https://github.com/deepbiolab/gans-research

🎨 A step-by-step implementation of key GAN architectures from Vanilla GAN to StyleGAN3, focusing on training stability, conditional control, and high-quality image generation.
https://github.com/deepbiolab/gans-research

deep-learning gans image-generation progressive-gan stylegan wgan

Last synced: 6 months ago
JSON representation

🎨 A step-by-step implementation of key GAN architectures from Vanilla GAN to StyleGAN3, focusing on training stability, conditional control, and high-quality image generation.

Host: GitHub
URL: https://github.com/deepbiolab/gans-research
Owner: deepbiolab
License: mit
Created: 2025-04-15T18:18:50.000Z (6 months ago)
Default Branch: main
Last Pushed: 2025-04-23T13:59:28.000Z (6 months ago)
Last Synced: 2025-04-24T00:12:32.096Z (6 months ago)
Topics: deep-learning, gans, image-generation, progressive-gan, stylegan, wgan
Language: Python
Homepage:
Size: 9.41 MB
Stars: 0
Watchers: 1
Forks: 0
Open Issues: 0
Metadata Files:
- Readme: README.md
- License: LICENSE

Awesome Lists containing this project

README

# GANs Research Reproduction from Scratch

```mermaid
graph LR
subgraph Basic_Models["Basic Models (2014-2015)"]
VanillaGAN[Vanilla GAN 2014] --> DCGAN[DCGAN 2015]
end

subgraph Training_Stability["Training Stability Path"]
DCGAN --> WGAN[WGAN 2017]
WGAN --> WGAN_GP[WGAN-GP 2017]
end

subgraph Control_Mechanisms["Control Mechanisms"]
VanillaGAN --> CGAN[Conditional GAN 2014]
note1[Explicit Control]
end

subgraph Quality_Improvement["Quality Improvement Path"]
DCGAN --> ProgGAN[Progressive GAN 2017]
WGAN --> StyleGAN1[StyleGAN 2018]
ProgGAN --> StyleGAN1
WGAN_GP -.-> StyleGAN1
StyleGAN1 --> StyleGAN2[StyleGAN2 2019]
StyleGAN2 --> StyleGAN3[StyleGAN3 2021]
note2[Implicit Style Control]
end

%% Relationship Notes
Training_Stability -. "Stability Focus" .-> Quality_Improvement
Control_Mechanisms -. "Control Evolution" .-> Quality_Improvement

%% Styling
classDef done fill:#A9B7C0,stroke:#555,stroke-width:2px;
classDef pending fill:#D7C4BB,stroke:#555,stroke-width:2px;
classDef current fill:#B8C4B8,stroke:#555,stroke-width:2px;
classDef note fill:#fff,stroke:#555,stroke-width:1px,stroke-dasharray: 5 5;

%% Apply styles
class VanillaGAN done;
class DCGAN,WGAN,WGAN_GP,CGAN,ProgGAN,StyleGAN1,StyleGAN2,StyleGAN3 pending;
class note1,note2 note;

%% Subgraph styling
style Basic_Models fill:#f5f5f5,stroke:#666,stroke-width:2px;
style Training_Stability fill:#e8f4f8,stroke:#666,stroke-width:2px;
style Control_Mechanisms fill:#f8f4e8,stroke:#666,stroke-width:2px;
style Quality_Improvement fill:#f0f8f0,stroke:#666,stroke-width:2px;

```

This repository contains implementations of key Generative Adversarial Network (GAN) architectures, organized in an incremental learning approach.

## Implementation Progress

1. **Basic GANs**
- [x] **Vanilla GAN**: Original formulation with fully connected layers for both generator and discriminator networks.
- [x] **DCGAN**: Introduced convolutional architectures and established key architectural guidelines for stable GAN training.

2. **Training Stability Improvements**
- [x] **WGAN**: Improved training stability using Wasserstein distance.
- [x] **WGAN-GP**: Enhanced WGAN with gradient penalty.

3. **Conditional Control**
- [x] **Conditional GAN**: Added ability to control generation with explicit labels.

4. **High-Quality Generation**
- [x] **Progressive GAN**: Introduced progressive growing methodology for generating high-resolution images.
- [ ] **StyleGAN v1**: Combined progressive growing with style-based generation.
- [ ] **StyleGAN v2**: Improved architecture and removing progressive growing.
- [ ] **StyleGAN v3**: Added alias-free generation techniques.

