cast(tf.range(9), dtype), (3, 3))
+    rotation = 90.
+    transformed = augment.rotate(image=image, degrees=rotation)
+    expected = [[2, 5, 8],
+                [1, 4, 7],
+                [0, 3, 6]]
+    self.assertAllEqual(transformed, expected)
+
+  def test_rotate_shapes(self, dtype):
+    degrees = 0.
+    for shape in [(3, 3), (5, 5), (224, 224, 3)]:
+      image = tf.zeros(shape, dtype=dtype)
+      self.assertAllEqual(image, augment.rotate(image, degrees))
+
+
+class AutoaugmentTest(tf.test.TestCase):
+
+  def test_autoaugment(self):
+    """Smoke test to be sure there are no syntax errors."""
+    image = tf.zeros((224, 224, 3), dtype=tf.uint8)
+
+    augmenter = augment.AutoAugment()
+    aug_image = augmenter.distort(image)
+
+    self.assertEqual((224, 224, 3), aug_image.shape)
+
+  def test_randaug(self):
+    """Smoke test to be sure there are no syntax errors."""
+    image = tf.zeros((224, 224, 3), dtype=tf.uint8)
+
+    augmenter = augment.RandAugment()
+    aug_image = augmenter.distort(image)
+
+    self.assertEqual((224, 224, 3), aug_image.shape)
+
+  def test_all_policy_ops(self):
+    """Smoke test to be sure all augmentation functions can execute."""
+
+    prob = 1
+    magnitude = 10
+    replace_value = [128] * 3
+    cutout_const = 100
+    translate_const = 250
+
+    image = tf.ones((224, 224, 3), dtype=tf.uint8)
+
+    for op_name in augment.NAME_TO_FUNC:
+      func, _, args = augment._parse_policy_info(op_name,
+                                                 prob,
+                                                 magnitude,
+                                                 replace_value,
+                                                 cutout_const,
+                                                 translate_const)
+      image = func(image, *args)
+
+    self.assertEqual((224, 224, 3), image.shape)
+
+if __name__ == '__main__':
+  assert tf.version.VERSION.startswith('2.')
+  tf.test.main()

official/vision/image_classification/callbacks.py

+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Common modules for callbacks."""
+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 logging
+
+import tensorflow as tf
+from typing import Any, List, MutableMapping, Text
+
+
+def get_callbacks(model_checkpoint: bool = True,
+                  include_tensorboard: bool = True,
+                  track_lr: bool = True,
+                  write_model_weights: bool = True,
+                  initial_step: int = 0,
+                  model_dir: Text = None) -> List[tf.keras.callbacks.Callback]:
+  """Get all callbacks."""
+  model_dir = model_dir or ''
+  callbacks = []
+  if model_checkpoint:
+    ckpt_full_path = os.path.join(model_dir, 'model.ckpt-{epoch:04d}')
+    callbacks.append(tf.keras.callbacks.ModelCheckpoint(
+        ckpt_full_path, save_weights_only=True, verbose=1))
+  if include_tensorboard:
+    callbacks.append(CustomTensorBoard(
+        log_dir=model_dir,
+        track_lr=track_lr,
+        initial_step=initial_step,
+        write_images=write_model_weights))
+  return callbacks
+
+
+def get_scalar_from_tensor(t: tf.Tensor) -> int:
+  """Utility function to convert a Tensor to a scalar."""
+  t = tf.keras.backend.get_value(t)
+  if callable(t):
+    return t()
+  else:
+    return t
+
+
+class CustomTensorBoard(tf.keras.callbacks.TensorBoard):
+  """A customized TensorBoard callback that tracks additional datapoints.
+
+  Metrics tracked:
+  - Global learning rate
+
+  Attributes:
+    log_dir: the path of the directory where to save the log files to be
+      parsed by TensorBoard.
+    track_lr: `bool`, whether or not to track the global learning rate.
+    initial_step: the initial step, used for preemption recovery.
+    **kwargs: Additional arguments for backwards compatibility. Possible key
+      is `period`.
+  """
+  # TODO(b/146499062): track params, flops, log lr, l2 loss,
+  # classification loss
+
+  def __init__(self,
+               log_dir: Text,
+               track_lr: bool = False,
+               initial_step: int = 0,
+               **kwargs):
+    super(CustomTensorBoard, self).__init__(log_dir=log_dir, **kwargs)
+    self.step = initial_step
+    self._track_lr = track_lr
+
+  def on_batch_begin(self,
+                     epoch: int,
+                     logs: MutableMapping[Text, Any] = None) -> None:
+    self.step += 1
+    if logs is None:
+      logs = {}
+    logs.update(self._calculate_metrics())
+    super(CustomTensorBoard, self).on_batch_begin(epoch, logs)
+
+  def on_epoch_begin(self,
+                     epoch: int,
+                     logs: MutableMapping[Text, Any] = None) -> None:
+    if logs is None:
+      logs = {}
+    metrics = self._calculate_metrics()
+    logs.update(metrics)
+    for k, v in metrics.items():
+      logging.info('Current %s: %f', k, v)
+    super(CustomTensorBoard, self).on_epoch_begin(epoch, logs)
+
+  def on_epoch_end(self,
+                   epoch: int,
+                   logs: MutableMapping[Text, Any] = None) -> None:
+    if logs is None:
+      logs = {}
+    metrics = self._calculate_metrics()
+    logs.update(metrics)
+    super(CustomTensorBoard, self).on_epoch_end(epoch, logs)
+
+  def _calculate_metrics(self) -> MutableMapping[Text, Any]:
+    logs = {}
+    if self._track_lr:
+      logs['learning_rate'] = self._calculate_lr()
+    return logs
+
+  def _calculate_lr(self) -> int:
+    """Calculates the learning rate given the current step."""
+    lr = self._get_base_optimizer().lr
+    if callable(lr):
+      lr = lr(self.step)
+    return get_scalar_from_tensor(lr)
+
+  def _get_base_optimizer(self) -> tf.keras.optimizers.Optimizer:
+    """Get the base optimizer used by the current model."""
+
+    optimizer = self.model.optimizer
+
+    # The optimizer might be wrapped by another class, so unwrap it
+    while hasattr(optimizer, '_optimizer'):
+      optimizer = optimizer._optimizer  # pylint:disable=protected-access
+
+    return optimizer

official/vision/image_classification/callbacks_test.py

+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Tests for callbacks."""
+
+from __future__ import absolute_import
+from __future__ import division
+# from __future__ import google_type_annotations
+from __future__ import print_function
+
+import collections
+import functools
+import os
+
+from absl.testing import parameterized
+
+import numpy as np
+import tensorflow as tf
+
+from tensorflow.python.keras import callbacks_test
+from tensorflow.python.keras import keras_parameterized
+from official.vision.image_classification import callbacks
+
+_ObservedSummary = collections.namedtuple('_ObservedSummary', ('logdir', 'tag'))
+
+
+def _trivial_function(a):
+  return a
+
+
+class UtilFunctionTests(tf.test.TestCase, parameterized.TestCase):
+  """Tests to check utility functions provided in callbacks.py."""
+
+  @parameterized.named_parameters(
+      ('integer', 1),
+      ('float', 1.),
+      ('lambda', lambda: 1),
+      ('partial', functools.partial(_trivial_function, 1)))
+  def test_scalar_from_tensors(self, t):
+    t = tf.Variable(t)
+    value = callbacks.get_scalar_from_tensor(t)
+    print (value)
+    self.assertTrue(np.isscalar(value))
+
+
+@keras_parameterized.run_with_all_model_types
+@keras_parameterized.run_all_keras_modes(always_skip_v1=True)
+class CustomTensorBoardTest(callbacks_test.TestTensorBoardV2):
+
+  def test_custom_tb_learning_rate(self):
+    os.chdir(self.get_temp_dir())
+    model = self._get_model()
+    x, y = np.ones((10, 10, 10, 1)), np.ones((10, 1))
+    tb_cbk = callbacks.CustomTensorBoard(log_dir=self.logdir,
+                                         track_lr=True)
+
+    model.fit(
+        x,
+        y,
+        batch_size=2,
+        epochs=2,
+        validation_data=(x, y),
+        callbacks=[tb_cbk])
+
+    summary_file = callbacks_test.list_summaries(logdir=self.logdir)
+    self.assertEqual(
+        summary_file.scalars, {
+            _ObservedSummary(logdir=self.train_dir, tag='epoch_loss'),
+            _ObservedSummary(logdir=self.train_dir, tag='epoch_learning_rate'),
+            _ObservedSummary(logdir=self.validation_dir, tag='epoch_loss'),
+        })
+
+
+if __name__ == '__main__':
+  tf.test.main()

