Naming .pth checkpoints using PyTorch Ignite ModelCheckpoint handler

ignite
1

I use the ModelCheckpoint handler in Ignite to save the model using these lines

ckpt_handler = ModelCheckpoint(f'./experiments/{log_num}/checkpoints', filename_prefix='epoch_', n_saved=None, create_dir=True)
trainer.add_event_handler(Events.EPOCH_COMPLETED(every=2), ckpt_handler, to_save)

but the .pth filename follow this pattern --> epoch__checkpoint_344.pth (344 refers to iterations).

I would to include the number of epoch into the filename… Any idea how’s that done?

2

@xen0f0n it can be done by adding a global_step_transform=lambda e, _: e.state.epoch

checkpointer = ModelCheckpoint(*args, **kwargs, global_step_transform=lambda e, _: e.state.epoch)
1 Like
3

Thank you @vfdev-5! Not sure how this works… I tried accessing trainer.state.epoch (before trainer.run) and state was None. Is this what that lambda does? Accessing trainer engine state after trainer.run?

4

Yes, state at the begining of the training is None as it is not defined.
When you attached

trainer.add_event_handler(Events.EPOCH_COMPLETED(every=2), ckpt_handler, to_save)

once every 2 epoch, ckpt_handler is triggered to save what to save. Its argument global_step_transform is an optional callable that

global step transform function to output a desired global step. Input of the function is (engine, event_name). Output of function should be an integer. Default is None, global_step based on attached engine. If provided, uses function output as global_step.

So it is just executed during the training, where trainer’s state is initialized and can be used.

Please, see the docs for info : ignite.handlers — PyTorch-Ignite v0.4.13 Documentation

Hope this helps :slight_smile: