VQ-VAE for Tiny ImageNet (ImageNet-200)

🧠 Model Details

• Model Type: Vector Quantized Variational Autoencoder (VQ-VAE)
• Dataset: Tiny ImageNet (ImageNet-200)
• Epochs: 35
• Latent Space: Discrete codebook (vector quantization)
• Input Size: 64×64 RGB
• Loss Function: Mean Squared Error (MSE) + VQ commitment loss
• Final Training Loss: ~0.0292
• FID Score: ~102.87
• Architecture: 3-layer CNN Encoder & Decoder with quantization bottleneck

📦 Files

• generator.pt — Trained VQ-VAE model weights
• loss_curve.png — Plot of training loss across 35 epochs
• fid_score.json — FID evaluation result on 1000 generated samples
• fid_real/ — 1000 real Tiny ImageNet samples used for FID
• fid_fake/ — 1000 VQ-VAE reconstructions used for FID

🔧 Usage