def experimental_tpu_fit_loop(model,
dataset,
epochs=100,
verbose=1,
callbacks=None,
initial_epoch=0,
steps_per_epoch=None,
val_dataset=None,
validation_steps=None,
validation_freq=1):
"""Fit loop for training with TPU DistributionStrategy.
Arguments:
model: Keras Model instance.
dataset: Dataset that returns inputs and targets
epochs: Number of times to iterate over the data
verbose: Integer, Verbosity mode, 0, 1 or 2
callbacks: List of callbacks to be called during training
initial_epoch: Epoch at which to start training
(useful for resuming a previous training run)
steps_per_epoch: Total number of steps (batches of samples)
before declaring one epoch finished and starting the
next epoch. Ignored with the default value of `None`.
val_dataset: Dataset for validation data.
validation_steps: Number of steps to run validation for
(only if doing validation from data tensors).
Ignored with the default value of `None`.
validation_freq: Only relevant if validation data is provided. Integer or
`collections.Container` instance (e.g. list, tuple, etc.). If an
integer, specifies how many training epochs to run before a new
validation run is performed, e.g. `validation_freq=2` runs
validation every 2 epochs. If a Container, specifies the epochs on
which to run validation, e.g. `validation_freq=[1, 2, 10]` runs
validation at the end of the 1st, 2nd, and 10th epochs.
Returns:
Returns `None`.
Raises:
ValueError: in case of invalid arguments.
"""
mode = ModeKeys.TRAIN
# TODO(fchollet): add support for `steps_per_epoch=None` in TPU loops.
current_strategy = model._distribution_strategy
iterator = distributed_training_utils.get_iterator(dataset,
current_strategy)
steps_per_epoch = training_utils.infer_steps_for_dataset(
dataset, steps_per_epoch, epochs, steps_name='steps_per_epoch')
if (current_strategy.extended.steps_per_run != 1
and steps_per_epoch is None):
raise ValueError('`steps_per_epoch` should be specified when calling '
'`fit` on the model with TPUStrategy when '
'`steps_per_run` != 1 .')
scope = distributed_training_utils.distributed_scope(
strategy=current_strategy, learning_phase=1)
scope.__enter__()
def _per_device_fit_function(model):
model._make_fit_function()
return (model._fit_function.inputs, model._fit_function.outputs,
model._fit_function.updates_op,
model._fit_function.session_kwargs)
out_labels = model.metrics_names or []
def step_fn(ctx, inputs):
"""Clones the model and calls make_fit_function."""
inputs, targets = inputs
if model._compile_distribution:
distributed_training_utils.clone_model_on_replicas(
model, current_strategy, mode, inputs=inputs, targets=targets)
else:
distributed_training_utils._build_distributed_network(
model, current_strategy, mode, inputs, targets)
(grouped_inputs, grouped_outputs, grouped_updates, grouped_session_args
) = current_strategy.extended.call_for_each_replica(
_per_device_fit_function,
args=(distributed_training_utils.get_distributed_model(
model, ModeKeys.TRAIN), ))
(all_inputs, all_outputs, all_updates,
all_session_args) = distributed_training_utils.unwrap_values(
current_strategy, grouped_inputs, grouped_outputs,
grouped_updates, grouped_session_args)
combined_fn = K.function(all_inputs,
all_outputs,
updates=all_updates,
name='distributed_fit_function',
**all_session_args)
for label, output in zip(out_labels, combined_fn.outputs):
if label == 'loss':
reduce_op = ds_reduce_util.ReduceOp.SUM
else:
# We reduce all other metrics using mean for now. This is temporary
# workaround until new metrics are in place.
reduce_op = ds_reduce_util.ReduceOp.MEAN
ctx.set_last_step_output(label, output, reduce_op)
# TODO(priyag, sourabhbajaj): Ignoring these things from the combined_fn:
# feed_dict, session kwargs, run options, run_metadata for now. These should
# be handled appropriately
return combined_fn.updates_op
# Add initial dummy values for loss and other metric tensors.
initial_loop_values = {}
initial_loop_values['loss'] = constant_op.constant(1e7)
for name in model.metrics_names[1:]:
tensor = model._all_stateful_metrics_tensors[name]
initial_loop_values[name] = array_ops.zeros(tensor.shape, tensor.dtype)
use_steps = steps_per_epoch is not None
if use_steps:
iteration_value = min(steps_per_epoch,
current_strategy.extended.steps_per_run)
else:
iteration_value = current_strategy.extended.steps_per_run
steps_per_run = K.variable(value=iteration_value,
dtype='int32',
name='steps_per_run')
ctx = current_strategy.extended.experimental_run_steps_on_iterator(
step_fn,
iterator,
iterations=steps_per_run,
initial_loop_values=initial_loop_values)
train_op = ctx.run_op
output_tensors = ctx.last_step_outputs
do_validation = bool(validation_steps)
if model._compile_distribution:
distributed_training_utils._copy_weights_to_distributed_model(
model, mode)
callbacks = cbks.configure_callbacks(callbacks,
model,
do_validation=do_validation,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
verbose=verbose,
count_mode='steps',
mode=mode)