official/vision/image_classification/classifier_trainer.py

+# Lint as: python3
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Runs an Image Classification model."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import os
+
+import pprint
+from typing import Any, Tuple, Text, Optional, Mapping
+
+from absl import app
+from absl import flags
+from absl import logging
+import tensorflow.compat.v2 as tf
+
+from official.modeling import performance
+from official.modeling.hyperparams import params_dict
+from official.utils import hyperparams_flags
+from official.utils.logs import logger
+from official.utils.misc import distribution_utils
+from official.utils.misc import keras_utils
+from official.vision.image_classification import callbacks as custom_callbacks
+from official.vision.image_classification import dataset_factory
+from official.vision.image_classification import optimizer_factory
+from official.vision.image_classification.configs import base_configs
+from official.vision.image_classification.configs import configs
+from official.vision.image_classification.efficientnet import efficientnet_model
+from official.vision.image_classification.resnet import common
+from official.vision.image_classification.resnet import resnet_model
+
+MODELS = {
+    'efficientnet': efficientnet_model.EfficientNet.from_name,
+    'resnet': resnet_model.resnet50,
+}
+
+
+def _get_metrics(one_hot: bool) -> Mapping[Text, Any]:
+  """Get a dict of available metrics to track."""
+  if one_hot:
+    return {
+        # (name, metric_fn)
+        'acc': tf.keras.metrics.CategoricalAccuracy(name='accuracy'),
+        'accuracy': tf.keras.metrics.CategoricalAccuracy(name='accuracy'),
+        'top_1': tf.keras.metrics.CategoricalAccuracy(name='accuracy'),
+        'top_5': tf.keras.metrics.TopKCategoricalAccuracy(
+            k=5,
+            name='top_5_accuracy'),
+    }
+  else:
+    return {
+        # (name, metric_fn)
+        'acc': tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy'),
+        'accuracy': tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy'),
+        'top_1': tf.keras.metrics.SparseCategoricalAccuracy(name='accuracy'),
+        'top_5': tf.keras.metrics.SparseTopKCategoricalAccuracy(
+            k=5,
+            name='top_5_accuracy'),
+    }
+
+
+def get_image_size_from_model(
+    params: base_configs.ExperimentConfig) -> Optional[int]:
+  """If the given model has a preferred image size, return it."""
+  if params.model_name == 'efficientnet':
+    efficientnet_name = params.model.model_params.model_name
+    if efficientnet_name in efficientnet_model.MODEL_CONFIGS:
+      return efficientnet_model.MODEL_CONFIGS[efficientnet_name].resolution
+  return None
+
+
+def _get_dataset_builders(params: base_configs.ExperimentConfig,
+                          strategy: tf.distribute.Strategy,
+                          one_hot: bool
+                         ) -> Tuple[Any, Any, Any]:
+  """Create and return train, validation, and test dataset builders."""
+  if one_hot:
+    logging.warning('label_smoothing > 0, so datasets will be one hot encoded.')
+  else:
+    logging.warning('label_smoothing not applied, so datasets will not be one '
+                    'hot encoded.')
+
+  num_devices = strategy.num_replicas_in_sync
+  image_size = get_image_size_from_model(params)
+
+  dataset_configs = [
+      params.train_dataset, params.validation_dataset, params.test_dataset
+  ]
+  builders = []
+
+  for config in dataset_configs:
+    if config is not None and config.has_data:
+      builder = dataset_factory.DatasetBuilder(
+          config,
+          image_size=image_size or config.image_size,
+          num_devices=num_devices,
+          one_hot=one_hot)
+    else:
+      builder = None
+    builders.append(builder)
+
+  return builders
+
+
+def get_loss_scale(params: base_configs.ExperimentConfig,
+                   fp16_default: float = 128.) -> float:
+  """Returns the loss scale for initializations."""
+  loss_scale = params.model.loss.loss_scale
+  if loss_scale == 'dynamic':
+    return loss_scale
+  elif loss_scale is not None:
+    return float(loss_scale)
+  elif params.train_dataset.dtype == 'float32':
+    return 1.
+  else:
+    assert params.train_dataset.dtype == 'float16'
+    return fp16_default
+
+
+def _get_params_from_flags(flags_obj: flags.FlagValues):
+  """Get ParamsDict from flags."""
+  model = flags_obj.model_type.lower()
+  dataset = flags_obj.dataset.lower()
+  params = configs.get_config(model=model, dataset=dataset)
+
+  flags_overrides = {
+      'model_dir': flags_obj.model_dir,
+      'mode': flags_obj.mode,
+      'model': {
+          'name': model,
+      },
+      'runtime': {
+          'enable_eager': flags_obj.enable_eager,
+          'tpu': flags_obj.tpu,
+      },
+      'train_dataset': {
+          'data_dir': flags_obj.data_dir,
+      },
+      'validation_dataset': {
+          'data_dir': flags_obj.data_dir,
+      },
+      'test_dataset': {
+          'data_dir': flags_obj.data_dir,
+      },
+  }
+
+  overriding_configs = (flags_obj.config_file,
+                        flags_obj.params_override,
+                        flags_overrides)
+
+  pp = pprint.PrettyPrinter()
+
+  logging.info('Base params: %s', pp.pformat(params.as_dict()))
+
+  for param in overriding_configs:
+    logging.info('Overriding params: %s', param)
+    # Set is_strict to false because we can have dynamic dict parameters.
+    params = params_dict.override_params_dict(params, param, is_strict=False)
+
+  params.validate()
+  params.lock()
+
+  logging.info('Final model parameters: %s', pp.pformat(params.as_dict()))
+  return params
+
+
+def resume_from_checkpoint(model: tf.keras.Model,
+                           model_dir: str,
+                           train_steps: int) -> int:
+  """Resumes from the latest checkpoint, if possible.
+
+  Loads the model weights and optimizer settings from a checkpoint.
+  This function should be used in case of preemption recovery.
+
+  Args:
+    model: The model whose weights should be restored.
+    model_dir: The directory where model weights were saved.
+    train_steps: The number of steps to train.
+
+  Returns:
+    The epoch of the latest checkpoint, or 0 if not restoring.
+
+  """
+  logging.info('Load from checkpoint is enabled.')
+  latest_checkpoint = tf.train.latest_checkpoint(model_dir)
+  logging.info('latest_checkpoint: %s', latest_checkpoint)
+  if not latest_checkpoint:
+    logging.info('No checkpoint detected.')
+    return 0
+
+  logging.info('Checkpoint file %s found and restoring from '
+               'checkpoint', latest_checkpoint)
+  model.load_weights(latest_checkpoint)
+  initial_epoch = model.optimizer.iterations // train_steps
+  logging.info('Completed loading from checkpoint.')
+  logging.info('Resuming from epoch %d', initial_epoch)
+  return int(initial_epoch)
+
+
+def initialize(params: base_configs.ExperimentConfig):
+  """Initializes backend related initializations."""
+  keras_utils.set_session_config(
+      enable_eager=params.runtime.enable_eager,
+      enable_xla=params.runtime.enable_xla)
+  if params.runtime.gpu_threads_enabled:
+    keras_utils.set_gpu_thread_mode_and_count(
+        per_gpu_thread_count=params.runtime.per_gpu_thread_count,
+        gpu_thread_mode=params.runtime.gpu_thread_mode,
+        num_gpus=params.runtime.num_gpus,
+        datasets_num_private_threads=params.runtime.dataset_num_private_threads)
+
+  dataset = params.train_dataset or params.validation_dataset
+  performance.set_mixed_precision_policy(dataset.dtype)
+
+  if dataset.data_format:
+    data_format = dataset.data_format
+  elif tf.config.list_physical_devices('GPU'):
+    data_format = 'channels_first'
+  else:
+    data_format = 'channels_last'
+  tf.keras.backend.set_image_data_format(data_format)
+  distribution_utils.configure_cluster(
+      params.runtime.worker_hosts,
+      params.runtime.task_index)
+  if params.runtime.enable_eager:
+    # Enable eager execution to allow step-by-step debugging
+    tf.config.experimental_run_functions_eagerly(True)
+
+
+def define_classifier_flags():
+  """Defines common flags for image classification."""
+  hyperparams_flags.initialize_common_flags()
+  flags.DEFINE_string(
+      'data_dir',
+      default=None,
+      help='The location of the input data.')
+  flags.DEFINE_string(
+      'mode',
+      default=None,
+      help='Mode to run: `train`, `eval`, `train_and_eval` or `export`.')
+  flags.DEFINE_bool(
+      'enable_eager',
+      default=None,
+      help='Use eager execution and disable autograph for debugging.')
+  flags.DEFINE_string(
+      'model_type',
+      default=None,
+      help='The type of the model, e.g. EfficientNet, etc.')
+  flags.DEFINE_string(
+      'dataset',
+      default=None,
+      help='The name of the dataset, e.g. ImageNet, etc.')
+
+
+def serialize_config(params: base_configs.ExperimentConfig,
+                     model_dir: str):
+  """Serializes and saves the experiment config."""
+  params_save_path = os.path.join(model_dir, 'params.yaml')
+  logging.info('Saving experiment configuration to %s', params_save_path)
+  tf.io.gfile.makedirs(model_dir)
+  params_dict.save_params_dict_to_yaml(params, params_save_path)
+
+
+def train_and_eval(
+    params: base_configs.ExperimentConfig,
+    strategy_override: tf.distribute.Strategy) -> Mapping[str, Any]:
+  """Runs the train and eval path using compile/fit."""
+  logging.info('Running train and eval.')
+
+  # Note: for TPUs, strategy and scope should be created before the dataset
+  strategy = strategy_override or distribution_utils.get_distribution_strategy(
+      distribution_strategy=params.runtime.distribution_strategy,
+      all_reduce_alg=params.runtime.all_reduce_alg,
+      num_gpus=params.runtime.num_gpus,
+      tpu_address=params.runtime.tpu)
+
+  strategy_scope = distribution_utils.get_strategy_scope(strategy)
+
+  logging.info('Detected %d devices.', strategy.num_replicas_in_sync)
+
+  label_smoothing = params.model.loss.label_smoothing
+  one_hot = label_smoothing and label_smoothing > 0
+
+  builders = _get_dataset_builders(params, strategy, one_hot)
+  datasets = [builder.build() if builder else None for builder in builders]
+
+  # Unpack datasets and builders based on train/val/test splits
+  train_builder, validation_builder, test_builder = builders  # pylint: disable=unbalanced-tuple-unpacking
+  train_dataset, validation_dataset, test_dataset = datasets
+
+  train_epochs = params.train.epochs
+  train_steps = params.train.steps or train_builder.num_steps
+  validation_steps = params.evaluation.steps or validation_builder.num_steps
+
+  logging.info('Global batch size: %d', train_builder.global_batch_size)
+
+  with strategy_scope:
+    model_params = params.model.model_params.as_dict()
+    model = MODELS[params.model.name](**model_params)
+    learning_rate = optimizer_factory.build_learning_rate(
+        params=params.model.learning_rate,
+        batch_size=train_builder.global_batch_size,
+        train_steps=train_steps)
+    optimizer = optimizer_factory.build_optimizer(
+        optimizer_name=params.model.optimizer.name,
+        base_learning_rate=learning_rate,
+        params=params.model.optimizer.as_dict())
+
+    metrics_map = _get_metrics(one_hot)
+    metrics = [metrics_map[metric] for metric in params.train.metrics]
+
+    if one_hot:
+      loss_obj = tf.keras.losses.CategoricalCrossentropy(
+          label_smoothing=params.model.loss.label_smoothing)
+    else:
+      loss_obj = tf.keras.losses.SparseCategoricalCrossentropy()
+    model.compile(optimizer=optimizer,
+                  loss=loss_obj,
+                  metrics=metrics,
+                  run_eagerly=params.runtime.enable_eager)
+
+    initial_epoch = 0
+    if params.train.resume_checkpoint:
+      initial_epoch = resume_from_checkpoint(model=model,
+                                             model_dir=params.model_dir,
+                                             train_steps=train_steps)
+
+  serialize_config(params=params, model_dir=params.model_dir)
+  # TODO(dankondratyuk): callbacks significantly slow down training
+  callbacks = custom_callbacks.get_callbacks(
+      model_checkpoint=params.train.callbacks.enable_checkpoint_and_export,
+      include_tensorboard=params.train.callbacks.enable_tensorboard,
+      track_lr=params.train.tensorboard.track_lr,
+      write_model_weights=params.train.tensorboard.write_model_weights,
+      initial_step=initial_epoch * train_steps,
+      model_dir=params.model_dir)
+
+  history = model.fit(
+      train_dataset,
+      epochs=train_epochs,
+      steps_per_epoch=train_steps,
+      initial_epoch=initial_epoch,
+      callbacks=callbacks,
+      validation_data=validation_dataset,
+      validation_steps=validation_steps,
+      validation_freq=params.evaluation.epochs_between_evals)
+
+  validation_output = model.evaluate(
+      validation_dataset, steps=validation_steps, verbose=2)
+
+  # TODO(dankondratyuk): eval and save final test accuracy
+
+  stats = common.build_stats(history,
+                             validation_output,
+                             callbacks)
+  return stats
+
+
+def export(params: base_configs.ExperimentConfig):
+  """Runs the model export functionality."""
+  logging.info('Exporting model.')
+  model_params = params.model.model_params.as_dict()
+  model = MODELS[params.model.name](**model_params)
+  checkpoint = params.export.checkpoint
+  if checkpoint is None:
+    logging.info('No export checkpoint was provided. Using the latest '
+                 'checkpoint from model_dir.')
+    checkpoint = tf.train.latest_checkpoint(params.model_dir)
+
+  model.load_weights(checkpoint)
+  model.save(params.export.destination)
+
+
+def run(flags_obj: flags.FlagValues,
+        strategy_override: tf.distribute.Strategy = None) -> Mapping[str, Any]:
+  """Runs Image Classification model using native Keras APIs.
+
+  Args:
+    flags_obj: An object containing parsed flag values.
+    strategy_override: A `tf.distribute.Strategy` object to use for model.
+
+  Returns:
+    Dictionary of training/eval stats
+  """
+  params = _get_params_from_flags(flags_obj)
+  initialize(params)
+
+  if params.mode == 'train_and_eval':
+    return train_and_eval(params, strategy_override)
+  elif params.mode == 'export_only':
+    export(params)
+  else:
+    raise ValueError('{} is not a valid mode.'.format(params.mode))
+
+
+def main(_):
+  with logger.benchmark_context(flags.FLAGS):
+    stats = run(flags.FLAGS)
+  if stats:
+    logging.info('Run stats:\n%s', stats)
+
+
+if __name__ == '__main__':
+  logging.set_verbosity(logging.INFO)
+  define_classifier_flags()
+  flags.mark_flag_as_required('data_dir')
+  flags.mark_flag_as_required('mode')
+  flags.mark_flag_as_required('model_type')
+  flags.mark_flag_as_required('dataset')
+
+  assert tf.version.VERSION.startswith('2.')
+  app.run(main)

