NervanaSystems/ngraph-mxnet

Verify MaskRCNN accuracy

louisfeng opened this issue · 2 comments

Check inference results vs. baseline.

There was a bug in ngraph that's causing incorrect results. After some exhaustive debugging, I have found the sigmoid was not handling mkldnn input layout correctly. I'll make a PR to fix it. With that fixed, I think we have reasonable accuracy compared to the CPU baseline. I'll follow up with @lnguyen-nvn on testing.

Verify