HAN

README.md

@@ -1,2 +1,72 @@
 # pytorch_HAN
-Heterogeneous Graph Attention Network (HAN) with pytorch
+Paper address:<br>
+<a href="https://github.com/Jhy1993/Representation-Learning-on-Heterogeneous-Graph">https://github.com/Jhy1993/Representation-Learning-on-Heterogeneous-Graph</a><br>
+Heterogeneous Graph Attention Network (HAN) with pytorch. If you want to pursue the performance in the original paper, 
+this may not be suitable for you, because there is still a problem: training loss decreases, but verification loss increases.<br>
+
+If you just want to figure out the basic principles of HAN and how to change tensorflow code to pytorch code, then this is for you.
+I implemented it according to the original tensorflow code structure.<br>
+
+If you want to pursue higher performance, please refer to:<br>
+Official tensorflow implementation:<a href="https://github.com/Jhy1993/HAN">https://github.com/Jhy1993/HAN</a><br>
+DGL implementation:<a href="https://github.com/dmlc/dgl/tree/master/examples/pytorch/han">https://github.com/dmlc/dgl/tree/master/examples/pytorch/han</a><br>
+
+# The result
+Address of ACM data set：
+Preprocessed ACM can be found in: <a href="https://pan.baidu.com/s/1V2iOikRqHPtVvaANdkzROw">https://pan.baidu.com/s/1V2iOikRqHPtVvaANdkzROw</a> 提取码：50k2<br>
+
+You can use the following command to run:
+```python
+python main.py
+```
+
+Training result:
+```python
+600 300 2125
+y_train:(3025, 3), y_val:(3025, 3), y_test:(3025, 3), train_idx:(1, 600), val_idx:(1, 300), test_idx:(1, 2125)
+2
+model: pre_trained/acm/acm_allMP_multi_fea_.ckpt
+fea_list[0].shape torch.Size([1, 1870, 3025])
+biases_list[0].shape: torch.Size([1, 3025, 3025])
+3
+2
+torch.Size([1, 1870, 3025]) torch.Size([1, 3025, 3025])
+torch.Size([1, 1870, 3025]) torch.Size([1, 3025, 3025])
+训练节点个数： 600
+验证节点个数： 300
+测试节点个数： 2125
+epoch:001, loss:1.1004, TrainAcc:0.3517, ValLoss:1.1022, ValAcc:0.4000
+epoch:002, loss:1.0762, TrainAcc:0.4250, ValLoss:1.1980, ValAcc:0.0533
+epoch:003, loss:1.0007, TrainAcc:0.6300, ValLoss:1.4572, ValAcc:0.0533
+epoch:004, loss:0.8876, TrainAcc:0.6583, ValLoss:2.0040, ValAcc:0.0500
+epoch:005, loss:0.8145, TrainAcc:0.6350, ValLoss:2.7091, ValAcc:0.0500
+epoch:006, loss:0.7897, TrainAcc:0.6267, ValLoss:3.2186, ValAcc:0.0500
+epoch:007, loss:0.7804, TrainAcc:0.6150, ValLoss:3.4550, ValAcc:0.0500
+epoch:008, loss:0.7527, TrainAcc:0.6150, ValLoss:3.5096, ValAcc:0.0500
+epoch:009, loss:0.7404, TrainAcc:0.6117, ValLoss:3.5125, ValAcc:0.0600
+epoch:010, loss:0.7329, TrainAcc:0.6633, ValLoss:3.5349, ValAcc:0.0400
+epoch:011, loss:0.7169, TrainAcc:0.6983, ValLoss:3.5743, ValAcc:0.0133
+epoch:012, loss:0.6934, TrainAcc:0.6917, ValLoss:3.6612, ValAcc:0.0033
+epoch:013, loss:0.6711, TrainAcc:0.6750, ValLoss:3.7738, ValAcc:0.0033
+epoch:014, loss:0.6645, TrainAcc:0.6733, ValLoss:3.9418, ValAcc:0.0200
+epoch:015, loss:0.6652, TrainAcc:0.6833, ValLoss:4.0934, ValAcc:0.0300
+epoch:016, loss:0.6515, TrainAcc:0.6883, ValLoss:4.2498, ValAcc:0.0300
+epoch:017, loss:0.6238, TrainAcc:0.7050, ValLoss:4.4304, ValAcc:0.0300
+epoch:018, loss:0.6082, TrainAcc:0.7317, ValLoss:4.5820, ValAcc:0.0333
+epoch:019, loss:0.6030, TrainAcc:0.7517, ValLoss:4.7110, ValAcc:0.0367
+epoch:020, loss:0.5933, TrainAcc:0.7850, ValLoss:4.8053, ValAcc:0.0400
+epoch:021, loss:0.5824, TrainAcc:0.8267, ValLoss:4.8781, ValAcc:0.0333
+epoch:022, loss:0.5655, TrainAcc:0.8017, ValLoss:4.9006, ValAcc:0.0267
+epoch:023, loss:0.5333, TrainAcc:0.8083, ValLoss:4.9148, ValAcc:0.0167
+epoch:024, loss:0.5175, TrainAcc:0.8050, ValLoss:4.8788, ValAcc:0.0100
+epoch:025, loss:0.4994, TrainAcc:0.8117, ValLoss:4.7670, ValAcc:0.0033
+epoch:026, loss:0.4888, TrainAcc:0.8333, ValLoss:4.5965, ValAcc:0.0033
+```
+This is where the problem lies.<br>
+If you know how to solve this problem, please don't hesitate to tell me.
+
+
+
+
+
+

