ByteTensor to FloatTensor is slow?

1

Hey,
So I when I try to convert the resulting ByteTensor out of torch.ge() to FloatTensor by:
my_tensor.type(torch.FloatTensor)
it took around 2.5e-4
while when I check the converting of other types to FloatTensor they all took on the magnitude of e-5, why is this converting from Byte to Float relatively slow? Are there any other ways to do this?
Thanks a lot!

2

right now there’s no faster way than just typecasting (as you did).

3

With the latest 0.4 release, you can use the .to function along with device objects to make this much cleaner.

# on top of your script somewhere
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")

# in your code, use .to(device)
(x_u == 0).to(device, dtype=torch.float32)
2 Likes