[Port HIL-SERL] Add HF vision encoder option in SAC #651
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
ChorntonYoel
commented
Jan 21, 2025
| @@ -74,7 +74,23 @@ def make_dataset(cfg, split: str = "train") -> LeRobotDataset | MultiLeRobotData | |||
|
|
|||
| image_transforms = None | |||
| if cfg.training.image_transforms.enable: | |||
| cfg_tf = cfg.training.image_transforms | |||
| default_tf = OmegaConf.create( | |||
There was a problem hiding this comment.
Added this to have more flexibility when creating transforms. Lmk if you don't think we need it, or if you have a better idea
helper2424
reviewed
Jan 21, 2025
| @@ -27,6 +27,7 @@ | |||
| import torch.nn.functional as F # noqa: N812 | |||
| from huggingface_hub import PyTorchModelHubMixin | |||
| from torch import Tensor | |||
| from transformers import AutoModel | |||
There was a problem hiding this comment.
Need to move it under the class definition as with the Reward Classifier - https://github.com/huggingface/lerobot/pull/565/files#diff-160b98695ab8295e4ac586d8b9e50cb8e849b2bc31b2daceb28ded10580ab574R46. The reason is that by default we don't install transformers, so the library will crash without hil-serl deps installation.
| @@ -476,7 +505,13 @@ def forward(self, obs_dict: dict[str, Tensor]) -> Tensor: | |||
| # Concatenate all images along the channel dimension. | |||
| image_keys = [k for k in self.config.input_shapes if k.startswith("observation.image")] | |||
| for image_key in image_keys: | |||
| feat.append(flatten_forward_unflatten(self.image_enc_layers, obs_dict[image_key])) | |||
| if self.config.vision_encoder_name is not None: | |||
There was a problem hiding this comment.
Just a small refactoring - let's extract to a separate method has_pretrained_visual_encoder
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.
What this does
Adds a pretrained vision encoder from HF to the sac modeling + allows for appropriate transforms when training.
How it was tested
Ran a sac training with moss.
My sac training script and associated modeling have slight differences, so it would be good to double check with a training loop incorporating real images and the current sac training script.