LODE: Deep Local Deblurring and A New Benchmark arxiv

1. Code structure

|-LocalBlurNet         
  |-config             #config files
  |-data               #dataloader
    |-dataloader.py    
  |-model              
  |-network            #network structure
    |-DeblurNet.py     
    |-layers.py        
    |-MaskNet.py       
  |-result             
  |-criterion.py       
  |-eval.py            
  |-train.py           

2. requirements

• Python=3.6
• pytorch=1.0.1
• visdom=0.1.8.9
• opencv-contrib-python=4.4.0.44
• opencv-python=4.1.2.30

3. training configuration

customizing your configuration file as follows:
```yaml
date: 1217               
gpu_available: '1,2'     #GPU IDS
gpu_num: 2               #number of GPU in-use
onlyTrainMask: false     
useMask: false           
fixMask: false           
usingSA: true            
usingMaskLoss: false     
usingSALoss: true        
finetune: false          
local: false             
other: ''                

dataset: real            

train_sharp: /path       
train_blur: /path        

train_sharp2: None       
train_blur2: None        
train_mask: None         

test_sharp: /path        
test_blur: /path         

resize_size: 0           
crop_size: 192           
model_dir: ./model       
result_dir: ./result     

batchsize: 56            

save_epoch: 1            
lr: 0.001                
step: [150, 200, 300]    

mask_pretrained_model: None 
sanet_pretrianed_model: None 

pretrained_model: ./path 

best_psnr: 0             
startEpoch: 0            
```

• module selection

  the relationship between configuration file and networks in paper

  configurations	UNet(BladeNet-)	BSA only	LBP only	BladeNet
  onlyTrainMask	×	×	×	×
  useMask	×	×	√	√
  fixMask	×	×	×	×
  usingSA	×	√	×	√
  usingMaskLoss	×	√	√	√
  usingSALoss	×	√	×	√

• experiment scripts

  for training：python train.py --c [PATH TO YAML FILE]

  for testing：python eval.py --c [PATH TO YAML FILE]