def run(flags_obj): """Run ResNet Cifar-10 training and eval loop using native Keras APIs. Args: flags_obj: An object containing parsed flag values. Raises: ValueError: If fp16 is passed as it is not currently supported. Returns: Dictionary of training and eval stats. """ if flags_obj.enable_eager: tf.enable_eager_execution() dtype = flags_core.get_tf_dtype(flags_obj) if dtype == 'fp16': raise ValueError( 'dtype fp16 is not supported in Keras. Use the default ' 'value(fp32).') data_format = flags_obj.data_format if data_format is None: data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last') tf.keras.backend.set_image_data_format(data_format) if flags_obj.use_synthetic_data: input_fn = keras_common.get_synth_input_fn( height=cifar_main.HEIGHT, width=cifar_main.WIDTH, num_channels=cifar_main.NUM_CHANNELS, num_classes=cifar_main.NUM_CLASSES, dtype=flags_core.get_tf_dtype(flags_obj)) else: input_fn = cifar_main.input_fn train_input_dataset = input_fn(is_training=True, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) eval_input_dataset = input_fn(is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) strategy = distribution_utils.get_distribution_strategy( num_gpus=flags_obj.num_gpus, turn_off_distribution_strategy=flags_obj.turn_off_distribution_strategy ) strategy_scope = keras_common.get_strategy_scope(strategy) with strategy_scope: optimizer = keras_common.get_optimizer() model = resnet_cifar_model.resnet56(classes=cifar_main.NUM_CLASSES) model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['categorical_accuracy']) time_callback, tensorboard_callback, lr_callback = keras_common.get_callbacks( learning_rate_schedule, cifar_main.NUM_IMAGES['train']) train_steps = cifar_main.NUM_IMAGES['train'] // flags_obj.batch_size train_epochs = flags_obj.train_epochs if flags_obj.train_steps: train_steps = min(flags_obj.train_steps, train_steps) train_epochs = 1 num_eval_steps = (cifar_main.NUM_IMAGES['validation'] // flags_obj.batch_size) validation_data = eval_input_dataset if flags_obj.skip_eval: tf.keras.backend.set_learning_phase(1) num_eval_steps = None validation_data = None history = model.fit( train_input_dataset, epochs=train_epochs, steps_per_epoch=train_steps, callbacks=[time_callback, lr_callback, tensorboard_callback], validation_steps=num_eval_steps, validation_data=validation_data, verbose=2) eval_output = None if not flags_obj.skip_eval: eval_output = model.evaluate(eval_input_dataset, steps=num_eval_steps, verbose=1) stats = keras_common.build_stats(history, eval_output, time_callback) return stats
def run(flags_obj): """Run ResNet Cifar-10 training and eval loop using native Keras APIs. Args: flags_obj: An object containing parsed flag values. Raises: ValueError: If fp16 is passed as it is not currently supported. Returns: Dictionary of training and eval stats. """ keras_utils.set_session_config(enable_eager=flags_obj.enable_eager, enable_xla=flags_obj.enable_xla) dtype = flags_core.get_tf_dtype(flags_obj) if dtype == 'fp16': raise ValueError( 'dtype fp16 is not supported in Keras. Use the default ' 'value(fp32).') data_format = flags_obj.data_format if data_format is None: data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last') tf.keras.backend.set_image_data_format(data_format) strategy = distribution_utils.get_distribution_strategy( distribution_strategy=flags_obj.distribution_strategy, num_gpus=flags_obj.num_gpus, num_workers=distribution_utils.configure_cluster(), all_reduce_alg=flags_obj.all_reduce_alg, num_packs=flags_obj.num_packs) strategy_scope = distribution_utils.get_strategy_scope(strategy) if flags_obj.use_synthetic_data: distribution_utils.set_up_synthetic_data() input_fn = keras_common.get_synth_input_fn( height=cifar_main.HEIGHT, width=cifar_main.WIDTH, num_channels=cifar_main.NUM_CHANNELS, num_classes=cifar_main.NUM_CLASSES, dtype=flags_core.get_tf_dtype(flags_obj)) else: distribution_utils.undo_set_up_synthetic_data() input_fn = cifar_main.input_fn train_input_dataset = input_fn( is_training=True, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=cifar_main.parse_record, datasets_num_private_threads=flags_obj.datasets_num_private_threads, dtype=dtype) eval_input_dataset = None if not flags_obj.skip_eval: eval_input_dataset = input_fn(is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=cifar_main.parse_record) with strategy_scope: optimizer = keras_common.get_optimizer() model = resnet_cifar_model.resnet56(classes=cifar_main.NUM_CLASSES) model.compile(loss='sparse_categorical_crossentropy', optimizer=optimizer, metrics=(['sparse_categorical_accuracy'] if flags_obj.report_accuracy_metrics else None), run_eagerly=flags_obj.run_eagerly) callbacks = keras_common.get_callbacks(learning_rate_schedule, cifar_main.NUM_IMAGES['train']) train_steps = cifar_main.NUM_IMAGES['train'] // flags_obj.batch_size train_epochs = flags_obj.train_epochs if flags_obj.train_steps: train_steps = min(flags_obj.train_steps, train_steps) train_epochs = 1 num_eval_steps = (cifar_main.NUM_IMAGES['validation'] // flags_obj.batch_size) validation_data = eval_input_dataset if flags_obj.skip_eval: if flags_obj.set_learning_phase_to_train: # TODO(haoyuzhang): Understand slowdown of setting learning phase when # not using distribution strategy. tf.keras.backend.set_learning_phase(1) num_eval_steps = None validation_data = None if not strategy and flags_obj.explicit_gpu_placement: # TODO(b/135607227): Add device scope automatically in Keras training loop # when not using distribition strategy. no_dist_strat_device = tf.device('/device:GPU:0') no_dist_strat_device.__enter__() history = model.fit(train_input_dataset, epochs=train_epochs, steps_per_epoch=train_steps, callbacks=callbacks, validation_steps=num_eval_steps, validation_data=validation_data, validation_freq=flags_obj.epochs_between_evals, verbose=2) eval_output = None if not flags_obj.skip_eval: eval_output = model.evaluate(eval_input_dataset, steps=num_eval_steps, verbose=2) if not strategy and flags_obj.explicit_gpu_placement: no_dist_strat_device.__exit__() stats = keras_common.build_stats(history, eval_output, callbacks) return stats
def run(flags_obj): """Run ResNet Cifar-10 training and eval loop using native Keras APIs. Args: flags_obj: An object containing parsed flag values. Raises: ValueError: If fp16 is passed as it is not currently supported. Returns: Dictionary of training and eval stats. """ keras_utils.set_session_config(enable_eager=flags_obj.enable_eager, enable_xla=flags_obj.enable_xla) # Execute flag override logic for better model performance if flags_obj.tf_gpu_thread_mode: keras_common.set_gpu_thread_mode_and_count(flags_obj) keras_common.set_cudnn_batchnorm_mode() dtype = flags_core.get_tf_dtype(flags_obj) if dtype == 'fp16': raise ValueError( 'dtype fp16 is not supported in Keras. Use the default ' 'value(fp32).') data_format = flags_obj.data_format if data_format is None: data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last') tf.keras.backend.set_image_data_format(data_format) strategy = distribution_utils.get_distribution_strategy( distribution_strategy=flags_obj.distribution_strategy, num_gpus=flags_obj.num_gpus, num_workers=distribution_utils.configure_cluster(), all_reduce_alg=flags_obj.all_reduce_alg, num_packs=flags_obj.num_packs) if strategy: # flags_obj.enable_get_next_as_optional controls whether enabling # get_next_as_optional behavior in DistributedIterator. If true, last # partial batch can be supported. strategy.extended.experimental_enable_get_next_as_optional = ( flags_obj.enable_get_next_as_optional) strategy_scope = distribution_utils.get_strategy_scope(strategy) if flags_obj.use_synthetic_data: distribution_utils.set_up_synthetic_data() input_fn = keras_common.get_synth_input_fn( height=cifar_preprocessing.HEIGHT, width=cifar_preprocessing.WIDTH, num_channels=cifar_preprocessing.NUM_CHANNELS, num_classes=cifar_preprocessing.NUM_CLASSES, dtype=flags_core.get_tf_dtype(flags_obj), drop_remainder=True) else: distribution_utils.undo_set_up_synthetic_data() input_fn = cifar_preprocessing.input_fn train_input_dataset = input_fn( is_training=True, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=cifar_preprocessing.parse_record, datasets_num_private_threads=flags_obj.datasets_num_private_threads, dtype=dtype, # Setting drop_remainder to avoid the partial batch logic in normalization # layer, which triggers tf.where and leads to extra memory copy of input # sizes between host