official/vision/image_classification/classifier_trainer_test.py

+# Lint as: python3
+# Copyright 2017 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Unit tests for the classifier trainer models."""
+
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import copy
+import functools
+import json
+
+import os
+import sys
+
+from typing import Any, Callable, Iterable, Mapping, MutableMapping, Optional, Tuple
+
+from absl import flags
+from absl.testing import parameterized
+import tensorflow.compat.v2 as tf
+
+from tensorflow.python.distribute import combinations
+from tensorflow.python.distribute import strategy_combinations
+from official.utils.flags import core as flags_core
+from official.vision.image_classification import classifier_trainer
+from official.vision.image_classification import dataset_factory
+from official.vision.image_classification import test_utils
+from official.vision.image_classification.configs import base_configs
+
+classifier_trainer.define_classifier_flags()
+
+
+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.tpu_strategy,
+          strategy_combinations.one_device_strategy_gpu,
+      ],
+      model=[
+          'efficientnet',
+          'resnet',
+      ],
+      mode='eager',
+      dataset=[
+          'imagenet',
+      ],
+  )
+
+
+def get_params_override(params_override: Mapping[str, Any]) -> str:
+  """Converts params_override dict to string command."""
+  return '--params_override=' + json.dumps(params_override)
+
+
+def basic_params_override() -> MutableMapping[str, Any]:
+  """Returns a basic parameter configuration for testing."""
+  return {
+      'train_dataset': {
+          'builder': 'synthetic',
+          'use_per_replica_batch_size': True,
+          'batch_size': 1,
+          'image_size': 224,
+      },
+      'validation_dataset': {
+          'builder': 'synthetic',
+          'batch_size': 1,
+          'use_per_replica_batch_size': True,
+          'image_size': 224,
+      },
+      'test_dataset': {
+          'builder': 'synthetic',
+          'batch_size': 1,
+          'use_per_replica_batch_size': True,
+          'image_size': 224,
+      },
+      'train': {
+          'steps': 1,
+          'epochs': 1,
+          'callbacks': {
+              'enable_checkpoint_and_export': True,
+              'enable_tensorboard': False,
+          },
+      },
+      'evaluation': {
+          'steps': 1,
+      },
+  }
+
+
+def get_trivial_model(num_classes: int) -> tf.keras.Model:
+  """Creates and compiles trivial model for ImageNet dataset."""
+  model = test_utils.trivial_model(num_classes=num_classes)
+  lr = 0.01
+  optimizer = tf.keras.optimizers.SGD(learning_rate=lr)
+  loss_obj = tf.keras.losses.SparseCategoricalCrossentropy()
+  model.compile(optimizer=optimizer,
+                loss=loss_obj,
+                run_eagerly=True)
+  return model
+
+
+def get_trivial_data() -> tf.data.Dataset:
+  """Gets trivial data in the ImageNet size."""
+  def generate_data(_) -> tf.data.Dataset:
+    image = tf.zeros(shape=(224, 224, 3), dtype=tf.float32)
+    label = tf.zeros([1], dtype=tf.int32)
+    return image, label
+
+  dataset = tf.data.Dataset.range(1)
+  dataset = dataset.repeat()
+  dataset = dataset.map(generate_data,
+                        num_parallel_calls=tf.data.experimental.AUTOTUNE)
+  dataset = dataset.prefetch(buffer_size=1).batch(1)
+  return dataset
+
+
+def run_end_to_end(main: Callable[[Any], None],
+                   extra_flags: Optional[Iterable[str]] = None,
+                   model_dir: Optional[str] = None):
+  """Runs the classifier trainer end-to-end."""
+  extra_flags = [] if extra_flags is None else extra_flags
+  args = [sys.argv[0], '--model_dir', model_dir] + extra_flags
+  flags_core.parse_flags(argv=args)
+  main(flags.FLAGS)
+
+
+class ClassifierTest(tf.test.TestCase, parameterized.TestCase):
+  """Unit tests for Keras models."""
+  _tempdir = None
+
+  @classmethod
+  def setUpClass(cls):  # pylint: disable=invalid-name
+    super(ClassifierTest, cls).setUpClass()
+
+  def tearDown(self):
+    super(ClassifierTest, self).tearDown()
+    tf.io.gfile.rmtree(self.get_temp_dir())
+
+  @combinations.generate(distribution_strategy_combinations())
+  def test_end_to_end_train_and_eval_export(self, distribution, model, dataset):
+    """Test train_and_eval and export for Keras classifier models."""
+    # Some parameters are not defined as flags (e.g. cannot run
+    # classifier_train.py --batch_size=...) by design, so use
+    # "--params_override=..." instead
+    model_dir = self.get_temp_dir()
+    base_flags = [
+        '--data_dir=not_used',
+        '--model_type=' + model,
+        '--dataset=' + dataset,
+    ]
+    train_and_eval_flags = base_flags + [
+        get_params_override(basic_params_override()),
+        '--mode=train_and_eval',
+    ]
+
+    export_params = basic_params_override()
+    export_path = os.path.join(model_dir, 'export')
+    export_params['export'] = {}
+    export_params['export']['destination'] = export_path
+    export_flags = base_flags + [
+        '--mode=export_only',
+        get_params_override(export_params)
+    ]
+
+    run = functools.partial(classifier_trainer.run,
+                            strategy_override=distribution)
+    run_end_to_end(main=run,
+                   extra_flags=train_and_eval_flags,
+                   model_dir=model_dir)
+    run_end_to_end(main=run,
+                   extra_flags=export_flags,
+                   model_dir=model_dir)
+    self.assertTrue(os.path.exists(export_path))
+
+  @combinations.generate(distribution_strategy_combinations())
+  def test_end_to_end_invalid_mode(self, distribution, model, dataset):
+    """Test the Keras EfficientNet model with `strategy`."""
+    model_dir = self.get_temp_dir()
+    extra_flags = [
+        '--data_dir=not_used',
+        '--mode=invalid_mode',
+        '--model_type=' + model,
+        '--dataset=' + dataset,
+        get_params_override(basic_params_override()),
+    ]
+
+    run = functools.partial(classifier_trainer.run,
+                            strategy_override=distribution)
+    with self.assertRaises(ValueError):
+      run_end_to_end(main=run, extra_flags=extra_flags, model_dir=model_dir)
+
+
+class UtilTests(parameterized.TestCase, tf.test.TestCase):
+  """Tests for individual utility functions within classifier_trainer.py."""
+
+  @parameterized.named_parameters(
+      ('efficientnet-b0', 'efficientnet', 'efficientnet-b0', 224),
+      ('efficientnet-b1', 'efficientnet', 'efficientnet-b1', 240),
+      ('efficientnet-b2', 'efficientnet', 'efficientnet-b2', 260),
+      ('efficientnet-b3', 'efficientnet', 'efficientnet-b3', 300),
+      ('efficientnet-b4', 'efficientnet', 'efficientnet-b4', 380),
+      ('efficientnet-b5', 'efficientnet', 'efficientnet-b5', 456),
+      ('efficientnet-b6', 'efficientnet', 'efficientnet-b6', 528),
+      ('efficientnet-b7', 'efficientnet', 'efficientnet-b7', 600),
+      ('resnet', 'resnet', '', None),
+  )
+  def test_get_model_size(self, model, model_name, expected):
+    config = base_configs.ExperimentConfig(
+        model_name=model,
+        model=base_configs.ModelConfig(
+            model_params={
+                'model_name': model_name,
+            },
+        )
+    )
+    size = classifier_trainer.get_image_size_from_model(config)
+    self.assertEqual(size, expected)
+
+  @parameterized.named_parameters(
+      ('dynamic', 'dynamic', None, 'dynamic'),
+      ('scalar', 128., None, 128.),
+      ('float32', None, 'float32', 1),
+      ('float16', None, 'float16', 128),
+  )
+  def test_get_loss_scale(self, loss_scale, dtype, expected):
+    config = base_configs.ExperimentConfig(
+        model=base_configs.ModelConfig(
+            loss=base_configs.LossConfig(loss_scale=loss_scale)),
+        train_dataset=dataset_factory.DatasetConfig(dtype=dtype))
+    ls = classifier_trainer.get_loss_scale(config, fp16_default=128)
+    self.assertEqual(ls, expected)
+
+  @parameterized.named_parameters(
+      ('float16', 'float16'),
+      ('bfloat16', 'bfloat16')
+  )
+  def test_initialize(self, dtype):
+    config = base_configs.ExperimentConfig(
+        runtime=base_configs.RuntimeConfig(
+            enable_eager=False,
+            enable_xla=False,
+            gpu_threads_enabled=True,
+            per_gpu_thread_count=1,
+            gpu_thread_mode='gpu_private',
+            num_gpus=1,
+            dataset_num_private_threads=1,
+        ),
+        train_dataset=dataset_factory.DatasetConfig(dtype=dtype),
+        model=base_configs.ModelConfig(
+            loss=base_configs.LossConfig(loss_scale='dynamic')),
+    )
+    classifier_trainer.initialize(config)
+
+  def test_resume_from_checkpoint(self):
+    """Tests functionality for resuming from checkpoint."""
+    # Set the keras policy
+    policy = tf.keras.mixed_precision.experimental.Policy('mixed_bfloat16')
+    tf.keras.mixed_precision.experimental.set_policy(policy)
+
+    # Get the model, datasets, and compile it.
+    model = get_trivial_model(10)
+
+    # Create the checkpoint
+    model_dir = self.get_temp_dir()
+    train_epochs = 1
+    train_steps = 10
+    ds = get_trivial_data()
+    callbacks = [
+        tf.keras.callbacks.ModelCheckpoint(
+            os.path.join(model_dir, 'model.ckpt-{epoch:04d}'),
+            save_weights_only=True)
+    ]
+    model.fit(
+        ds,
+        callbacks=callbacks,
+        epochs=train_epochs,
+        steps_per_epoch=train_steps)
+
+    # Test load from checkpoint
+    clean_model = get_trivial_model(10)
+    weights_before_load = copy.deepcopy(clean_model.get_weights())
+    initial_epoch = classifier_trainer.resume_from_checkpoint(
+        model=clean_model,
+        model_dir=model_dir,
+        train_steps=train_steps)
+    self.assertEqual(initial_epoch, 1)
+    self.assertNotAllClose(weights_before_load, clean_model.get_weights())
+
+    tf.io.gfile.rmtree(model_dir)
+
+  def test_serialize_config(self):
+    """Tests functionality for serializing data."""
+    config = base_configs.ExperimentConfig()
+    model_dir = self.get_temp_dir()
+    classifier_trainer.serialize_config(params=config, model_dir=model_dir)
+    saved_params_path = os.path.join(model_dir, 'params.yaml')
+    self.assertTrue(os.path.exists(saved_params_path))
+    tf.io.gfile.rmtree(model_dir)
+
+if __name__ == '__main__':
+  assert tf.version.VERSION.startswith('2.')
+  tf.test.main()

