how to calculate the loss?

[Chrisable]

Usually when we calculate the loss, we use the training node, such as the code below：

import torch.nn.functional as F

data = dataset[0]
optimizer = torch.optim.Adam(model.parameters(), lr=0.01)

for epoch in range(200):
    pred = model(data.x, data.edge_index)
    loss = F.cross_entropy(pred[data.train_mask], data.y[data.train_mask])

    # Backpropagation
    optimizer.zero_grad()
    loss.backward()
    optimizer.step()

But in this paper, they use the nodes of the test set and the validation set to calculate the loss：

if conf['model_name'] == 'PLP':
        loss = my_loss(logp[idx_no_train], cas[-1][idx_no_train])
    else:
        loss = F.kl_div(logp, cas[-1], reduction='batchmean')
    acc_train = accuracy(logp[idx_train], labels[idx_train])

The complete code is as follows: https://github.com/BUPT-GAMMA/CPF/blob/master/distill_dgl.py

Is it reasonable to use test sets to participate in calculating loss?

[EdisonLeeeee]

If I understand it correctly, the test/validation set, i.e., idx_no_train was used for knowledge distillation, which does not involve any label information. This seems reasonable to me.

[Chrisable]

Does that cause overfitting problems?

[EdisonLeeeee]

No. Overfitting typically occurs when a model fits the training data too closely.

[Chrisable]

Thank you very much.

[Chrisable]

I still don't understand, can you please explain more specifically？Thank you

[EdisonLeeeee]

Sorry if my answer was not clear. Can you please let me know which part you're still unclear about?

[Chrisable]

I still don't understand, can you please explain more specifically？Thank you   抓住努力的尾巴 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "Jintang ***@***.***>; 发送时间: 2023年6月20日(星期二) 晚上11:08 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [pyg-team/pytorch_geometric] how to calculate the loss? (Discussion #7612) No. Overfitting typically occurs when a model fits the training data too closely. — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[Chrisable]

是因为GCN是半监督学习，所以可以利用测试集和验证集的节点的特征信息吗  抓住努力的尾巴 ***@***.***  

…

------------------ 原始邮件 ------------------ 发件人: "Jintang ***@***.***>; 发送时间: 2023年6月21日(星期三) 下午3:44 收件人: ***@***.***>; 抄送: ***@***.***>; ***@***.***>; 主题: Re: [pyg-team/pytorch_geometric] how to calculate the loss? (Discussion #7612) Sorry if my answer was not clear. Can you please let me know which part you're still unclear about? — Reply to this email directly, view it on GitHub, or unsubscribe. You are receiving this because you authored the thread.Message ID: ***@***.***>

[EdisonLeeeee]

Yes but not exactly. This paper uses the predicted results (logits) from the teacher model to guide the training of the student model on unlabeled nodes (i.e., idx_no_train). Such an approach is feasible for both supervised and semi-supervised learning.