and GPU. drop_remainder=(not flags_obj.enable_get_next_as_optional)) eval_input_dataset = None if not flags_obj.skip_eval: eval_input_dataset = input_fn( is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=cifar_preprocessing.parse_record) with strategy_scope: optimizer = keras_common.get_optimizer() model = resnet_cifar_model.resnet56( classes=cifar_preprocessing.NUM_CLASSES) # TODO(b/138957587): Remove when force_v2_in_keras_compile is on longer # a valid arg for this model. Also remove as a valid flag. if flags_obj.force_v2_in_keras_compile is not None: model.compile( loss='categorical_crossentropy', optimizer=optimizer, metrics=(['categorical_accuracy'] if flags_obj.report_accuracy_metrics else None), run_eagerly=flags_obj.run_eagerly, experimental_run_tf_function=flags_obj. force_v2_in_keras_compile) else: model.compile( loss='categorical_crossentropy', optimizer=optimizer, metrics=(['categorical_accuracy'] if flags_obj.report_accuracy_metrics else None), run_eagerly=flags_obj.run_eagerly) callbacks = keras_common.get_callbacks( learning_rate_schedule, cifar_preprocessing.NUM_IMAGES['train']) train_steps = cifar_preprocessing.NUM_IMAGES[ 'train'] // flags_obj.batch_size train_epochs = flags_obj.train_epochs if flags_obj.train_steps: train_steps = min(flags_obj.train_steps, train_steps) train_epochs = 1 num_eval_steps = (cifar_preprocessing.NUM_IMAGES['validation'] // flags_obj.batch_size) validation_data = eval_input_dataset if flags_obj.skip_eval: if flags_obj.set_learning_phase_to_train: # TODO(haoyuzhang): Understand slowdown of setting learning phase when # not using distribution strategy. tf.keras.backend.set_learning_phase(1) num_eval_steps = None validation_data = None if not strategy and flags_obj.explicit_gpu_placement: # TODO(b/135607227): Add device scope automatically in Keras training loop # when not using distribition strategy. no_dist_strat_device = tf.device('/device:GPU:0') no_dist_strat_device.__enter__() history = model.fit(train_input_dataset, epochs=train_epochs, steps_per_epoch=train_steps, callbacks=callbacks, validation_steps=num_eval_steps, validation_data=validation_data, validation_freq=flags_obj.epochs_between_evals, verbose=2) eval_output = None if not flags_obj.skip_eval: eval_output = model.evaluate(eval_input_dataset, steps=num_eval_steps, verbose=2) if not strategy and flags_obj.explicit_gpu_placement: no_dist_strat_device.__exit__() stats = keras_common.build_stats(history, eval_output, callbacks) return stats
def run(flags_obj): """Run ResNet Cifar-10 training and eval loop using native Keras APIs. Args: flags_obj: An object containing parsed flag values. Raises: ValueError: If fp16 is passed as it is not currently supported. Returns: Dictionary of training and eval stats. """ if flags_obj.enable_eager: tf.enable_eager_execution() dtype = flags_core.get_tf_dtype(flags_obj) if dtype == 'fp16': raise ValueError('dtype fp16 is not supported in Keras. Use the default ' 'value(fp32).') data_format = flags_obj.data_format if data_format is None: data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last') tf.keras.backend.set_image_data_format(data_format) if flags_obj.use_synthetic_data: input_fn = keras_common.get_synth_input_fn( height=cifar_main.HEIGHT, width=cifar_main.WIDTH, num_channels=cifar_main.NUM_CHANNELS, num_classes=cifar_main.NUM_CLASSES, dtype=flags_core.get_tf_dtype(flags_obj)) else: input_fn = cifar_main.input_fn train_input_dataset = input_fn( is_training=True, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) eval_input_dataset = input_fn( is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) strategy = distribution_utils.get_distribution_strategy( num_gpus=flags_obj.num_gpus, turn_off_distribution_strategy=flags_obj.turn_off_distribution_strategy) strategy_scope = keras_common.get_strategy_scope(strategy) with strategy_scope: optimizer = keras_common.get_optimizer() model = resnet_cifar_model.resnet56(classes=cifar_main.NUM_CLASSES) model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=['categorical_accuracy']) time_callback, tensorboard_callback, lr_callback = keras_common.get_callbacks( learning_rate_schedule, cifar_main.NUM_IMAGES['train']) train_steps = cifar_main.NUM_IMAGES['train'] // flags_obj.batch_size train_epochs = flags_obj.train_epochs if flags_obj.train_steps: train_steps = min(flags_obj.train_steps, train_steps) train_epochs = 1 num_eval_steps = (cifar_main.NUM_IMAGES['validation'] // flags_obj.batch_size) validation_data = eval_input_dataset if flags_obj.skip_eval: tf.keras.backend.set_learning_phase(1) num_eval_steps = None validation_data = None history = model.fit(train_input_dataset, epochs=train_epochs, steps_per_epoch=train_steps, callbacks=[ time_callback, lr_callback, tensorboard_callback ], validation_steps=num_eval_steps, validation_data=validation_data, verbose=1) eval_output = None if not flags_obj.skip_eval: eval_output = model.evaluate(eval_input_dataset, steps=num_eval_steps, verbose=1) stats = keras_common.build_stats(history, eval_output, time_callback) return stats
def run(flags_obj): """Run ResNet Cifar-10 training and eval loop using native Keras APIs. Args: flags_obj: An object containing parsed flag values. Raises: ValueError: If fp16 is passed as it is not currently supported. Returns: Dictionary of training and eval stats. """ # TODO(tobyboyd): Remove eager flag when tf 1.0 testing ends. # Eager is default in tf 2.0 and should not be toggled if keras_common.is_v2_0(): keras_common.set_config_v2() else: config = keras_common.get_config_proto_v1() if flags_obj.enable_eager: tf.compat.v1.enable_eager_execution(config=config) else: sess = tf.Session(config=config) tf.keras.backend.set_session(sess) dtype = flags_core.get_tf_dtype(flags_obj) if dtype == 'fp16': raise ValueError( 'dtype fp16 is not supported in Keras. Use the default ' 'value(fp32).') data_format = flags_obj.data_format if data_format is None: data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last') tf.keras.backend.set_image_data_format(data_format) strategy = distribution_utils.get_distribution_strategy( distribution_strategy=flags_obj.distribution_strategy, num_gpus=flags_obj.num_gpus) strategy_scope = distribution_utils.get_strategy_scope(strategy) if flags_obj.use_synthetic_data: distribution_utils.set_up_synthetic_data() input_fn = keras_common.get_synth_input_fn( height=cifar_main.HEIGHT, width=cifar_main.WIDTH, num_channels=cifar_main.NUM_CHANNELS, num_classes=cifar_main.NUM_CLASSES, dtype=flags_core.get_tf_dtype(flags_obj)) else: distribution_utils.undo_set_up_synthetic_data() input_fn = cifar_main.input_fn train_input_dataset = input_fn(is_training=True, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) eval_input_dataset = input_fn(is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) with strategy_scope: optimizer = keras_common.get_optimizer() model = resnet_cifar_model.resnet56(classes=cifar_main.NUM_CLASSES) model.compile(loss='categorical_crossentropy', optimizer=optimizer, run_eagerly=flags_obj.run_eagerly, metrics=['categorical_accuracy']) callbacks = keras_common.get_callbacks(learning_rate_schedule, cifar_main.NUM_IMAGES['train']) train_steps = cifar_main.NUM_IMAGES['train'] // flags_obj.batch_size train_epochs = flags_obj.train_epochs if flags_obj.train_steps: train_steps = min(flags_obj.train_steps, train_steps) train_epochs = 1 num_eval_steps = (cifar_main.NUM_IMAGES['validation'] // flags_obj.batch_size) validation_data = eval_input_dataset if flags_obj.skip_eval: tf.keras.backend.set_learning_phase(1) num_eval_steps = None validation_data = None history = model.fit(train_input_dataset, epochs=train_epochs, steps_per_epoch=train_steps, callbacks=callbacks, validation_steps=num_eval_steps, validation_data=validation_data, validation_freq=flags_obj.epochs_between_evals, verbose=2) eval_output = None if not flags_obj.skip_eval: eval_output = model.evaluate(eval_input_dataset, steps=num_eval_steps, verbose=2) stats = keras_common.build_stats(history, eval_output, callbacks) return stats
