pytorch_tao.plugins.Checkpoint

class pytorch_tao.plugins.Checkpoint(metric_name: str, objects: Dict, score_sign: int = 1, n_saved=3)

Save models or any other states. This is just a wrapper of ignite.handlers.checkpoint.Checkpoint.

Parameters

• metric_name – the name of the metric to determine which is the best model to save.

• objects – objects to save.

• score_sign – 1 higher the better or -1 lower the better.

• n_saved – how many file to keep.

Hooks

Hook Point	Logic
EPOCH_COMPLETED	saving the states

import pytorch_tao as tao

model = ...
trainer = tao.Trainer()

# save the top-3 models of accuracy
trainer.use(Checkpoint("accuracy", {"model": model}))

__init__(metric_name: str, objects: Dict, score_sign: int = 1, n_saved=3)

Methods

__init__(metric_name, objects[, score_sign, ...])
after_use()
attach(engine)
set_engine(engine)

Attributes