face.evoLVe icon indicating copy to clipboard operation
face.evoLVe copied to clipboard

Published 20 hours ago • verified publisher icon ZhaoJ9014

BatchNorm1d in last layer of backbone cause tensorRT conversion failed

Open flyingmrwang opened this issue 5 years ago • 1 comments

I am trying to convert model to tensorRT. In my exported onnx model, BatchNorm1d always comes along with unsqueeze before it, which is not supported by tensorRT5. I noticed there is BatchNorm1d in output layer of backbones. From my understanding, it is not a must to get face embedding. While the loss failed to converge once it is removed. Could anyone explain it a little bit?

flyingmrwang avatar Jan 09 '20 03:01 flyingmrwang

Hey I'm sorry but did you successfully convert this model to tensorRT?

khoapham2801 avatar Nov 19 '21 11:11 khoapham2801