ddpm_pt — PyTorch DDPM (Ho et al., 2020) + trajectory visualization

This repo provides:

• DDPM U-Net (time embedding, ResBlocks, optional attention)
• Gaussian diffusion utilities (q, p, sampling)
• Training script with AMP + EMA
• Sampling script
• Trajectory visualization (forward noising + reverse denoising) to MP4/GIF

Design choice: torch is installed explicitly (not in pyproject.toml)

We intentionally DO NOT list torch, torchvision, or torchaudio in pyproject.toml.

Reason: CUDA wheels for PyTorch are platform- and GPU-sensitive (and for very new GPUs, you may need a specific CUDA index and/or nightly wheels). Installing torch-related packages explicitly from the correct PyTorch index URL avoids dependency resolver issues and version mismatches.

uv.lock will pin only the non-torch Python dependencies.

---

1) Prerequisites

• Python 3.11 recommended (strongest PyTorch wheel coverage)
• NVIDIA driver installed and working (nvidia-smi should run)
• uv installed (see uv docs)

---

2) Create the environment (uv venv)

From the repo root:

uv venv --python 3.11
UV_HTTP_TIMEOUT=1000 uv pip install torch torchvision torchaudio --index-url https://download.pytorch.org/whl/cu128

3) Train the model (uv run)

From the repo root:

uv run python tools/train.py --config configs/cifar10.yaml --workdir runs/cifar10