fix: small fixes used when finetuning a model #45
Conversation
avelinoapheris
commented
Nov 18, 2025
avelinoapheris
commented
- Allow training without templates. The current setup will crash if the templates directories are not given, whoever we find use cases where we will like to train without them.
- Allows running a validation dataset whose cache do not have "token_count" or "use_metric" fields. This is useful if one wants to run .validate on the training dataset.
- Fix a bug that requires the "ema" weights to be present when loading a checkpoint manually. There are cases where we want to start from a public checkpoint with only the weights of the model.
- Expose all the arguments for the lightning checkpoint callback.
…oint without ema weights + add extra options for the Checkpoint callback + allows running .validate with caches that do not have token size or use metric flag
jnwei
left a comment
jnwei
left a comment
There was a problem hiding this comment.
Thank you for adding the fixes for supporting fine tuning and enabling computation of metrics on the training set.
One suggestion: would it be possible to add a test to check that the case where dataset_config.dataset_paths.template_structure_array_directory is None works as intended? Perhaps a test that is a variation of the tests that are used to test training dataset runner yamls would be suitable, e.g. https://github.com/aqlaboratory/openfold-3/blob/main/openfold3/tests/test_entry_points.py#L63
| k = np.min([np.random.randint(0, l), n_templates]) | ||
|
|
||
| if k > 0: | ||
| if (k > 0) and (template_cache_directory is not None): |
There was a problem hiding this comment.
Could the parentheses around the individual statements be removed? I believe this will cause an issue for our ruff formatter.
| class CheckpointConfig(BaseModel): | ||
| """Settings for training checkpoint writing.""" | ||
|
|
||
| model_config = PydanticConfigDict(extra="allow") |
There was a problem hiding this comment.
Is extra="allow" required to support backwards compatibility with old versions of the code?
We generally prefer to set extra="forbid" to help catch instances where a field might be ignored if it is not recognized by the Model (e.g. in the case of a typo or a new field).
2 participants
