CryoGEM Reproduction

End-to-end reproduction of CryoGEM (NeurIPS 2024) for physics-informed cryo-EM image synthesis.

Quick Start (Google Colab)

1. Open notebooks/CryoGEM_Reproduction.ipynb in Google Colab
2. Select GPU runtime (T4 or better)
3. Run all cells

Project Structure

cryogem-reproduction/
├── src/                    # Core modules
│   ├── physics/            # CTF, FFT, noise simulation
│   ├── models/             # U-Net, PatchGAN, NCE networks
│   ├── losses/             # GAN + NCE losses
│   ├── data/               # Dataset loading
│   └── utils/              # Visualization, rotations
├── scripts/                # Pipeline scripts
├── notebooks/              # Colab notebook
├── config/                 # Dataset configs
├── data/                   # Downloaded data
├── outputs/                # Generated outputs
└── logs/                   # Training logs

Pipeline

1. gen_data - Generate synthetic training micrographs
2. esti_ice - Estimate ice gradients from real data
3. train - Train GAN with mask-aware NCE loss
4. test - Generate realistic synthetic dataset

Requirements

• Python >= 3.8
• PyTorch >= 1.7
• CUDA GPU (15GB+ recommended)

License

Apache 2.0 (following original CryoGEM)