How to configure experiments

Our experiments are managed through Hydra configs.

The hydra configuration hierarchy looks like:

configs/
    config.yaml          # the default config file that summarizes all configs
    data/default.yaml    # the configs related to data
    eval/default.yaml    # the configs related to evaluation
    lifelong
        base.yaml        # the sequential finetuning baseline configs
        agem.yaml        # the agem configs
        er.yaml          # the er configs
        ewc.yaml         # the ewc configs
        packnet.yaml     # the packnet configs
        multitask.yaml   # the multitask learning configs
        single_task.yaml # the single task learning configs
    policy
        data_augmentation
        image_encoder
        language_encoder
        policy_head
        position_encoding
        bc_rnn_policy.yaml          # the config for ResNet-LSTM
        bc_transformer_policy.yaml  # the config for ResNet-Transformer
        bc_vilt_policy.yaml         # the config for ViT-Transformer
    train
        optimizer
        shceduler
        default.yaml      # the configs related to training

If you want to modify any existing configuration, you can directly do that in command line instead of modifying the original yaml file. For instance, consider

export CUDA_VISIBLE_DEVICES=GPU_ID && \
export MUJOCO_EGL_DEVICE_ID=GPU_ID && \
python lifelong/main.py seed=SEED \
                        benchmark_name=BENCHMARK 
                        policy=POLICY \
                        lifelong=ewc \
                        lifelong.e_lambda=100000 \
                        lifelong.gamma=0.95

This will change the e_lambda and gamma configs of ewc to 100000 and 0.95 respectively.