how to finetune this smolvla with my own dataset?

#2
by Xiaoyan97 - opened

Has your model worked? In my attempt, I recorded 40 action records at the same position and trained 20,000 steps using the official fine-tuning code. However, my model was unable to complete the actions.

python lerobot/scripts/train.py   --policy.path=lerobot/smolvla_base   --dataset.repo_id=Xiaoyan97/pick_up_scissors_v1   --dataset.root=Xiaoyan97/pick_up_scissors_v1   --batch_size=64   --steps=20000   --output_dir=outputs/smolvla_v1   --job_name=my_smolvla_training   --policy.device=cuda

Is it that my training steps are insufficient or for some other reason?

Owner
This comment has been hidden (marked as Spam)

Has your model worked? In my attempt, I recorded 40 action records at the same position and trained 20,000 steps using the official fine-tuning code. However, my model was unable to complete the actions.

python lerobot/scripts/train.py   --policy.path=lerobot/smolvla_base   --dataset.repo_id=Xiaoyan97/pick_up_scissors_v1   --dataset.root=Xiaoyan97/pick_up_scissors_v1   --batch_size=64   --steps=20000   --output_dir=outputs/smolvla_v1   --job_name=my_smolvla_training   --policy.device=cuda

Is it that my training steps are insufficient or for some other reason?

@Xiaoyan97
・Training steps should be sufficient, as the policy trained for 4000 steps can already somewhat grasp the candy.
・Also, the cameras on wrist and top seems appropriate

Did you also provide language instructions during data collection and when running the policy?

I used top and wrist cameras, and the details can be found in my dataset. Before you reply, I increased the number of training steps to see if it worked. Once it works, I will give you feedback.

Sign up or log in to comment