Deep Image Deblurring

For the project details, see:

• Relation: https://docs.google.com/document/d/1LiOAaUSIEWBpxRx6_HTnsfB5-Y9ZCqZPKQxvQJNx9RQ/edit?usp=sharing
• Presentation: https://docs.google.com/presentation/d/1uKaZ98-PazS-LcbIWVa0RbZZUyG4oiORhSBVfOCI5so/edit?usp=sharing

Download the project

git clone https://github.com/albertobagnacani/Deblur.git

and cd it:

cd Deblur/

Deployment

Run

conda env create -f environment.yml

to install the needed dependencies.

Activate the environment:

conda activate deblur

Dataset preparation

CIFAR-10

Create the directory that will contain the dataset:

mkdir -p res/datasets/cifar-10/

Download the dataset and unzip it:

wget https://www.cs.toronto.edu/~kriz/cifar-10-python.tar.gz
tar -xf cifar-10-python.tar.gz

Move the downloaded files inside the dataset folder:

mv cifar-10-batches-py/* res/datasets/cifar-10/

REDS

Create the directories that will contain the dataset:

mkdir -p res/datasets/REDS/train/train_sharp
mkdir -p res/datasets/REDS/train/train_blur
mkdir -p res/datasets/REDS/val/val_sharp
mkdir -p res/datasets/REDS/val/val_blur

Download the following files from https://seungjunnah.github.io/Datasets/reds.html and then unzip them:

• train_sharp
• val_sharp
• train_blur
• val_blur

Move the folders inside the dataset folder:

mv train_sharp/* res/datasets/REDS/train/train_sharp
mv train_blur/* res/datasets/REDS/train/train_blur
mv val_sharp/* res/datasets/REDS/val/val_sharp
mv val_blur/* res/datasets/REDS/val/val_blur

The structure should be:

res/datasets/reds/train/train_sharp/000
res/datasets/reds/train/train_sharp/001
...

Download models

Create the directories that will contain the models:

mkdir -p res/models/reds/checkpoints/
mkdir -p res/models/cifar/checkpoints/

Download the models from here:

• REDS-srn: https://drive.google.com/file/d/1nlkpi-Q3U7RGKczhXdrwtbt-HbkNHItU/view?usp=sharing
• CIFAR-10-srn: https://drive.google.com/file/d/1hx1-D4ogGWjbXjdBDZvo7RsANlFbCnl_/view?usp=sharing
• CIFAR-10-fcn: https://drive.google.com/file/d/1JNxLsSgLPNn_diqfMNTJHNBbNY-o-Qu6/view?usp=sharing
• CIFAR-10-unet: https://drive.google.com/file/d/1EbckBTXt0N3KS5yFVR4eT9GXzzM5lyI7/view?usp=sharing
• CIFAR-10-rednet: https://drive.google.com/file/d/1thLvmVehKwrccgdJlBYUx10wmn3tyrjO/view?usp=sharing

Move the downloaded files in the following paths (parent of the created checkpoints folders; be aware to respect this naming):

• REDS-srn: res/models/reds/
• CIFAR-10-(srn/fcn/unet/rednet): res/models/cifar/

The structure should be:

res/models/reds/model-reds-srn-100.h5
res/models/cifar/model-cifar-srn-100.h5
res/models/cifar/model-cifar-fcn-100.h5
res/models/cifar/model-cifar-unet-100.h5
res/models/cifar/model-cifar-rednet-100.h5

Run

Enter the src folder:

cd src

and run the main:

python3 main.py

A subset of the (hyper)parameters are defined in the src/params.json file. Note that this file is actually used by the src/main.py script.

Note that the first run may need some time to prepare the datasets in the structure needed by the project.

params.json file

The src/params.json contains a list of (hyper)parameters to run the script:

• "task": "cifar" or "reds". Dataset on which perform the training/prediction/evaluation
• "model": "srn" or "fcn" or "unet" or "rednet". The architecture to load
• "epochs": int. Number of epochs to train
• "batch_size": int. Batch size (on NVIDIA GEFORCE RTX 2060, batch size = 8 for the REDS task and batch size = 32 for the CIFAR task (up to 512 in this case))
• "initial_lr": float. Initial Learning Rate of the Neural Network
• "load_epoch": int. Epoch to load to resume training or making prediction/evaluation. If 0, it doesn't load any model/weights (new training)
• "action": int. Can be 0, 1 or 2 for training, predicting and evaluating the Neural Network, respectively
• "subset": boolean. true if the training/validation set must be a subset of the original one (for fast testing. Note that you have to create a subset manually. This option is usually left to false); false otherwise
• "seed": int. Seed
• "mc_period": int. ModelCheckpoint saving period (frequency in epochs of the the model/weights saving). If set to 1, the model/weights are saved at each epoch

The REDS test set (the downloaded validation set is actually used as a test set) predictions can be found in the res/datasets/REDS/out/val/folder/ folder.

Examples for params.json file

For training a new model, set:

• "load_epoch": 0 (no checkpoints to be resumed, new training)
• "action": 0

For predicting, set:

• "load_epoch": 100 (the model file must be named model-reds-100.h5 (or model-cifar-100.h5 for CIFAR-10 task) if you want to load the model trained after e.g. 100 epochs)
• "action": 1

For evaluating, set:

• "load_epoch": 100
• "action": 2