How to write custom callback with monitor

[ZENGYIMING-EAMON]

I am using PL-1.6.1.

I am using the official pl.callbacks.ModelCheckpoint with monitor: 'some_lss/dataloader_idx_1', mode: 'min' and it works fine.

Now I write a custom callback class CustomCallback(pl.Callback): to save some images at the end of train epoch,
by writing def on_train_epoch_end(self,trainer,pl_module): under class CustomCallback(pl.Callback):.

Now I save it every 10 epochs using if trainer.current_epoch%50==0: under def on_train_epoch_end(self,trainer,pl_module):.

But I want to change it to save every time the model meets a lower loss, i.e., every time the model meets a min 'some_lss/dataloader_idx_1'. How can I incorporate the monitor: 'some_lss/dataloader_idx_1', mode: 'min' into my custom callback?

[rohitgr7]

you can access this some_lss/dataloader_idx_1 using trainer.callback_metrics['some_lss/dataloader_idx_1'] and use on_validation_end hook to write your logic to checkpoint if the new value is lower than the current lowest.

class CustomCallback(pl.Callback):

    def __init__(self, mode='min', monitor='some_lss/dataloader_idx_1'):
        ...
        torch_inf = torch.tensor(np.Inf)
        mode_dict = {"min": torch_inf, "max": -torch_inf}
        self.monitor_op = {"min": torch.lt, "max": torch.gt}[mode]
        self.best_metric_value = mode_dict[mode]

    def on_validation_end(self, trainer, pl_module):
        current_metric = trainer.callback_metrics[self.monitor]
        if self.monitor_op(current_metric, self.best_metric_value):
            self.best_metric_value = current_metric
            # do whatever you want