{}
Model Description
This repository contains the trained checkpoints (.pt) and generated samples (.png) for two generative models trained on a subset of the STL-10 dataset.
Pixel DDPM: Standard diffusion on 96x96 RGB images.
Latent DDPM (LDM): Diffusion on a compressed 12x12x4 latent space (using a VAE).
Note: This repository only contains the weights. To load and use these models, you must have the original model definitions (DDPM.py and VAE.py) in your local Python environment.
- Developed by: Linoy Geva & Ron Chernoguz
- Model type: DDPM/LPM
Uses
This model is designed to generate synthetic data that mimics the characteristics of the STL-10 dataset. It is intended for use in experimental research comparing the effectiveness of diffusion-based data augmentation against traditional/manual augmentation techniques for CNN classification.
Recommendations
Training Details
Training Data
[More Information Needed]
Training Procedure
Training Hyperparameters
========== PIXEL DDPM (Baseline Model) =========================================================== • Optimization Strategy:
- Optimizer: AdamW
- Learning Rate: 2e-4 (with Cosine Annealing)
- Weight Decay: 1e-3
- Batch Size: 32
- Total Epochs: 225
• Model Architecture:
- Input: 96x96 RGB Images
- Base Channels: 64
- Channel Mults: (1, 2, 4, 8)
- EMA: Disabled
- Diffusion: 1000 Timesteps (Linear Schedule)
========== LATENT DDPM (LDM) =========================================================== • Optimization Strategy:
- Optimizer: AdamW
- Learning Rate: 1e-4 (with Cosine Annealing)
- Weight Decay: 1e-3
- Batch Size: 64
- Total Epochs: 2500 (Target)
• Model Architecture:
- Input: 12x12x4 Latents (via VAE)
- Base Channels: 128
- Channel Mults: (2, 2)
- EMA: Enabled (Decay: 0.99)
• Conditioning (CFG): - Label Dropout: 15% (during training) - Guidance Scale: 5.0 (during inference)
Testing Data
[More Information Needed]
Results
[More Information Needed]
Summary
Compute Infrastructure
[More Information Needed]
Hardware
[More Information Needed]