Open-source codebase for forward-inverse cycle consistency (FICC), from "Become a Proficient Player with Limited Data through Watching Pure Videos" at ICLR 2023.
This response use Atari replay dataset as pretrain data.
Pretrain: bash pretrain.sh breakout
| Arguments | Description |
|---|---|
--env |
Name of the pretrain environment |
--dataset_path |
Path of the dataset. |
--device |
Select working GPU. |
--batch_size |
Mini-batch size. |
--lr |
Initiate learning rate for pretrain. |
--latent_action_dim |
Dimension for latent action |
--num_embeddings |
Num of embeddings of latent action generator (LAG) |
--l1_penalty |
L1 penalty for output of dynamics function |
--weight_decay |
L2 penalty for model parameters |
@inproceedings{ye2023become,
title={Become a Proficient Player with Limited Data through Watching Pure Videos},
author={Weirui Ye, and Yunsheng Zhang, and Pieter Abbeel, and Yang Gao},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023}
}