Unconditional DDPM (Standalone)

Minimal PyTorch unconditional DDPM project (MNIST default) with:

• U-Net denoiser + timestep embedding
• Linear beta schedule
• Epsilon-prediction MSE training
• Ancestral sampling script

Setup

cd /Users/jiayibaobei/Desktop/unconditional-ddpm
python -m pip install -r requirements.txt

Train

python train.py --outdir ./runs --epochs 20 --batch_size 128

Outputs:

• Checkpoints: ./runs/checkpoints/
• Sample grids: ./runs/samples/

Sample from checkpoint

python sample.py \
  --checkpoint ./runs/checkpoints/model_epoch_20.pt \
  --out ./runs/final_samples.png \
  --num_samples 64

Notes

• This is unconditional: class labels are not used.
• Inputs are normalized to [-1, 1].
• Default dataset is MNIST resized to 32x32.