UMSN Files

README.md

@@ -21,7 +21,11 @@ We propose a novel multi-stream architecture and training methodology that explo
 3. python test_face_deblur.py --dataroot ./facades/github/ --valDataroot <path_to_test_data> --netG ./pretrained_models/Deblur_epoch_Best.pth
 
 ## To train UMRL:
-python train_face_deblur.py --dataroot <path_to_train_data> --valDataroot ./facades/github/ --exp ./face_deblur --batchSize 10
+1. Kernels are generated using,
+     - [Boracchi and Foi, 2012]	Modeling the Performance of Image Restoration from Motion Blur Giacomo Boracchi and Alessandro Foi, Image Processing, IEEE Transactions on. vol.21, no.8, pp. 3502 - 3517, Aug. 2012,
+     - 25000 kernels with size ranging from 13 to 29 are generated and saved as ".mat" file
+2. Clean face images from Helen and CelebA are aligned and used as input to train UMSN 
+3. python train_face_deblur.py --dataroot <path_to_train_data> --valDataroot ./facades/github/ --exp ./face_deblur --batchSize 10
 
 - input should be clean image. blurry images for training are generated by the code it self.
 

train_face_deblur.py

@@ -356,7 +356,7 @@ for epoch in range(opt.niter):
 
     #x1 = xmask1*class_msk1*x_hat+(1-xmask1)*class_msk1*target
     #smaps_hat,smaps64_hat = netS(x_hat1,x_hat64)
-    if epoch>4 or (epoch<4 and ganIterations%10 ~= 9):
+    if epoch>0:
       with torch.no_grad():
         smaps,smaps64 = netS(target,target_256)
       L_img_ = 0.33*criterionCAE(x_hat64, target_256) #+ 0.5*criterionCAE(smaps_hat, smaps)