Placing a condition on ModelCheckpoint Callback

[skrish13]

For example: How to save checkpoints ONLY when valset accuracy is greater than some specific threshold?

If this is not possible using parameters in ModelCheckpoint callback, how can we extend ModelCheckpoint to achieve this?

[rohitgr7]

you can do:

class CustomModelCheckpoint(ModelCheckpoint):
    def on_validation_end(self, trainer, pl_module):
        score = trainer.callback_metrics[self.monitor]
        if score > some_threshold:
            super().on_validation_end(trainer, pl_module)


trainer = Trainer(callbacks=[CustomModelCheckpoint(...)])