A Naive Approach to Deblur Photos Using Pix2Pix

(WARNING: Only a demonstrative project. Far from production ready.)

Blog post

Pre-requisites

• Nvidia GPU (Sorry no CPU mode)
• pytorch 0.2.0
• The dataset should be organized as:
  • dataset_root/
    • train/
      • |arbitrary_naming|/
        • train images
    • val/
      • |arbitrary_naming|/
        • validation images

The easiest way to reproduce results is to use nvidia-docker. This project is tested on ceshine/cuda-pytorch:0.2.0 image. There is also a Dockerfile in the root folder. Use docker build -t deblur . to build the docker image.

Training

The following assumes you have built a Docker image named deblur.

Start the docker image (replace /mnt/SSD_Data/mirflickr/ with the path to your dataset):

nvidia-docker run -ti --init -v /mnt/SSD_Data/mirflickr/:/data \
                  --name deblur-c --ipc=host deblur bash

Start training (use python train.py -h to see the list of available command-line parameters):

python train.py --dataset /data --batchSize 16 --nEpochs 200 \
                --cuda --testBatchSize 32 --lamb 5 --lrG 5e-5 --lrD 5e-5

Debug images will be write to debug folder. Modek checkpoints will be write to checkpoint.

Predicting / Deblurring

Work In Progress (deblur.py)

Currently only photos with longer edge shorter than 512 px are supported (Given that your GPU has enough RAM).