Training is very slow on Intel Xeon 2.3Ghz, CUDA not working

[sharan21]

Training is very slow on Intel Xeon 2.3Ghz, CUDA not working

[tirtharajdash]

I figured this out.
If you have just one GPU card (assuming it's working and well-set, of course), you can change the default ID in the argument "gpu_device" from 1 to 0. This is the standard PCI-BUS ID of the first GPU.

Alternatively, you can run it as:
python main.py --gpu_device=0

However, before you could run the code without error, two changes are needed in main.py.

In the function training():

for batch in train_iter:
    #rec_loss, kl_loss = train_batch(x, G_inp, step, train=True)
    rec_loss, kl_loss = train_batch(get_cuda(x), get_cuda(G_inp), step, train=True) #TD

Similarly, in the validation iteration:

#rec_loss, kl_loss = train_batch(x, G_inp, step, train=False)
rec_loss, kl_loss = train_batch(get_cuda(x), get_cuda(G_inp), step, train=False) #TD

It works fine afterwards. I think this was a bug.