def __init__(
self,
model: tf.keras.models.Model,
session: tf.compat.v1.ConfigProto,
train_config: keras.TFKerasTrainConfig,
*args: Any,
**kwargs: Any,
) -> None:
super().__init__(*args, **kwargs)
self.model = model
self.session = session
# Configure optimizers, done for backwards compatibility.
self.context._select_optimizers()
keras._check_if_aggregation_frequency_will_work(
model=self.model, hvd_config=self.hvd_config)
self._train_input_manager, self._validation_input_manager = keras._init_input_managers(
context=self.context, train_config=train_config)
# If a load path is provided, load weights and restore the data location.
self._load()
self._configure_callbacks(train_config.callbacks)
self.train_response_func = None # type: Optional[workload.ResponseFunc]
self.train_workload_metrics = [] # type: List[Dict[str, Any]]
self.train_workload_batches = 0
self.train_workload_inputs = 0
self.train_workload_len = 0
self.test_inputs = 0
def __init__(
self,
model: tf.keras.models.Model,
session: tf.compat.v1.ConfigProto,
train_config: keras.TFKerasTrainConfig,
*args: Any,
**kwargs: Any,
) -> None:
super().__init__(*args, **kwargs)
self.model = model
self.session = session
# Configure optimizers, done for backwards compatibility.
self.context._select_optimizers()
keras._check_if_aggregation_frequency_will_work(
model=self.model, hvd_config=self.hvd_config
)
self.training_data = train_config.training_data
self.validation_data = train_config.validation_data
# Support the deprecated SequenceAdapter API.
if isinstance(self.training_data, keras.SequenceAdapter):
self.context._configure_fit(
workers=self.training_data.workers,
use_multiprocessing=self.training_data.use_multiprocessing,
max_queue_size=self.training_data.max_queue_size,
)
# Use the provided Sequence directly.
self.training_data = self.training_data.sequence
if isinstance(self.validation_data, keras.SequenceAdapter):
# Ignore these settings and use the same settings as for the fit call.
self.validation_data = self.validation_data.sequence
self._check_training_data()
self._check_validation_data()
self.enqueuers = [] # type: List[keras._Enqueuer]
# If a load path is provided, load weights and restore the data location.
self._load()
self._configure_callbacks(train_config.callbacks)
self.train_response_func = None # type: Optional[workload.ResponseFunc]
self.train_workload_metrics = [] # type: List[Dict[str, Any]]
self.train_workload_batches = 0
self.train_workload_inputs = 0
self.train_workload_len = 0
self.test_inputs = 0
def __init__(
self,
model: tf.keras.models.Model,
session: tf.compat.v1.ConfigProto,
train_config: keras.TFKerasTrainConfig,
trial: "TFKerasTrial",
*args: Any,
**kwargs: Any,
) -> None:
super().__init__(*args, **kwargs)
self.model = model
self.session = session
self.trial = trial
# Configure optimizers, done for backwards compatibility.
self.context._select_optimizers()
keras._check_if_aggregation_frequency_will_work(
model=self.model,
use_horovod=self.use_horovod,
aggregation_frequency=self.context._aggregation_frequency,
)
self.training_data = train_config.training_data
self.validation_data = train_config.validation_data
# Support the deprecated SequenceAdapter API.
if isinstance(self.training_data, keras.SequenceAdapter):
self.context._configure_fit(
workers=self.training_data.workers,
use_multiprocessing=self.training_data.use_multiprocessing,
max_queue_size=self.training_data.max_queue_size,
)
# Use the provided Sequence directly.
self.training_data = self.training_data.sequence
if isinstance(self.validation_data, keras.SequenceAdapter):
# Ignore these settings and use the same settings as for the fit call.
self.validation_data = self.validation_data.sequence
if self.context.distributed.size > 1:
assert self.use_horovod, (
"TF Keras trial must be launched with a horovod backend if "
"doing distributed training"
)
self._check_training_data()
self._check_validation_data()
self.enqueuers = [] # type: List[keras._Enqueuer]
self.wlsq = None # type: Optional[layers.WorkloadSequencer]
if self.workloads is None:
self.workloads, self.wlsq = layers.make_compatibility_workloads(
self.context._core,
self.env,
self.context.get_global_batch_size(),
)
# If a load path is provided, load weights and restore the data location.
self.multiplexer_load_state = None # type: Optional[Dict]
if self.env.latest_checkpoint is not None:
logging.info(f"Restoring trial from checkpoint {self.env.latest_checkpoint}")
with self.context._core.checkpoint.restore_path(
self.env.latest_checkpoint
) as load_path:
self._load(load_path)
self._configure_callbacks(train_config.callbacks)
self.train_response_func = None # type: Optional[workload.ResponseFunc]
self.train_workload_metrics = [] # type: List[Dict[str, Any]]
self.train_workload_batches = 0
self.train_workload_inputs = 0
self.train_workload_len = 0
self.test_inputs = 0
self.steps_completed = self.env.steps_completed
