Conditional diffusion part1 #49
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
…re network base class to do conditional forward randomly
rousseab
reviewed
Jun 3, 2024
| @@ -42,6 +43,9 @@ class ScoreNetworkParameters: | |||
| """Base Hyper-parameters for score networks.""" | |||
| architecture: str | |||
| spatial_dimension: int = 3 # the dimension of Euclidean space where atoms live. | |||
| conditional_prob: float = 0. # probability of making an conditional forward - else, do a unconditional forward | |||
There was a problem hiding this comment.
'an conditional' -> 'a conditional'
'a unconditional' -> 'an unconditional'
rousseab
reviewed
Jun 3, 2024
| """Model forward. | ||
|
|
||
| Args: | ||
| batch : dictionary containing the data to be processed by the model. | ||
| conditional: if True, do an conditional forward, if False, do a unconditional forward. If None, choose |
There was a problem hiding this comment.
'a unconditional' -> 'an unconditional'
rousseab
reviewed
Jun 3, 2024
rousseab
requested changes
Jun 3, 2024
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Outdated suggestions cannot be applied.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.
in this PR, I implement a conditional version of the score network based on the classifier-free guidance approach. In short, the modified score becomes:
where I used \gamma = 1 + w instead to match the convention in MatterGen (hyper-parameter in the config file - set to 2 by default as per MatterGen). The second term on the right-hand side is the usual score network. The first one is a modified version that takes a condition c - here the cartesian forces - as an input. This modification takes the form of an added value in MatterGen:
where H is the node features at a given layer in the GNN used (GemNet in the case of MatterGen). Here, I implemented the added value as a linear layer on top of each layer of the MLP approach. I didn't tackle MACE yet.
During training, a forward for a batch can be done unconditionally (as if the property was null) or conditionally. The probability is set by an hyper-parameter in the config file. This is different from MatterGen where training was first done unconditionally, then fine-tuned on the conditions.