def model_iteration(model,
data,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
validation_freq=1,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=False,
initial_epoch=0,
mode=ModeKeys.TRAIN,
batch_size=None,
steps_name='steps',
**kwargs):
"""Loop function for arrays of data with modes TRAIN/TEST/PREDICT.
Arguments:
model: Keras Model instance.
data: Either a tuple of NumPy/Tensor inputs (i.e. `(x,)` or `(x, y)` or
`(x, y, sample_weights)`) or a generator or
`keras.utils.data_utils.Sequence` object or Eager Iterator or Dataset.
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`.
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.
validation_data: Either a tuple of NumPy/Tensor inputs (i.e. `(x,)` or
`(x, y)` or `(x, y, sample_weights)`) or a generator or
`keras.utils.data_utils.Sequence` object or Eager Iterator or Dataset.
validation_steps: Total number of steps (batches of samples) before
declaring validation finished.
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.
class_weight: Dictionary mapping class indices to a weight for the class.
max_queue_size: Integer. Maximum size for the generator queue. If
unspecified, `max_queue_size` will default to 10.
workers: Integer. Maximum number of processes to spin up when using
process-based threading. If unspecified, `workers` will default to 1. If
0, will execute the generator on the main thread.
use_multiprocessing: Boolean. If `True`, use process-based threading. If
unspecified, `use_multiprocessing` will default to `False`. Note that
because this implementation relies on multiprocessing, you should not
pass non-picklable arguments to the generator as they can't be passed
easily to children processes.
shuffle: Boolean. Whether to shuffle the order of the batches at the
beginning of each epoch. Only used with instances of `Sequence`
(`keras.utils.Sequence`). Has no effect when `steps_per_epoch` is not
`None`.
initial_epoch: Epoch at which to start training (useful for resuming a
previous training run).
mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT.
batch_size: Integer batch size or None if unknown. Will only be used if
`data` is in NumPy/Tensor format.
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. `steps` is
accepted as an alias for `steps_per_epoch`.
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.
"""
if 'steps' in kwargs:
steps_per_epoch = kwargs['steps']
# Determine the number of steps per epoch and whether we should reset the
# dataset at the end of each epoch.
reset_dataset_after_each_epoch = False
original_dataset = None
is_dataset = isinstance(data, (dataset_ops.DatasetV2, dataset_ops.DatasetV1))
if is_dataset:
original_dataset = data
if steps_per_epoch is None:
reset_dataset_after_each_epoch = True
steps_per_epoch = training_utils.infer_steps_for_dataset(
data, steps_per_epoch, epochs=epochs, steps_name=steps_name)
# Convert to a format that supports `next(generator)`.
generator, steps_per_epoch = convert_to_generator_like(
data,
steps_per_epoch=steps_per_epoch,
batch_size=batch_size,
epochs=epochs - initial_epoch,
shuffle=shuffle)
do_validation = validation_data is not None
is_sequence = isinstance(generator, data_utils.Sequence)
_validate_arguments(is_sequence, is_dataset, use_multiprocessing, workers,
steps_per_epoch, validation_data, validation_steps, mode,
kwargs)
batch_function = _make_execution_function(
model, mode, class_weight=class_weight)
# Create the queue for the generator.
enqueuer = None
if not is_dataset:
generator, enqueuer = _make_enqueued_generator(
generator,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
shuffle=shuffle)
num_samples_or_steps, use_steps = _get_num_samples_or_steps(
data, steps_per_epoch)
count_mode = 'steps' if use_steps else 'samples'
callbacks = cbks.configure_callbacks(
callbacks,
model,
do_validation=do_validation,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
batch_size=batch_size,
samples=num_samples_or_steps,
verbose=0, # Handle ProgBar as part of Callbacks once hooks are ready.
