fix lr, add relu to the last conv layer, add l2 norm, add dropout. da result to 60%, source only result to 52%.

datasets/gtsrb.py

@@ -15,7 +15,7 @@ def get_gtsrb(dataset_root, batch_size, train):
 
     # image pre-processing
     pre_process = transforms.Compose([
-        transforms.Resize((40, 40)),
+        transforms.Resize((48, 48)),
         transforms.ToTensor(),
         transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
     ])

datasets/synsigns.py

@@ -43,7 +43,7 @@ def get_synsigns(dataset_root, batch_size, train):
     """Get Synthetic Signs datasets loader."""
     # image pre-processing
     pre_process = transforms.Compose([
-        transforms.Resize((40, 40)),
+        transforms.Resize((48, 48)),
         transforms.ToTensor(),
         transforms.Normalize(mean=(0.5, 0.5, 0.5), std=(0.5, 0.5, 0.5))
     ])

experiments/synsigns_gtsrb.py

@@ -13,7 +13,7 @@ class Config(object):
     # params for path
     model_name = "synsigns-gtsrb"
     model_base = '/home/wogong/models/pytorch-dann'
-    note = 'fix_lr'
+    note = '48-fixrelu-l2-dropout'
     model_root = os.path.join(model_base, model_name, note + '_' + datetime.datetime.now().strftime('%m%d_%H%M%S'))
     os.makedirs(model_root)
     config = os.path.join(model_root, 'config.txt')
@@ -51,6 +51,7 @@ class Config(object):
     # params for SGD optimizer
     lr = 0.01
     momentum = 0.9
+    weight_decay = 1e-6
 
     def __init__(self):
         public_props = (name for name in dir(self) if not name.startswith('_'))

experiments/synsigns_gtsrb_src_only.py

@@ -13,7 +13,7 @@ class Config(object):
     # params for path
     model_name = "synsigns-gtsrb"
     model_base = '/home/wogong/models/pytorch-dann'
-    note = 'src-only'
+    note = 'src-only-48-fixrelu-l2-dropout'
     model_root = os.path.join(model_base, model_name, note + '_' + datetime.datetime.now().strftime('%m%d_%H%M%S'))
     os.makedirs(model_root)
     config = os.path.join(model_root, 'config.txt')
@@ -23,6 +23,7 @@ class Config(object):
 
     # params for datasets and data loader
     batch_size = 128
+    img_size = 40
 
     # params for source dataset
     src_dataset = "synsigns"
@@ -45,12 +46,13 @@ class Config(object):
     save_step = 100
     eval_step = 5
 
-    manual_seed = None
+    manual_seed = 42
     alpha = 0
 
     # params for optimizing models
     lr = 0.01
     momentum = 0.9
+    weight_decay = 1e-6
 
     def __init__(self):
         """save config to model root"""

models/model.py

@@ -213,24 +213,26 @@ class GTSRBmodel(nn.Module):
         self.restored = False
 
         self.feature = nn.Sequential(
-            nn.Conv2d(in_channels=3, out_channels=96, kernel_size=(5, 5)),  # 36
+            nn.Conv2d(in_channels=3, out_channels=96, kernel_size=(5, 5)),  # 36 ; 44
             nn.ReLU(inplace=True),
-            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),  # 18
-            nn.Conv2d(in_channels=96, out_channels=144, kernel_size=(3, 3)),  # 16
+            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),  # 18 ; 22
+            nn.Conv2d(in_channels=96, out_channels=144, kernel_size=(3, 3)),  # 16 ; 20
             nn.ReLU(inplace=True),
-            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),  # 8
-            nn.Conv2d(in_channels=144, out_channels=256, kernel_size=(5, 5)),  # 4
-            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),  # 2
+            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),  # 8 ; 10
+            nn.Conv2d(in_channels=144, out_channels=256, kernel_size=(5, 5)),  # 4 ; 6
+            nn.Dropout2d(),
+            nn.ReLU(inplace=True),
+            nn.MaxPool2d(kernel_size=(2, 2), stride=(2, 2)),  # 2 ; 3
         )
 
         self.classifier = nn.Sequential(
-            nn.Linear(256 * 2 * 2, 512),
+            nn.Linear(256 * 3 * 3, 512),
             nn.ReLU(inplace=True),
             nn.Linear(512, 43),
         )
 
         self.discriminator = nn.Sequential(
-            nn.Linear(256 * 2 * 2, 1024),
+            nn.Linear(256 * 3 * 3, 1024),
             nn.ReLU(inplace=True),
             nn.Linear(1024, 1024),
             nn.ReLU(inplace=True),
@@ -238,9 +240,9 @@ class GTSRBmodel(nn.Module):
         )
 
     def forward(self, input_data, alpha = 1.0):
-        input_data = input_data.expand(input_data.data.shape[0], 3, 40, 40)
+        input_data = input_data.expand(input_data.data.shape[0], 3, 48, 48)
         feature = self.feature(input_data)
-        feature = feature.view(-1, 256 * 2 * 2)
+        feature = feature.view(-1, 256 * 3 * 3)
         reverse_feature = ReverseLayerF.apply(feature, alpha)
         class_output = self.classifier(feature)
         domain_output = self.discriminator(reverse_feature)