Pytorch - nn.CrossEntropyLoss

Question

I want to apply nn.CrossEntropyLoss我想申请 nn.CrossEntropyLoss

output = model(data)
output.shape -> 1,30,7 (batch, frame, class)
label.shape -> 1,30 (batch, frame)

In this case,在这种情况下，

label = label.squeeze(0)
output = output.squeeze(0)
criterion = nn.CrossEntropyLoss()
loss = criterion(outputs, targets)

can I solve this?我能解决这个问题吗？

But if the batch size is 2但如果批量大小为 2

output.shape is 2,30,7 and label.shape is 2, 30 output.shape 是 2,30,7 和 label.shape 是 2, 30

How to apply loss = criterion(outputs, targets)如何应用loss = criterion(outputs, targets)

Answer 1

The loss function nn.CrossEntropyLoss can be applied to multi-dim predictions:损失 function nn.CrossEntropyLoss可以应用于多维预测：

All you need is to make sure your C dimension (7 in your case) is the second:您所需要的只是确保您的C维度（在您的情况下为 7）是第二个：

output = output.transpose(1, 2)  # B,30,7 -> B,7,30
loss = criterion(outputs, targets)

You do not need to change targets at all.您根本不需要更改targets 。