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
if mode == ModeKeys.PREDICT:
aggregator = training_utils.OutputsAggregator(True, steps_per_epoch)
else:
aggregator = training_utils.MetricsAggregator(True, steps_per_epoch)
should_set_learning_phase = context.executing_eagerly() and model.run_eagerly
if should_set_learning_phase:
old_learning_phase = backend.learning_phase()
backend.set_eager_learning_phase(1 if mode == ModeKeys.TRAIN else 0)
callbacks.model.stop_training = False
callbacks._call_begin_hook(mode)
progbar.on_train_begin()
initial_epoch = model._maybe_load_initial_epoch_from_ckpt(initial_epoch, mode)
for epoch in range(initial_epoch, epochs):
if callbacks.model.stop_training:
break
# Setup work for each epoch.
model.reset_metrics()
epoch_logs = {}
if mode == ModeKeys.TRAIN:
callbacks.on_epoch_begin(epoch, epoch_logs)
progbar.on_epoch_begin(epoch, epoch_logs)
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_data = _get_next_batch(generator, mode)
if batch_data is None:
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
# `batch_size` used for validation data if validation
# data is NumPy/EagerTensors.
batch_size = int(nest.flatten(batch_data)[0].shape[0])
# Callbacks batch begin.
batch_logs = {'batch': step, 'size': batch_size}
callbacks._call_batch_hook(mode, 'begin', step, batch_logs)
progbar.on_batch_begin(step, batch_logs)
is_deferred = not model._is_compiled
batch_outs = batch_function(*batch_data)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if step == 0:
aggregator.create(batch_outs)
if is_deferred:
# Set callbacks params. We do this here when model is compiled only
# in the first iteration of this loop (deferred build scenario).
cbks.set_callback_parameters(
callbacks,
model,
do_validation=do_validation,
batch_size=batch_size,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
samples=num_samples_or_steps,
verbose=verbose,
mode=mode)
progbar.params = callbacks.params
progbar.params['verbose'] = verbose
# Aggregate results.
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
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
training_utils.should_run_validation(validation_freq, epoch) and
not callbacks.model.stop_training):
val_results = model_iteration(
model,
validation_data,
steps_per_epoch=validation_steps,
batch_size=batch_size,
class_weight=class_weight,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
callbacks=callbacks,
verbose=0,
mode=ModeKeys.TEST,
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 mode == ModeKeys.TRAIN:
# Epochs only apply to `fit`.
callbacks.on_epoch_end(epoch, epoch_logs)
progbar.on_epoch_end(epoch, epoch_logs)
# Recreate dataset iterator for the next epoch.
if reset_dataset_after_each_epoch and epoch < epochs - 1:
generator = dataset_ops.make_one_shot_iterator(original_dataset)
callbacks._call_end_hook(mode)
if enqueuer is not None:
enqueuer.stop()
if should_set_learning_phase:
backend.set_eager_learning_phase(old_learning_phase)
if mode == ModeKeys.TRAIN:
return model.history
return results
def model_iteration(model,
data,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
validation_freq=1,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=False,
initial_epoch=0,
mode=ModeKeys.TRAIN,
batch_size=None,
steps_name='steps',
**kwargs):
"""Loop function for arrays of data with modes TRAIN/TEST/PREDICT.
Arguments:
model: Keras Model instance.
data: Either a tuple of NumPy/Tensor inputs (i.e. `(x,)` or `(x, y)` or
`(x, y, sample_weights)`) or a generator or
`keras.utils.data_utils.Sequence` object or Eager Iterator or Dataset.
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`.
epochs: Number of times to iterate over the data.
verbose: 0, 1, or 2. Verbosity mode.
0 = silent, 1 = progress bar, 2 = one line per epoch.
