def fit(self,
summary: Optional[str] = None,
warmup: bool = True,
eager: bool = False) -> Optional[Summary]:
"""Train the network for the number of epochs specified by the estimator's constructor.
Args:
summary: A name for the experiment. If provided, the log history will be recorded in-memory and returned as
a summary object at the end of training.
warmup: Whether to perform warmup before training begins. The warmup procedure will test one step at every
epoch where schedulers cause the execution graph to change. This can take some time up front, but can
also save significant heartache on epoch 300 when the training unexpectedly fails due to a tensor size
mismatch.
eager: Whether to run the training in eager mode. This is only related to TensorFlow training because
PyTorch by nature is always in eager mode.
Returns:
A summary object containing the training history for this session iff a `summary` name was provided.
"""
os.environ[
'TF_CPP_MIN_LOG_LEVEL'] = '1' # Prevent tf from constantly printing useless information
draw()
self.system.reset(summary, self.fe_summary())
self._prepare_traces(run_modes={"train", "eval"})
if warmup:
self._warmup(eager=eager)
self._start(run_modes={"train", "eval"}, eager=eager)
return self.system.summary or None
def fit(self, summary: Optional[str] = None, warmup: Union[bool, str] = True) -> Optional[Summary]:
"""Train the network for the number of epochs specified by the estimator's constructor.
Args:
summary: A name for the experiment. If provided, the log history will be recorded in-memory and returned as
a summary object at the end of training.
warmup: Whether to perform warmup before training begins. The warmup procedure will test one step at every
epoch where schedulers cause the execution graph to change. This can take some time up front, but can
also save significant heartache on epoch 300 when the training unexpectedly fails due to a tensor size
mismatch. When set to "debug", the warmup will be performed in eager execution for easier debugging.
Returns:
A summary object containing the training history for this session iff a `summary` name was provided.
"""
draw()
self.system.reset(summary, self.fe_summary())
self._prepare_traces(run_modes={"train", "eval"})
if warmup:
self._warmup(warmup=warmup)
self._start(run_modes={"train", "eval"})
return self.system.summary or None