frameworks: Use log_artifact

[daavoo]

Once #378 is done, it would be nice to automatically use it under the hood in the frameworks where we currently support saving the model, i.e.:

dvclive/src/dvclive/keras.py

Lines 49 to 53 in 41d12d7

	if self.model_file:
	if self.save_weights_only:
	self.model.save_weights(self.model_file)
	else:
	self.model.save(self.model_file)

To:

        if self.model_file:
            if self.save_weights_only:
                self.model.save_weights(self.model_file)
            else:
                self.model.save(self.model_file)
            self.live.log_artifact(self.model_file)

[dberenbaum]

Even better would be if DVCLive can inspect if other callbacks are being used to save the model and use log_artifact to track those paths.

[dberenbaum]

See also #472: if we have a log_model method, obviously let's use that here instead of the generic log_artifact.

[dberenbaum]

@daavoo Can you take this as a follow-up to #464?

[daavoo]

@daavoo Can you take this as a follow-up to #464?

Yep! was already planning on it 🚀
I was also counting on including the type=model meta field as it doesn't really break anything to include it

[dberenbaum]

I was also counting on including the type=model meta field as it doesn't really break anything to include it

Thanks! First need to settle the discussion on how to handle type=model from iterative/mlem.ai#323 (comment) / https://iterativeai.slack.com/archives/CB41NAL8H/p1678197700155189.

[SH2282000]

If the file is part of the outputs of a stage, then the self.live.log_artifact(self.model_file) will cause the following warning to be thrown:

WARNING:dvclive:Failed to dvc add model.h5: cannot update 'model.h5': not a data source

Therefore, what should be the best practice in terms of staging? Does the model.h5 should be stored as a standalone file, or how the model_file can be used without throwing warnings while still being part of the outs:

stages:
  train:
    [...]
    outs:
      - model.h5
      - [...]

Solutions found ()

In those solution, a model.h5.dvc is added (this is not what I want):

• Adding the model.h5 via dvc add as a standalone (generating the model.h5.dvc).
• Not saving the model (will automatically call the dvc add).

In this other solution, the file is being saved only after the training (this is not what I want either):

• Using keras model.save() function only at the end of the training (losing the ability to save the model for each epoch).

[daavoo]

r how the model_file can be used without throwing warnings while still being part of the outs:

Hi @SH2282000 , the dvc.yaml you shared is correct and the warnings should not appear in the latest version of DVCLive.