Merge pull request #10449 from miguelCalado:vgg

PiperOrigin-RevId: 421667465

Author: tensorflower-gardener

official/legacy/image_classification/README.md

@@ -152,6 +152,20 @@ python3 classifier_trainer.py \
   --config_file=configs/examples/resnet/imagenet/tpu.yaml
 ```
 
+### VGG-16
+
+#### On GPU:
+```bash
+python3 classifier_trainer.py \
+  --mode=train_and_eval \
+  --model_type=vgg \
+  --dataset=imagenet \
+  --model_dir=$MODEL_DIR \
+  --data_dir=$DATA_DIR \
+  --config_file=configs/examples/vgg/imagenet/gpu.yaml \
+  --params_override='runtime.num_gpus=$NUM_GPUS'
+```
+
 ### EfficientNet
 **Note: EfficientNet development is a work in progress.**
 #### On GPU:

official/legacy/image_classification/classifier_trainer.py

@@ -32,6 +32,7 @@
 from official.legacy.image_classification.efficientnet import efficientnet_model
 from official.legacy.image_classification.resnet import common
 from official.legacy.image_classification.resnet import resnet_model
+from official.legacy.image_classification.vgg import vgg_model
 from official.modeling import hyperparams
 from official.modeling import performance
 from official.utils import hyperparams_flags
@@ -43,6 +44,7 @@ def get_models() -> Mapping[str, tf.keras.Model]:
   return {
       'efficientnet': efficientnet_model.EfficientNet.from_name,
       'resnet': resnet_model.resnet50,
+      'vgg': vgg_model.vgg16,
   }
 
 

official/legacy/image_classification/classifier_trainer_test.py

@@ -15,10 +15,6 @@
 # Lint as: python3
 """Unit tests for the classifier trainer models."""
 
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
-
 import functools
 import json
 
@@ -53,6 +49,7 @@ def distribution_strategy_combinations() -> Iterable[Tuple[Any, ...]]:
       model=[
           'efficientnet',
           'resnet',
+          'vgg',
       ],
       dataset=[
           'imagenet',
@@ -149,6 +146,7 @@ def test_end_to_end_train_and_eval(self, distribution, model, dataset):
           model=[
               'efficientnet',
               'resnet',
+              'vgg',
           ],
           dataset='imagenet',
           dtype='float16',
@@ -193,6 +191,7 @@ def test_gpu_train(self, distribution, model, dataset, dtype):
           model=[
               'efficientnet',
               'resnet',
+              'vgg',
           ],
           dataset='imagenet',
           dtype='bfloat16',

official/legacy/image_classification/configs/configs.py

@@ -14,16 +14,14 @@
 
 # Lint as: python3
 """Configuration utils for image classification experiments."""
-from __future__ import absolute_import
-from __future__ import division
-from __future__ import print_function
 
 import dataclasses
 
 from official.legacy.image_classification import dataset_factory
 from official.legacy.image_classification.configs import base_configs
 from official.legacy.image_classification.efficientnet import efficientnet_config
 from official.legacy.image_classification.resnet import resnet_config
+from official.legacy.image_classification.vgg import vgg_config
 
 
 @dataclasses.dataclass
@@ -92,12 +90,38 @@ class ResNetImagenetConfig(base_configs.ExperimentConfig):
   model: base_configs.ModelConfig = resnet_config.ResNetModelConfig()
 
 
+@dataclasses.dataclass
+class VGGImagenetConfig(base_configs.ExperimentConfig):
+  """Base configuration to train vgg-16 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)
+  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'],
+      time_history=base_configs.TimeHistoryConfig(log_steps=100),
+      tensorboard=base_configs.TensorBoardConfig(
+          track_lr=True, write_model_weights=False),
+      set_epoch_loop=False)
+  evaluation: base_configs.EvalConfig = base_configs.EvalConfig(
+      epochs_between_evals=1, steps=None)
+  model: base_configs.ModelConfig = vgg_config.VGGModelConfig()
+
+
 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(),
+          'vgg': VGGImagenetConfig(),
       }
   }
   try:

official/legacy/image_classification/configs/examples/vgg16/imagenet/gpu.yaml

@@ -0,0 +1,46 @@
+# Training configuration for VGG-16 trained on ImageNet on GPUs.
+# Reaches > 72.8% within 90 epochs.
+# Note: This configuration uses a scaled per-replica batch size based on the number of devices.
+runtime:
+  distribution_strategy: 'mirrored'
+  num_gpus: 1
+  batchnorm_spatial_persistent: true
+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:
+  name: 'vgg'
+  optimizer:
+    name: 'momentum'
+    momentum: 0.9
+    epsilon: 0.001
+  loss:
+    label_smoothing: 0.0
+train:
+  resume_checkpoint: true
+  epochs: 90
+evaluation:
+  epochs_between_evals: 1

official/legacy/image_classification/vgg/__init__.py

@@ -0,0 +1,15 @@
+# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+

official/legacy/image_classification/vgg/vgg_config.py

@@ -0,0 +1,45 @@
+# Copyright 2021 The TensorFlow Authors. All Rights Reserved.
+#
+# Licensed under the Apache License, Version 2.0 (the "License");
+# you may not use this file except in compliance with the License.
+# You may obtain a copy of the License at
+#
+#     http://www.apache.org/licenses/LICENSE-2.0
+#
+# Unless required by applicable law or agreed to in writing, software
+# distributed under the License is distributed on an "AS IS" BASIS,
+# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+# See the License for the specific language governing permissions and
+# limitations under the License.
+
+# Lint as: python3
+"""Configuration definitions for VGG losses, learning rates, and optimizers."""
+
+import dataclasses
+from official.legacy.image_classification.configs import base_configs
+from official.modeling.hyperparams import base_config
+
+
+@dataclasses.dataclass
+class VGGModelConfig(base_configs.ModelConfig):
+  """Configuration for the VGG model."""
+  name: str = 'VGG'
+  num_classes: int = 1000
+  model_params: base_config.Config = dataclasses.field(default_factory=lambda: {   # pylint:disable=g-long-lambda
+      'num_classes': 1000,
+      'batch_size': None,
+      'use_l2_regularizer': True
+  })
+  loss: base_configs.LossConfig = base_configs.LossConfig(
+      name='sparse_categorical_crossentropy')
+  optimizer: base_configs.OptimizerConfig = base_configs.OptimizerConfig(
+      name='momentum', epsilon=0.001, momentum=0.9, moving_average_decay=None)
+  learning_rate: base_configs.LearningRateConfig = (
+      base_configs.LearningRateConfig(
+          name='stepwise',
+          initial_lr=0.01,
+          examples_per_epoch=1281167,
+          boundaries=[30, 60],
+          warmup_epochs=0,
+          scale_by_batch_size=1. / 256.,
+          multipliers=[0.01 / 256, 0.001 / 256, 0.0001 / 256]))