This repository was archived by the owner on Nov 17, 2023. It is now read-only.
Repository navigation
Replies: 0 comments
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
Uh oh!
There was an error while loading. Please reload this page.
I am using the mxnet-cu101mkl version. Learning is in progress with the gpu version. Inference must inevitably go with the cpu version. I want to change the model architecture trained with float32 to float16. How can I change the datatype of nn.Conv2d?

(Note that the cpu version cannot use cast.)
All reactions