speedblitz/lowlightpsnrkaggle

Low-Light Image Super-Resolution (HAT)

This repository contains the trained weights and architecture for a Hybrid Attention Transformer (HAT) fine-tuned for joint low-light enhancement and 4x super-resolution.

Model Details

• Architecture: HAT (Hybrid Attention Transformer)
• Task: 4x Super-Resolution & Denoising on low-light images
• Evaluation Metric: PSNR
• Performance: Achieved ~39.57 PSNR on a blind Kaggle test set.

Files Included

• hat_phase2.pth: The final fine-tuned PyTorch model weights.
• hat_arch.py: The architecture definition used to instantiate the model.
• config.json: The exact hyperparameters required to construct the architecture.

Evaluation

Note on Reproducibility: The reported PSNR of 39.57 is achieved using 8-fold Test-Time Augmentation (TTA) combined with overlapping patch inference (tile size 128) to prevent memory constraints.

Running a single raw forward pass without TTA will result in a slightly lower PSNR. Please refer to the provided inference script/notebook for the exact TTA implementation.