DeMemWM

This repo is forked from Boyuan Chen's research template repo. By its MIT license, you must keep this sentence in README.md and the LICENSE file to credit the author.

DeMemWM is a Memory-DiT video prediction project built on the local research template. The primary algorithm entry point is DeMemWMMinecraft, registered through the Hydra algorithm config dememwm_memory_dit.

Quick Start

python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
python -m pytest tests

Run a local offline experiment after setting the dataset path in configurations/dataset/video_minecraft.yaml:

python main.py +name=dememwm_debug algorithm=dememwm_memory_dit wandb.mode=offline

Use resume_ckpt_path=/path/to/checkpoint.ckpt for deterministic checkpoint resume, or keep auto_resume=true to resume from output_dir/checkpoints when available.

Layout

• algorithms/worldmem/dememwm/: DeMemWM memory construction, retrieval, scheduling, diagnostics, and injection code.
• algorithms/worldmem/dememwm_memory_dit.py: primary DeMemWM algorithm class.
• configurations/algorithm/dememwm_memory_dit.yaml: consumed DeMemWM training and evaluation contract.
• scripts/: Slurm and inspection scripts using the DeMemWM naming.
• tests/: static and unit coverage for DeMemWM config, retrieval, compression, schedules, and training behavior.

Reproducibility Notes

• Keep wandb.mode=offline for local reproducible runs that do not depend on network access.
• Set seed=<int> on the command line to seed Lightning and dataloader workers.
• Runtime artifacts such as outputs/, slurm_logs/, Python caches, checkpoints, and local datasets are ignored by git.
• The default Hydra training config selects algorithm: dememwm_memory_dit.