Commit @3690a9e3555e148f80134cf4afb7174c58a4a9a6

윤영준 2023-06-23

updating loss function

@3690a9e3555e148f80134cf4afb7174c58a4a9a6

dbcc026

3690a9e

data/example.yaml

--- data/example.yaml

+++ data/example.yaml


     num_link: 1
     path_source: /path/to/source/files/file1.jpg
     path_raindrop1: /path/to/source/files/file1.jpg
+    data_source: #url or origin
+    license:
   def456: # hash key
     path: /path/to/source/files/file2.jpg
   ghi789: # hash key

dbcc026

3690a9e

model/AttentiveRNN.py

--- model/AttentiveRNN.py

+++ model/AttentiveRNN.py


 from torch import nn
 from torch.nn import functional as F
 
-# nn.Sequential does not handel multiple input by design
+# nn.Sequential does not handle multiple input by design, and this is a workaround
 # https://github.com/pytorch/pytorch/issues/19808#
 class mySequential(nn.Sequential):
     def forward(self, *input):

 
     def forward(self, x):
         cell_state = None
-        attention_map = None
+        attention_map = []
+        lstm_feats = []
         for generator_block in self.generator_blocks:
-            x, attention_map, cell_state, lstm_feats = generator_block(x, cell_state)
-        return x, attention_map
+            x, attention_map_i, cell_state, lstm_feats_i = generator_block(x, cell_state)
+            attention_map.append(attention_map_i)
+            lstm_feats.append(lstm_feats_i)
+        ret = {
+            'x' : x,
+            'attention_map_list' : attention_map,
+            'lstm_feats' : lstm_feats
+        }
+        return x, attention_map, lstm_feats
 
 # need fixing
 class AttentiveRNNLoss(nn.Module):
-    def __init__(self):
+    def __init__(self, theta=0.8):
         super(AttentiveRNNLoss, self).__init__()
-
+        self.theta = theta
     def forward(self, input_tensor, label_tensor):
         # Initialize attentive rnn model
         attentive_rnn = AttentiveRNN

         loss = 0.0
         n = len(inference_ret['attention_map_list'])
         for index, attention_map in enumerate(inference_ret['attention_map_list']):
-            mse_loss = (0.8 ** (n - index + 1)) * nn.MSELoss()(attention_map, label_tensor)
+            mse_loss = (self.theta ** (n - index + 1)) * nn.MSELoss()(attention_map, label_tensor)
             loss += mse_loss
 
         return loss, inference_ret['final_attention_map']

Add a comment

Open 0
Closed 0

List

...	...	@@ -2,7 +2,7 @@
2	2	from torch import nn
3	3	from torch.nn import functional as F
4	4
5		-# nn.Sequential does not handel multiple input by design
	5	+# nn.Sequential does not handle multiple input by design, and this is a workaround
6	6	# https://github.com/pytorch/pytorch/issues/19808#
7	7	class mySequential(nn.Sequential):
8	8	def forward(self, *input):
...	...	@@ -193,16 +193,24 @@
193	193
194	194	def forward(self, x):
195	195	cell_state = None
196		- attention_map = None
	196	+ attention_map = []
	197	+ lstm_feats = []
197	198	for generator_block in self.generator_blocks:
198		- x, attention_map, cell_state, lstm_feats = generator_block(x, cell_state)
199		- return x, attention_map
	199	+ x, attention_map_i, cell_state, lstm_feats_i = generator_block(x, cell_state)
	200	+ attention_map.append(attention_map_i)
	201	+ lstm_feats.append(lstm_feats_i)
	202	+ ret = {
	203	+ 'x' : x,
	204	+ 'attention_map_list' : attention_map,
	205	+ 'lstm_feats' : lstm_feats
	206	+ }
	207	+ return x, attention_map, lstm_feats
200	208
201	209	# need fixing
202	210	class AttentiveRNNLoss(nn.Module):
203		- def __init__(self):
	211	+ def __init__(self, theta=0.8):
204	212	super(AttentiveRNNLoss, self).__init__()
205		-
	213	+ self.theta = theta
206	214	def forward(self, input_tensor, label_tensor):
207	215	# Initialize attentive rnn model
208	216	attentive_rnn = AttentiveRNN
...	...	@@ -211,7 +219,7 @@
211	219	loss = 0.0
212	220	n = len(inference_ret['attention_map_list'])
213	221	for index, attention_map in enumerate(inference_ret['attention_map_list']):
214		- mse_loss = (0.8 ** (n - index + 1)) * nn.MSELoss()(attention_map, label_tensor)
	222	+ mse_loss = (self.theta ** (n - index + 1)) * nn.MSELoss()(attention_map, label_tensor)
215	223	loss += mse_loss
216	224
217	225	return loss, inference_ret['final_attention_map']

...	...	@@ -3,6 +3,8 @@
3	3	num_link: 1
4	4	path_source: /path/to/source/files/file1.jpg
5	5	path_raindrop1: /path/to/source/files/file1.jpg
	6	+ data_source: #url or origin
	7	+ license:
6	8	def456: # hash key
7	9	path: /path/to/source/files/file2.jpg
8	10	ghi789: # hash key

Delete comment