[Model]fix model bugs (#71)

* [Model]fix init

* [Model]fix model bugs

Author: dddg617

examples/simplehgn/simplehgn_trainer.py

@@ -118,7 +118,7 @@ def main(args):
             val_micro_f1, val_macro_f1 = calculate_f1_score(val_logits, val_y)
             print("Epoch [{:0>3d}]  ".format(epoch + 1),
                "   train loss: {:.4f}".format(train_loss.item()),
-               "   val loss: {:.4f}".format(val_loss),
+               "   val loss: {:.4f}".format(val_loss.item()),
                "   val micro: {:.4f}".format(val_micro_f1),
                "   val macro: {:.4f}".format(val_macro_f1),)
             if val_loss < best_val_loss:
@@ -132,7 +132,7 @@ def main(args):
 
         model.load_weights(args.best_model_path+model.name+".npz", format='npz_dict')
         if tlx.BACKEND == 'torch':
-            model.to(data["x"].device)
+            model.to(data["x"][0].device)
         model.set_eval()
         logits = model(data['x'], data['edge_index'], data['e_feat'])
         test_logits = tlx.gather(logits, data['test_idx'])

gammagl/layers/conv/__init__.py

@@ -30,7 +30,7 @@
     'JumpingKnowledge',
     'HANConv',
     'ChebConv',
-    'SimpleHGNConv'
+    'SimpleHGNConv',
     'FAGCNConv',
     'GPRConv',
 ]

gammagl/layers/conv/simplehgn_conv.py

@@ -85,8 +85,8 @@ def __init__(self,
         self.out_feats = out_feats
         self.edge_embedding = tlx.nn.Embedding(num_etypes, edge_feats)
 
-        self.fc_node = tlx.nn.Linear(out_feats * heads, in_features=in_feats, b_init=None, W_init=tlx.initializers.XavierNormal(gain=1.414), name='fc_node')
-        self.fc_edge = tlx.nn.Linear(edge_feats * heads, in_features=edge_feats, b_init=None, W_init=tlx.initializers.XavierNormal(gain=1.414), name='fc_edge')
+        self.fc_node = tlx.nn.Linear(out_feats * heads, in_features=in_feats, b_init=None, W_init=tlx.initializers.XavierNormal(gain=1.414))
+        self.fc_edge = tlx.nn.Linear(edge_feats * heads, in_features=edge_feats, b_init=None, W_init=tlx.initializers.XavierNormal(gain=1.414))
 
         self.attn_src = self._get_weights('attn_l', shape=(1, heads, out_feats), init=tlx.initializers.XavierNormal(gain=1.414), order=True)
         self.attn_dst = self._get_weights('attn_r', shape=(1, heads, out_feats), init=tlx.initializers.XavierNormal(gain=1.414), order=True)
@@ -96,7 +96,7 @@ def __init__(self,
         self.attn_drop = tlx.nn.Dropout(attn_drop)
         self.leaky_relu = tlx.nn.LeakyReLU(negative_slope)
 
-        self.fc_res = tlx.nn.Linear(heads * out_feats, in_features=in_feats, b_init=None, W_init=tlx.initializers.XavierNormal(gain=1.414), name='fc_res') if residual else None
+        self.fc_res = tlx.nn.Linear(heads * out_feats, in_features=in_feats, b_init=None, W_init=tlx.initializers.XavierNormal(gain=1.414)) if residual else None
 
         self.activation = activation
 

tests/layers/conv/test_simplehgn_conv.py

@@ -5,7 +5,7 @@
 def test_simplehgn_conv():
     x = tlx.random_normal(shape=(4, 64))
     edge_index = tlx.convert_to_tensor([[0, 0, 0, 1, 2, 3], [1, 2, 3, 0, 0, 0]])
-    edge_feat = [0, 1, 2, 3, 4, 5]
+    edge_feat = tlx.convert_to_tensor([0, 1, 2, 3, 4, 5])
 
     conv = SimpleHGNConv(in_feats=64, out_feats=128, num_etypes=6, edge_feats=64, heads=8)
     out, _ = conv(x, edge_index, edge_feat=edge_feat)