domhel/studytable_open_drawer_depth_no_aug

Weights with supposedly no data augmentation.

WARNING: not recommended for use: has image crop 84x84 and for some reason exact same loss as with data augmentation

Train script: https://github.com/huggingface/lerobot/blob/8e2a39444255d62869d2fd63ea4f8a157236a029/lerobot/scripts/train.py

Train command (incomplete): python /workspace/.venv/lib/python3.13/site-packages/lerobot/scripts/train.py --dataset.repo_id=${repo_id} --dataset.root=${repo_id} --policy.type=diffusion --output_dir=outputs/train/${job_name} --job_name=${job_name} --policy.device=cuda --wandb.enable=true and then add transforms.none or something like that (see source code for diffusion config / pretrained config params)

Wandb: https://wandb.ai/vsaucemichaelhere00-university-of-stuttgart/lerobot/runs/c7n113qa/workspace?nw=nwuservsaucemichaelhere00

Loss: