modify attentionblock

main.py

@@ -198,10 +198,14 @@
                     else:
                         inference_array.extend(outputs.cpu().numpy())
                         target_array.extend(target.cpu().numpy())
+                    if tepoch.n == 0 and __TIME__:
+                        save_time = time.time()
+
             if model_params["name"] == "DeepPhys":
                 inference_array = scipy.signal.detrend(np.cumsum(inference_array))
                 target_array = scipy.signal.detrend(np.cumsum(target_array))
 
+            if __TIME__ and epoch == 0:
+                log_info_time("inference time \t: ", datetime.timedelta(seconds=save_time - start_time))
+
             plot_graph(0, 300, target_array, inference_array)
-        if __TIME__ and epoch == 0:
-            log_info_time("inference time \t: ", datetime.timedelta(seconds=time.time() - start_time))

nets/blocks/attentionBlocks.py

@@ -4,7 +4,7 @@
 from ..modules.modules import DAModule
 
 
-class AttentionBlock(nn.Module):
+class AttentionBlock_DA(nn.Module):
     def __init__(self, in_channels):
         super().__init__()
         self.attention = DAModule(in_channels)
@@ -19,3 +19,19 @@ def forward(self, input):
         mask = torch.div(mask * H * W, norm)
         mask = self.conv1x1(mask)
         return mask
+
+
+
+class AttentionBlock(torch.nn.Module):
+    def __init__(self, in_channels):
+        super().__init__()
+        self.attention = torch.nn.Conv2d(in_channels, 1, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, input):
+        mask = self.attention(input)
+        mask = torch.sigmoid(mask)
+        B, _, H, W = input.shape
+        norm = 2 * torch.norm(mask, p=1, dim=(1, 2, 3))
+        norm = norm.reshape(B, 1, 1, 1)
+        mask = torch.div(mask * H * W, norm)
+        return mask

nets/models/DeepPhys.py

@@ -14,7 +14,7 @@ def __init__(self):
         self.attention_mask1 = None
         self.attention_mask2 = None
 
-        self.appearance_model = AppearanceModel_2D(in_channels=self.in_channels, out_channels=self.out_channels * 4,
+        self.appearance_model = AppearanceModel_2D(in_channels=self.in_channels, out_channels=self.out_channels,
                                                    kernel_size=self.kernel_size)
         self.motion_model = MotionModel(in_channels=self.in_channels, out_channels=self.out_channels,
                                         kernel_size=self.kernel_size)

nets/models/sub_models/MotionModel.py

@@ -28,7 +28,7 @@ def forward(self, inputs, mask1, mask2):
         M1 = torch.tanh(self.m_batch_Normalization1(self.m_conv1(inputs)))
         M2 = self.m_batch_Normalization2(self.m_conv2(M1))
         # element wise multiplication Mask1
-        g1 = torch.tanh(torch.mul(1 * mask1, M2))
+        g1 = torch.tanh(torch.mul(torch.ones(size= mask1.shape).to('cuda')@mask1, M2))
         M3 = self.m_dropout1(g1)
         # pooling
         M4 = self.m_avg1(M3)

params.json

@@ -1,7 +1,7 @@
 {
     "__TIME__" : 1,
     "__PREPROCESSING__" : 0,
-    "__MODEL_SUMMARY__" : 0,
+    "__MODEL_SUMMARY__" : 1,
     "options":{
         "parallel_criterion" : 1,
         "parallel_criterion_comment" : "TODO need to verification"
@@ -18,7 +18,7 @@
         "validation_ratio_comment" : "split train dataset using validation_ratio",
         "train_batch_size" : 32,
         "train_shuffle" : 0,
-        "test_batch_size" : 1,
+        "test_batch_size" : 32,
         "test_shuffle" : 0
     },
     "hyper_params":
@@ -38,16 +38,16 @@
                     "adam","sgd","rms_prop","ada_delta","ada_grad","ada_max",
                     "ada_mw","a_sgd","lbfgs","n_adam","r_adam","rprop","sparse_adam"
                 ],
-        "learning_rate": 1,
+        "learning_rate": 0.001,
         "learning_rate_comment": [
             "DeepPhys : lr = 1",
             "PhysNet : lr = 0.001"
         ],
-        "epochs" : 1
+        "epochs" : 30
     },
     "model_params":
     {
-        "name": "PhysNet",
+        "name": "DeepPhys",
         "name_comment":
                 [
                     "DeepPhys",