cannizzaro12/archive-594f8629

CMU 10799 Diffusion & Flow Matching Homework Starter Code

Welcome to CMU 10799 Spring 2026: Diffusion & Flow Matching!

This repository contains your starter code for all four homework assignments. You'll build a working diffusion model from scratch, starting with DDPM fundamentals and progressing through flow matching to your chosen specialization track. The codebase is designed to grow with you, and what you implement in HW1 becomes the foundation for everything that follows.

Class Website | Homework Handouts

Project Structure

cmu-10799-diffusion/
├── train.py                  # Training script
├── sample.py                 # Sampling script
├── download_dataset.py       # Download CelebA dataset
├── modal_app.py              # Modal cloud setup
├── setup.sh                  # Setup script (pip + venv)
├── setup-uv.sh               # Setup script (uv - faster!)
├── pyproject.toml            # Python package configuration
│
├── src/                      # Source code
│   ├── models/
│   │   ├── blocks.py         # U-Net components (ResBlock, Attention, etc.)
│   │   └── unet.py           # Complete U-Net architecture
│   ├── methods/
│   │   ├── base.py           # Base method class
│   │   └── ddpm.py           # DDPM implementation
│   ├── data/
│   │   └── celeba.py         # CelebA dataset loading, TODO: fill in your data transforms functions
│   └── utils/
│       ├── ema.py            # EMA helper
│       └── logging_utils.py  # Logging utilities
│
├── configs/                  # Hyperparameter configurations
│   ├── ddpm_modal.yaml       # DDPM config for Modal
│   └── ddpm_babel.yaml       # DDPM config for Babel cluster
│
├── scripts/                  # Shell scripts for Modal/cluster
│   ├── train.sh              # SLURM job template (tested on babel)
│   ├── evaluate_torch_fidelity.sh        # torch-fidelity evaluation on cluster or locally (tested on babel)
│   ├── evaluate_modal_torch_fidelity.sh  # torch-fidelity evaluation on Modal
│   └── list_checkpoints.sh   # List Modal checkpoints
│
├── notebooks/                # Jupyter notebooks
│   ├── 01_1d_playground.ipynb
│   ├── 02_dataset_exploration.ipynb
│   └── 03_sampling_visualization.ipynb
│
├── environments/             # Environment configurations
│   ├── requirements.txt      # Base dependencies
│   ├── requirements-cpu.txt  # CPU-only PyTorch
│   ├── requirements-cuda*.txt  # PyTorch + CUDA versions
│   └── requirements-rocm.txt # PyTorch + ROCm (AMD)
│
└── docs/                     # Documentation
    ├── SETUP.md              # Platform-specific setup guide
    ├── QUICKSTART-MODAL.md   # Quick start for Modal users
    └── DIRECTORY-STRUCTURE.md # Directory structure guide

List of TODOs

1. src/data/celeba.py: fill in your data transforms functions
2. src/methods/ddpm.py: implement everything in this file
3. src/models/unet.py: implement the unet model architecture and its forward pass
4. configs/: create your own model configs
5. train.py: incorporate your sampling scheme to the training pipeline and save generated samples as images for logging
6. sample.py: incorporate your sampling scheme to the training pipeline and save generated samples as images

Example scripts & Helper skeleton notebooks

Besides the parts that you need to implement, the starter code also provides several example scripts & helper notebook skeleton to help you get started.

The scripts are stored in scripts/ and the notebooks are stored in notebooks/. Feel free to use (or not use), add, delete or modify any and all parts of these scripts and notebooks.

Quick Start

1. Setup Environment

git clone <repo-url>
cd cmu-10799-diffusion

# Run setup script (auto-detects your GPU)
./setup-uv.sh                 # Using uv (faster)
# or
./setup.sh                    # Using standard pip

# Activate the environment (name depends on detected hardware)
source .venv-cpu/bin/activate        # If CPU was detected
source .venv-cuda121/bin/activate    # If CUDA 12.1 was detected

2. Download Dataset

python download_dataset.py

3. Train (SLURM Cluster)

# Train DDPM (uses configs/ddpm.yaml by default)
sbatch scripts/train.sh ddpm

# Train with custom config
sbatch scripts/train.sh ddpm configs/ddpm_babel.yaml

# Resume from checkpoint
sbatch scripts/train.sh ddpm --resume checkpoints/ddpm_50000.pt

4. Train (Modal Cloud GPU)

For students without a local GPU:

# First time: cache dataset to Modal volume
modal run modal_app.py --action download

# Train DDPM
modal run modal_app.py --action train --method ddpm

See docs/QUICKSTART-MODAL.md for complete Modal setup.

