First run of gradient descent

.gitignore

@@ -3,4 +3,6 @@
 *.csv
 *.ipynb
 *Logs*
-*.log
\ No newline at end of file
+*.log
+*test*
+*.ipynb*

1508.06576/feature_maps.py

 import torch
+import torch.nn as NN
 from logger import Logger
 LOGGER = Logger().logger()
 LOGGER.info("Started Feature Maps")
-device=torch.device( "cuda" if (torch.cude.is_available()) else 'cpu')
+device=torch.device( "cuda" if (torch.cuda.is_available()) else 'cpu')
+LOGGER.info("Running the model cuda_available = "+str(torch.cuda.is_available()))
 #Author: @meetdoshi
-class FeatureMaps:
+class FeatureMaps():
     def __init__(self,arch="vgg19"):
         '''
         Init function
         @params
         arch: str {vgg11,vgg13,vgg16,vgg19,vgg19bn}
         '''
+        super()
         try:
             self.model = torch.hub.load('pytorch/vision:v0.10.0',arch,pretrained=True)
         except:
@@ -49,7 +52,7 @@ class FeatureMaps:
                 fmaps.append(img)
             layer_num+=1
         return fmaps
-
+'''
 if __name__ == "__main__":
     fmap = FeatureMaps()
     model = fmap.get_model()
@@ -58,3 +61,4 @@ if __name__ == "__main__":
     print(len(weights))
     for weight in weights:
         print(type(weight),weight.shape)
+'''

1508.06576/loss.py

@@ -16,10 +16,10 @@ class Loss:
         l2_norm_sq = None
         try:
             diff = F-P
-            l2_norm_sq = np.sum(diff**2)
+            l2_norm_sq = torch.norm(diff)**2
         except Exception as e:
             LOGGER.error("Error computing loss",e)
-        return l2_norm_sq/2.0
+        return l2_norm_sq
 
     @staticmethod
     def gram_matrix(F):
@@ -40,21 +40,22 @@ class Loss:
         @params
         Author: @soumyagupta
         '''
-        num_channels = F[1]
-        h = F[2]
-        w = F[3]
+        num_channels = F.shape[1]
+        h = F.shape[2]
+        w = F.shape[3]
         style_gram_matrix = Loss.gram_matrix(F)
         target_gram_matrix = Loss.gram_matrix(A)
-        loss_s = np.sum((style_gram_matrix-target_gram_matrix)**2)
+        loss_s = torch.norm(style_gram_matrix-target_gram_matrix)**2
         constant = 1/(4.0*(num_channels**2)*((h*w)**2))
         return constant*loss_s
 
     @staticmethod
-    def total_loss(alpha,beta,cont_fmap_real,cont_fmap_noise,style_fmap_real,style_fmap_noise):
+    def total_loss(alpha,beta,cont_fmap_real,style_fmap_real,content_fmap_gen):
         '''
         Function which computes total loss and returns it
         @params
         Author: @jiteshg
         '''
-        loss_t = alpha*Loss.content_loss(cont_fmap_real,cont_fmap_noise) + beta*Loss.style_loss(style_fmap_real,style_fmap_noise)
+        for gen,cont,sty in zip(content_fmap_gen,cont_fmap_real,style_fmap_real):
+            loss_t = alpha*Loss.content_loss(cont,gen) + beta*Loss.style_loss(sty,gen)
         return loss_t
\ No newline at end of file

1508.06576/optimizer.py

 from loss import Loss
-from feature_maps import FeatureMaps
+from feature_maps import LOGGER, FeatureMaps
 import torch.optim as optim
 from torchvision.utils import save_image
-
+import matplotlib.pyplot as plt
+plt.ion()
 class Optimizer:
     @staticmethod
     def gradient_descent(content_img, style_img, content_img_clone):
@@ -14,24 +15,25 @@ class Optimizer:
         content_img_clone: Copy of Original Image
         Author: @gaurangathavale
         '''
-        epoch = 1000
+        LOGGER.info("Running gradient descent with the following parameters")
+        epoch = 5000
         learning_rate = 0.001
         alpha = 10
         beta = 100
+        LOGGER.info(f"{epoch},{learning_rate},{alpha},{beta}")
 
         optimizer=optim.Adam([content_img_clone],lr=learning_rate)
-        print(optimizer)
-
+        LOGGER.info("Optimizer = " + str(optimizer))
+        #fig = plt.figure()
+        #ax = fig.add_subplot(111)
+        feature_maps = FeatureMaps()
         for e in range(epoch):
     
