Simple image classification model, why this error?

1

I try to output internal layers’ output of a cnn, but why the “preds” is not a Variable? How to solve this?

class CNN(nn.Module):
    def __init__(self, channel):
        super(CNN, self).__init__()
        self.pretrained = models.resnet18(pretrained=True)
        self.modules = list(self.pretrained.children())[:-3]
        self.resnet = nn.Sequential(*self.modules)
        self.layer1 =nn.AvgPool2d(14)
        self.layer2 = nn.Linear(channel,4)
    def forward(self,input):
        o0 = self.resnet(input)
        o1 = self.layer1(o0)
        o2 = o1.squeeze(2)
        o2 = o2.squeeze(2)
        o3 = self.layer2(o2)
        return o0,o2,o3
input=Variable(torch.FloatTensor(3,3,224,224)).cuda()
cnn=CNN(256).cuda()

att_feats,fc_feats,outputs = cnn(input)
print(outputs)
_ ,preds = torch.max(outputs.data,1)
print(preds)
target = Variable(torch.LongTensor(3,4)).cuda()
criterion=nn.CrossEntropyLoss().cuda()
loss= criterion(outputs,target)

the error looks like:

Variable containing:
-0.1038 0.0071 0.0303 0.0366
-0.0857 -0.0365 0.0138 0.0473
-0.0920 -0.0510 0.0131 0.0482
[torch.cuda.FloatTensor of size 3x4 (GPU 0)]

3
3
3
[torch.cuda.LongTensor of size 3x1 (GPU 0)]
RuntimeError: multi-target not supported at /home/jcc/pytorch/torch/lib/THCUNN/generic/ClassNLLCriterion.cu:17

2

The docs say that the second return value of torch.max is indeed a LongTensor. Does wrapping it in a Variable work for you? ie
loss = criterion(preds, Variable(target))

4

loss= criterion(Variable(preds),target)

cause this error:

KeyError: <class ‘torch.cuda.LongTensor’>

5

sorry, it should be loss= criterion(outputs,target), but still error

6

Sorry, I didn’t read the error message in your above post. The error message suggests that target should be a vector, not a 3x1 tensor – you can fix that with target.squeeze(1)

9

Hello, am receiving this error with some fastai custom classes.
Have not discovered where to try to squeeze the target.
In case you are familiar with the package it involves ColumnModelData and
a StructuredLearner with CrossEntropyLoss.

Not sure this is the error type but it is my top suspect.

10

What is the error message?

11

So sorry, seems I forgot it:

-> 1052 return torch._C._nn.nll_loss(input, target, weight, size_average, ignore_index, reduce)
1053 elif dim == 4:
1054 return torch._C._nn.nll_loss2d(input, target, weight, size_average, ignore_index, reduce)

RuntimeError: multi-target not supported at /opt/conda/conda-bld/pytorch_1518244421288/work/torch/lib/THNN/generic/ClassNLLCriterion.c:22