Wrong result while training using GPU

[ln1equals0]

Hi, we have a simple CNN network, which works fine on CPU, but since we have a new GPU installed (TITAN X), we want to train the network using GPU, but it gives us wrong results.

This is the code for training on CPU, it works fine

This is the result, I made it display the average loss every 100 iterations.

So then I modified it and try to make it work on GPU, this are the changes I made

It is indeed much faster, but the result is totally wrong, you can see the loss of first 100 - 300 iterations became 3.4+, and it doesn't converge at all, the loss will stay at 3.3-3.4 till the end of training.

Did I miss something? How can I make the training work on GPU. Thanks

[csxmli2016]

What is the result if you run test.sh in ~/torch file? Can everything pass successfully?