Checkpointing on structure factor #79
Merged
Conversation
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
This refactor lets the lightning module compute the various ks metrics, which used to be computed in a callback. The callback now just records to disck / plots the results. The configuration has also been modified to reflect these changes.
Merged
sblackburn86
approved these changes
Sep 25, 2024
|
|
||
| def training_step(self, batch, batch_idx): | ||
| """Runs a prediction step for training, returning the loss.""" | ||
| logger.info(f" - Starting training step with batch index {batch_idx}") |
There was a problem hiding this comment.
as discussed - probably better to remove
| return output | ||
|
|
||
| if self.metrics_parameters.compute_energies: | ||
| logger.info(" * registering reference energies") |
There was a problem hiding this comment.
same as for training step logger.info
| self.energy_ks_metric.register_reference_samples(reference_energies.cpu()) | ||
|
|
||
| if self.metrics_parameters.compute_structure_factor: | ||
| logger.info(" * registering reference distances") |
There was a problem hiding this comment.
remove,debug or optional (i.e. remove)
| reference_distances.cpu() | ||
| ) | ||
|
|
||
| logger.info(f" Done validation step with batch index {batch_idx}") |
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
2 participants
Add this suggestion to a batch that can be applied as a single commit.
This suggestion is invalid because no changes were made to the code.
Suggestions cannot be applied while the pull request is closed.
Suggestions cannot be applied while viewing a subset of changes.
Only one suggestion per line can be applied in a batch.
Add this suggestion to a batch that can be applied as a single commit.
Applying suggestions on deleted lines is not supported.
You must change the existing code in this line in order to create a valid suggestion.
Outdated suggestions cannot be applied.
This suggestion has been applied or marked resolved.
Suggestions cannot be applied from pending reviews.
Suggestions cannot be applied on multi-line comments.
Suggestions cannot be applied while the pull request is queued to merge.
Suggestion cannot be applied right now. Please check back later.
I refactored how the Kolmogorov Smirnov (KS) values are computed. This is now done in the model lightning module instead of in a callback. This is necessary in order to log these values and use them for checkpoint monitoring and/or early stopping (the call order of callbacks / model hooks is indertermined, so we can't rely on a callback to log metrics). We now get more telemetry during training on just what the heck is going on (Kudos on the structure factor, btw, it's very nice).
What used to be the sampling callback has been refactored to visualize and write to disk what is already calculated in the model lightning module. This relies on internal properties (ie, "states"), which is not super clean, but it's the PL way.
I also refactor the configuration file in a way that breaks backward compatibility. I didn't not fix every example / sanity check / off-broadway code bits we have. See the following file for an example:
examples/config_files/diffusion/config_diffusion_mlp.yaml
Future possible improvements might be to compute the validation set distances and energies only once (write them in a tmp cache file perhaps) instead of wasting time redoing this at every epoch.