delu.random.get_state

delu.random.get_state()

Aggregate global random states from random, numpy and torch.

The function is useful for creating checkpoints and allows resuming activities dependent on global random states (e.g. data streams) in a reproducible manner. Also, see the note below. The result of this function can be passed to set_state.

Returns:

state

Return type:

Dict[str, Any]

See also

delu.random.set_state

Example:

# Resuming from a checkpoint:
checkpoint_path = ...
if checkpoint_path.exists():
    checkpoint = torch.load(checkpoint_path)
    delu.random.set_state(checkpoint['random_state'])
...
# Training:
for batch in batches:
    ...
    if step % checkpoint_frequency == 0:
        torch.save(
            {
                'model': ...,
                'optimizer': ...,
                'random_state': delu.random.get_state(),
            },
            checkpoint_path,
        )