You signed in with another tab or window. Reload to refresh your session.You signed out in another tab or window. Reload to refresh your session.You switched accounts on another tab or window. Reload to refresh your session.Dismiss alert
XLA supports grouped convolutions with it's feature_group_count option (right now we always set it to 1). We will just need to add support for groups in the binary implementation of Nx.conv. This involves an additional iteration in the binary implementation as well as some additional shape checks.
XLA also supports batch_group_count for grouping parts of the batch. I see less value in adding this option as I haven't been able to find any applications that are all that common.
The text was updated successfully, but these errors were encountered:
Grouped Convolutions are common enough that we should support them. Here's some discussion and papers on grouped convolutions: https://paperswithcode.com/method/grouped-convolution
XLA supports grouped convolutions with it's
feature_group_count
option (right now we always set it to 1). We will just need to add support for groups in the binary implementation ofNx.conv
. This involves an additional iteration in the binary implementation as well as some additional shape checks.XLA also supports
batch_group_count
for grouping parts of the batch. I see less value in adding this option as I haven't been able to find any applications that are all that common.The text was updated successfully, but these errors were encountered: