Does PL support customizing indicators in checkpoint callback？

[morestart]

this is my val func:

    def validation_step(self, batch, batch_idx):
        data = batch

        output = self.forward(data['img'])
        boxes, scores = self.postprocess(output.cpu().numpy(), batch['shape'])
        raw_metric = self.metric(batch, (boxes, scores))

        return raw_metric

    def validation_epoch_end(self, outputs):

        metric = self.metric.gather_measure(outputs)
        self.log('recall', value=metric['recall'].avg)
        self.log('precision', value=metric['precision'].avg)
        self.log('hmean', value=metric['fmeasure'].avg)
        return {'hmean': metric['fmeasure'].avg}

this is my checkpoint callback:

checkpoint_callback = ModelCheckpoint(
    monitor='hmean',
    mode='max',
    dirpath='../det_weights',
    filename='{epoch:02d}-{hmean:.2f}',
    save_last=True,
)
checkpoint_callback.FILE_EXTENSION = '.pt'

use this callback, i can only get a epoch=00-hmean=0.00.pt file, It does not record the value of heman

[morestart]

i solvd this problem, thanks...

[tchaton]

Hey @morestart. Mind sharing what was the solution for people coming on this discussion ?

[morestart]

thanks for your reply, question description is the answer... I ran a wrong project.... hahaha, i am so stupid😂

[morestart]

emmm, remeber set mode... the default value is min, if your metric ues max value, it will be very sad. 😉