Can I merge multiple ModelCheckpoints?

[ShaneTian]

This is the pseudo code about my multiple ModelCheckpoints:

model = "T5"
save_steps = 100
monitor = "micro_f1"

# setup callbacks
callbacks = []
val_ckpt_callback = pl.callbacks.ModelCheckpoint(
    filename=model + "-val_end-{epoch}-{step}-{val_loss:.4f}-{val_ppl:.4f}",
    save_top_k=-1,
    every_n_epochs=1
)
callbacks.append(val_ckpt_callback)

if save_steps > 0:
    step_ckpt_callback = pl.callbacks.ModelCheckpoint(
        filename=model + "-train_step-{epoch}-{step}-{val_loss:.4f}-{val_ppl:.4f}",
        save_last=True,
        save_top_k=-1,
        every_n_train_steps=save_steps
    )
    callbacks.append(step_ckpt_callback)

if monitor:
    monitor_ckpt_callback = pl.callbacks.ModelCheckpoint(
        filename=model + "-monitor-{epoch}-{step}-{val_loss:.4f}-{val_ppl:.4f}-{" + monitor + ":.4f}",
        monitor=monitor,
        save_top_k=1,
        mode="max"
    )
    callbacks.append(monitor_ckpt_callback)

trainer = pl.Trainer(
    # ...
    callbacks=callbacks,
    # ...
)

This looks wordy. I want to know if there is a better way to achieve the exact same functionality.
Thanks :)

[rohitgr7]

hey @ShaneTian !

yes, this should work. Modelcheckpoint is a callback that is independent of other callbacks, so individual instances for it will work.

[ShaneTian]

Thank you @rohitgr7 !

Yes, the above code works. So I want to know can I merge them to a ModelCheckpoint maybe?

[rohitgr7]

sorry, what do you mean by merging here?

[ShaneTian]

I mean can I use one ModelCheckpoint to merge the first two ModelCheckpoints:

model = "T5"
save_steps = 100
monitor = "micro_f1"

# setup callbacks
callbacks = []
val_and_step_ckpt_callback = pl.callbacks.ModelCheckpoint(
    filename=model + "-val_and_step-{epoch}-{step}-{val_loss:.4f}-{val_ppl:.4f}",
    save_last=True,
    save_top_k=-1,
    every_n_epochs=1,
    every_n_train_steps=save_steps if save_steps > 0 else 0
)
callbacks.append(val_and_step_ckpt_callback)

if monitor:
    monitor_ckpt_callback = pl.callbacks.ModelCheckpoint(
        filename=model + "-monitor-{epoch}-{step}-{val_loss:.4f}-{val_ppl:.4f}-{" + monitor + ":.4f}",
        monitor=monitor,
        save_top_k=1,
        mode="max"
    )
    callbacks.append(monitor_ckpt_callback)

trainer = pl.Trainer(
    # ...
    callbacks=callbacks,
    # ...
)

[rohitgr7]

filename, monitor is different in both, so they will run as independent checkpoint callbacks.

[ShaneTian]

OK, thank you