layer.py

@@ -18,11 +18,6 @@ def __init__(self,in_channel, out_sz, bias_mat, in_drop=0.0, coef_drop=0.0, acti
         self.coef_dropout = nn.Dropout(coef_drop)
         self.activation = activation
 
-        """
-        for m in self.modules():
-            if isinstance(m, nn.Conv1d):
-                nn.init.kaiming_normal_(m.weight, mode="fan_out")
-        """
 
     def forward(self,x):
         seq = x.float()
@@ -67,9 +62,9 @@ def __init__(self, inputs, attention_size, time_major=False, return_alphas=False
         self.reset_parameters()
 
     def reset_parameters(self):
-       nn.init.kaiming_uniform_(self.w_omega)
+       nn.init.xavier_uniform_(self.w_omega)
        nn.init.zeros_(self.b_omega)
-       nn.init.kaiming_uniform_(self.u_omega)
+       nn.init.xavier_uniform_(self.u_omega)
 
     def forward(self,x):
         #print("x.shape:",x.shape)

main.py

@@ -2,7 +2,35 @@
 import torch
 import torch.nn as nn
 import torch.optim as optim
+from model import *
+from utils import *
 
+device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')
+#device = "cpu"
+
+adj_list, fea_list, y_train, y_val, y_test, train_mask, val_mask, test_mask, my_data = load_data_dblp()
+nb_nodes = fea_list[0].shape[0] #节点数目 3025
+ft_size = fea_list[0].shape[1] #特征的维度 1870
+nb_classes = y_train.shape[1]  #标签的数目 3
+
+fea_list = [torch.transpose(torch.from_numpy(fea[np.newaxis]),2,1).to(device) for fea in fea_list]
+#fea_list = torch.from_numpy(np.array(fea_list)).to(device)
+adj_list = [adj[np.newaxis] for adj in adj_list]
+y_train = y_train[np.newaxis]
+y_val = y_val[np.newaxis]
+y_test = y_test[np.newaxis]
+#train_mask = train_mask[np.newaxis]
+#val_mask = val_mask[np.newaxis]
+#test_mask = test_mask[np.newaxis]
+
+my_labels = my_data['my_labels']
+train_my_labels = my_data['train_my_labels']
+val_my_labels = my_data['val_my_labels']
+test_my_labels = my_data['test_my_labels']
+
+
+biases_list = [torch.transpose(torch.from_numpy(adj_to_bias(adj, [nb_nodes], nhood=1)),2,1).to(device) for adj in adj_list]
+print(len(biases_list))
 
 dataset = 'acm'
 featype = 'fea'

model.py

@@ -1,5 +1,6 @@
 import torch
-import trch.nn as nn
+import torch.nn as nn
+from layer import *
 
 class HeteGAT_multi(nn.Module):
     def __init__(self, inputs_list, nb_classes, nb_nodes, attn_drop, ffd_drop,

utils.py

@@ -24,7 +24,7 @@ def adj_to_bias(adj, sizes, nhood=1):
                     mt[g][i][j] = 1.0
     return -1e9 * (1.0 - mt)
 
-def load_data_dblp(path='/data02/gob/ACM3025.mat'):
+def load_data_dblp(path='data/ACM3025.mat'):
     data = sio.loadmat(path)
     #truelabels：[3025,3] truefeatures:[3025,1870]
     truelabels, truefeatures = data['label'], data['feature'].astype(float)