iVideoGPT: Interactive VideoGPTs are Scalable World Models (NeurIPS 2024)

[Website] [Paper] [Model]

This repo provides official code and checkpoints for iVideoGPT, a generic and efficient world model architecture that has been pre-trained on millions of human and robotic manipulation trajectories.

News

🚩 2024.09.26: iVideoGPT has been accepted by NeurIPS 2024, congrats!
🚩 2024.08.31: Training code is released (Work in progress 🚧 and please stay tuned!)
🚩 2024.05.31: Project website with video samples is released.
🚩 2024.05.30: Model pre-trained on Open X-Embodiment and inference code are released.
🚩 2024.05.27: Our paper is released on arXiv.

Installation

conda create -n ivideogpt python==3.9
conda activate ivideogpt
pip install -r requirements.txt

Models

At the moment we provide the following models:

Model	Resolution	Action	Tokenizer Size	Transformer Size
ivideogpt-oxe-64-act-free	64x64	No	114M	138M

If no network connection to Hugging Face, you can manually download from Tsinghua Cloud.

Inference Examples

Action-free Video Prediction on Open X-Embodiment

python inference/predict.py --pretrained_model_name_or_path "thuml/ivideogpt-oxe-64-act-free" --input_path inference/samples/fractal_sample.npz --dataset_name fractal20220817_data

To try more samples, download the dataset from the Open X-Embodiment Dataset and extract single episodes as follows:

python oxe_data_converter.py --dataset_name {dataset_name, e.g. bridge} --input_path {path to OXE} --output_path samples --max_num_episodes 10

Training Video Prediction

Pretrained Models

To finetune our pretrained iVideoGPT, download it into pretrained_models/ivideogpt-oxe-64-act-free.

To evaluate the FVD metric, download pretrained I3D model into pretrained_models/i3d/i3d_torchscript.pt.

Data Preprocessing

BAIR Robot Pushing: Download the dataset and preprocess with the following script:

wget http://rail.eecs.berkeley.edu/datasets/bair_robot_pushing_dataset_v0.tar -P .
tar -xvf ./bair_robot_pushing_dataset_v0.tar -C .

python datasets/preprocess_bair.py --input_path bair_robot_pushing_dataset_v0/softmotion30_44k --save_path bair_preprocessed

Then modify the saved paths (e.g. bair_preprocessed/train and bair_preprocessed/test) in DATASET.yaml.

Finetuning Tokenizer

accelerate launch train_tokenizer.py \
    --exp_name bair_tokenizer_ft --output_dir log_vqgan --seed 0 --mixed_precision bf16 \
    --model_type ctx_vqgan \
    --train_batch_size 16 --gradient_accumulation_steps 1 --disc_start 1000005 \
    --oxe_data_mixes_type bair --resolution 64 --dataloader_num_workers 16 \
    --rand_select --video_stepsize 1 --segment_horizon 16 --segment_length 8 --context_length 1 \
    --pretrained_model_name_or_path pretrained_models/ivideogpt-oxe-64-act-free/tokenizer

Finetuning Transformer

For action-conditioned video prediction, run the following:

accelerate launch train_gpt.py \
    --exp_name bair_llama_ft --output_dir log_trm --seed 0 --mixed_precision bf16 \
    --vqgan_type ctx_vqgan \
    --pretrained_model_name_or_path {log directory of finetuned tokenizer}/unwrapped_model \
    --config_name configs/llama/config.json --load_internal_llm --action_conditioned --action_dim 4 \
    --pretrained_transformer_path pretrained_models/ivideogpt-oxe-64-act-free/transformer \
    --per_device_train_batch_size 16 --gradient_accumulation_steps 1 \
    --learning_rate 1e-4 --lr_scheduler_type cosine \
    --oxe_data_mixes_type bair --resolution 64 --dataloader_num_workers 16 \
    --video_stepsize 1 --segment_length 16 --context_length 1 \
    --use_eval_dataset --use_fvd --use_frame_metrics \
    --weight_decay 0.01 --llama_attn_drop 0.1 --embed_no_wd

For action-free video prediction, remove --load_internal_llm --action_conditioned.

Training Visual Model-based RL

Preparation

Install the Metaworld version we used:

pip install git+https://github.com/Farama-Foundation/Metaworld.git@83ac03ca3207c0060112bfc101393ca794ebf1bd

Modify paths in mbrl/cfgs/mbpo_config.yaml to your own paths (currently only support absolute paths).

MBRL with iVideoGPT

python mbrl/train_metaworld_mbpo.py task=plate_slide num_train_frames=100002 demo=true

Showcases

Citation

If you find this project useful, please cite our paper as:

@article{wu2024ivideogpt,
    title={iVideoGPT: Interactive VideoGPTs are Scalable World Models}, 
    author={Jialong Wu and Shaofeng Yin and Ningya Feng and Xu He and Dong Li and Jianye Hao and Mingsheng Long},
    journal={arXiv preprint arXiv:2405.15223},
    year={2024},
}

Contact

If you have any question, please contact wujialong0229@gmail.com.

Acknowledgement

Our codebase is heavily built upon huggingface/diffusers and facebookresearch/drqv2.

thuml / iVideoGPT