def evaluate_distributed(model,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None,
callbacks=None):
"""Evaluate loop for Distribution Strategies."""
distributed_training_utils.validate_inputs(x, y,
model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps, batch_size)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
dataset = model._distribution_standardize_user_data(
x, y, sample_weight=sample_weight, batch_size=batch_size)
if distributed_training_utils.is_tpu_strategy(
model._distribution_strategy):
return experimental_tpu_test_loop(model,
dataset,
verbose=verbose,
steps=steps,
callbacks=callbacks)
else:
return training_arrays.test_loop(model,
inputs=dataset,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def predict_distributed(model,
x=None,
batch_size=None,
verbose=0,
steps=None,
callbacks=None):
"""Predict loop for Distribution Strategies."""
distributed_training_utils.validate_inputs(
x, None, model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps, batch_size)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
iterator = model._distribution_standardize_user_data(
x,
batch_size=batch_size,
check_steps=True,
steps_name='steps',
steps=steps)
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
# TODO(fchollet): why aren't callbacks supported here?
return experimental_tpu_predict_loop(
model, iterator, verbose=verbose, steps=steps)
else:
return training_arrays.predict_loop(
model,
iterator,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def predict_distributed(model,
x=None,
batch_size=None,
verbose=0,
steps=None,
callbacks=None):
"""Predict loop for Distribution Strategies."""
distributed_training_utils.validate_inputs(
x, None, model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps, batch_size)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
dataset = model._distribution_standardize_user_data(
x,
batch_size=batch_size,
check_steps=True,
steps_name='steps',
steps=steps)
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
# TODO(fchollet): why aren't callbacks supported here?
return experimental_tpu_predict_loop(
model, dataset, verbose=verbose, steps=steps)
else:
return training_arrays.predict_loop(
model,
dataset,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def predict_distributed(model,
x=None,
batch_size=None,
verbose=0,
steps=None,
callbacks=None):
"""Predict loop for Distribution Strategies."""
distributed_training_utils.validate_inputs(
x, None, model._distribution_strategy, allow_partial_batch=True)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps,
batch_size, mode=ModeKeys.PREDICT)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
dataset = model._distribution_standardize_user_data(
x,
batch_size=batch_size,
repeat=False,
allow_partial_batch=True)
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
return experimental_tpu_predict_loop(
model, dataset, verbose=verbose, steps=steps, callbacks=callbacks)
else:
return training_arrays.predict_loop(
model,
dataset,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def predict_distributed(model,
x=None,
batch_size=None,
verbose=0,
steps=None,
callbacks=None):
"""Predict loop for Distribution Strategies."""
distributed_training_utils.validate_inputs(
x, None, model._distribution_strategy, allow_partial_batch=True)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps,
batch_size, mode=ModeKeys.PREDICT)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
dataset = model._distribution_standardize_user_data(
x,
batch_size=batch_size,
repeat=False,
allow_partial_batch=True)
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
return experimental_tpu_predict_loop(
model, dataset, verbose=verbose, steps=steps, callbacks=callbacks)
else:
return training_arrays.predict_loop(
model,
dataset,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def evaluate_distributed(model,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None,
callbacks=None):
"""Evaluate loop for Distribution Strategies."""
distributed_training_utils.validate_inputs(x, y, model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps, batch_size)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
dataset = model._distribution_standardize_user_data(
x, y,
sample_weight=sample_weight,
batch_size=batch_size)
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
return experimental_tpu_test_loop(
model, dataset, verbose=verbose, steps=steps, callbacks=callbacks)
else:
return training_arrays.test_loop(
model,
inputs=dataset,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def test_on_dataset_with_unknown_cardinality_without_steps(
self, distribution):
with self.cached_session():
with distribution.scope():
model = get_model()
optimizer = gradient_descent.GradientDescentOptimizer(0.001)
loss = 'mse'
metrics = ['mae', keras.metrics.CategoricalAccuracy()]
model.compile(optimizer, loss, metrics=metrics)
inputs = np.zeros((1000, 3), dtype=np.float32)
targets = np.zeros((1000, 4), dtype=np.float32)
