世界モデルにおける時空間情報の理解に関する検証 / Investigating the Understanding of Spatio-temporal Information in World Models

Getting Started

1. After cloning this repo, clone DreamerV3 repository in the root project directory

• DreamerV3

2. Replace the following file paths in the DreamerV3 repository with files in our dreamerv3_edits directory.

• /dreamerv3/configs.yaml
• /embodied/envs/crafter.py

3. execute the command ./training_dreamer.sh to train DreamerV3. To restart existing training loop, refer to ./training_dreamer.sh for more information.

4. After training DreamerV3, open probing/utils.py and edit the paths to log directory and target directory. Then execute the command python ./probing/utils.py to migrate collected stats files to probing folder

5. run probing/save_activations.ipynb to collect activations.

6. Finally, probing is performed by running probing/probing.ipynb.