gyes00205 / VRDL_HW4

VRDL HW4

1. Install dependencies

Install required libraries.

pip install -r requirements.txt

2. Training code

Train IMDN model by following command.

• root: the training data directory
• scale: scale factor
• nEpochs: total epochs
• step_size: learning rate descreased every step epoch

python train_IMDN.py --root train_decoded/ --scale 3 -nEpochs 200 --step_size 50

3. Pre-trained models

The pretrained model is in checkpoint_x3. You clone the whole github to inference.

4. Inference code

You can reproduce the results by the following command.

• test_lr_folder: testing data directory
• output_folder: output directory
• checkpoint: model weights
• upscale_factor: upscale faactor

python inference.py  \
--test_lr_folder testing_lr_images/ \
--output_folder results_30epoch \
--checkpoint checkpoint_x3/epoch_30.pth \
--upscale_factor 3

Result

• Low reaolution images
  

• High resolution images (IMDN)
  

Reference

1. IMDN: https://github.com/Zheng222/IMDN
2. IMDN paper: https://arxiv.org/pdf/1909.11856v1.pdf
3. VDSR: https://github.com/twtygqyy/pytorch-vdsr
4. Seven ways to improve example-based single image super resolution, https://arxiv.org/abs/1511.02228