hgt hidden

autogl/module/model/dgl/hetero/hgt.py

@@ -157,16 +157,20 @@ class HGT(nn.Module):
 
         if not self.num_layers == len(self.args["hidden"]):
             LOGGER.warn("layer size {} does not match the length of hidden units {}".format(self.num_layers, len(self.args["hidden"])))
+        
+        hidden_size = self.args["hidden"][0]
+        hidden_size = hidden_size//self.args["heads"]*self.args["heads"]
+        LOGGER.warn('only use the first hidden size={} (divided exactly the number of heads) for all HGT layers'.format(hidden_size))
 
         self.adapt_ws  = nn.ModuleList()
         for t in range(len(self.node_dict)):
-            self.adapt_ws.append(nn.Linear(self.args["features_num"], self.args["hidden"][0]))
+            self.adapt_ws.append(nn.Linear(self.args["features_num"], hidden_size))
 
-        for i in range(1, self.num_layers):
-            self.gcs.append(HGTLayer(self.args["hidden"][i - 1], self.args["hidden"][i], self.node_dict, self.edge_dict, \
+        for i in range(self.num_layers):
+            self.gcs.append(HGTLayer(hidden_size, hidden_size, self.node_dict, self.edge_dict, \
                 self.args["heads"], use_norm = self.args["use_norm"], dropout = self.args["dropout"]))
             
-        self.out = nn.Linear(self.args["hidden"][-1], self.args["num_class"])
+        self.out = nn.Linear(hidden_size, self.args["num_class"])
 
     def forward(self, G):
         h = {}
@@ -250,7 +254,7 @@ class AutoHGT(BaseHeteroModelMaintainer):
 
         self.hyper_parameters = {
             "num_layers": 2,
-            "hidden": [256],
+            "hidden": [256,256,256],
             "heads": 4,
             "dropout": 0.2,
             "act": "gelu",

test/performance/heterogeneous/dgl/hgt_main.py

@@ -83,7 +83,7 @@ if __name__=='__main__':
                 init=False
             ).from_hyper_parameter({
                 "num_layers": 2,
-                "hidden": [256,256,256],
+                "hidden": [256,256],
                 "heads": 4,
                 "dropout": 0.2,
                 "act": "gelu",