After 1000 epoch, the training loss and test accuracy remains the same?

Hi @zhichunguo

I am facing a challenge with MetaMGNN. I trained the model for 1000 epoch, but the loss and accuracy remains the same. This means it is not training the meta model at all. I am applying your code on my data with 79 node features, however, I tried with your papers features too and the result is similar. I am wondering I am doing mistakes somewhere....

DEBUG:root:Epoch: 0 Train Loss: 0.8704835772514343  Test Accuracy: 43.87436384313247    
DEBUG:root:Epoch: 1 Train Loss: 0.32417869567871094 Test Accuracy: 49.572593688964844   
DEBUG:root:Epoch: 2 Train Loss: 0.09464379400014877 Test Accuracy: 42.63082504272461    
DEBUG:root:Epoch: 3 Train Loss: 0.07533042877912521 Test Accuracy: 54.545582715202784   
DEBUG:root:Epoch: 4 Train Loss: 0.11637484282255173 Test Accuracy: 44.320564943201404   
DEBUG:root:Epoch: 5 Train Loss: 0.034378036856651306    Test Accuracy: 54.640758626601276   
DEBUG:root:Epoch: 6 Train Loss: 0.03435908630490303 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 7 Train Loss: 0.03385530039668083 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 8 Train Loss: 0.03292354196310043 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 9 Train Loss: 0.0315357968211174  Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 10    Train Loss: 0.03256123140454292 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 11    Train Loss: 0.03294479474425316 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 12    Train Loss: 0.032359760254621506    Test Accuracy: 50.21079287809484    
.
.
.

DEBUG:root:Epoch: 994   Train Loss: 0.031872864812612534    Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 995   Train Loss: 0.03162331134080887 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 996   Train Loss: 0.03111858479678631 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 997   Train Loss: 0.03308967500925064 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 998   Train Loss: 0.031367577612400055    Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 999   Train Loss: 0.03169362619519234 Test Accuracy: 50.21079287809484    
DEBUG:root:Epoch: 1000  Train Loss: 0.0316644087433815  Test Accuracy: 50.21079287809484    
INFO:root:Training has finished.

What I did include the followings:

I have extracted 79 node features and 10 edge features from my data

I want to run on MAML firts, so I set

parser.add_argument('--add_similarity', type=bool, default=False)
parser.add_argument('--add_selfsupervise', type=bool, default=False)
parser.add_argument('--interact', type=bool, default=False)

and

parser.add_argument('--num_tasks', type=int, default= 33, help='number of tasks')
parser.add_argument('--num_train_tasks', type=int, default= 16, help='number of meta-training task')
parser.add_argument('--num_test_tasks', type=int, default= 17, help='number of meta-testing tasks')
parser.add_argument('--n_ways', type=int, default=2, help='n-ways of dataset')
parser.add_argument('--m_support', type=int, default=10, help='size of support set')
parser.add_argument('--k_query', type=int, default=20, help='size of query set')
parser.add_argument('--meta_lr', type=float, default=0.001,
                help='meta learning rate (i.e., outer lr:  beta) (default: 0.001)')
parser.add_argument('--update_lr', type=float, default=0.005,
                help='task-specific learning rate (inner lr: alpha) (default: 0.04)')
parser.add_argument('--batch_size', type=int, default= 32,
                help='input batch size for training (default: 32)')
parser.add_argument('--epochs', type=int, default= 1000,
                help='number of epochs to train (default: 100)')
parser.add_argument('--lr', type=float, default=0.001,
                help='learning rate (default: 0.001')
parser.add_argument('--lr_scale',type=float, default=0.1,
                help='relative learning rate for feature extraction layer (default: 1)')
parser.add_argument('--decay', type=float, default=0,
                help='weight decay (default: 0)')
parser.add_argument('--num_layer' ,type=int, default=7,
                help='number of GNN message passing layer (default: 5)')
parser.add_argument('--emb_dim', type=int, default = 256,
                help='embedding dimension (default: 300)', choices=[64, 300])

zhichunguo / Meta-MGNN

After 1000 epoch, the training loss and test accuracy remains the same? #6