official/vision/image_classification/configs/__init__.py

+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================

official/vision/image_classification/configs/base_configs.py

+# Lint as: python3
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Definitions for high level configuration groups.."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+
+from typing import Any, List, Mapping, Optional
+
+import dataclasses
+
+from official.modeling.hyperparams import base_config
+
+
+CallbacksConfig = base_config.CallbacksConfig
+TensorboardConfig = base_config.TensorboardConfig
+RuntimeConfig = base_config.RuntimeConfig
+
+
+@dataclasses.dataclass
+class ExportConfig(base_config.Config):
+  """Configuration for exports.
+
+  Attributes:
+    checkpoint: the path to the checkpoint to export.
+    destination: the path to where the checkpoint should be exported.
+
+  """
+  checkpoint: str = None
+  destination: str = None
+
+
+@dataclasses.dataclass
+class MetricsConfig(base_config.Config):
+  """Configuration for Metrics.
+
+  Attributes:
+    accuracy: Whether or not to track accuracy as a Callback. Defaults to None.
+    top_5: Whether or not to track top_5_accuracy as a Callback. Defaults to
+      None.
+
+  """
+  accuracy: bool = None
+  top_5: bool = None
+
+
+@dataclasses.dataclass
+class TrainConfig(base_config.Config):
+  """Configuration for training.
+
+  Attributes:
+    resume_checkpoint: Whether or not to enable load checkpoint loading.
+      Defaults to None.
+    epochs: The number of training epochs to run. Defaults to None.
+    steps: The number of steps to run per epoch. If None, then this will be
+      inferred based on the number of images and batch size. Defaults to None.
+    callbacks: An instance of CallbacksConfig.
+    metrics: An instance of MetricsConfig.
+    tensorboard: An instance of TensorboardConfig.
+
+  """
+  resume_checkpoint: bool = None
+  epochs: int = None
+  steps: int = None
+  callbacks: CallbacksConfig = CallbacksConfig()
+  metrics: List[str] = None
+  tensorboard: TensorboardConfig = TensorboardConfig()
+
+
+@dataclasses.dataclass
+class EvalConfig(base_config.Config):
+  """Configuration for evaluation.
+
+  Attributes:
+    epochs_between_evals: The number of train epochs to run between evaluations.
+      Defaults to None.
+    steps: The number of eval steps to run during evaluation. If None, this will
+      be inferred based on the number of images and batch size. Defaults to
+      None.
+
+  """
+  epochs_between_evals: int = None
+  steps: int = None
+
+
+@dataclasses.dataclass
+class LossConfig(base_config.Config):
+  """Configuration for Loss.
+
+  Attributes:
+    name: The name of the loss. Defaults to None.
+    loss_scale: The type of loss scale
+    label_smoothing: Whether or not to apply label smoothing to the loss. This
+      only applies to 'categorical_cross_entropy'.
+
+  """
+  name: str = None
+  loss_scale: str = None
+  label_smoothing: float = None
+
+
+@dataclasses.dataclass
+class OptimizerConfig(base_config.Config):
+  """Configuration for Optimizers.
+
+  Attributes:
+    name: The name of the optimizer. Defaults to None.
+    decay: Decay or rho, discounting factor for gradient. Defaults to None.
+    epsilon: Small value used to avoid 0 denominator. Defaults to None.
+    momentum: Plain momentum constant. Defaults to None.
+    nesterov: Whether or not to apply Nesterov momentum. Defaults to None.
+    moving_average_decay: The amount of decay to apply. If 0 or None, then
+      exponential moving average is not used. Defaults to None.
+    lookahead: Whether or not to apply the lookahead optimizer. Defaults to
+      None.
+    beta_1: The exponential decay rate for the 1st moment estimates. Used in
+      the Adam optimizers. Defaults to None.
+    beta_2: The exponential decay rate for the 2nd moment estimates. Used in
+      the Adam optimizers. Defaults to None.
+    epsilon: Small value used to avoid 0 denominator. Defaults to 1e-7.
+
+  """
+  name: str = None
+  decay: float = None
+  epsilon: float = None
+  momentum: float = None
+  nesterov: bool = None
+  moving_average_decay: Optional[float] = None
+  lookahead: Optional[bool] = None
+  beta_1: float = None
+  beta_2: float = None
+  epsilon: float = None
+
+
+@dataclasses.dataclass
+class LearningRateConfig(base_config.Config):
+  """Configuration for learning rates.
+
+  Attributes:
+    name: The name of the learning rate. Defaults to None.
+    initial_lr: The initial learning rate. Defaults to None.
+    decay_epochs: The number of decay epochs. Defaults to None.
+    decay_rate: The rate of decay. Defaults to None.
+    warmup_epochs: The number of warmup epochs. Defaults to None.
+    batch_lr_multiplier: The multiplier to apply to the base learning rate,
+      if necessary. Defaults to None.
+    examples_per_epoch: the number of examples in a single epoch.
+      Defaults to None.
+    boundaries: boundaries used in piecewise constant decay with warmup.
+    multipliers: multipliers used in piecewise constant decay with warmup.
+    scale_by_batch_size: Scale the learning rate by a fraction of the batch
+      size. Set to 0 for no scaling (default).
+
+  """
+  name: str = None
+  initial_lr: float = None
+  decay_epochs: float = None
+  decay_rate: float = None
+  warmup_epochs: int = None
+  examples_per_epoch: int = None
+  boundaries: List[int] = None
+  multipliers: List[float] = None
+  scale_by_batch_size: float = 0.
+
+
+@dataclasses.dataclass
+class ModelConfig(base_config.Config):
+  """Configuration for Models.
+
+  Attributes:
+    name: The name of the model. Defaults to None.
+    model_params: The parameters used to create the model. Defaults to None.
+    num_classes: The number of classes in the model. Defaults to None.
+    loss: A `LossConfig` instance. Defaults to None.
+    optimizer: An `OptimizerConfig` instance. Defaults to None.
+
+  """
+  name: str = None
+  model_params: Mapping[str, Any] = None
+  num_classes: int = None
+  loss: LossConfig = None
+  optimizer: OptimizerConfig = None
+
+
+@dataclasses.dataclass
+class ExperimentConfig(base_config.Config):
+  """Base configuration for an image classification experiment.
+
+  Attributes:
+    model_dir: The directory to use when running an experiment.
+    mode: e.g. 'train_and_eval', 'export'
+    runtime: A `RuntimeConfig` instance.
+    train: A `TrainConfig` instance.
+    evaluation: An `EvalConfig` instance.
+    model: A `ModelConfig` instance.
+    export: An `ExportConfig` instance.
+
+  """
+  model_dir: str = None
+  model_name: str = None
+  mode: str = None
+  runtime: RuntimeConfig = None
+  train_dataset: Any = None
+  validation_dataset: Any = None
+  test_dataset: Any = None
+  train: TrainConfig = None
+  evaluation: EvalConfig = None
+  model: ModelConfig = None
+  export: ExportConfig = None

official/vision/image_classification/configs/configs.py

+# Lint as: python3
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Configuration utils for image classification experiments."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+import dataclasses
+
+from official.vision.image_classification import dataset_factory
+from official.vision.image_classification.configs import base_configs
+from official.vision.image_classification.efficientnet import efficientnet_config
+from official.vision.image_classification.resnet import resnet_config
+
+
+@dataclasses.dataclass
+class EfficientNetImageNetConfig(base_configs.ExperimentConfig):
+  """Base configuration to train efficientnet-b0 on ImageNet.
+
+  Attributes:
+    export: An `ExportConfig` instance
+    runtime: A `RuntimeConfig` instance.
+    dataset: A `DatasetConfig` instance.
+    train: A `TrainConfig` instance.
+    evaluation: An `EvalConfig` instance.
+    model: A `ModelConfig` instance.
+
+  """
+  export: base_configs.ExportConfig = base_configs.ExportConfig()
+  runtime: base_configs.RuntimeConfig = base_configs.RuntimeConfig()
+  train_dataset: dataset_factory.DatasetConfig = \
+      dataset_factory.ImageNetConfig(split='train')
+  validation_dataset: dataset_factory.DatasetConfig = \
+      dataset_factory.ImageNetConfig(split='validation')
+  test_dataset: dataset_factory.DatasetConfig = \
+      dataset_factory.ImageNetConfig(split='validation')
+  train: base_configs.TrainConfig = base_configs.TrainConfig(
+      resume_checkpoint=True,
+      epochs=500,
+      steps=None,
+      callbacks=base_configs.CallbacksConfig(enable_checkpoint_and_export=True,
+                                             enable_tensorboard=True),
+      metrics=['accuracy', 'top_5'],
+      tensorboard=base_configs.TensorboardConfig(track_lr=True,
+                                                 write_model_weights=False))
+  evaluation: base_configs.EvalConfig = base_configs.EvalConfig(
+      epochs_between_evals=1,
+      steps=None)
+  model: base_configs.ModelConfig = \
+    efficientnet_config.EfficientNetModelConfig()
+
+
+@dataclasses.dataclass
+class ResNetImagenetConfig(base_configs.ExperimentConfig):
+  """Base configuration to train resnet-50 on ImageNet."""
+  export: base_configs.ExportConfig = base_configs.ExportConfig()
+  runtime: base_configs.RuntimeConfig = base_configs.RuntimeConfig()
+  train_dataset: dataset_factory.DatasetConfig = \
+      dataset_factory.ImageNetConfig(split='train',
+                                     one_hot=False,
+                                     mean_subtract=True,
+                                     standardize=True)
+  validation_dataset: dataset_factory.DatasetConfig = \
+      dataset_factory.ImageNetConfig(split='validation',
+                                     one_hot=False,
+                                     mean_subtract=True,
+                                     standardize=True)
+  test_dataset: dataset_factory.DatasetConfig = \
+      dataset_factory.ImageNetConfig(split='validation',
+                                     one_hot=False,
+                                     mean_subtract=True,
+                                     standardize=True)
+  train: base_configs.TrainConfig = base_configs.TrainConfig(
+      resume_checkpoint=True,
+      epochs=90,
+      steps=None,
+      callbacks=base_configs.CallbacksConfig(enable_checkpoint_and_export=True,
+                                             enable_tensorboard=True),
+      metrics=['accuracy', 'top_5'],
+      tensorboard=base_configs.TensorboardConfig(track_lr=True,
+                                                 write_model_weights=False))
+  evaluation: base_configs.EvalConfig = base_configs.EvalConfig(
+      epochs_between_evals=1,
+      steps=None)
+  model: base_configs.ModelConfig = resnet_config.ResNetModelConfig()
+
+
+def get_config(model: str, dataset: str) -> base_configs.ExperimentConfig:
+  """Given model and dataset names, return the ExperimentConfig."""
+  dataset_model_config_map = {
+      'imagenet': {
+          'efficientnet': EfficientNetImageNetConfig(),
+          'resnet': ResNetImagenetConfig(),
+      }
+  }
+  try:
+    return dataset_model_config_map[dataset][model]
+  except KeyError:
+    if dataset not in dataset_model_config_map:
+      raise KeyError('Invalid dataset received. Received: {}. Supported '
+                     'datasets include: {}'.format(
+                         dataset,
+                         ', '.join(dataset_model_config_map.keys())))
+    raise KeyError('Invalid model received. Received: {}. Supported models for'
+                   '{} include: {}'.format(
+                       model,
+                       dataset,
+                       ', '.join(dataset_model_config_map[dataset].keys())))

official/vision/image_classification/configs/examples/efficientnet/imagenet/efficientnet-b0-gpu.yaml

+# Training configuration for EfficientNet-b0 trained on ImageNet on GPUs.
+# Takes ~32 minutes per epoch for 8 V100s.
+# Reaches ~76.1% within 350 epochs.
+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  model_dir: null
+  mode: 'train_and_eval'
+  distribution_strategy: 'mirrored'
+  num_gpus: 1
+train_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'train'
+  num_classes: 1000
+  num_examples: 1281167
+  batch_size: 32
+  use_per_replica_batch_size: True
+  dtype: 'float32'
+  augmenter:
+    name: 'autoaugment'
+validation_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'validation'
+  num_classes: 1000
+  num_examples: 50000
+  batch_size: 32
+  use_per_replica_batch_size: True
+  dtype: 'float32'
+model:
+  model_params:
+    model_name: 'efficientnet-b0'
+    overrides:
+      num_classes: 1000
+      batch_norm: 'default'
+      dtype: 'float32'
+  optimizer:
+    name: 'rmsprop'
+    momentum: 0.9
+    decay: 0.9
+  learning_rate:
+    name: 'exponential'
+  loss:
+    label_smoothing: 0.1
+train:
+  resume_checkpoint: True
+  epochs: 500
+evaluation:
+  epochs_between_evals: 1

official/vision/image_classification/configs/examples/efficientnet/imagenet/efficientnet-b0-tpu.yaml

