[Optional] Training Your Own Home Pretrained Representations

Completely optional step, if you want to tinker with the pretrained representation or train your own model.

This step assumes you have already downloaded the HoNY RGB + actions dataset to somewhere on your machine, and have updated your configs/env_vars/env_vars.yaml accordingly.

  • Single GPU: To reproduce our HPR encoder, run in terminal:

    python train.py --config-name=train_moco
  • Multi-GPU: We use huggingface 🤗 accelerate to run multi-GPU training. Run the following commands

    accelerate config # Interactively walk you through setting up multi-GPU training
    accelerate launch train.py --config-name=train_moco

Last updated