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)
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)
if steps_per_epoch is None:
raise ValueError('`steps_per_epoch` should be specified when calling '
'`fit` on the model.')
steps_per_run = K.variable(
value=min(steps_per_epoch, current_strategy.extended.steps_per_run),
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.
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)
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
for step_count in steps_to_run:
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:
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)
break
batch_logs.update(outputs)
callbacks._call_batch_hook(mode, 'end', step_index, batch_logs)
step_index = step_index + step_count
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 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.
"""
# 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)
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,
ModeKeys.TRAIN,
inputs=inputs,
targets=targets)
else:
distributed_training_utils._build_distributed_network(
model, current_strategy, ModeKeys.TRAIN, inputs, targets)
(grouped_inputs, grouped_outputs, grouped_updates, grouped_session_args
) = current_strategy.extended.call_for_each_replica(
_per_device_fit_function, args=(model._distributed_model_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 = distribute_lib.get_loss_reduction()
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)
if steps_per_epoch is None:
raise ValueError('`steps_per_epoch` should be specified when calling '
'`fit` on the model.')
steps_per_run = K.variable(value=min(
steps_per_epoch, current_strategy.extended.steps_per_run),
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, ModeKeys.TRAIN)
callbacks = cbks.configure_callbacks(callbacks,
model,
do_validation=do_validation,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
verbose=verbose)
# Calculate the steps each time on the device.
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)
callbacks.on_train_begin()
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
for step_count in steps_to_run:
batch_logs = {
'batch': step_index,
'size': 1,
'num_steps': step_count
}
callbacks.on_batch_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:
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)
break
batch_logs.update(outputs)
callbacks.on_batch_end(step_index, batch_logs)
step_index = step_index + step_count
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)
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.on_train_end()
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 model_iteration(model,
inputs,
targets=None,
sample_weights=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
val_inputs=None,
val_targets=None,
val_sample_weights=None,
shuffle=True,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
validation_freq=1,
mode=ModeKeys.TRAIN,
validation_in_fit=False,
prepared_feed_values_from_dataset=False,
steps_name='steps',
**kwargs):
"""Loop function for arrays of data with modes TRAIN/TEST/PREDICT.
Arguments:
model: Keras Model instance.
inputs: Either a list or dictionary of arrays, or a dataset instance.
targets: List/dictionary of input arrays.
sample_weights: Optional list of sample weight arrays.
batch_size: Integer batch size or None if unknown.
epochs: Number of times to iterate over the data
verbose: Verbosity mode, 0, 1 or 2
callbacks: List of callbacks to be called during training
val_inputs: Either a list or dictionary of arrays, or a dataset instance.
val_targets: List/dictionary of target arrays.
val_sample_weights: Optional list of sample weight arrays.
shuffle: Whether to shuffle the data at the beginning of each epoch
concatenation of list the display names of the outputs of `f` and the
list of display names of the outputs of `f_val`.
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`.
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.
mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT.
validation_in_fit: if true, then this method is invoked from within
training iteration (for validation). In the case where `val_inputs` is a
dataset, this flag indicates that its iterator and feed values are
already created so should properly reuse resources.
prepared_feed_values_from_dataset: if True, `inputs` is a list of feed
tensors returned from `_prepare_feed_values` call on the validation
dataset, so do not call it again on `inputs`. Should only be used for
inline validation (i.e., only if `validation_in_fit` is also True).
steps_name: The string name of the steps argument, either `steps`,
`validation_steps`, or `steps_per_epoch`. Only used for error message
formatting.
**kwargs: Additional arguments for backwards compatibility.
Returns:
- In TRAIN mode: `History` object.
- In TEST mode: Evaluation metrics.
- In PREDICT mode: Outputs of the Model called on inputs.
Raises:
ValueError: in case of invalid arguments.
"""
# Backwards compatibility.
if 'steps' in kwargs:
steps_per_epoch = kwargs.pop('steps')
if kwargs:
raise TypeError('Unknown arguments: %s' % (kwargs,))
# In case we were passed a dataset, we extract symbolic tensors from it.
reset_dataset_after_each_epoch = False
input_iterator = None
is_dataset = isinstance(inputs,
(dataset_ops.DatasetV1, dataset_ops.DatasetV2))
