MAE – Masked Autoencoder Image Reconstruction
AI Assignment 02 | Generative AI AI4009 | FAST-NUCES Spring 2026
This Space demonstrates a self-supervised Masked Autoencoder (MAE) trained on TinyImageNet.
How it works
- Upload any image.
- Adjust the masking ratio slider (default 75 %).
- Click Reconstruct to see:
- The original (resized to 224 × 224)
- The masked input (grey patches = hidden)
- The MAE reconstruction
Architecture
| Component | Spec |
|---|---|
| Encoder | ViT-Base B/16, embed=768, depth=12, heads=12, ~86M params |
| Decoder | ViT-Small S/16, embed=384, depth=12, heads=6, ~22M params |
| Patch size | 16 × 16 |
| Visible patches | 25 % (49 of 196) |
| Loss | MSE on masked patches only |
| Training | AdamW + CosineAnnealing, Mixed Precision, 50 epochs |
Setup
Upload mae_tiny_imagenet.pth (the trained weights) to the root of this Space.
The file is generated by running the Kaggle notebook AI_ASS02_XXF_YYYY.ipynb.
Inference Providers NEW
This model isn't deployed by any Inference Provider. 🙋 Ask for provider support