+# Training configuration for EfficientNet-b0 trained on ImageNet on TPUs.
+# Takes ~2 minutes, 50 seconds per epoch for v3-32.
+# Reaches ~76.1% within 350 epochs.
+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  model_dir: null
+  mode: 'train_and_eval'
+  distribution_strategy: 'tpu'
+train_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'train'
+  num_classes: 1000
+  num_examples: 1281167
+  batch_size: 128
+  use_per_replica_batch_size: True
+  dtype: 'bfloat16'
+  augmenter:
+    name: 'autoaugment'
+validation_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'validation'
+  num_classes: 1000
+  num_examples: 50000
+  batch_size: 128
+  use_per_replica_batch_size: True
+  dtype: 'bfloat16'
+model:
+  model_params:
+    model_name: 'efficientnet-b0'
+    overrides:
+      num_classes: 1000
+      batch_norm: 'tpu'
+      dtype: 'bfloat16'
+  optimizer:
+    name: 'rmsprop'
+    momentum: 0.9
+    decay: 0.9
+    moving_average_decay: 0.
+    lookahead: false
+  learning_rate:
+    name: 'exponential'
+  loss:
+    label_smoothing: 0.1
+train:
+  resume_checkpoint: True
+  epochs: 500
+evaluation:
+  epochs_between_evals: 1

official/vision/image_classification/configs/examples/efficientnet/imagenet/efficientnet-b1-gpu.yaml

+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  model_dir: null
+  mode: 'train_and_eval'
+  distribution_strategy: 'mirrored'
+  num_gpus: 1
+train_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'train'
+  num_classes: 1000
+  num_examples: 1281167
+  batch_size: 32
+  dtype: 'float32'
+validation_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'validation'
+  num_classes: 1000
+  num_examples: 50000
+  batch_size: 32
+  dtype: 'float32'
+model:
+  model_params:
+    model_name: 'efficientnet-b1'
+    overrides:
+      num_classes: 1000
+      batch_norm: 'default'
+      dtype: 'float32'
+  optimizer:
+    name: 'rmsprop'
+    momentum: 0.9
+    decay: 0.9
+  learning_rate:
+    name: 'exponential'
+  loss:
+    label_smoothing: 0.1
+train:
+  resume_checkpoint: True
+  epochs: 500
+evaluation:
+  epochs_between_evals: 1

official/vision/image_classification/configs/examples/efficientnet/imagenet/efficientnet-b1-tpu.yaml

+# Training configuration for EfficientNet-b1 trained on ImageNet on TPUs.
+# Takes ~3 minutes, 15 seconds per epoch for v3-32.
+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  model_dir: null
+  mode: 'train_and_eval'
+  distribution_strategy: 'tpu'
+train_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'train'
+  num_classes: 1000
+  num_examples: 1281167
+  batch_size: 128
+  use_per_replica_batch_size: True
+  dtype: 'bfloat16'
+  augmenter:
+    name: 'autoaugment'
+validation_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'validation'
+  num_classes: 1000
+  num_examples: 50000
+  batch_size: 128
+  use_per_replica_batch_size: True
+  dtype: 'bfloat16'
+model:
+  model_params:
+    model_name: 'efficientnet-b1'
+    overrides:
+      num_classes: 1000
+      batch_norm: 'tpu'
+      dtype: 'bfloat16'
+  optimizer:
+    name: 'rmsprop'
+    momentum: 0.9
+    decay: 0.9
+  learning_rate:
+    name: 'exponential'
+  loss:
+    label_smoothing: 0.1
+train:
+  resume_checkpoint: True
+  epochs: 500
+evaluation:
+  epochs_between_evals: 1

official/vision/image_classification/configs/examples/resnet/imagenet/gpu.yaml

+# Training configuration for ResNet trained on ImageNet on GPUs.
+# Takes ~3 minutes, 15 seconds per epoch for 8 V100s.
+# Reaches ~76.1% within 90 epochs.
+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  model_dir: null
+  mode: 'train_and_eval'
+  distribution_strategy: 'mirrored'
+  num_gpus: 1
+train_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'train'
+  image_size: 224
+  num_classes: 1000
+  num_examples: 1281167
+  batch_size: 128
+  use_per_replica_batch_size: True
+  dtype: 'float32'
+  mean_subtract: True
+  standardize: True
+validation_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'validation'
+  image_size: 224
+  num_classes: 1000
+  num_examples: 50000
+  batch_size: 128
+  use_per_replica_batch_size: True
+  dtype: 'float32'
+  mean_subtract: True
+  standardize: True
+model:
+  model_name: 'resnet'
+  model_params:
+    rescale_inputs: False
+  optimizer:
+    name: 'momentum'
+    momentum: 0.9
+    decay: 0.9
+    epsilon: 0.001
+  learning_rate:
+    name: 'piecewise_constant_with_warmup'
+  loss:
+    label_smoothing: 0.1
+train:
+  resume_checkpoint: True
+  epochs: 90
+evaluation:
+  epochs_between_evals: 1

official/vision/image_classification/configs/examples/resnet/imagenet/tpu.yaml

+# Training configuration for ResNet trained on ImageNet on TPUs.
+# Takes ~2 minutes, 43 seconds per epoch for a v3-32.
+# Reaches ~76.1% within 90 epochs.
+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  model_dir: null
+  mode: 'train_and_eval'
+  distribution_strategy: 'tpu'
+train_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'train'
+  one_hot: False
+  image_size: 224
+  num_classes: 1000
+  num_examples: 1281167
+  batch_size: 128
+  use_per_replica_batch_size: True
+  mean_subtract: True
+  standardize: True
+  dtype: 'bfloat16'
+validation_dataset:
+  name: 'imagenet2012'
+  data_dir: null
+  builder: 'records'
+  split: 'validation'
+  one_hot: False
+  image_size: 224
+  num_classes: 1000
+  num_examples: 50000
+  batch_size: 128
+  use_per_replica_batch_size: True
+  mean_subtract: True
+  standardize: True
+  dtype: 'bfloat16'
+model:
+  model_name: 'resnet'
+  model_params:
+    rescale_inputs: False
+  optimizer:
+    name: 'momentum'
+    momentum: 0.9
+    decay: 0.9
+    epsilon: 0.001
+    moving_average_decay: 0.
+    lookahead: false
+  learning_rate:
+    name: 'piecewise_constant_with_warmup'
+  loss:
+    label_smoothing: 0.1
+train:
+  callbacks:
+    enable_checkpoint_and_export: True
+  resume_checkpoint: True
+  epochs: 90
+evaluation:
+  epochs_between_evals: 1