# TODO(fchollet): consider moving `steps_per_epoch` inference to
# _standardize_user_data and set reset_dataset_after_each_epoch as an
# attribute on the dataset instance.
if is_dataset:
if steps_per_epoch is None:
reset_dataset_after_each_epoch = True
steps_per_epoch = training_utils.infer_steps_for_dataset(
inputs, steps_per_epoch, epochs=epochs, steps_name=steps_name)
input_iterator = _get_iterator(inputs, model._distribution_strategy)
if mode == ModeKeys.TRAIN:
_print_train_info(inputs, val_inputs, steps_per_epoch, verbose)
# Enter DistributionStrategy scope.
if model._distribution_strategy:
scope = distributed_training_utils.distributed_scope(
strategy=model._distribution_strategy,
learning_phase=(1 if mode == ModeKeys.TRAIN else 0))
scope.__enter__()
# Get step function and loop type.
f = _make_execution_function(model, mode)
use_steps = is_dataset or steps_per_epoch is not None
do_validation = val_inputs is not None
# Convert Eager Tensors to NumPy arrays to support batching/shuffling.
inputs, targets, sample_weights = training_utils. \
convert_eager_tensors_to_numpy((inputs, targets, sample_weights))
# Prepare input data.
inputs = input_iterator or inputs
if validation_in_fit and prepared_feed_values_from_dataset:
# When invoking validation in training loop, avoid creating iterator and
# list of feed values for the same validation dataset multiple times (which
# essentially would call `iterator.get_next()` that slows down execution and
# leads to OOM errors eventually.
ins = inputs
else:
ins = _prepare_feed_values(model, inputs, targets, sample_weights, mode)
if not is_dataset:
num_samples_or_steps = _get_num_samples_or_steps(ins, batch_size,
steps_per_epoch)
else:
num_samples_or_steps = steps_per_epoch
# Prepare validation data. Hold references to the iterator and the input list
# to properly reinitialize and reuse in multiple validation passes.
val_iterator = None
if isinstance(val_inputs, (dataset_ops.DatasetV1, dataset_ops.DatasetV2)):
if validation_steps is None:
# Because we pass an iterator feed instead of a Dataset to the eval
# model_iteration() call, it will not trigger the dataset-input path
# that determines the number of steps required. To avoid this issue,
# set validation_steps here if validation_steps is None.
validation_steps = training_utils.infer_steps_for_dataset(
val_inputs,
validation_steps,
epochs=epochs,
steps_name='validation_steps')
val_iterator = _get_iterator(val_inputs, model._distribution_strategy)
val_inputs = _prepare_feed_values(
model, val_iterator, val_targets, val_sample_weights, ModeKeys.TEST)
# Configure callbacks.
count_mode = 'steps' if use_steps else 'samples'
callbacks = cbks.configure_callbacks(
callbacks,
model,
do_validation=do_validation,
batch_size=batch_size,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
samples=num_samples_or_steps,
verbose=0, # Handle ProgBarLogger separately in this loop.
mode=mode)
# TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready.
progbar = training_utils.get_progbar(model, count_mode)
progbar.params = callbacks.params
progbar.params['verbose'] = verbose
# Find beforehand arrays that need sparse-to-dense conversion.
if issparse is not None and not use_steps:
indices_for_conversion_to_dense = []
feed = _get_model_feed(model, mode)
for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)):
if issparse(input_data) and not K.is_sparse(feed_tensor):
indices_for_conversion_to_dense.append(i)
# Select aggregation method.
if mode == ModeKeys.PREDICT:
aggregator = training_utils.OutputsAggregator(use_steps,
num_samples_or_steps)
else:
aggregator = training_utils.MetricsAggregator(use_steps,
num_samples_or_steps)
if model._compile_distribution:
distributed_training_utils._copy_weights_to_distributed_model(model, mode)
callbacks.model.stop_training = False
callbacks._call_begin_hook(mode)
progbar.on_train_begin()
for epoch in range(initial_epoch, epochs):
if callbacks.model.stop_training:
break
# Setup work for each epoch
epoch_logs = {}
model.reset_metrics()
if mode == ModeKeys.TRAIN:
callbacks.on_epoch_begin(epoch, epoch_logs)
progbar.on_epoch_begin(epoch, epoch_logs)
if use_steps:
# Step-wise loop.
if steps_per_epoch is None:
# Loop over dataset until `OutOfRangeError` is raised.
target_steps = np.inf
else:
# Loop over dataset for the specified number of steps.
target_steps = steps_per_epoch
step = 0
while step < target_steps:
batch_logs = {'batch': step, 'size': 1}
callbacks._call_batch_hook(mode, 'begin', step, batch_logs)
progbar.on_batch_begin(step, batch_logs)
# Get outputs.
try:
# `ins` can be callable in DistributionStrategy + eager case.
actual_inputs = ins() if callable(ins) else ins
batch_outs = f(actual_inputs)
except errors.OutOfRangeError:
if is_dataset:
# The dataset passed by the user ran out of batches.
# Now we know the cardinality of the dataset.
# If steps_per_epoch was specified, then running out of data is
# unexpected, so we stop training and inform the user.
if steps_per_epoch:
callbacks.model.stop_training = True
logging.warning(
'Your dataset ran out of data; interrupting training. '
'Make sure that your dataset can generate at least '
'`%s * epochs` batches (in this case, %d batches). '
'You may need to use the repeat() function when '
'building your dataset.'
