| license: mit | |
| tags: | |
| - vqvae | |
| - image-generation | |
| - unsupervised-learning | |
| - pytorch | |
| - cifar10 | |
| - generative-model | |
| datasets: | |
| - cifar10 | |
| library_name: pytorch | |
| model-index: | |
| - name: VQ-VAE-CIFAR10 | |
| results: [] | |
| # VQ-VAE for CIFAR-10 | |
| This is a **Vector Quantized Variational Autoencoder (VQ-VAE)** trained on the CIFAR-10 dataset using PyTorch. It is part of an image augmentation pipeline for generative modeling and unsupervised learning research. | |
| ## 🧠 Model Details | |
| - **Model Type**: VQ-VAE | |
| - **Dataset**: CIFAR-10 | |
| - **Epochs**: 35 | |
| - **Latent Space**: Discrete (quantized vectors) | |
| - **Input Size**: 64×64 | |
| - **Reconstruction Loss**: MSE-based | |
| - **Implementation**: Custom PyTorch, 3-layer Conv Encoder/Decoder | |
| - **FID Score**: **71.11** | |
| - **Loss Curve**: [`loss_curve.png`](./loss_curve.png) | |
| > This model was trained to learn a compact representation of CIFAR-10 images via vector quantization and used for downstream data augmentation. | |
| ## 📁 Files | |
| - `generator.pt`: Trained VQ-VAE model weights. | |
| - `loss_curve.png`: Visual plot of training loss across 35 epochs. | |
| - `fid_score.json`: Stored Fréchet Inception Distance (FID) evaluation result. | |
| - `fid_real/` and `fid_fake/`: 1000 real and generated samples used for FID computation. | |
| ## 📦 How to Use | |
| ```python | |
| import torch | |
| from models.vqvae.model import VQVAE | |
| model = VQVAE() | |
| model.load_state_dict(torch.load("generator.pt", map_location="cpu")) | |
| model.eval() | |