Stateless scheduler #8
Open
Conversation
This file contains 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
Also, make it compatible with any number of parallel devices instead of assuming 8. This is useful to me while testing in my computer with two different GPUs.
|
|
||
| sample = state.sample | ||
| model_output = state.model_output | ||
| prev_sample = self._get_prev_sample(state, sample, timestep, prev_timestep, model_output) | ||
|
|
||
| prev_sample = self._get_prev_sample(sample, timestep, prev_timestep, model_output) | ||
| self.counter += 1 |
There was a problem hiding this comment.
do we still use the self.counter variable?
There was a problem hiding this comment.
Also we should return the state somewhere no?
|
That's a great PR @pcuenca! Think some clean-ups are left to make it functional, but overall the design seems like the correct design to me |
|
@patrickvonplaten You are right! I stashed my changes to apply them in a new branch and messed up somehow. I have just pushed the correct version, I hope. Sorry about that! |
patrickvonplaten
approved these changes
Sep 2, 2022
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
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Outdated suggestions cannot be applied.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.
This is the biggest change derived from #5.
In that branch I was previously using a standard Python dataclass and a
dictto manage the scheduler state, and I converted between them when needed. I have now adoptedflax.struct.dataclass, which is essentially the same but:I'm not super happy with the result for these reasons:
scheduler.set_timesteps()requires more information to initialize the state, such as the shape of the expected inputs. This makes the code a bit uglier because you need to prepare that in advance.betasand thenum_training_steps, and then thereplacemethod doesn't work. I had to create a specific class method to simplify the initial creation of the instance.stepare harder to follow than the PyTorch logic.Note also that the initial PRK steps have not been implemented yet (since we are skipping them by default). If we go with this approach they should be easy to add.