fix bugs

Author: Gaoyiminggithub

networks/deeplab_xception_transfer.py

@@ -775,7 +775,7 @@ def forward(self, input,adj1_target=None, adj2_source=None,adj3_transfer=None ):
         source_graph = self.source_featuremap_2_graph(x)
         source_graph1 = self.source_graph_conv1.forward(source_graph,adj=adj2_source, relu=True)
         source_graph2 = self.source_graph_conv2.forward(source_graph1, adj=adj2_source, relu=True)
-        source_graph3 = self.source_graph_conv3.forward(source_graph2, adj=adj2_source, relu=True)
+        source_graph3 = self.source_graph_conv2.forward(source_graph2, adj=adj2_source, relu=True)
 
         source_2_target_graph1_v5 = self.transpose_graph.forward(source_graph1, adj=adj3_transfer, relu=True)
         source_2_target_graph2_v5 = self.transpose_graph.forward(source_graph2, adj=adj3_transfer, relu=True)

networks/gcn.py

@@ -27,10 +27,10 @@ def reset_parameters(self):
         #     self.bias.data.uniform_(-stdv,stdv)
 
     def forward(self, input,adj=None,relu=False):
-        support = torch.matmul(input,self.weight)
+        support = torch.matmul(input, self.weight)
         # print(support.size(),adj.size())
         if adj is not None:
-            output = torch.matmul(adj,support)
+            output = torch.matmul(adj, support)
         else:
             output = support
         # print(output.size())
@@ -97,7 +97,7 @@ def forward(self, input, source_pre_fea):
         fea_node = torch.matmul(input1,self.pre_fea) # n x hw x n_classes
         weight_node = torch.matmul(input1,self.weight) # n x hw x hidden_layer
         # softmax fea_node
-        fea_node = F.softmax(fea_node,dim=-1)
+        fea_node = F.softmax(fea_node,dim=1)
         # print(fea_node.size(),weight_node.size())
         graph_node = F.relu(torch.matmul(fea_node.transpose(1,2),weight_node))
         return graph_node # n x n_class x hidden_layer
@@ -145,6 +145,9 @@ def forward(self, input, res_feature):
         new_node = torch.matmul(new_fea, self.node_fea) # batch x hw x nodes x 1
         new_weight = torch.matmul(input, self.weight)  # batch x node x channel
         new_node = new_node.view(batch, hi*wi, nodes)
+        # 0721
+        new_node = F.softmax(new_node, dim=-1)
+        #
         feature_out = torch.matmul(new_node,new_weight)
         # print(feature_out.size())
         feature_out = feature_out.transpose(2,3).contiguous().view(res_feature.size())
@@ -194,6 +197,9 @@ def forward(self, input, res_feature):
         # new_node = torch.matmul(new_fea, self.node_fea) # batch x hw x nodes x 1
         new_weight = torch.matmul(input, self.weight) # batch x node x channel
         new_node = new_node.view(batch, hi*wi, nodes)
+        # 0721
+        new_node = F.softmax(new_node, dim=-1)
+        #
         feature_out = torch.matmul(new_node,new_weight)
         # print(feature_out.size())
         feature_out = feature_out.transpose(2,3).contiguous().view(res_feature.size())
@@ -217,7 +223,7 @@ def __init__(self,in_features,out_features,begin_nodes=7,end_nodes=2,bias=False,
             self.bias = Parameter(torch.FloatTensor(out_features))
         else:
             self.register_parameter('bias',None)
-        self.reset_parameters()
+        # self.reset_parameters()
 
     def reset_parameters(self):
         # stdv = 1./math.sqrt(self.weight(1))
@@ -266,7 +272,8 @@ def norm_trans_adj(self,adj):  # maybe can use softmax
 if __name__ == '__main__':
 
     graph = torch.randn((7,128))
-    en = GraphConvolution(128,12)
-    # pred = (torch.rand((7,7))*7).int()
+    en = GraphConvolution(128,128)
     a = en.forward(graph)
-    print(a)
+    print(a)
+    # a = en.forward(graph,pred)
+    # print(a.size())