## Installation

### Option 1: Install dependencies only
```bash
pip install -r requirements.txt
```

### Option 2: Install as a package (recommended for development)
```bash
# If you have a previous installation, remove it first
pip uninstall gans-research

# Install in editable mode
pip install -e .
```

## Usage

### Training
Each GAN implementation has its own training script in the `experiments` directory:

```bash
# Example: Train a basic GAN on MNIST
python experiments/stage1_basic/train_vanilla_gan.py
```

### Evaluation
Evaluation scripts are located in the `experiments` directory:

```bash
# Evaluate a trained model
python experiments/evaluate.py --config configs/vanilla_gan.yaml

# Evaluate specific metrics
python experiments/evaluate.py --config configs/vanilla_gan.yaml --metrics quality
python experiments/evaluate.py --config configs/vanilla_gan.yaml --metrics coverage
python experiments/evaluate.py --config configs/vanilla_gan.yaml --metrics speed

# Customize evaluation parameters
python experiments/evaluate.py --config configs/vanilla_gan.yaml --num_samples 5000 --batch_size 32
```

**Configuration Example**

Evaluation configuration is part of the model's config YAML file:

```yaml
evaluation:
batch_size: 64
num_samples: 10000
```

### Inference

#### Unconditional Generation

Example: generate images using a trained Vanilla GAN model on MINIST:

```
python experiments/inference.py --config configs/vanilla_gan.yaml
```

#### Conditional Generation

Example: generate images using a trained Conditional GAN model on MINIST:

- generate multiple images for each label:

```python
python experiments/inference.py --config configs/cgan.yaml --labels 0,1,2,3,4,5,6,7,8,9 --num_samples 5
```

- generate one image for each label:

```
python experiments/inference.py --config configs/cgan.yaml --labels 2,5
```

**Configuration Example**

YAML config should include the following fields:

```yaml
inference: # or dcgan, wgan, etc.
checkpoint_path: outputs/vanilla_gan/checkpoints/final_model.pth
num_samples: 16
model:
name: vanilla_gan
experiment:
output_dir: outputs/vanilla_gan
```

## Project Structure

The project is organized as follows:

```
gans-research/
├── configs/ # Configuration files for each GAN variant
├── data/ # Dataset loading and processing utilities
├── evaluation/ # Evaluation metrics and visualization tools
│ ├── metrics/ # Implementation of evaluation metrics (FID, etc.)
│ └── visualization/ # Visualization utilities for generated images
├── experiments/ # Training and evaluation scripts
│ └── stage1_basic/ # Training scripts for basic GAN models
├── src/ # Core source code
│ ├── data/ # Dataset handling
│ ├── losses/ # Loss functions for different GAN variants
│ ├── metrics/ # Evaluation metrics (FID, precision-recall)
│ ├── models/ # Model implementations
│ │ ├── base/ # Base classes for generators and discriminators
│ │ ├── vanilla_gan/ # Vanilla GAN implementation
│ │ ├── dcgan/ # Deep Convolutional GAN implementation
│ │ ├── cgan/ # Conditional GAN implementation
│ │ ├── wgan/ # Wasserstein GAN implementation
│ │ ├── wgan_gp/ # Wasserstein GAN with Gradient Penalty implementation
│ │ ├── prog_gan/ # Progressive GAN implementation
│ │ ├── stylegan1/ # StyleGAN v1 implementation
│ │ ├── stylegan2/ # StyleGAN v2 implementation
│ │ └── stylegan3/ # StyleGAN v3 implementation
│ ├── training/ # Training utilities
│ └── utils/ # Helper utilities
└── pyproject.toml # Package configuration
```

## References

- [Generative Adversarial Networks (2014)](https://arxiv.org/abs/1406.2661)
- [DCGAN (2015)](https://arxiv.org/abs/1511.06434)
- [Wasserstein GAN (2017)](https://arxiv.org/abs/1701.07875)
- [WGAN-GP (2017)](https://arxiv.org/abs/1704.00028)
- [Conditional GANs (2014)](https://arxiv.org/abs/1411.1784)
- [Progressive GAN (2017)](https://arxiv.org/abs/1710.10196)
- [StyleGAN (2018)](https://arxiv.org/abs/1812.04948)
- [StyleGAN2 (2019)](https://arxiv.org/abs/1912.04958)
- [StyleGAN3 (2021)](https://arxiv.org/abs/2106.12423)

ecosyste.ms

Data

Tools

Indexes

Applications

Experiments

Awesome

https://github.com/deepbiolab/gans-research

Awesome Lists containing this project

README