bstdenis/VisionCourseTutorial

DenseUNet

An image-to-image model for upscaling blurry MNIST-like images.

Table of contents

• Model Summary
• Use Cases & Limitations
• How to Use
• Model Details
• Training Procedure
• Evaluation
• Ethical Considerations & Limitations
• Caveats and Recommendations
• Citation
• License
• Contact

---

Model Summary

• Task: Image to image (upscaling)
• Architecture: UNet with Dense layers
• Inputs: MNIST image
• Outputs: Upscaled images of numbers
• Intended Users: Students learning image-to-image tasks

Use Cases & Limitations

Intended uses

• learning image-to-image model behavior

Out-of-scope uses / limitations

• Not for production or deployment

How to Use

# ToDo

Model Details

• Developers: Blaise Gauvin St-Denis (bstdenis)
• Organization: Ouranos
• Version: v1.0 (2025-08-13)
• Dataset: mnist (trainsplit)
• Evaluation dataset: mnist (test split)
• Compute: 1x L40 (ToDo training time)
• Framework: Dense UNet / PyTorch 2.8
• Hardware/Software dependencies: CUDA 12.8, Python 3.12

Training Procedure

Data preprocessing

• Resize to 32 x 32 with all possible locations for the 28 x 28 original data
• Downcale 4x (to 8 x 8)

Hyperparameters

• Epochs:
• Batch size:
• Learnings rate:
• Weight decay:
• Seed: 0

Evaluation

Ethical Considerations & Limitations

• For educational purpose only

Caveats and Recommendations

Citation

If you use this model, please cite

ToDo

License

This model is released under the Apache 2.0 License. See LICENSE file.

Contact

• Maintainer: @bstdenis