-            feature_maps = FeatureMaps()
             content_fmaps = feature_maps.get_fmaps(content_img)
             style_fmaps = feature_maps.get_fmaps(style_img)
-            # content_clone_fmaps = feature_maps.get_fmaps(content_img_clone)
-            content_white_noise_fmaps = feature_maps.get_fmaps(content_img, [21])
-            style_white_noise_fmaps = feature_maps.get_fmaps(style_img, [21])
+            content_generated_fmaps = feature_maps.get_fmaps(content_img_clone)
             
-            total_loss = Loss.total_loss(alpha, beta, content_fmaps, content_white_noise_fmaps, style_fmaps, style_white_noise_fmaps)
+            total_loss = Loss.total_loss(alpha, beta, content_fmaps, style_fmaps, content_generated_fmaps)
             
             # clears x.grad for every parameter x in the optimizer. 
             # It’s important to call this before total_loss.backward(), otherwise it will accumulate the gradients from multiple passes.
@@ -42,8 +44,8 @@ class Optimizer:
             
             # Optimization Step / Update Rule
             optimizer.step()
-            
-            if(not (e%100)):
-                print(total_loss)
-                
+            #plt.clf()
+            #plt.plot(content_img_clone)
+            if(e%10):
+                LOGGER.info(f"Epoch = {e} Total Loss = {total_loss}")
                 save_image(content_img_clone,"styled.png")
\ No newline at end of file

1508.06576/preprocess.py

@@ -5,7 +5,7 @@ import torchvision.transforms as transforms
 from PIL import Image
 import numpy as np
 LOGGER = Logger().logger()
-device=torch.device( "cuda" if (torch.cude.is_available()) else 'cpu')
+device=torch.device( "cuda" if (torch.cuda.is_available()) else 'cpu')
 #Author: @meetdoshi
 class Preprocessor:
     @staticmethod
@@ -35,9 +35,9 @@ class Preprocessor:
         @params
         img: 3d numpy array
         '''
-        loader = transforms.Compose([transforms.ToTensor(),transforms.Resize([224,224]),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225],),])
+        loader = transforms.Compose([transforms.ToTensor(),transforms.Resize([512,512]),transforms.Normalize(mean=[0.485, 0.456, 0.406],std=[0.229, 0.224, 0.225],),])
         img = loader(img).unsqueeze(0)
-        assert img.shape == (1,3,224,224)
+        assert img.shape == (1,3,512,512)
         return img.to(device,torch.float)
 
 

1508.06576/style_transfer.py

+ 
 import os
+import warnings
+from optimizer import Optimizer
+from loss import Loss
+from preprocess import Preprocessor
+from feature_maps import FeatureMaps
 import numpy as np
 import time 
 import torch
 import argparse
+import torchvision.models as models
+import torch.optim as optim
+from torchvision.utils import save_image
+warnings.filterwarnings('ignore')
 from logger import Logger
 LOGGER = Logger().logger()
 LOGGER.info("Started Style Transfer")
-
 class StyleTransfer:
     '''
     Style Transfer Base Class
@@ -21,6 +30,22 @@ class StyleTransfer:
         @params: None
         Author: @gaurangathavale
         '''
+
+        device = torch.device( "cuda" if (torch.cuda.is_available()) else 'cpu')
+
+        content_img_path = 'test/content.jpg'
+        style_img_path = 'test/style.jpg'
+
+        content_img = Preprocessor.process(content_img_path)
+        style_img = Preprocessor.process(style_img_path)
+
+        content_img_clone = content_img.clone().requires_grad_(True)
+
+        Optimizer.gradient_descent(content_img, style_img, content_img_clone)
+
+if __name__ == "__main__":
+    stf = StyleTransfer()
+    stf.pipeline()
         
 
     
\ No newline at end of file