Deep Generative Models

Selected implementations and experiments from a graduate Deep Generative Models course, covering latent-variable models, adversarial translation, energy-based modeling, score-based generation, and diffusion models.

Included experiments

1. β-VAE and disentanglement on dSprites

A convolutional VAE / β-VAE experiment with quantitative representation analysis using the Mutual Information Gap (MIG).

β	Overall MIG
1	0.0148
6	0.2076
12	0.1548

The experiment illustrates the trade-off between reconstruction pressure and disentanglement: in these runs, β=6 produced the strongest overall MIG.

2. CycleGAN for unpaired image-to-image translation

A CycleGAN implementation with:

• residual-block generators
• PatchGAN discriminators
• adversarial, cycle-consistency, and identity losses
• image replay buffer
• Horse↔Zebra and Summer↔Winter experiments

3. Energy-Based Model on MNIST

An energy-based generative model trained with negative samples produced by Langevin dynamics, including experiments with and without a replay buffer.

Key components:

• convolutional energy network
• Langevin sampling
• contrastive energy objective
• replay-buffer comparison
• generation and denoising experiments

4. Score-based generative modeling / NCSN

A score network trained with weighted denoising score matching on multiple Gaussian noise scales.

Key components:

• Gaussian/Fourier noise embeddings
• adaptive residual blocks
• weighted denoising score matching
• annealed Langevin dynamics
• conditional generation experiments

5. DDPM and DDIM on FashionMNIST

A diffusion pipeline built around a U-Net noise predictor, with both DDPM and DDIM sampling.

Selected results from the original experiment:

• approximate FID (500 generated samples): 81.75
• DDIM sampling (50 samples): ~1.15 s
• DDPM sampling (50 samples): ~23.7 s

The FID value is reported as an approximate course experiment rather than a benchmark result.

Repository structure

deep-generative-models/
├── notebooks/
│   ├── beta_vae_disentanglement.ipynb
│   ├── cyclegan_unpaired_translation.ipynb
│   ├── energy_based_model_mnist.ipynb
│   ├── score_based_ncsn_mnist.ipynb
│   └── ddpm_ddim_fashionmnist.ipynb
├── assets/
├── data/
│   └── README.md
├── README.md
├── requirements.txt
└── .gitignore

Setup

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

Open the notebooks with JupyterLab or Jupyter Notebook.

Data

The notebooks use public datasets such as dSprites, MNIST, FashionMNIST, horse2zebra, and summer2winter_yosemite. Large datasets and model checkpoints are intentionally not included in this repository. See data/README.md for notes.

Notes

• Notebook outputs were cleaned for the public portfolio version; representative figures and the reported metrics are preserved in the README.
• Environment-specific Kaggle/Colab paths were replaced with portable relative paths where practical.
• The original coursework also included a normalizing-flow/MAF experiment and DreamBooth-LoRA work. They are not included in this public version because those implementations need additional cleanup/validation before publication.
• A separate repository is used for the Flow Matching financial time-series project because it forms a coherent standalone project.

Topics

deep-learning generative-models vae cyclegan energy-based-models score-based-models diffusion-models ddpm ddim pytorch