README.md

Transformer-Based Diffusion for Game Generation

This project applies transformer-based Diffusion models to simulate Doom (the find my way home scenario).

See the final report here.

Local Setup

1. Create the conda environment

conda env create -f env.yml --prefix ./penv

2. Activate the conda environment:

conda activate ./penv

3. Install python dependencies

pip install -r requirements.txt

4. Install the root package

pip install -e .

5. Login to wandb

wandb login

To Train locally

1. Train the VAE

   1. Collect frames data for VAE training

   python scripts/collect_doom_data.py --observations_only --save_path {vae_train_data}
   

   2. Fine tune the VAE decoder

   python src/vae/train_vae.py --data_dir {vae_train_data}
   

2. Collect game play data using a random policy

   1. Train data

   python scripts/collect_doom_data.py --save_path {data_dir}/original/train
   

   2. Test data

   python scripts/collect_doom_data.py --save_path {data_dir}/original/test --num_episodes 200
   

3. Preprocess gameplay data with the VAE

python scripts/preprocess_data_with_vae.py --data_path {data_dir}/original --save_path {data_dir}/latent

4. Train the diffusion model. You will need to update project_storage_base_path for the output directory, the vae decoder path, and data_path to {data_dir} in hydra configs

• To train the concatenation conditioning single-frame prediction model:

  python src/main.py diffusion_model=DiT_B_4_concat experiment_name=DiT_B_4_concat
  

• To train the cross-attention conditioning single-frame prediction model:

  python src/main.py diffusion_model=DiT_B_4_cross_attn experiment_name=DiT_B_4_cross_attn
  

• To train the video generation multi-frame prediction model:

  git checkout vg-model && python src/main.py experiment_name=DiT_VG
  

Inference

To evaluate your single-frame prediction model on a trajectory:

python scripts/evaluate_trajectory --run_dir {training_run_dir} --vae_decoder_path {finetuned_decoder} --episode_path {test_episode}

To evaluate your video-generation model on a trajectory:

git checkout vg-model && python scripts/evaluate_trajectory_vg.py --run_dir {training_run_dir} --model_version {model_file_name} --episodes_path {file_with_episode_paths}