Method val

models/SIEN_model.py:238–300

(self)

                ind=ind+1
        return np.array(image) 
    def val(self):
        self.netG.eval()
        with torch.no_grad():
            # deg_image = self.var_L# /255.0
            
            deg_image = np.array(cv2.imread(self.de_path[0],cv2.IMREAD_UNCHANGED))/255
            print('read:',np.max(deg_image), np.min(deg_image))
            h,w,_ = deg_image.shape
            if(h<384):
                cv2.resize(deg_image, (w, 384))
                h = 384
            if(w<384):
                cv2.resize(deg_image, (384,h))
                w = 384
            while w%4!=0:
                w+=1
            while h%4!=0:
                h+=1
            deg_image = cv2.resize(deg_image,(w,h))
            print(deg_image.shape)
            print('resize:',np.max(deg_image), np.min(deg_image))

            test_image = deg_image
            print('deg_images.shape:', test_image.shape)
            h =  ((test_image.shape [0] // 256) +1)*256 
            w =  ((test_image.shape [1] // 256 ) +1)*256

            test_padding = np.zeros((h,w,3))+1
            test_padding[:test_image.shape[0],:test_image.shape[1]]=test_image

            test_image_p=self.split2(test_padding.reshape(1,h,w,3),1,h,w)
            #print('p:', test_image_p.shape)
            predicted_list=[]
            for l in range(test_image_p.shape[0]):
                #print("patch shape:",np.transpose(test_image_p[l], [2,0,1]).shape)
                tmp = np.transpose(test_image_p[l], [2,0,1])
                tmp = torch.from_numpy(tmp).reshape(1,3,256,256).float()
                #print(tmp.shape)
                self.netG(tmp)
                
                # for torch : 1.7.1
                predict = self.netG(tmp)[0].squeeze().permute((1,2,0)).cpu().numpy()
                # else :
                # predict = torch.permute(self.netG(tmp)[0].squeeze(), (1,2,0)).cpu().numpy()

                #print(predict.shape)
                predicted_list.append(predict)

            predicted_image = np.array(predicted_list)#.reshape()
            predicted_image=self.merge_image2(predicted_image,h,w)
            
            # if training with threshold
            # predicted_image = (predicted_image[:,:]>0.95)*1
            
            predicted_image=(predicted_image[:test_image.shape[0],:test_image.shape[1]]*255).round()
            print('predicted shape:', predicted_image.shape)
            print(predicted_image)