Rum69/nppe-3

NPPE Super-Resolution Model

This repository contains a PyTorch model trained for the NPPE image super-resolution task.

Model

Architecture:

• StrongSRNet
• Residual learning over bicubic upsampling
• Residual Channel Attention Blocks (RCAB)
• 6 residual groups
• 10 RCAB blocks per group
• 96 feature channels
• 4x super-resolution

Training

Training patch size: 128x128 LR

Upscaled target patch: 512x512 HR

Optimizer: AdamW

Learning rate: 1e-4

Training epochs: 200

Validation

Best validation PSNR:

39.137919607422496

Best epoch:

91

Inference

The final predictions were generated using x8 test-time self-ensemble.

Files

• best_model_v2.pth - trained PyTorch checkpoint
• config.json - model/training configuration
• submission.csv - competition submission
• README.md - model documentation