request for the training code for action tokenizer

[dzj441]

Hi GalaxeaVLA team,

Thank you for releasing the G0.5 codebase and checkpoints. I am currently studying the ActionCodec / Action Tokenizer component and would like to reproduce or further research the tokenizer training process.

From the current repository, I can see that the ActionCodecV2 implementation includes the model architecture, encode/decode logic, reconstruction loss,
commitment loss, consistency loss, and checkpoint loading support:

• configs/tokenizer/actioncodec.yaml
• src/g05/tokenizer/models/actioncodec2_v2/modeling_actioncodec2v2.py
• src/g05/tokenizer/models/actioncodec2_v2/wrapper.py
• src/g05/tokenizer/interface/vq_base.py

I also noticed a few references suggesting that an internal training entrypoint may have existed, for example:

• processor_utils.py mentions train_vq
• tests/test_dataloader_batch.py mentions a collate function from scripts/train_vq.py
• ActionCodecV2Wrapper.load_model() warns: Re-save with train_vq.py to embed it

However, I could not find the actual tokenizer training script, such as scripts/train_vq.py, or the corresponding training configs in the public
repository.

Would it be possible to release or document the missing files needed to train action_tokenizer.pt? In particular, it would be very helpful to have:

1. The tokenizer training entrypoint, e.g. scripts/train_vq.py
2. Example Hydra/config files for ActionCodecV2 training
3. Expected dataset/processor setup for tokenizer training
4. Checkpoint format details, especially tokenizer_meta
5. Recommended hyperparameters and training schedule

This would make it much easier for researchers to reproduce the ActionCodec tokenizer, adapt it to new embodiments, and study the effect of tokenizer design on G0.5 performance.

Thanks again for the release and for any guidance you can provide.

[anqiiyang]

@dzj441 Hi, we plan to released it in a couple of days, and please stay tuned

[RexGPT]

👀