% (steps_name, steps_per_epoch * epochs))
elif step > 0:
steps_per_epoch = step
aggregator.num_samples_or_steps = steps_per_epoch
if mode == ModeKeys.TRAIN:
progbar.params['steps'] = steps_per_epoch
progbar.progbar.target = steps_per_epoch
else:
# We ran out of batches while the user passed an iterator (legacy).
callbacks.model.stop_training = True
logging.warning(
'Your dataset iterator ran out of data; '
'interrupting training. Make sure that your iterator '
'can generate at least `%s * epochs` '
'batches (in this case, %d batches). You may need to'
'use the repeat() function when building your '
'dataset.' % (steps_name, steps_per_epoch * epochs))
break
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if model._distribution_strategy:
batch_outs = distributed_training_utils._per_device_aggregate_batch(
batch_outs, model, mode)
# Aggregate results.
if step == 0:
aggregator.create(batch_outs)
aggregator.aggregate(batch_outs)
# Callbacks batch end.
batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode)
callbacks._call_batch_hook(mode, 'end', step, batch_logs)
progbar.on_batch_end(step, batch_logs)
step += 1
if callbacks.model.stop_training:
break
else:
# Sample-wise loop.
index_array = np.arange(num_samples_or_steps)
if shuffle == 'batch':
index_array = training_utils.batch_shuffle(index_array, batch_size)
elif shuffle:
np.random.shuffle(index_array)
batches = make_batches(num_samples_or_steps, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_ids = index_array[batch_start:batch_end]
# Slice into a batch.
try:
if ins and isinstance(ins[-1], int):
# Do not slice the training phase flag.
ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
else:
ins_batch = slice_arrays(ins, batch_ids)
except TypeError:
raise TypeError('TypeError while preparing batch. '
'If using HDF5 input data, '
'pass shuffle="batch".')
# Sparse to dense conversion.
if issparse is not None:
for i in indices_for_conversion_to_dense:
ins_batch[i] = ins_batch[i].toarray()
# Callbacks batch_begin.
batch_logs = {'batch': batch_index, 'size': len(batch_ids)}
callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs)
progbar.on_batch_begin(batch_index, batch_logs)
# Get outputs.
batch_outs = f(ins_batch)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
# Aggregate results.
if batch_index == 0:
aggregator.create(batch_outs)
aggregator.aggregate(batch_outs, batch_start, batch_end)
# Callbacks batch end.
batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode)
callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs)
progbar.on_batch_end(batch_index, batch_logs)
if callbacks.model.stop_training:
break
aggregator.finalize()
results = aggregator.results
epoch_logs = cbks.make_logs(model, epoch_logs, results, mode)
if len(results) == 1:
results = results[0]
# Run the test loop every `validation_freq` epochs during training.
if (do_validation and
training_utils.should_run_validation(validation_freq, epoch) and
not callbacks.model.stop_training):
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_results = model_iteration(
model,
val_inputs,
targets=val_targets,
sample_weights=val_sample_weights,
batch_size=batch_size,
steps_per_epoch=validation_steps,
callbacks=callbacks,
verbose=0,
mode=ModeKeys.TEST,
validation_in_fit=True,
prepared_feed_values_from_dataset=(val_iterator is not None),
steps_name='validation_steps')
if not isinstance(val_results, list):
val_results = [val_results]
epoch_logs = cbks.make_logs(
model, epoch_logs, val_results, mode, prefix='val_')
if val_iterator and epoch < epochs - 1:
_reinitialize_iterator(val_iterator, model._distribution_strategy)
if mode == ModeKeys.TRAIN:
# Epochs only apply to `fit`.
callbacks.on_epoch_end(epoch, epoch_logs)
progbar.on_epoch_end(epoch, epoch_logs)
