Commit @744883d07d3d7028072d6a3fc46551eade5f2f87

윤영준 2023-06-27

Loss function for Autoencoder, also we may have to use other pretrained network for feature extractor for the loss function or even not using feature extractor. I ... honestly do not think using neural network for VGG is strictly nesccesary, and may have to be replaced with other image preprocessing like MSCN which was implimented before.

@744883d07d3d7028072d6a3fc46551eade5f2f87

dd1c901

744883d

model/Autoencoder.py

--- model/Autoencoder.py

+++ model/Autoencoder.py


         self.skip_output3 = nn.Conv2d(in_channels=32, out_channels=3, kernel_size=3, stride=1, padding=1, bias=False)
 
         # Loss specific definitions
+        # the paper uses vgg16 for features extraction,
+        # however, since vgg16 is not a light model, we may consider it to be replaced
         self.vgg = vgg16(pretrained=True).features
         self.vgg.eval()
         for param in self.vgg.parameters():

             lm_loss += mse_loss
 
         # Compute lp_loss
-        src_vgg_feats = self.extract_vgg_feats(label_tensor)
-        pred_vgg_feats = self.extract_vgg_feats(output_list[-1])
+        src_vgg_feats = self.vgg(label_tensor)
+        pred_vgg_feats = self.vgg(output_list[-1])
 
         lp_losses = []
         for index in range(len(src_vgg_feats)):

 
         loss = lm_loss + lp_loss
 
-        return loss, inference_ret['skip_3']
-
-    def extract_vgg_feats(self, input_tensor):
-        # Extract features from the input tensor using the VGG network
-        feats = []
-        x = input_tensor
-        for layer_num, layer in enumerate(self.vgg):
-            x = layer(x)
-            if layer_num in {3, 8, 15, 22, 29}:
-                feats.append(x)
-        return feats
+        return loss, inference_ret['skip_3']

Add a comment

Open 0
Closed 0

List

...	...	@@ -39,6 +39,8 @@
39	39	self.skip_output3 = nn.Conv2d(in_channels=32, out_channels=3, kernel_size=3, stride=1, padding=1, bias=False)
40	40
41	41	# Loss specific definitions
	42	+ # the paper uses vgg16 for features extraction,
	43	+ # however, since vgg16 is not a light model, we may consider it to be replaced
42	44	self.vgg = vgg16(pretrained=True).features
43	45	self.vgg.eval()
44	46	for param in self.vgg.parameters():
...	...	@@ -106,8 +108,8 @@
106	108	lm_loss += mse_loss
107	109
108	110	# Compute lp_loss
109		- src_vgg_feats = self.extract_vgg_feats(label_tensor)
110		- pred_vgg_feats = self.extract_vgg_feats(output_list[-1])
	111	+ src_vgg_feats = self.vgg(label_tensor)
	112	+ pred_vgg_feats = self.vgg(output_list[-1])
111	113
112	114	lp_losses = []
113	115	for index in range(len(src_vgg_feats)):
...	...	@@ -116,14 +118,4 @@
116	118
117	119	loss = lm_loss + lp_loss
118	120
119		- return loss, inference_ret['skip_3']
120		-
121		- def extract_vgg_feats(self, input_tensor):
122		- # Extract features from the input tensor using the VGG network
123		- feats = []
124		- x = input_tensor
125		- for layer_num, layer in enumerate(self.vgg):
126		- x = layer(x)
127		- if layer_num in {3, 8, 15, 22, 29}:
128		- feats.append(x)
129		- return feats (파일 끝에 줄바꿈 문자 없음)
	121	+ return loss, inference_ret['skip_3'] (파일 끝에 줄바꿈 문자 없음)

Delete comment