DLMBL 2023 excercise

[ziw-liu]

No description provided.

[mattersoflight]

@ziw-liu I've tested the whole exercise on dlmbl1 node. In addition to prototyping 3 solutions above, can you suggest how to log all validation samples? At this point, the subsample of validation set comes out empty in multiple runs.

With solutions in hand, I'll write the exercise prompt. Earlier exercise prompts are commented out so that you can just run the script by clicking Run Above in the last cell.

[ziw-liu]

can you suggest how to log all validation samples? At this point, the subsample of validation set comes out empty in multiple runs.

I think I know what's happening here. The validation samples are the first sample of the first log_num_samples batches. So if total validation steps is 1 it will only have 1 sample. I can modify the behavior to only use the first batch in the epoch, with the caveat that log_num_samples will then have to be smaller than batch size.

Edit: see f7229a8

[ziw-liu]

Minor language edit: these 3 metrics are actually differentiable because the developers of torchmetrics has made sure that all the operators used in the implementation are differentiable.

[mattersoflight]

@ziw-liu the conflicts in engine.py seem to be related to how you are logging the loss in two branches. Please resolve them and merge this branch into #37. I can start a smaller PR with updates from the course.

<<<<<<< dlmbl2023
        self.log("loss/validate", loss, batch_size=self.batch_size, sync_dist=True)
        if batch_idx == 0:
            self.validation_step_outputs.extend(
=======
        self.log("loss/validate", loss, sync_dist=True)
        if batch_idx < self.log_num_samples:
            self.validation_step_outputs.append(
>>>>>>> 21d-upscale-decoder
                self._detach_sample((source, target, pred))
            )