def resume_training(out: ModelDir,
notes: str = None,
dry_run=False,
start_eval=False):
""" Resume training an existing model """
train_params = out.get_last_train_params()
model = out.get_model()
train_data = train_params["data"]
evaluators = train_params["evaluators"]
params = train_params["train_params"]
params.num_epochs = 24 * 3
if isinstance(train_data, PreprocessedData):
# TODO don't hard code # of processes
train_data.preprocess(6, 1000)
latest = tf.train.latest_checkpoint(out.save_dir)
if latest is None:
raise ValueError("No checkpoint to resume from found in " +
out.save_dir)
_train(model, train_data, latest, None, False, params, evaluators, out,
notes, dry_run, start_eval)
def resume_training(model_to_resume: str,
dataset_oversampling: Dict[str, int],
checkpoint: Optional[str] = None,
epochs: Optional[int] = None):
"""Resume training on a partially trained model (or finetune an existing model)
:param model_to_resume: path to the model directory of the model to resume training
:param dataset_oversampling: dictionary mapping dataset names to integer counts of how much
to oversample them
:param checkpoint: optional string to specify which checkpoint to resume from. Uses the latest
if not specified
:param epochs: Optional int specifying how many epochs to train for. If not detailed, runs for 24
"""
out = ModelDir(model_to_resume)
train_params = out.get_last_train_params()
evaluators = train_params["evaluators"]
params = train_params["train_params"]
params.num_epochs = epochs if epochs is not None else 24
model = out.get_model()
notes = None
dry_run = False
data = prepare_data(model, TrainConfig(), dataset_oversampling)
if checkpoint is None:
checkpoint = tf.train.latest_checkpoint(out.save_dir)
_train_async(model=model,
data=data,
checkpoint=checkpoint,
parameter_checkpoint=None,
save_start=False,
train_params=params,
evaluators=evaluators,
out=out,
notes=notes,
dry_run=dry_run,
start_eval=False)
