Loss instability when resuming pretraining from converted HF weights #60
Description
Thanks for your brilliant work.
I am currently trying to run Stage 2 - Pretraining of the Generalist Policy on the Bridge V2 dataset. Since the weights of univla-7b-bridge-pt in UniVLA’s Prismatic format were not released (only the Hugging Face format was made available), I implemented a script convert_hf_to_univla_weights.py (referring to vla-scripts/extern/convert_openvla_weights_to_hf.py) to convert the released HF weights back into UniVLA Prismatic format.
Using the converted weights, I resumed training with settings --resume_step 0 --resume_epoch 0.
Theoretically, since the weights are already pretrained, I would expect the training loss to start from a well-converged state.
However, I observed that the loss is quite unstable. For first 400 steps, most of the loss values lie in the 0.5–2.0 range, some steps even exceed 3.
I would like to confirm:
1. Is this kind of loss behavior normal in your experience?
2. When you resumed from the official UniVLA Prismatic weights with resume_step / resume_epoch specified, what range of loss values did you typically observe?
Any clarification would be greatly appreciated.