Note that the progress bar is not particularly useful when
logged to a file, so verbose=2 is recommended when not running
interactively (eg, in a production environment).
callbacks: List of callbacks to be called during training.
validation_data: Either a tuple of NumPy/Tensor inputs (i.e. `(x,)` or
`(x, y)` or `(x, y, sample_weights)`) or a generator or
`keras.utils.data_utils.Sequence` object or Eager Iterator or Dataset.
validation_steps: Total number of steps (batches of samples) before
declaring validation finished.
validation_freq: Only relevant if validation data is provided. Integer or
`collections.abc.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.
class_weight: Dictionary mapping class indices to a weight for the class.
max_queue_size: Integer. Maximum size for the generator queue. If
unspecified, `max_queue_size` will default to 10.
workers: Integer. Maximum number of processes to spin up when using
process-based threading. If unspecified, `workers` will default to 1. If
0, will execute the generator on the main thread.
use_multiprocessing: Boolean. If `True`, use process-based threading. If
unspecified, `use_multiprocessing` will default to `False`. Note that
because this implementation relies on multiprocessing, you should not
pass non-picklable arguments to the generator as they can't be passed
easily to children processes.
shuffle: Boolean. Whether to shuffle the order of the batches at the
beginning of each epoch. Only used with instances of `Sequence`
(`keras.utils.Sequence`). Has no effect when `steps_per_epoch` is not
`None`.
initial_epoch: Epoch at which to start training (useful for resuming a
previous training run).
mode: One of ModeKeys.TRAIN/ModeKeys.TEST/ModeKeys.PREDICT.
batch_size: Integer batch size or None if unknown. Will only be used if
`data` is in NumPy/Tensor format.
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. `steps` is
accepted as an alias for `steps_per_epoch`.
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.
"""
if 'steps' in kwargs:
steps_per_epoch = kwargs['steps']
# Determine the number of steps per epoch and whether we should reset the
# dataset at the end of each epoch.
reset_dataset_after_each_epoch = False
original_dataset = None
is_dataset = isinstance(data,
(dataset_ops.DatasetV2, dataset_ops.DatasetV1))
if is_dataset:
original_dataset = data
if steps_per_epoch is None:
reset_dataset_after_each_epoch = True
steps_per_epoch = training_utils.infer_steps_for_dataset(
model,
data,
steps_per_epoch,
epochs=epochs,
steps_name=steps_name)
# Convert to a format that supports `next(generator)`.
generator, steps_per_epoch = convert_to_generator_like(
data,
steps_per_epoch=steps_per_epoch,
batch_size=batch_size,
epochs=epochs - initial_epoch,
shuffle=shuffle)
do_validation = validation_data is not None
is_sequence = isinstance(generator, data_utils.Sequence)
_validate_arguments(is_sequence, is_dataset, use_multiprocessing, workers,
steps_per_epoch, validation_data, validation_steps,
mode, kwargs)
batch_function = _make_execution_function(model,
mode,
class_weight=class_weight)
# Create the queue for the generator.
enqueuer = None
if not is_dataset:
generator, enqueuer = _make_enqueued_generator(
generator,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
shuffle=shuffle)
num_samples_or_steps, use_steps = _get_num_samples_or_steps(
data, steps_per_epoch)
count_mode = 'steps' if use_steps else 'samples'
callbacks = cbks.configure_callbacks(callbacks,
model,
do_validation=do_validation,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
batch_size=batch_size,
samples=num_samples_or_steps,
count_mode=count_mode,
verbose=verbose,
mode=mode)
if mode == ModeKeys.PREDICT:
aggregator = training_utils.OutputsAggregator(True,
steps=steps_per_epoch)
else:
aggregator = training_utils.MetricsAggregator(True,
steps=steps_per_epoch)
should_set_learning_phase = context.executing_eagerly(
) and model.run_eagerly
if should_set_learning_phase:
learning_phase_scope = backend.eager_learning_phase_scope(
1 if mode == ModeKeys.TRAIN else 0)
learning_phase_scope.__enter__()
callbacks.model.stop_training = False
callbacks._call_begin_hook(mode)
initial_epoch = model._maybe_load_initial_epoch_from_ckpt(
initial_epoch, mode)
for epoch in range(initial_epoch, epochs):
if callbacks.model.stop_training:
break
# Setup work for each epoch.
model.reset_metrics()
epoch_logs = {}
if mode == ModeKeys.TRAIN:
callbacks.on_epoch_begin(epoch, epoch_logs)
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_data = _get_next_batch(generator)
if batch_data is None:
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.steps = 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