official/vision/image_classification/dataset_factory.py

+# Lint as: python3
+# Copyright 2018 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Dataset utilities for vision tasks using TFDS and tf.data.Dataset."""
+from __future__ import absolute_import
+from __future__ import division
+# from __future__ import google_type_annotations
+from __future__ import print_function
+
+import os
+from typing import Any, List, Optional, Tuple, Mapping, Union
+from absl import logging
+from dataclasses import dataclass
+import tensorflow.compat.v2 as tf
+import tensorflow_datasets as tfds
+
+from official.modeling.hyperparams import base_config
+from official.vision.image_classification import augment
+from official.vision.image_classification import preprocessing
+
+
+AUGMENTERS = {
+    'autoaugment': augment.AutoAugment,
+    'randaugment': augment.RandAugment,
+}
+
+
+@dataclass
+class AugmentConfig(base_config.Config):
+  """Configuration for image augmenters.
+
+  Attributes:
+    name: The name of the image augmentation to use. Possible options are
+      None (default), 'autoaugment', or 'randaugment'.
+    params: Any paramaters used to initialize the augmenter.
+  """
+  name: Optional[str] = None
+  params: Optional[Mapping[str, Any]] = None
+
+  def build(self) -> augment.ImageAugment:
+    """Build the augmenter using this config."""
+    params = self.params or {}
+    augmenter = AUGMENTERS.get(self.name, None)
+    return augmenter(**params) if augmenter is not None else None
+
+
+@dataclass
+class DatasetConfig(base_config.Config):
+  """The base configuration for building datasets.
+
+  Attributes:
+    name: The name of the Dataset. Usually should correspond to a TFDS dataset.
+    data_dir: The path where the dataset files are stored, if available.
+    filenames: Optional list of strings representing the TFRecord names.
+    builder: The builder type used to load the dataset. Value should be one of
+      'tfds' (load using TFDS), 'records' (load from TFRecords), or 'synthetic'
+      (generate dummy synthetic data without reading from files).
+    split: The split of the dataset. Usually 'train', 'validation', or 'test'.
+    image_size: The size of the image in the dataset. This assumes that
+      `width` == `height`. Set to 'infer' to infer the image size from TFDS
+      info. This requires `name` to be a registered dataset in TFDS.
+    num_classes: The number of classes given by the dataset. Set to 'infer'
+      to infer the image size from TFDS info. This requires `name` to be a
+      registered dataset in TFDS.
+    num_channels: The number of channels given by the dataset. Set to 'infer'
+      to infer the image size from TFDS info. This requires `name` to be a
+      registered dataset in TFDS.
+    num_examples: The number of examples given by the dataset. Set to 'infer'
+      to infer the image size from TFDS info. This requires `name` to be a
+      registered dataset in TFDS.
+    batch_size: The base batch size for the dataset.
+    use_per_replica_batch_size: Whether to scale the batch size based on
+      available resources. If set to `True`, the dataset builder will return
+      batch_size multiplied by `num_devices`, the number of device replicas
+      (e.g., the number of GPUs or TPU cores).
+    num_devices: The number of replica devices to use. This should be set by
+      `strategy.num_replicas_in_sync` when using a distribution strategy.
+    data_format: The data format of the images. Should be 'channels_last' or
+      'channels_first'.
+    dtype: The desired dtype of the dataset. This will be set during
+      preprocessing.
+    one_hot: Whether to apply one hot encoding. Set to `True` to be able to use
+      label smoothing.
+    augmenter: The augmenter config to use. No augmentation is used by default.
+    download: Whether to download data using TFDS.
+    shuffle_buffer_size: The buffer size used for shuffling training data.
+    file_shuffle_buffer_size: The buffer size used for shuffling raw training
+      files.
+    skip_decoding: Whether to skip image decoding when loading from TFDS.
+    deterministic_train: Whether the examples in the training set should output
+      in a deterministic order.
+    use_slack: whether to introduce slack in the last prefetch. This may reduce
+      CPU contention at the start of a training step.
+    cache: whether to cache to dataset examples. Can be used to avoid re-reading
+      from disk on the second epoch. Requires significant memory overhead.
+    mean_subtract: whether or not to apply mean subtraction to the dataset.
+    standardize: whether or not to apply standardization to the dataset.
+  """
+  name: Optional[str] = None
+  data_dir: Optional[str] = None
+  filenames: Optional[List[str]] = None
+  builder: str = 'tfds'
+  split: str = 'train'
+  image_size: Union[int, str] = 'infer'
+  num_classes: Union[int, str] = 'infer'
+  num_channels: Union[int, str] = 'infer'
+  num_examples: Union[int, str] = 'infer'
+  batch_size: int = 128
+  use_per_replica_batch_size: bool = False
+  num_devices: int = 1
+  data_format: str = 'channels_last'
+  dtype: str = 'float32'
+  one_hot: bool = True
+  augmenter: AugmentConfig = AugmentConfig()
+  download: bool = False
+  shuffle_buffer_size: int = 10000
+  file_shuffle_buffer_size: int = 1024
+  skip_decoding: bool = True
+  deterministic_train: bool = False
+  use_slack: bool = True
+  cache: bool = False
+  mean_subtract: bool = False
+  standardize: bool = False
+
+  @property
+  def has_data(self):
+    """Whether this dataset is has any data associated with it."""
+    return self.name or self.data_dir or self.filenames
+
+
+@dataclass
+class ImageNetConfig(DatasetConfig):
+  """The base ImageNet dataset config."""
+  name: str = 'imagenet2012'
+  # Note: for large datasets like ImageNet, using records is faster than tfds
+  builder: str = 'records'
+  image_size: int = 224
+  batch_size: int = 128
+
+
+@dataclass
+class Cifar10Config(DatasetConfig):
+  """The base CIFAR-10 dataset config."""
+  name: str = 'cifar10'
+  image_size: int = 224
+  batch_size: int = 128
+  download: bool = True
+  cache: bool = True
+
+
+class DatasetBuilder:
+  """An object for building datasets.
+
+  Allows building various pipelines fetching examples, preprocessing, etc.
+  Maintains additional state information calculated from the dataset, i.e.,
+  training set split, batch size, and number of steps (batches).
+  """
+
+  def __init__(self, config: DatasetConfig, **overrides: Any):
+    """Initialize the builder from the config."""
+    self.config = config.replace(**overrides)
+    self.builder_info = None
+
+    if self.config.augmenter is not None:
+      logging.info('Using augmentation: %s', self.config.augmenter.name)
+      self.augmenter = self.config.augmenter.build()
+    else:
+      self.augmenter = None
+
+  @property
+  def is_training(self) -> bool:
+    """Whether this is the training set."""
+    return self.config.split == 'train'
+
+  @property
+  def batch_size(self) -> int:
+    """The batch size, multiplied by the number of replicas (if configured)."""
+    if self.config.use_per_replica_batch_size:
+      return self.global_batch_size
+    else:
+      return self.config.batch_size
+
+  @property
+  def global_batch_size(self):
+    """The global batch size across all replicas."""
+    return self.config.batch_size * self.config.num_devices
+
+  @property
+  def num_steps(self) -> int:
+    """The number of steps (batches) to exhaust this dataset."""
+    # Always divide by the global batch size to get the correct # of steps
+    return self.num_examples // self.global_batch_size
+
+  @property
+  def image_size(self) -> int:
+    """The size of each image (can be inferred from the dataset)."""
+
+    if self.config.image_size == 'infer':
+      return self.info.features['image'].shape[0]
+    else:
+      return int(self.config.image_size)
+
+  @property
+  def num_channels(self) -> int:
+    """The number of image channels (can be inferred from the dataset)."""
+    if self.config.num_channels == 'infer':
+      return self.info.features['image'].shape[-1]
+    else:
+      return int(self.config.num_channels)
+
+  @property
+  def num_examples(self) -> int:
+    """The number of examples (can be inferred from the dataset)."""
+    if self.config.num_examples == 'infer':
+      return self.info.splits[self.config.split].num_examples
+    else:
+      return int(self.config.num_examples)
+
+  @property
+  def num_classes(self) -> int:
+    """The number of classes (can be inferred from the dataset)."""
+    if self.config.num_classes == 'infer':
+      return self.info.features['label'].num_classes