# Reinitialize dataset iterator for the next epoch.
if reset_dataset_after_each_epoch and epoch < epochs - 1:
_reinitialize_iterator(input_iterator, model._distribution_strategy)
callbacks._call_end_hook(mode)
if model._distribution_strategy:
if model._compile_distribution:
# TODO(priyag, psv): Copy back metrics to the original model as well?
distributed_training_utils._copy_weights_to_original_model(model, mode)
scope.__exit__(None, None, None)
if mode == ModeKeys.TRAIN:
return model.history
return results
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_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
def model_iteration(model,
inputs,
targets=None,
sample_weights=None,
batch_size=None,
epochs=1,
verbose=1,
callbacks=None,
val_inputs=None,
val_targets=None,
val_sample_weights=None,
shuffle=True,
initial_epoch=0,
steps_per_epoch=None,
validation_steps=None,
mode=ModeKeys.TRAIN,
validation_in_fit=False,
**kwargs):
"""Loop function for arrays of data with modes TRAIN/TEST/PREDICT.
Arguments:
model: Keras Model instance.
inputs: Either a list of arrays or a dictionary.
targets: List of target arrays.
sample_weights: Optional list of sample weight arrays.
batch_size: Integer batch size or None if unknown.
epochs: Number of times to iterate over the data
verbose: Verbosity mode, 0, 1 or 2
callbacks: List of callbacks to be called during training
val_inputs: List of input arrays.
val_targets: List of target arrays.
val_sample_weights: Optional list of sample weight arrays.
shuffle: Whether to shuffle the data at the beginning of each epoch
concatenation of list the display names of the outputs of `f` and the
list of display names of the outputs of `f_val`.
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`.
validation_steps: Number of steps to run validation for (only if doing
validation from data tensors). Ignored with the default value of `None`.
mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT.
validation_in_fit: DEPRECATED: if true, then this method is invoked from
within training iteration (for validation). In this case, do not copy
weights when using a tf.distribute.Strategy. The input is deprecated as
it is not required if the user creates a distributed model under the
distribution strategy scope rather than passing it to compile.
**kwargs: Additional arguments for backwards compatibility.
Returns:
- In TRAIN mode: `History` object.
- In TEST mode: Evaluation metrics.
- In PREDICT mode: Outputs of the Model called on inputs.
Raises:
ValueError: in case of invalid arguments.
"""
# Backwards compatibility.
if 'steps' in kwargs:
steps_per_epoch = kwargs['steps']
_validate_arguments(steps_per_epoch, validation_steps, kwargs)
if mode == ModeKeys.TRAIN:
_print_train_info(inputs, val_inputs, steps_per_epoch, verbose)
# Enter DistributionStrategy scope.
if model._distribution_strategy:
scope = model._distribution_strategy.scope()
scope.__enter__()
# Get step function and loop type.
f = _make_execution_function(model, mode)
use_steps = steps_per_epoch is not None
do_validation = val_inputs is not None
# Convert Eager Tensors to NumPy arrays to support batching/shuffling.
inputs, targets, sample_weights = training_utils. \
convert_eager_tensors_to_numpy((inputs, targets, sample_weights))
# Prepare input data.
ins = _prepare_feed_values(model, inputs, targets, sample_weights, mode)
num_samples_or_steps = _get_num_samples_or_steps(ins, batch_size,
steps_per_epoch)
# Configure callbacks.
count_mode = 'steps' if use_steps else 'samples'
callbacks = cbks.configure_callbacks(
callbacks,
model,
do_validation=do_validation,
batch_size=batch_size,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
samples=num_samples_or_steps,
verbose=0, # Handle ProgBarLogger separately in this loop.
mode=mode)
# TODO(omalleyt): Handle ProgBar as part of Callbacks once hooks are ready.
progbar = training_utils.get_progbar(model, count_mode)
progbar.params = callbacks.params
progbar.params['verbose'] = verbose
# Find beforehand arrays that need sparse-to-dense conversion.
if issparse is not None and not use_steps:
indices_for_conversion_to_dense = []
feed = _get_model_feed(model, mode)
for i, (input_data, feed_tensor) in enumerate(zip(ins, feed)):
if issparse(input_data) and not K.is_sparse(feed_tensor):
indices_for_conversion_to_dense.append(i)
