kakaobrain/trident

Low performance of Conv2d

Closed this issue ยท 2 comments

๐Ÿž Describe the bug

According to https://github.com/kakaobrain/trident/wiki/Performance#conv2d, trident is up to 40x slower than torch. Is the graph (and the table below) there correct? Or could it be a bug in the graph?

@jaedochung The graph is correct. Conv2d isn't optimized well. We will optimze Conv2d as soon as possible.

Sorry, we have decided to deprecate Conv2d. See #118. We will back Conv2d with performance and features.