Thers is a question about cross_entropy

autograd
1

loss=F.cross_entropy(output,image)

When I use the loss function cross_entropy, there is a erro:
Traceback (most recent call last):
** File “cnn.py”, line 131, in **
** optimizer.step()**

RuntimeError: cuda runtime error (59) : device-side assert triggered at /opt/conda/conda-bld/pytorch_1512386481460/work/torch/lib/THC/generated/…/generic/THCTensorMathPointwise.cu:301

I add a sigmoid function before the net output,that can correctly run.But my input value is not only 0~1 , I use autoencoder,so I think the output by add sigmoid is not close the input

Why? How can I correctly use the cross_entropy?

2

The CrossEntropyLoss is used for classification with class labels.
I think it might be not the right criterion for your autoencoder, since the labels will most likely be floats.

As a workaround you could use nn.MSELoss as a criterion.
Would this work for you?

3

nn.L1Loss is good than nn.MSELoss(),but I think nn.L1Loss is not the best loss function for autoencoder

4

so, for CrossEntropyLoss ,the labels can not be floats? Then it must be what type value?

5

The targets should be of dtype=torch.long for CrossEntropyLoss.

6

Thank you for your help