+    else:
+      return int(self.config.num_classes)
+
+  @property
+  def info(self) -> tfds.core.DatasetInfo:
+    """The TFDS dataset info, if available."""
+    if self.builder_info is None:
+      self.builder_info = tfds.builder(self.config.name).info
+    return self.builder_info
+
+  def build(self, input_context: tf.distribute.InputContext = None
+           ) -> tf.data.Dataset:
+    """Construct a dataset end-to-end and return it.
+
+    Args:
+      input_context: An optional context provided by `tf.distribute` for
+        cross-replica training. This isn't necessary if using Keras
+        compile/fit.
+
+    Returns:
+      A TensorFlow dataset outputting batched images and labels.
+    """
+
+    builders = {
+        'tfds': self.load_tfds,
+        'records': self.load_records,
+        'synthetic': self.load_synthetic,
+    }
+
+    builder = builders.get(self.config.builder, None)
+
+    if builder is None:
+      raise ValueError('Unknown builder type {}'.format(self.config.builder))
+
+    dataset = builder()
+    dataset = self.pipeline(dataset, input_context)
+
+    return dataset
+
+  def load_tfds(self) -> tf.data.Dataset:
+    """Return a dataset loading files from TFDS."""
+
+    logging.info('Using TFDS to load data.')
+
+    builder = tfds.builder(self.config.name,
+                           data_dir=self.config.data_dir)
+
+    if self.config.download:
+      builder.download_and_prepare()
+
+    decoders = {}
+
+    if self.config.skip_decoding:
+      decoders['image'] = tfds.decode.SkipDecoding()
+
+    read_config = tfds.ReadConfig(
+        interleave_parallel_reads=64,
+        interleave_block_length=1)
+
+    dataset = builder.as_dataset(
+        split=self.config.split,
+        as_supervised=True,
+        shuffle_files=True,
+        decoders=decoders,
+        read_config=read_config)
+
+    return dataset
+
+  def load_records(self) -> tf.data.Dataset:
+    """Return a dataset loading files with TFRecords."""
+    logging.info('Using TFRecords to load data.')
+
+    if self.config.filenames is None:
+      if self.config.data_dir is None:
+        raise ValueError('Dataset must specify a path for the data files.')
+
+      file_pattern = os.path.join(self.config.data_dir,
+                                  '{}*'.format(self.config.split))
+      dataset = tf.data.Dataset.list_files(file_pattern, shuffle=True)
+    else:
+      dataset = tf.data.Dataset.from_tensor_slices(self.config.filenames)
+      if self.is_training:
+        # Shuffle the input files.
+        dataset.shuffle(buffer_size=self.config.file_shuffle_buffer_size)
+
+    return dataset
+
+  def load_synthetic(self) -> tf.data.Dataset:
+    """Return a dataset generating dummy synthetic data."""
+    logging.info('Generating a synthetic dataset.')
+
+    def generate_data(_):
+      image = tf.zeros([self.image_size, self.image_size, self.num_channels],
+                       dtype=self.config.dtype)
+      label = tf.zeros([1], dtype=tf.int32)
+      return image, label
+
+    dataset = tf.data.Dataset.range(1)
+    dataset = dataset.repeat()
+    dataset = dataset.map(generate_data,
+                          num_parallel_calls=tf.data.experimental.AUTOTUNE)
+    return dataset
+
+  def pipeline(self,
+               dataset: tf.data.Dataset,
+               input_context: tf.distribute.InputContext = None
+              ) -> tf.data.Dataset:
+    """Build a pipeline fetching, shuffling, and preprocessing the dataset.
+
+    Args:
+      dataset: A `tf.data.Dataset` that loads raw files.
+      input_context: An optional context provided by `tf.distribute` for
+        cross-replica training. This isn't necessary if using Keras
+        compile/fit.
+
+    Returns:
+      A TensorFlow dataset outputting batched images and labels.
+    """
+    if input_context and input_context.num_input_pipelines > 1:
+      dataset = dataset.shard(input_context.num_input_pipelines,
+                              input_context.input_pipeline_id)
+
+    if self.is_training and not self.config.cache:
+      dataset = dataset.repeat()
+
+    if self.config.builder == 'records':
+      # Read the data from disk in parallel
+      buffer_size = 8 * 1024 * 1024  # Use 8 MiB per file
+      dataset = dataset.interleave(
+          lambda name: tf.data.TFRecordDataset(name, buffer_size=buffer_size),
+          cycle_length=16,
+          num_parallel_calls=tf.data.experimental.AUTOTUNE)
+
+    dataset = dataset.prefetch(self.global_batch_size)
+
+    if self.config.cache:
+      dataset = dataset.cache()
+
+    if self.is_training:
+      dataset = dataset.shuffle(self.config.shuffle_buffer_size)
+      dataset = dataset.repeat()
+
+    # Parse, pre-process, and batch the data in parallel
+    if self.config.builder == 'records':
+      preprocess = self.parse_record
+    else:
+      preprocess = self.preprocess
+    dataset = dataset.map(preprocess,
+                          num_parallel_calls=tf.data.experimental.AUTOTUNE)
+
+    dataset = dataset.batch(self.batch_size, drop_remainder=self.is_training)
+
+    # Note: we could do image normalization here, but we defer it to the model
+    # which can perform it much faster on a GPU/TPU
+    # TODO(dankondratyuk): if we fix prefetching, we can do it here
+
+    if self.is_training and self.config.deterministic_train is not None:
+      options = tf.data.Options()
+      options.experimental_deterministic = self.config.deterministic_train
+      options.experimental_slack = self.config.use_slack
+      options.experimental_optimization.parallel_batch = True
+      options.experimental_optimization.map_fusion = True
+      options.experimental_optimization.map_vectorization.enabled = True
+      options.experimental_optimization.map_parallelization = True
+      dataset = dataset.with_options(options)
+
+    # Prefetch overlaps in-feed with training
+    # Note: autotune here is not recommended, as this can lead to memory leaks.
+    # Instead, use a constant prefetch size like the the number of devices.
+    dataset = dataset.prefetch(self.config.num_devices)
+
+    return dataset
+
+  def parse_record(self, record: tf.Tensor) -> Tuple[tf.Tensor, tf.Tensor]:
+    """Parse an ImageNet record from a serialized string Tensor."""
+    keys_to_features = {
+        'image/encoded':
+            tf.io.FixedLenFeature((), tf.string, ''),
+        'image/format':
+            tf.io.FixedLenFeature((), tf.string, 'jpeg'),
+        'image/class/label':
+            tf.io.FixedLenFeature([], tf.int64, -1),
+        'image/class/text':
+            tf.io.FixedLenFeature([], tf.string, ''),
+        'image/object/bbox/xmin':
+            tf.io.VarLenFeature(dtype=tf.float32),
+        'image/object/bbox/ymin':
+            tf.io.VarLenFeature(dtype=tf.float32),
+        'image/object/bbox/xmax':
+            tf.io.VarLenFeature(dtype=tf.float32),
+        'image/object/bbox/ymax':
+            tf.io.VarLenFeature(dtype=tf.float32),
+        'image/object/class/label':
+            tf.io.VarLenFeature(dtype=tf.int64),
+    }
+
+    parsed = tf.io.parse_single_example(record, keys_to_features)
+
+    label = tf.reshape(parsed['image/class/label'], shape=[1])
+    label = tf.cast(label, dtype=tf.int32)
+
+    # Subtract one so that labels are in [0, 1000)
+    label -= 1
+
+    image_bytes = tf.reshape(parsed['image/encoded'], shape=[])
+    image, label = self.preprocess(image_bytes, label)
+
+    return image, label
+
+  def preprocess(self, image: tf.Tensor, label: tf.Tensor
+                ) -> Tuple[tf.Tensor, tf.Tensor]:
+    """Apply image preprocessing and augmentation to the image and label."""
+    if self.is_training:
+      image = preprocessing.preprocess_for_train(
+          image,
+          image_size=self.image_size,
+          mean_subtract=self.config.mean_subtract,
+          standardize=self.config.standardize,
+          dtype=self.config.dtype,
+          augmenter=self.augmenter)
+    else:
+      image = preprocessing.preprocess_for_eval(
+          image,
+          image_size=self.image_size,
+          num_channels=self.num_channels,
+          mean_subtract=self.config.mean_subtract,
+          standardize=self.config.standardize,
+          dtype=self.config.dtype)
+
+    label = tf.cast(label, tf.int32)
+    if self.config.one_hot:
+      label = tf.one_hot(label, self.num_classes)
+      label = tf.reshape(label, [self.num_classes])
+
+    return image, label
+
+  @classmethod
+  def from_params(cls, *args, **kwargs):
+    """Construct a dataset builder from a default config and any overrides."""
+    config = DatasetConfig.from_args(*args, **kwargs)
+    return cls(config)