# steps/steps_per_epoch are calculated when using numpy arrays as
# input data.
fit_with_numpy = model.fit(inputs, targets, epochs=1,
batch_size=10).history
fit_with_numpy_multiple_epochs = model.fit(
inputs, targets, epochs=2, batch_size=10).history
eval_with_numpy = model.evaluate(inputs, targets, batch_size=10)
predict_with_numpy = model.predict(inputs, batch_size=10)
dataset = convert_numpy_to_dataset_with_unknown_cardinality(
inputs, targets)
predict_dataset = convert_numpy_to_dataset_with_unknown_cardinality(
inputs)
self.assertEqual(keras.backend.get_value(cardinality.cardinality(
dataset)), cardinality.UNKNOWN)
self.assertEqual(keras.backend.get_value(cardinality.cardinality(
predict_dataset)), cardinality.UNKNOWN)
eval_with_ds = model.evaluate(dataset)
predict_with_ds = model.predict(predict_dataset)
self.assertAllClose(
eval_with_numpy, eval_with_ds, atol=1e-4, rtol=1e-4)
self.assertAllClose(
predict_with_numpy, predict_with_ds, atol=1e-4, rtol=1e-4)
if (distributed_training_utils.is_tpu_strategy(distribution) and
distribution.extended.steps_per_run != 1):
with self.assertRaisesRegexp(ValueError, '`steps_per_epoch` '
'should be specified'):
fit_with_ds = model.fit(dataset, epochs=1)
else:
fit_with_ds = model.fit(dataset,
epochs=1).history
fit_with_ds_multiple_epochs = model.fit(dataset,
epochs=2).history
self.assertAllClose(
fit_with_numpy, fit_with_ds, atol=1e-4, rtol=1e-4)
self.assertAllClose(
fit_with_numpy_multiple_epochs,
fit_with_ds_multiple_epochs, atol=1e-4, rtol=1e-4)
def evaluate_distributed(model,
x=None,
y=None,
batch_size=None,
verbose=1,
sample_weight=None,
steps=None,
callbacks=None):
"""Evaluate loop for Distribution Strategies."""
# TODO(b/122314600): Remove the scope validate.
distributed_training_utils.validate_not_in_strategy_scope()
distributed_training_utils.validate_inputs(x, y,
model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps, batch_size = distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps, batch_size)
batch_size = model._validate_or_infer_batch_size(batch_size, steps, x)
iterator = model._distribution_standardize_user_data(
x,
y,
sample_weight=sample_weight,
batch_size=batch_size,
check_steps=True,
steps_name='steps',
steps=steps)
if distributed_training_utils.is_tpu_strategy(
model._distribution_strategy):
# TODO(fchollet): why aren't callbacks supported here?
return experimental_tpu_test_loop(model,
iterator=iterator,
verbose=verbose,
steps=steps)
else:
return training_arrays.test_loop(model,
inputs=iterator,
batch_size=batch_size,
verbose=verbose,
steps=steps,
callbacks=callbacks)
def fit_distributed(model,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
validation_freq=1):
"""Fit loop for Distribution Strategies."""
distributed_training_utils.validate_callbacks(callbacks, model.optimizer)