# `batch_size` used for validation data if validation
# data is NumPy/EagerTensors.
batch_size = int(nest.flatten(batch_data)[0].shape[0])
# Callbacks batch begin.
batch_logs = {'batch': step, 'size': batch_size}
callbacks._call_batch_hook(mode, 'begin', step, batch_logs)
is_deferred = not model._is_compiled
batch_outs = batch_function(*batch_data)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if step == 0:
aggregator.create(batch_outs)
if is_deferred:
# Set callbacks params. We do this here when model is compiled only
# in the first iteration of this loop (deferred build scenario).
cbks.set_callback_parameters(
callbacks,
model,
do_validation=do_validation,
batch_size=batch_size,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
samples=num_samples_or_steps,
verbose=verbose,
mode=mode)
# Aggregate results.
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)
step += 1
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 training_utils.should_run_validation(
validation_freq, epoch) and not callbacks.model.stop_training):
val_results = model_iteration(
model,
validation_data,
steps_per_epoch=validation_steps,
batch_size=batch_size,
class_weight=class_weight,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
callbacks=callbacks,
verbose=verbose,
mode=ModeKeys.TEST,
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 mode == ModeKeys.TRAIN:
# Epochs only apply to `fit`.
callbacks.on_epoch_end(epoch, epoch_logs)
# Recreate dataset iterator for the next epoch.
if reset_dataset_after_each_epoch and epoch < epochs - 1:
generator = dataset_ops.make_one_shot_iterator(original_dataset)
model._successful_loop_finish = True
callbacks._call_end_hook(mode)
if enqueuer is not None:
enqueuer.stop()
if should_set_learning_phase:
learning_phase_scope.__exit__(None, None, None)
if mode == ModeKeys.TRAIN:
return model.history
return results
def model_iteration(model,
data,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
validation_freq=1,
train_class_weight=None,
val_class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=False,
initial_epoch=0,
mode=ModeKeys.TRAIN,
batch_size=None,
steps_name='steps',
**kwargs):
if 'steps' in kwargs:
steps_per_epoch = kwargs['steps']
# Determine the number of steps per epoch and whether we should reset the
# dataset at the end of each epoch.
reset_dataset_after_each_epoch = False
original_dataset = None
is_dataset = isinstance(data,
(dataset_ops.DatasetV2, dataset_ops.DatasetV1))
if is_dataset:
original_dataset = data
if steps_per_epoch is None:
reset_dataset_after_each_epoch = True
steps_per_epoch = training_utils.infer_steps_for_dataset(
data, steps_per_epoch, epochs=epochs, steps_name=steps_name)
# Convert to a format that supports `next(generator)`.
generator, steps_per_epoch = convert_to_generator_like(
data,
steps_per_epoch=steps_per_epoch,
batch_size=batch_size,
epochs=epochs - initial_epoch,
shuffle=shuffle)
do_validation = validation_data is not None
is_sequence = isinstance(generator, data_utils.Sequence)
_validate_arguments(is_sequence, is_dataset, use_multiprocessing, workers,
steps_per_epoch, validation_data, validation_steps,
mode, kwargs)
# print(train_class_weight, 'before make execution')
######################################################################
batch_function = _make_execution_function(
model,
mode,
train_class_weight=train_class_weight,
val_class_weight=val_class_weight)
######################################################################
# Create the queue for the generator.
enqueuer = None
if not is_dataset:
generator, enqueuer = _make_enqueued_generator(
generator,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
shuffle=shuffle)
num_samples_or_steps, use_steps = _get_num_samples_or_steps(
data, steps_per_epoch)
count_mode = 'steps' if use_steps else 'samples'
callbacks = cbks.configure_callbacks(callbacks,
model,
do_validation=do_validation,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
batch_size=batch_size,
samples=num_samples_or_steps,
verbose=verbose,
count_mode=count_mode,
mode=mode)
if mode == ModeKeys.PREDICT:
aggregator = training_utils.OutputsAggregator(True,
steps=steps_per_epoch)
else:
aggregator = training_utils.MetricsAggregator(True,
steps=steps_per_epoch)
should_set_learning_phase = context.executing_eagerly(
) and model.run_eagerly
if should_set_learning_phase:
learning_phase_scope = backend.eager_learning_phase_scope(
1 if mode == ModeKeys.TRAIN else 0)
learning_phase_scope.__enter__()
callbacks.model.stop_training = False
callbacks._call_begin_hook(mode)
print(initial_epoch, mode)
# TODO: mode is a bug?
# https://github.com/tensorflow/tensorflow/blob/r2.2/tensorflow/python/keras/engine/training.py
initial_epoch = model._maybe_load_initial_epoch_from_ckpt(initial_epoch)
for epoch in range(initial_epoch, epochs):
if callbacks.model.stop_training:
break
# Setup work for each epoch.
model.reset_metrics()
epoch_logs = {}
if mode == ModeKeys.TRAIN:
callbacks.on_epoch_begin(epoch, epoch_logs)
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_data = _get_next_batch(generator)
if batch_data is None:
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.steps = 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
# `batch_size` used for validation data if validation
# data is NumPy/EagerTensors.
batch_size = int(nest.flatten(batch_data)[0].shape[0])
# Callbacks batch begin.
batch_logs = {'batch': step, 'size': batch_size}
callbacks._call_batch_hook(mode, 'begin', step, batch_logs)
is_deferred = not model._is_compiled
######################################################
batch_outs = batch_function(*batch_data)
######################################################
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if step == 0:
aggregator.create(batch_outs)
if is_deferred:
# Set callbacks params. We do this here when model is compiled only
# in the first iteration of this loop (deferred build scenario).
cbks.set_callback_parameters(
callbacks,
model,
do_validation=do_validation,
batch_size=batch_size,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
samples=num_samples_or_steps,
verbose=verbose,
mode=mode)
# Aggregate results.
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)
step += 1
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 training_utils.should_run_validation(
validation_freq, epoch) and not callbacks.model.stop_training):
############################################################################
val_results = model_iteration(
model,
validation_data,
steps_per_epoch=validation_steps,
batch_size=batch_size,
val_class_weight=val_class_weight, ######## HACK
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
callbacks=callbacks,
verbose=0,
mode=ModeKeys.TEST,
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 mode == ModeKeys.TRAIN:
# Epochs only apply to `fit`.
callbacks.on_epoch_end(epoch, epoch_logs)
# Recreate dataset iterator for the next epoch.
if reset_dataset_after_each_epoch and epoch < epochs - 1:
generator = dataset_ops.make_one_shot_iterator(original_dataset)
callbacks._call_end_hook(mode)
if enqueuer is not None:
enqueuer.stop()
if should_set_learning_phase:
learning_phase_scope.__exit__(None, None, None)
if mode == ModeKeys.TRAIN:
return model.history
return results
def _configure_callbacks(self, user_callbacks: Optional[List]) -> None:
"""
If we pass a callbacks parameter to model.fit() or model.evaluate() which is a
pre-constructed CallbackList, Keras will not alter it. We can use this property to
configure the exact callback order that we want in our system.
The implementation is based closely on from the real
tf.keras.callbacks.configure_callbacks(), with the following differences:
- We always assume we have the original Callbacks list.
- We prepend and append additional Determined and Horovod callbacks
- We create a det.keras.CallbackList instead of the normal tf.keras one.
"""
callbacks = user_callbacks or []
check.is_instance(
callbacks,
list,
"the callbacks parameter of model.fit() or model.eval() must be a list of Callbacks",
)
if self.env.experiment_config.get_records_per_epoch() is None:
for cb in callbacks:
if util.is_overridden(
cb.on_epoch_end,
tf.keras.callbacks.Callback) and not getattr(
cb, "_skip_epoch_end_check", False):
if isinstance(cb, keras.callbacks.Callback):
# New callbacks must obey the rules.
raise AssertionError(
"it is unsupported to use a Callback that defines on_epoch_end "
f"({type(cb).__name__}) without setting the records_per_epoch value "
"in the experiment config")
else:
# Pre-existing callbacks only get a warning.
logging.warning(
"It is unsupported to use a Callback that defines on_epoch_end "
f"({type(cb).__name__})without setting the records_per_epoch value in "
"the experiment config. Training will continue but on_epoch_end will "
"never be called.")
# Standard post-callback from the real configure_callbacks().
# Note that we are not including BaseLogger since it is only for averaging metrics over an
# entire epoch, and we don't report any metrics in on_epoch_end at all.
self.model.history = keras.callbacks._DeterminedHistory()
callbacks = callbacks + [self.model.history]
if self.context._fit_verbose:
# Our implementation of verbose=True.
callbacks = [keras.callbacks._DeterminedProgress()] + callbacks
# Calculate batches per epoch. We can only handle batches per epoch, not records per epoch,
# because we would have to communicate after every batch to know how many records were in
# each batch on each worker in order to trigger on_epoch_end callbacks correctly.
batches_per_epoch = None
records_per_epoch = self.env.experiment_config.get_records_per_epoch()
if records_per_epoch is not None:
batches_per_epoch = records_per_epoch // self.context.get_global_batch_size(
)
# We wrap all of the callbacks in a single Multiplexer.
self.multiplexer = TrialControllerMultiplexer(
self,
callbacks,
self.is_chief,
self.batch_size,
batches_per_epoch,
self.multiplexer_load_state,
)
callbacks = [self.multiplexer]
if self.hvd_config.use:
# Horovod synchronization of initial variables should happen even before we enter our
# control loop, in case we have an initial validation requested.
callbacks = [hvd.callbacks.BroadcastGlobalVariablesCallback(0)
] + callbacks
# The remainder of Determined control logic is done with a custom CallbackList
self.callback_list = CallbackList(callbacks)
# Disable timing of callbacks in some versions of keras. This can fail in some corner-cases
# because CallbackList is not designed to allow some callbacks to call other callbacks, and
# they can interact very poorly.
if hasattr(self.callback_list, "_timing"):
self.callback_list._timing["on_train_batch_begin"] = True
self.callback_list._timing["on_train_batch_end"] = True
self.callback_list._timing["on_test_batch_begin"] = True
self.callback_list._timing["on_test_batch_end"] = True
self.callback_list._timing["on_predict_batch_begin"] = True
self.callback_list._timing["on_predict_batch_end"] = True
# callback_model is the model given to callbacks, where we should be checking for
# stop_training. In horovod dtrain or non-dtrain, it should always be self.model.
callback_model = self.model._get_callback_model()
self.callback_list.set_model(callback_model)
# Fill in bogus values for most of these... some of them are very complex to calculate.
set_callback_parameters(
self.callback_list,
self.model,
do_validation=False,
batch_size=self.batch_size,
epochs=None,
steps_per_epoch=None,
samples=None,
verbose=False,
mode=ModeKeys.TRAIN,
)
self.callback_list.model.stop_training = False
def fit_generator(model,
generator,
steps_per_epoch=None,
epochs=1,
verbose=1,
callbacks=None,
validation_data=None,
validation_steps=None,
class_weight=None,
max_queue_size=10,
workers=1,
use_multiprocessing=False,
shuffle=True,
initial_epoch=0):
"""See docstring for `Model.fit_generator`."""
epoch = initial_epoch
do_validation = bool(validation_data)
if not context.executing_eagerly():
model._make_train_function()
if do_validation:
model._make_test_function()
is_sequence = isinstance(generator, Sequence)
if not is_sequence and use_multiprocessing and workers > 1:
logging.warning(
UserWarning('Using a generator with `use_multiprocessing=True`'
' and multiple workers may duplicate your data.'