official/vision/image_classification/efficientnet/common_modules.py

+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Common modeling utilities."""
+from __future__ import absolute_import
+from __future__ import division
+# from __future__ import google_type_annotations
+from __future__ import print_function
+
+import numpy as np
+
+import tensorflow.compat.v1 as tf1
+import tensorflow.compat.v2 as tf
+from typing import Text, Optional
+
+from tensorflow.python.tpu import tpu_function
+
+
+@tf.keras.utils.register_keras_serializable(package='Text')
+class TpuBatchNormalization(tf.keras.layers.BatchNormalization):
+  """Cross replica batch normalization."""
+
+  def __init__(self, fused: Optional[bool] = False, **kwargs):
+    if fused in (True, None):
+      raise ValueError('TpuBatchNormalization does not support fused=True.')
+    super(TpuBatchNormalization, self).__init__(fused=fused, **kwargs)
+
+  def _cross_replica_average(self, t: tf.Tensor, num_shards_per_group: int):
+    """Calculates the average value of input tensor across TPU replicas."""
+    num_shards = tpu_function.get_tpu_context().number_of_shards
+    group_assignment = None
+    if num_shards_per_group > 1:
+      if num_shards % num_shards_per_group != 0:
+        raise ValueError(
+            'num_shards: %d mod shards_per_group: %d, should be 0' %
+            (num_shards, num_shards_per_group))
+      num_groups = num_shards // num_shards_per_group
+      group_assignment = [[
+          x for x in range(num_shards) if x // num_shards_per_group == y
+      ] for y in range(num_groups)]
+    return tf1.tpu.cross_replica_sum(t, group_assignment) / tf.cast(
+        num_shards_per_group, t.dtype)
+
+  def _moments(self, inputs: tf.Tensor, reduction_axes: int, keep_dims: int):
+    """Compute the mean and variance: it overrides the original _moments."""
+    shard_mean, shard_variance = super(TpuBatchNormalization, self)._moments(
+        inputs, reduction_axes, keep_dims=keep_dims)
+
+    num_shards = tpu_function.get_tpu_context().number_of_shards or 1
+    if num_shards <= 8:  # Skip cross_replica for 2x2 or smaller slices.
+      num_shards_per_group = 1
+    else:
+      num_shards_per_group = max(8, num_shards // 8)
+    if num_shards_per_group > 1:
+      # Compute variance using: Var[X]= E[X^2] - E[X]^2.
+      shard_square_of_mean = tf.math.square(shard_mean)
+      shard_mean_of_square = shard_variance + shard_square_of_mean
+      group_mean = self._cross_replica_average(shard_mean, num_shards_per_group)
+      group_mean_of_square = self._cross_replica_average(
+          shard_mean_of_square, num_shards_per_group)
+      group_variance = group_mean_of_square - tf.math.square(group_mean)
+      return (group_mean, group_variance)
+    else:
+      return (shard_mean, shard_variance)
+
+
+def get_batch_norm(batch_norm_type: Text) -> tf.keras.layers.BatchNormalization:
+  """A helper to create a batch normalization getter.
+
+  Args:
+    batch_norm_type: The type of batch normalization layer implementation. `tpu`
+     will use `TpuBatchNormalization`.
+
+  Returns:
+    An instance of `tf.keras.layers.BatchNormalization`.
+  """
+  if batch_norm_type == 'tpu':
+    return TpuBatchNormalization
+
+  return tf.keras.layers.BatchNormalization
+
+
+def count_params(model, trainable_only=True):
+  """Returns the count of all model parameters, or just trainable ones."""
+  if not trainable_only:
+    return model.count_params()
+  else:
+    return int(np.sum([tf.keras.backend.count_params(p)
+                       for p in model.trainable_weights]))

official/vision/image_classification/efficientnet/efficientnet_config.py

+# Lint as: python3
+# Copyright 2019 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+# ==============================================================================
+"""Configuration definitions for EfficientNet losses, learning rates, and optimizers."""
+from __future__ import absolute_import
+from __future__ import division
+from __future__ import print_function
+
+from typing import Any, Mapping
+
+import dataclasses
+
+from official.vision.image_classification.configs import base_configs
+
+
+@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: Mapping[str, Any] = dataclasses.field(default_factory=lambda: {
+      'model_name': 'efficientnet-b0',
+      'model_weights_path': '',
+      'copy_to_local': False,
+      'overrides': {
+          'batch_norm': 'default',
+          'rescale_input': True,
+          'num_classes': 1000,
+      }
+  })
+  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.)