distributed_training_utils.validate_inputs(
x, y, model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
# Until support for partial batch is implemented across all
# functions and distribution strategy, we pass `mode` to selectively
# relax the costraint to consume all the training samples.
steps_per_epoch, batch_size = (
distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps_per_epoch,
batch_size, mode=ModeKeys.TRAIN))
batch_size = model._validate_or_infer_batch_size(
batch_size, steps_per_epoch, x)
steps_name = 'steps_per_epoch'
if isinstance(x, dataset_ops.DatasetV2):
steps_per_epoch = training_utils.infer_steps_for_dataset(
x, steps_per_epoch, steps_name=steps_name)
dataset = model._distribution_standardize_user_data(
x, y,
sample_weight=sample_weight,
class_weight=class_weight,
batch_size=batch_size,
check_steps=True,
steps_name=steps_name,
steps=steps_per_epoch,
validation_split=validation_split,
shuffle=shuffle)
val_dataset = None
if validation_data:
val_x, val_y, val_sample_weights = model._unpack_validation_data(
validation_data)
distributed_training_utils.validate_inputs(
val_x, val_y, model._distribution_strategy)
first_valx_value = nest.flatten(val_x)[0]
if isinstance(first_valx_value, np.ndarray):
validation_steps, _ = distributed_training_utils.get_input_params(
model._distribution_strategy, first_valx_value, validation_steps,
batch_size)
steps_name = 'validation_steps'
if isinstance(val_x, dataset_ops.DatasetV2):
validation_steps = training_utils.infer_steps_for_dataset(
val_x, validation_steps, steps_name=steps_name)
val_dataset = model._distribution_standardize_user_data(
val_x, val_y,
sample_weight=val_sample_weights,
class_weight=None,
batch_size=batch_size,
check_steps=True,
steps_name=steps_name,
steps=validation_steps,
validation_split=validation_split,
shuffle=shuffle)
elif validation_split:
raise ValueError('validation_split argument is not supported with '
'distribution strategies.')
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
return experimental_tpu_fit_loop(
model,
dataset,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_dataset=val_dataset,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=1)
else:
return training_arrays.fit_loop(
model,
dataset,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_inputs=val_dataset,
shuffle=shuffle,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=validation_freq)
def fit_distributed(model,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
validation_freq=1):
"""Fit loop for Distribution Strategies."""
distributed_training_utils.validate_callbacks(callbacks, model.optimizer)
distributed_training_utils.validate_inputs(x, y,
model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
# Until support for partial batch is implemented across all
# functions and distribution strategy, we pass `mode` to selectively
# relax the costraint to consume all the training samples.
steps_per_epoch, batch_size = (
distributed_training_utils.get_input_params(
model._distribution_strategy,
first_x_value,
steps_per_epoch,
batch_size,
mode=ModeKeys.TRAIN))
batch_size = model._validate_or_infer_batch_size(batch_size,
steps_per_epoch, x)
dataset = model._distribution_standardize_user_data(
x,
y,
sample_weight=sample_weight,
class_weight=class_weight,
batch_size=batch_size,
check_steps=True,
steps_name='steps_per_epoch',
steps=steps_per_epoch,
validation_split=validation_split,
shuffle=shuffle)
val_dataset = None
if validation_data:
val_x, val_y, val_sample_weights = model._unpack_validation_data(
validation_data)
distributed_training_utils.validate_inputs(
val_x, val_y, model._distribution_strategy)
first_valx_value = nest.flatten(val_x)[0]
if isinstance(first_valx_value, np.ndarray):
validation_steps, _ = distributed_training_utils.get_input_params(
model._distribution_strategy, first_valx_value,
validation_steps, batch_size)
val_dataset = model._distribution_standardize_user_data(
val_x,
val_y,
sample_weight=val_sample_weights,
class_weight=None,
batch_size=batch_size,
check_steps=True,
steps_name='validation_steps',
steps=validation_steps,
validation_split=validation_split,
shuffle=shuffle)
elif validation_split:
raise ValueError('validation_split argument is not supported with '
'distribution strategies.')
if distributed_training_utils.is_tpu_strategy(
model._distribution_strategy):
return experimental_tpu_fit_loop(model,
dataset,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_dataset=val_dataset,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=1)
else:
return training_arrays.fit_loop(model,
dataset,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_inputs=val_dataset,
shuffle=shuffle,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=validation_freq)
def fit_distributed(model,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None):
"""Fit loop for Distribution Strategies."""
distributed_training_utils.validate_callbacks(callbacks, model.optimizer)
distributed_training_utils.validate_inputs(
x, y, model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps_per_epoch, batch_size = (
distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps_per_epoch,
batch_size, is_training=True))
batch_size = model._validate_or_infer_batch_size(
batch_size, steps_per_epoch, x)
iterator = model._distribution_standardize_user_data(
x, y,
sample_weight=sample_weight,
class_weight=class_weight,
batch_size=batch_size,
check_steps=True,
steps_name='steps_per_epoch',
steps=steps_per_epoch,
validation_split=validation_split,
shuffle=shuffle)
val_iterator = None
if validation_data:
val_x, val_y, val_sample_weights = model._unpack_validation_data(
validation_data)
distributed_training_utils.validate_inputs(
val_x, val_y, model._distribution_strategy)
first_valx_value = nest.flatten(val_x)[0]
if isinstance(first_valx_value, np.ndarray):
validation_steps, _ = distributed_training_utils.get_input_params(
model._distribution_strategy, first_valx_value, validation_steps,
batch_size)
val_iterator = model._distribution_standardize_user_data(
val_x, val_y,
sample_weight=val_sample_weights,
class_weight=None,
batch_size=batch_size,
check_steps=True,
steps_name='validation_steps',
steps=validation_steps,
validation_split=validation_split,
shuffle=shuffle)
elif validation_split:
raise ValueError('validation_split argument is not supported with '
'distribution strategies.')
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
return experimental_tpu_fit_loop(
model,
iterator,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_iterator=val_iterator,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps)
else:
return training_arrays.fit_loop(
model,
iterator,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_inputs=val_iterator,
shuffle=shuffle,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps)
def fit_distributed(model,
x=None,
y=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
validation_split=0.,
validation_data=None,
shuffle=True,
class_weight=None,
sample_weight=None,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
validation_freq=1):
"""Fit loop for Distribution Strategies."""
distributed_training_utils.validate_callbacks(callbacks, model.optimizer)
distributed_training_utils.validate_inputs(
x, y, model._distribution_strategy)
first_x_value = nest.flatten(x)[0]
if isinstance(first_x_value, np.ndarray):
steps_per_epoch, batch_size = (
distributed_training_utils.get_input_params(
model._distribution_strategy, first_x_value, steps_per_epoch,
batch_size, is_training=True))
batch_size = model._validate_or_infer_batch_size(
batch_size, steps_per_epoch, x)
dataset = model._distribution_standardize_user_data(
x, y,
sample_weight=sample_weight,
class_weight=class_weight,
batch_size=batch_size,
check_steps=True,
steps_name='steps_per_epoch',
steps=steps_per_epoch,
validation_split=validation_split,
shuffle=shuffle)
val_dataset = None
if validation_data:
val_x, val_y, val_sample_weights = model._unpack_validation_data(
validation_data)
distributed_training_utils.validate_inputs(
val_x, val_y, model._distribution_strategy)
first_valx_value = nest.flatten(val_x)[0]
if isinstance(first_valx_value, np.ndarray):
validation_steps, _ = distributed_training_utils.get_input_params(
model._distribution_strategy, first_valx_value, validation_steps,
batch_size)
val_dataset = model._distribution_standardize_user_data(
val_x, val_y,
sample_weight=val_sample_weights,
class_weight=None,
batch_size=batch_size,
check_steps=True,
steps_name='validation_steps',
steps=validation_steps,
validation_split=validation_split,
shuffle=shuffle)
elif validation_split:
raise ValueError('validation_split argument is not supported with '
'distribution strategies.')
if distributed_training_utils.is_tpu_strategy(model._distribution_strategy):
return experimental_tpu_fit_loop(
model,
dataset,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_dataset=val_dataset,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=1)
else:
return training_arrays.fit_loop(
model,
dataset,
batch_size=batch_size,
epochs=epochs,
verbose=verbose,
callbacks=callbacks,
val_inputs=val_dataset,
shuffle=shuffle,
initial_epoch=initial_epoch,
steps_per_epoch=steps_per_epoch,
validation_steps=validation_steps,
validation_freq=validation_freq)