# Calculate the steps each time on the device.
if use_steps:
steps_to_run = (
[current_strategy.extended.steps_per_run] *
(steps_per_epoch // current_strategy.extended.steps_per_run))
if steps_per_epoch % current_strategy.extended.steps_per_run:
steps_to_run.append(steps_per_epoch %
current_strategy.extended.steps_per_run)
target_steps = len(steps_to_run)
else:
target_steps = np.inf
callbacks._call_begin_hook(mode)
for epoch in range(initial_epoch, epochs):
distributed_training_utils._reset_metrics(model)
callbacks.on_epoch_begin(epoch)
epoch_logs = {}
step_index = 0
prev_step_count = None
current_step = 0
while current_step < target_steps:
step_count = steps_to_run[current_step] if use_steps else 1
batch_logs = {
'batch': step_index,
'size': 1,
'num_steps': step_count
}
callbacks._call_batch_hook(mode, 'begin', step_index, batch_logs)
if prev_step_count is None or step_count != prev_step_count:
steps_per_run.load(step_count, K.get_session())
prev_step_count = step_count
try:
_, outputs = K.get_session().run([train_op, output_tensors])
except errors.OutOfRangeError:
if use_steps:
logging.warning(
'Your dataset iterator ran out of data; '
'interrupting training. Make sure that your dataset '
'can generate at least `steps_per_epoch * epochs` '
'batches (in this case, %d batches).' %
steps_per_epoch * epochs)
else:
target_steps = current_step
logging.info(
'Dataset iterator ran out of data. Inferring the '
'value of `steps_per_epoch` as %s .' % target_steps)
distributed_training_utils.initialize_iterator(
iterator, current_strategy)
break
batch_logs.update(outputs)
callbacks._call_batch_hook(mode, 'end', step_index, batch_logs)
step_index = step_index + step_count
current_step += 1
if callbacks.model.stop_training:
break
if (do_validation and training_utils.should_run_validation(
validation_freq, epoch)):
logging.info('Running validation at fit epoch: %s', epoch)
if model._compile_distribution:
# Since we create a new clone from the original model we need to copy
# the weights back to the original model before we can run validation.
distributed_training_utils._copy_weights_to_original_model(
model, ModeKeys.TRAIN)
val_outs = experimental_tpu_test_loop( # pylint: disable=undefined-variable
model,
val_dataset,
steps=validation_steps,
verbose=verbose,
callbacks=callbacks)
if not isinstance(val_outs, list):
val_outs = [val_outs]
# Same labels assumed.
for label, val_out in zip(out_labels, val_outs):
epoch_logs['val_' + label] = val_out
callbacks.on_epoch_end(epoch, epoch_logs)
if callbacks.model.stop_training:
break
callbacks._call_end_hook(mode)
if model._compile_distribution:
# Copy the weights back from the replicated model to the original model.
distributed_training_utils._copy_weights_to_original_model(
model, ModeKeys.TRAIN)
scope.__exit__(None, None, None)
return model.history
def _reinitialize_iterator(iterator, distribution_strategy=None):
if distribution_strategy:
distributed_training_utils.initialize_iterator(
iterator, distribution_strategy)
else:
training_utils.initialize_iterator(iterator)
def experimental_tpu_fit_loop(model,
dataset,
epochs=100,
verbose=1,
callbacks=None,
initial_epoch=0,
steps_per_epoch=None,
val_dataset=None,
validation_steps=None,
validation_freq=1):
"""Fit loop for training with TPU DistributionStrategy.
Arguments:
model: Keras Model instance.
dataset: Dataset that returns inputs and targets
epochs: Number of times to iterate over the data
verbose: Integer, Verbosity mode, 0, 1 or 2
callbacks: List of callbacks to be called during training
initial_epoch: Epoch at which to start training
(useful for resuming a previous training run)
steps_per_epoch: Total number of steps (batches of samples)
before declaring one epoch finished and starting the
next epoch. Ignored with the default value of `None`.
val_dataset: Dataset for validation data.
validation_steps: Number of steps to run validation for
(only if doing validation from data tensors).
Ignored with the default value of `None`.
validation_freq: Only relevant if validation data is provided. Integer or
`collections.Container` instance (e.g. list, tuple, etc.). If an
integer, specifies how many training epochs to run before a new
validation run is performed, e.g. `validation_freq=2` runs
validation every 2 epochs. If a Container, specifies the epochs on
which to run validation, e.g. `validation_freq=[1, 2, 10]` runs
validation at the end of the 1st, 2nd, and 10th epochs.
Returns:
Returns `None`.
Raises:
ValueError: in case of invalid arguments.
