Fine-tuning Policies

Fine-tuning policies with fresh demonstrations that you have collected.

Training Policies

The following assumes that the current working directory is this repository’s root folder.

Training a Behavior Cloning Policy

1. Modify include_task and include_env in finetune.yaml depending on the task and env you intend to finetune.

2. [Optional, non-default:] only if you're using torch encoder, set enc_weight_pth (path to pretrained encoder weights) in image_bc_depth.yaml. You can download the weights from https://dl.dobb-e.com/models/hpr_model.pt if you don't have them.

3. Run in terminal:

   python train.py --config-name=finetune

4. [Optional, experimental] If you want to take advantage of multi-GPU training using 🤗 accelerate, you can use the following command:

   accelerate config # Only the first time, to configure accelerate
   accelerate launch train.py --config-name=finetune