shreyashpc/swinir-dlp26t2-nppe3

SwinIR — DLP26T2 NPPE-3

Fine-tuned SwinIR-M model for low-light/noisy image denoising and 4x super-resolution.

Model

Architecture: SwinIR-M

• —Upscale factor: 4x
• —Input channels: 3
• —Patch size: 64x64 LR
• —Window size: 8
• —Depths: [6, 6, 6, 6, 6, 6]
• —Embedding dimension: 180
• —Attention heads: [6, 6, 6, 6, 6, 6]
• —MLP ratio: 2
• —Upsampler: nearest+conv
• —Residual connection: 1conv

Training

The model was initialized from the pretrained SwinIR-M x4 PSNR model and fine-tuned on the competition training pairs.

Training objective:

• —L1 loss
• —AdamW optimizer
• —Weight decay: 1e-4
• —Final fine-tuning learning rate: 2.5e-6
• —FP32 training
• —Gradient clipping: 1.0

Final checkpoint:

• —Epoch: 5
• —Training L1 loss: 0.00940543

Competition Result

Best Kaggle submission:

• —PSNR: 39.60561 dB

The best submission used an output ensemble of:

• —70% Epoch 5
• —20% Epoch 4
• —10% Epoch 3

with 8-way test-time augmentation.

The Hugging Face weights in this repository correspond to the final Epoch 5 trained model.

Files

• —pytorch_model.bin — trained model weights
• —config.json — architecture configuration
• —modeling_swinir.py — SwinIR architecture implementation
• —requirements.txt — Python dependencies

Input / Output

Input:

• —RGB low-resolution image
• —256x160 pixels

Output:

• —RGB 4x super-resolved image
• —1024x640 pixels

Loading

The architecture can be reconstructed using the configuration in config.json and the weights in pytorch_model.bin.