Issues running and changing the backend of the mnist example about burn HOT 6 CLOSED

For the question about the number provided to TchDevice::Cuda, I was curious as well and looked through the documentation. According to the backend, it is the device index.

from burn.

I think I saw that same source code which prompted me to try 0 and 1. It's still a bit unclear from the libraries what the value should be, is there a way to list valid device indices?

from burn.

I made a pull request that should fix the problem. The reason I did not see this before is that I always test using the --release flag, since it speeds up quite a bit the training without noticeable compilation time issue. I discovered that there was an underflow with the file checkpointer on debug builds.

I also added documentation on the TchDevice struct, the usize parameter is indeed the device index, which starts at zero. If you only have 1 GPU, it should just be zero.

I will work soon on error handling and proper logging to help understand those issues more clearly.

from burn.

Have you tried this trick in your toml to get the speed optimizations without needing --release and losing debug info:

[profile.dev]
opt-level = 0

[profile.dev.package."*"]
opt-level = 3

from burn.

I didn't think about this but yeah this is probably a good default and something to include in the example!

from burn.

In turns out that you can't set the optimization profile in packages under a workspace, I'm not sure how to make debug builds use a different optimization level (only for examples).

from burn.