' Please consider using the`keras.utils.Sequence'
' class.'))
if steps_per_epoch is None:
if is_sequence:
steps_per_epoch = len(generator)
else:
raise ValueError('`steps_per_epoch=None` is only valid for a'
' generator based on the `keras.utils.Sequence`'
' class. Please specify `steps_per_epoch` or use'
' the `keras.utils.Sequence` class.')
is_deferred = not model._is_compiled
batch_outs = batch_function(*batch_data)
if not isinstance(batch_outs, list):
batch_outs = [batch_outs]
if step == 0:
aggregator.create(batch_outs)
if is_deferred:
# Set callbacks params. We do this here when model is compiled only
# in the first iteration of this loop (deferred build scenario).
cbks.set_callback_parameters(
callbacks,
model,
do_validation=do_validation,
batch_size=batch_size,
epochs=epochs,
steps_per_epoch=steps_per_epoch,
samples=num_samples_or_steps,
verbose=verbose,
mode=mode)
progbar.params = callbacks.params
progbar.params['verbose'] = verbose
# Aggregate results.
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
callbacks.on_train_begin()
# Construct epoch logs.
epoch_logs = {}
while epoch < epochs:
for m in model.stateful_metric_functions:
m.reset_states()
callbacks.on_epoch_begin(epoch)
steps_done = 0
batch_index = 0
while steps_done < steps_per_epoch:
generator_output = next(output_generator)
if not hasattr(generator_output, '__len__'):
raise ValueError('Output of generator should be '
'a tuple `(x, y, sample_weight)` '
'or `(x, y)`. Found: ' +
str(generator_output))
if len(generator_output) == 2:
x, y = generator_output
sample_weight = None
elif len(generator_output) == 3:
x, y, sample_weight = generator_output
else:
raise ValueError('Output of generator should be '
'a tuple `(x, y, sample_weight)` '
'or `(x, y)`. Found: ' +
str(generator_output))
# build batch logs
batch_logs = {}
if isinstance(x, list):
batch_size = x[0].shape[0]
elif isinstance(x, dict):
batch_size = list(x.values())[0].shape[0]
else:
batch_size = x.shape[0]
batch_logs['batch'] = batch_index
batch_logs['size'] = batch_size
callbacks.on_batch_begin(batch_index, batch_logs)
outs = model.train_on_batch(x,
y,
sample_weight=sample_weight,
class_weight=class_weight)
if not isinstance(outs, list):
outs = [outs]
for l, o in zip(model.metrics_names, outs):
batch_logs[l] = o
callbacks.on_batch_end(batch_index, batch_logs)
batch_index += 1
steps_done += 1
# Epoch finished.
if steps_done >= steps_per_epoch and do_validation:
if val_gen:
val_outs = evaluate_generator(
model,
validation_data,
validation_steps,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size)
else:
# No need for try/except because
# data has already been validated.
val_outs = model.evaluate(
val_x,
val_y,
batch_size=batch_size,
sample_weight=val_sample_weights,
verbose=0)
if not isinstance(val_outs, list):
val_outs = [val_outs]
# Same labels assumed.
for l, o in zip(model.metrics_names, val_outs):
epoch_logs['val_' + l] = o
if callbacks.model.stop_training:
break
callbacks.on_epoch_end(epoch, epoch_logs)
epoch += 1
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 training_utils.should_run_validation(
validation_freq, epoch) and not callbacks.model.stop_training):
val_results = model_iteration(
model,
validation_data,
steps_per_epoch=validation_steps,
batch_size=batch_size,
class_weight=class_weight,
workers=workers,
use_multiprocessing=use_multiprocessing,
max_queue_size=max_queue_size,
callbacks=callbacks,
verbose=0,
mode=ModeKeys.TEST,
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 mode == ModeKeys.TRAIN:
# Epochs only apply to `fit`.
callbacks.on_epoch_end(epoch, epoch_logs)
progbar.on_epoch_end(epoch, epoch_logs)
# Recreate dataset iterator for the next epoch.
if reset_dataset_after_each_epoch and epoch < epochs - 1:
generator = dataset_ops.make_one_shot_iterator(original_dataset)
callbacks._call_end_hook(mode)
if enqueuer is not None:
enqueuer.stop()
if should_set_learning_phase:
backend.set_eager_learning_phase(old_learning_phase)
if mode == ModeKeys.TRAIN:
return model.history
return results
