Softmax and binary classification problem in MoleculeNet

[jiaruHithub]

In MoleculeNet, there is many binary classfication problem datasets. In general, BCE loss should be used during training on the datasets of MoleculeNet. But, I generated a generic representation g_rep for each class of data in a dataset, When a graph is represented by GNN, I want the representation to match the generic vector g_rep, and the class corresponding to the vector g_rep with the highest matching possibility is the label of the graph.

Thus I use softmax to output the matching possibility between current graph representation and every vector g_rep, then I use F.nll_loss to compute the loss. And the target is label, Because the order of g_rep for each class is the same as the order of label in one-hot label.

But surprisingly, the accuracy rate of the network is only 50%, and it stays the same no matter how hard it is trained. Is there something wrong with me? Help. Thank you very much.

[rusty1s]

I might be missing something, but if you want to train your GNN to output the representation of g_rep, wouldn't you need to train with MSE loss? Then, during inference you would pick the label of the closest embedding.

[jiaruHithub]

I m sure it's random initalized. Could you please tell me why it equal to random guess? Why this full-connected layer must be pretrained?

[jiaruHithub]

Doesn't that '''pro_head''' train with loss？
Another strange phenomenon if I comment pred = F.softmax(pred, dim = 1) , the networ will be trained, but acc will stay around 60%

[rusty1s]

Yes, it should train with loss (I thought you just run this during inference as shown in your example). One reason that the softmax might fail is that you apply a RELU at the end of pro_head. Is that the case?

[jiaruHithub]

if I use softmax, the acc will always stay 50.000%, like random guess. And I had not use relu function in the end of pro_head. You mean I can try relu ?

[rusty1s]

Mh, no you shouldn't use relu at the end (I just thought it might explain this behavior). Please note that if you are using nll_loss, you have to use log_softmax instead.