def run(flags_obj): """Run ResNet Cifar-10 training and eval loop using native Keras APIs. Args: flags_obj: An object containing parsed flag values. Raises: ValueError: If fp16 is passed as it is not currently supported. Returns: Dictionary of training and eval stats. """ if flags_obj.enable_eager: tf.enable_eager_execution() dtype = flags_core.get_tf_dtype(flags_obj) if dtype == 'fp16': raise ValueError( 'dtype fp16 is not supported in Keras. Use the default ' 'value(fp32).') data_format = flags_obj.data_format if data_format is None: data_format = ('channels_first' if tf.test.is_built_with_cuda() else 'channels_last') tf.keras.backend.set_image_data_format(data_format) if flags_obj.use_synthetic_data: input_fn = keras_common.get_synth_input_fn( height=jr_main.HEIGHT, width=jr_main.WIDTH, num_channels=jr_main.NUM_CHANNELS, num_classes=jr_main.NUM_CLASSES, dtype=flags_core.get_tf_dtype(flags_obj)) else: input_fn = jr_main.input_fn train_input_dataset = input_fn(mode='train', is_training=True, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) eval_input_dataset = input_fn(mode='val', is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) test_input_dataset = input_fn(mode='test', is_training=False, data_dir=flags_obj.data_dir, batch_size=flags_obj.batch_size, num_epochs=flags_obj.train_epochs, parse_record_fn=parse_record_keras) strategy = distribution_utils.get_distribution_strategy( num_gpus=flags_obj.num_gpus, turn_off_distribution_strategy=flags_obj.turn_off_distribution_strategy ) strategy_scope = keras_common.get_strategy_scope(strategy) with strategy_scope: #optimizer = keras_common.get_optimizer() optimizer = tf.keras.optimizers.SGD(lr=0.1, momentum=0.9) if flags_obj.resnet_size == '56': print('USING RESNET56') model = resnet_cifar_model.resnet56(classes=jr_main.NUM_CLASSES, size=jr_main.HEIGHT) model_to_save = resnet_cifar_model.resnet56( classes=jr_main.NUM_CLASSES) elif flags_obj.resnet_size == '18': print('USING RESNET18') model = ResNet18(input_shape=(jr_main.HEIGHT, jr_main.WIDTH, 3), classes=jr_main.NUM_CLASSES) model_to_save = ResNet18(input_shape=(jr_main.HEIGHT, jr_main.WIDTH, 3), classes=jr_main.NUM_CLASSES) elif flags_obj.resnet_size == '50': print('USING RESNET50') model = ResNet50(input_shape=(jr_main.HEIGHT, jr_main.WIDTH, 3), classes=jr_main.NUM_CLASSES) model_to_save = ResNet50(input_shape=(jr_main.HEIGHT, jr_main.WIDTH, 3), classes=jr_main.NUM_CLASSES) # model = model_zoo.ResNet50(include_top=False, weights='imagenet', input_tensor=None, input_shape=(224,224, 3), pooling='avg')) # model. # model_to_save = model_zoo.ResNet50(include_top=True, weights='imagenet', input_tensor=None, input_shape=(224,224, 3), classes=3) else: print('Need to specifcy resnet18 or 50!') sys.exit(0) #model = model_zoo.ResNet50(weights=None, classes=3, input_tensor=None, input_shape=(224,224, 3)) model.compile(loss='categorical_crossentropy', optimizer=optimizer, metrics=[ 'categorical_accuracy', single_class_accuracy_alt(0), single_class_accuracy_alt(1), single_class_accuracy_alt(2), mean_per_class_accuracy_alt ]) time_callback, tensorboard_callback, lr_callback = keras_common.get_callbacks( learning_rate_schedule, jr_main.NUM_IMAGES['train']) checkpoint_path = os.path.join(flags_obj.model_dir, 'checkpoint.h5') cp_callback = tf.keras.callbacks.ModelCheckpoint( checkpoint_path, monitor='val_mean_per_class_accuracy_alt', verbose=1, save_best_only=True, save_weights_only=True, period=1) train_steps = jr_main.NUM_IMAGES['train'] // flags_obj.batch_size train_epochs = flags_obj.train_epochs if flags_obj.train_steps: train_steps = min(flags_obj.train_steps, train_steps) train_epochs = 1 num_eval_steps = (jr_main.NUM_IMAGES['validation'] // flags_obj.batch_size) num_test_steps = (jr_main.NUM_IMAGES['test'] // flags_obj.batch_size) validation_data = eval_input_dataset if flags_obj.skip_eval: tf.keras.backend.set_learning_phase(1) num_eval_steps = None validation_data = None #print('saving model') history = model.fit(train_input_dataset, epochs=train_epochs, steps_per_epoch=train_steps, callbacks=[ time_callback, lr_callback, tensorboard_callback, cp_callback, ], validation_steps=num_eval_steps, validation_data=validation_data, verbose=2, workers=4) eval_output = None if not flags_obj.skip_eval: print('TESTING') eval_output = model.evaluate(test_input_dataset, steps=num_test_steps, verbose=1) stats = keras_common.build_stats(history, eval_output, time_callback) print('loading weights from best checkpoint') model_to_save.load_weights(checkpoint_path) print('saving final model') model_to_save.save('{}_final.h5'.format(checkpoint_path.split('/')[-2])) return stats