"""
mode = ModeKeys.TRAIN
# TODO(fchollet): add support for `steps_per_epoch=None` in TPU loops.
current_strategy = model._distribution_strategy
iterator = distributed_training_utils.get_iterator(dataset, current_strategy)
steps_per_epoch = training_utils.infer_steps_for_dataset(
dataset, steps_per_epoch, epochs, steps_name='steps_per_epoch')
if (current_strategy.extended.steps_per_run != 1 and
steps_per_epoch is None):
raise ValueError('`steps_per_epoch` should be specified when calling '
'`fit` on the model with TPUStrategy when '
'`steps_per_run` != 1 .')
scope = distributed_training_utils.distributed_scope(
strategy=current_strategy, learning_phase=1)
scope.__enter__()
out_labels = model.metrics_names or []
step_fn = _make_step_fn(model, ModeKeys.TRAIN, current_strategy, out_labels)
# Add initial dummy values for loss and other metric tensors.
initial_loop_values = {}
initial_loop_values['loss'] = constant_op.constant(1e7)
for name in model.metrics_names[1:]:
tensor = model._all_stateful_metrics_tensors[name]
initial_loop_values[name] = array_ops.zeros(tensor.shape, tensor.dtype)
use_steps = steps_per_epoch is not None
if use_steps:
iteration_value = min(steps_per_epoch,
current_strategy.extended.steps_per_run)
else:
iteration_value = current_strategy.extended.steps_per_run
steps_per_run = K.variable(
value=iteration_value,
dtype='int32',
name='steps_per_run')
ctx = current_strategy.extended.experimental_run_steps_on_iterator(
step_fn, iterator, iterations=steps_per_run,
initial_loop_values=initial_loop_values)
train_op = ctx.run_op
output_tensors = ctx.last_step_outputs
do_validation = bool(validation_steps)
if model._compile_distribution:
distributed_training_utils._copy_weights_to_distributed_model(model, mode)
callbacks = cbks.configure_callbacks(
callbacks,
model,
do_validation=do_validation,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
verbose=verbose,
count_mode='steps',
mode=mode)
# Calculate the steps each time on the device.
if use_steps:
steps_to_run = ([current_strategy.extended.steps_per_run] *
(steps_per_epoch //
current_strategy.extended.steps_per_run))
if steps_per_epoch % current_strategy.extended.steps_per_run:
steps_to_run.append(
steps_per_epoch % current_strategy.extended.steps_per_run)
target_steps = len(steps_to_run)
else:
target_steps = np.inf
callbacks._call_begin_hook(mode)
for epoch in range(initial_epoch, epochs):
distributed_training_utils._reset_metrics(model)
callbacks.on_epoch_begin(epoch)
epoch_logs = {}
step_index = 0
prev_step_count = None
current_step = 0
while current_step < target_steps:
step_count = steps_to_run[current_step] if use_steps else 1
batch_logs = {'batch': step_index, 'size': 1, 'num_steps': step_count}
callbacks._call_batch_hook(mode, 'begin', step_index, batch_logs)
if prev_step_count is None or step_count != prev_step_count:
steps_per_run.load(step_count, K.get_session())
prev_step_count = step_count
try:
_, outputs = K.batch_get_value([train_op, output_tensors])
except errors.OutOfRangeError:
if use_steps:
logging.warning('Your dataset iterator ran out of data; '
'interrupting training. Make sure that your dataset '
'can generate at least `steps_per_epoch * epochs` '
'batches (in this case, %d batches).' %
steps_per_epoch * epochs)
else:
target_steps = current_step
logging.info('Dataset iterator ran out of data. Inferring the '
'value of `steps_per_epoch` as %s .' % target_steps)
distributed_training_utils.initialize_iterator(iterator,
current_strategy)
break
batch_logs.update(outputs)
callbacks._call_batch_hook(mode, 'end', step_index, batch_logs)
step_index = step_index + step_count
current_step += 1
if callbacks.model.stop_training:
break
if (do_validation and
training_utils.should_run_validation(validation_freq, epoch)):
logging.info('Running validation at fit epoch: %s', epoch)
if model._compile_distribution:
# Since we create a new clone from the original model we need to copy
# the weights back to the original model before we can run validation.
distributed_training_utils._copy_weights_to_original_model(
model, ModeKeys.TRAIN)
val_outs = experimental_tpu_test_loop( # pylint: disable=undefined-variable
model,
val_dataset,
steps=validation_steps,
verbose=verbose,
callbacks=callbacks)
if not isinstance(val_outs, list):
val_outs = [val_outs]
# Same labels assumed.
for label, val_out in zip(out_labels, val_outs):
epoch_logs['val_' + label] = val_out
callbacks.on_epoch_end(epoch, epoch_logs)
if callbacks.model.stop_training:
break
callbacks._call_end_hook(mode)
if model._compile_distribution:
# Copy the weights back from the replicated model to the original model.
distributed_training_utils._copy_weights_to_original_model(
model, ModeKeys.TRAIN)
scope.__exit__(None, None, None)
return model.history