5. Evaluate

# Evaluate using torch-fidelity (local/cluster)
./scripts/evaluate_torch_fidelity.sh ddpm checkpoints/ddpm/ddpm_final.pt

# Evaluate on Modal
./scripts/evaluate_modal_torch_fidelity.sh ddpm checkpoints/ddpm/ddpm_final.pt

Jupyter Notebooks

• 01_1d_playground.ipynb - 1D diffusion experiments for building intuition
• 02_dataset_exploration.ipynb - Explore CelebA dataset
• 03_sampling_visualization.ipynb - Visualize your samples

---

Manual Installation

If the setup scripts don't work for your system, you can set up manually.

Option A: Manual setup with uv

uv is a fast Python package manager (10-100x faster than pip).

# 1. Install uv (one-time)
curl -LsSf https://astral.sh/uv/install.sh | sh   # macOS/Linux
# or: pip install uv

# 2. Clone and enter repo
git clone <repo-url>
cd cmu-10799-diffusion

# 3. Create virtual environment
uv venv .venv-cpu              # For CPU
uv venv .venv-cuda121          # For CUDA 12.1

# 4. Activate it
source .venv-cpu/bin/activate        # CPU
source .venv-cuda121/bin/activate    # GPU

# 5. Install dependencies (choose one)
uv pip install -r environments/requirements-cpu.txt      # No GPU / Modal users
uv pip install -r environments/requirements-cuda121.txt  # NVIDIA GPU (most common)

# 6. Verify
python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA: {torch.cuda.is_available()}')"

Option B: Manual setup with pip + venv

Check your CUDA version: nvidia-smi (top right shows "CUDA Version: XX.X")

# 1. Clone and enter repo
git clone <repo-url>
cd cmu-10799-diffusion

# 2. Create virtual environment
python -m venv .venv-cpu        # For CPU
python -m venv .venv-cuda121    # For CUDA 12.1

# 3. Activate it
source .venv-cpu/bin/activate        # CPU
source .venv-cuda121/bin/activate    # GPU

# 4. Install dependencies (choose one)
pip install -r environments/requirements-cpu.txt      # No GPU / Modal users
pip install -r environments/requirements-cuda121.txt  # NVIDIA GPU (most common)

# 5. Verify
python -c "import torch; print(f'PyTorch {torch.__version__}, CUDA: {torch.cuda.is_available()}')"

See docs/SETUP.md for detailed instructions for Modal, AWS, and SLURM clusters.

---

Running Customized Commands Manually

If you prefer to run Python commands directly instead of using the provided scripts:

Training

python train.py --method ddpm --config configs/ddpm.yaml

Generating Samples

python sample.py --checkpoint checkpoints/ddpm_final.pt --method ddpm --num_samples 64

# With custom number of sampling steps
python sample.py --checkpoint checkpoints/ddpm_final.pt --method ddpm --num_steps 500

Evaluation

Use the provided shell scripts that wrap torch-fidelity:

# Evaluate locally or on cluster
./scripts/evaluate_torch_fidelity.sh ddpm checkpoints/ddpm_final.pt

# Evaluate on Modal
./scripts/evaluate_modal_torch_fidelity.sh ddpm checkpoints/ddpm_final.pt

Modal (Advanced Options)

# Train with custom config
modal run modal_app.py --action train --method ddpm --config configs/custom.yaml

# Train with custom iterations
modal run modal_app.py --action train --method ddpm --iterations 50000

To enable Weights & Biases logging on Modal:

modal secret create wandb-api-key WANDB_API_KEY=your_real_key

Configuration

All hyperparameters are controlled via config files. Edit configs/ddpm*.yaml to configurate your model.

---

Authors

Yutong (Kelly) He, with assistance from Claude Code and OpenAI Codex.

References

• DDIM official code