Welcome to the Programming Ocean Academy's Deep Learning Repository! This project is a comprehensive educational suite showcasing a variety of generative models implemented with PyTorch, ranging from foundational architectures to modern, cutting-edge designs.
This repository serves as an academic and teaching-oriented resource for understanding, building, and visualizing deep generative models. It is designed to help students, researchers, and practitioners explore the diversity of generative learning approaches in a modular and clear format.
Each folder represents a specific category of generative or neural architecture:
| Folder Name | Description |
|---|---|
auto-regressive-models |
PixelCNN and related sequential density estimators |
cnn |
Basic CNN models for image recognition |
diffusion |
Denoising Diffusion Probabilistic Models (DDPM, DDIM) |
dit-models |
Diffusion Transformers (DiT) |
energy-based-models |
EBMs trained with Langevin dynamics |
flow-based-models |
RealNVP, Glow, and other invertible models |
gans |
GAN, DCGAN, WGAN, and conditional variants |
latent-manifold-auto-encoder |
Latent space exploration with VAEs and AEs |
multi-model |
Cross-modal tasks (e.g., text-to-image, image captioning) |
restricted-boltzmann-machine |
Contrastive Divergence and RBMs |
rnn |
Recurrent networks (LSTM, GRU) |
score-based-generative-convolution |
Score-matching models with CNN backbones |
score-based-generative-models |
Langevin and NCSN-style samplers |
time-series |
Forecasting models for temporal data |
transformer |
Sequence models and transformers (Vanilla, GPT) |
variational-auto-encoder |
VAEs and conditional variants |
vision-transformer |
ViT for image understanding |
"A Concise Implementation of Denoising Diffusion Probabilistic Models for Generative Image Synthesis in PyTorch"
- U-Net architecture with Gaussian noise scheduling
- Reverse sampling with denoising
"Adversarial Image Synthesis with Generative Networks: A PyTorch Implementation of GANs on MNIST"
- Generator + Discriminator loop
- Real vs. generated image comparison
"Latent Variable Modeling and Image Generation with Variational Autoencoders: A PyTorch-Based Study on MNIST"
- Reparameterization trick
- Sampling and interpolation
"Unsupervised Image Synthesis via Score Matching and Langevin Dynamics: A Score-Based Generative Framework on MNIST"
- Trainable score networks
- MCMC sampling
"Learning Discrete Visual Representations from Textual Descriptions: A Simplified VQ-VAE Framework for Text-to-Image Generation"
- VQ-VAE + Transformer
- Color/shape captioned image generation
"Visual Grounding through Language: A Minimalist Encoder-Decoder Framework for Image Captioning with Attention in PyTorch"
- ResNet + LSTM with soft attention
- Caption generation for synthetic scenes
All notebooks are written for clarity and modularity.
# Clone the repository
https://github.com/Programming-Ocean-Academy/deep-learning.gitOpen any .ipynb file in JupyterLab, Google Colab, or VSCode and run directly.
We welcome contributions to extend this educational repository:
- Add new generative model examples
- Improve visualizations or metrics
- Refactor notebooks into scripts or modules
MIT License. Free for personal, educational, and research use.
Inspired by work from:
- OpenAI, DeepMind, LucidRain
- PyTorch community and tutorials
- DALL·E, VQ-VAE, and DDPM original papers
Enjoy exploring generative deep learning! 🌊