# Select aggregation method.
if mode == ModeKeys.PREDICT:
aggregator = training_utils.OutputsAggregator(use_steps,
num_samples_or_steps)
else:
aggregator = training_utils.MetricsAggregator(use_steps,
num_samples_or_steps)
if model._compile_distribution and not validation_in_fit:
distributed_training_utils._copy_weights_to_distributed_model(
model, model._distributed_model)
callbacks.model.stop_training = False
callbacks._call_begin_hook(mode)
progbar.on_train_begin()
for epoch in range(initial_epoch, epochs):
if callbacks.model.stop_training:
break
# Setup work for each epoch
epoch_logs = {}
model.reset_metrics()
if mode == ModeKeys.TRAIN:
callbacks.on_epoch_begin(epoch, epoch_logs)
progbar.on_epoch_begin(epoch, epoch_logs)
if use_steps:
# Step-wise loop.
for step in range(steps_per_epoch):
batch_logs = {'batch': step, 'size': 1}
callbacks._call_batch_hook(mode, 'begin', step, batch_logs)
progbar.on_batch_begin(step, batch_logs)
# Get outputs.
try:
# `ins` can be callable in DistributionStrategy + eager case.
actual_inputs = ins() if callable(ins) else ins
batch_outs = f(actual_inputs)
except errors.OutOfRangeError:
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). You may need to'
'use the repeat() function when building your '
'dataset.' % steps_per_epoch * epochs)
break
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if model._distribution_strategy:
batch_outs = distributed_training_utils._per_device_aggregate_batch(
batch_outs, model, mode)
# Aggregate results.
if step == 0:
aggregator.create(batch_outs)
aggregator.aggregate(batch_outs)
# Callbacks batch end.
batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode)
callbacks._call_batch_hook(mode, 'end', step, batch_logs)
progbar.on_batch_end(step, batch_logs)
if callbacks.model.stop_training:
break
else:
# Sample-wise loop.
index_array = np.arange(num_samples_or_steps)
if shuffle == 'batch':
index_array = training_utils.batch_shuffle(index_array, batch_size)
elif shuffle:
np.random.shuffle(index_array)
batches = make_batches(num_samples_or_steps, batch_size)
for batch_index, (batch_start, batch_end) in enumerate(batches):
batch_ids = index_array[batch_start:batch_end]
# Slice into a batch.
try:
if ins and isinstance(ins[-1], int):
# Do not slice the training phase flag.
ins_batch = slice_arrays(ins[:-1], batch_ids) + [ins[-1]]
else:
ins_batch = slice_arrays(ins, batch_ids)
except TypeError:
raise TypeError('TypeError while preparing batch. '
'If using HDF5 input data, '
'pass shuffle="batch".')
# Sparse to dense conversion.
if issparse is not None:
for i in indices_for_conversion_to_dense:
ins_batch[i] = ins_batch[i].toarray()
# Callbacks batch_begin.
batch_logs = {'batch': batch_index, 'size': len(batch_ids)}
callbacks._call_batch_hook(mode, 'begin', batch_index, batch_logs)
progbar.on_batch_begin(batch_index, batch_logs)
# Get outputs.
batch_outs = f(ins_batch)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
# Aggregate results.
if batch_index == 0:
aggregator.create(batch_outs)
aggregator.aggregate(batch_outs, batch_start, batch_end)
# Callbacks batch end.
batch_logs = cbks.make_logs(model, batch_logs, batch_outs, mode)
callbacks._call_batch_hook(mode, 'end', batch_index, batch_logs)
progbar.on_batch_end(batch_index, batch_logs)
if callbacks.model.stop_training:
break
aggregator.finalize()
results = aggregator.results
epoch_logs = cbks.make_logs(model, epoch_logs, results, mode)
if len(results) == 1:
results = results[0]
# Run the test loop every epoch during training.
if do_validation and not callbacks.model.stop_training:
val_results = model_iteration(
model,
val_inputs,
targets=val_targets,
sample_weights=val_sample_weights,
batch_size=batch_size,
steps_per_epoch=validation_steps,
callbacks=callbacks,
verbose=0,
mode=ModeKeys.TEST,
validation_in_fit=True)
if not isinstance(val_results, list):
val_results = [val_results]
epoch_logs = cbks.make_logs(
model, epoch_logs, val_results, mode, prefix='val_')
if mode == ModeKeys.TRAIN:
# Epochs only apply to `fit`.
callbacks.on_epoch_end(epoch, epoch_logs)
progbar.on_epoch_end(epoch, epoch_logs)
callbacks._call_end_hook(mode)
if model._distribution_strategy:
if model._compile_distribution and not validation_in_fit:
# TODO(priyag, psv): Copy back metrics to the original model as well?
distributed_training_utils._copy_weights_to_original_model(
model, model._distributed_model, mode)
scope.__exit__(None, None, None)
if mode == ModeKeys.TRAIN:
return model.history
return results
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
