search

PaddlePaddle / models

Officially maintained, supported by PaddlePaddle, including CV, NLP, Speech, Rec, TS, big models and so on.
Apache License 2.0
6.92k stars 2.91k forks source link

PyramidBox face_detection: is it possible to transpile to float16? #1924

Open colojaro opened 5 years ago

colojaro commented 5 years ago

I have seen multiple references in PaddlePaddle code to transpiling to float16 with gains on Volta GPUs. This out of date doc gives instructions but I can't find anything for recent versions: http://www.paddlepaddle.org/documentation/docs/en/1.3/design/data_type/float16.html

I don't know Chinese but it seems as though the feature has been removed or moved. Is it possible to convert the face detector and will it be faster?

qingqing01 commented 5 years ago

We add example for image classfication on ImageNet, see: https://github.com/PaddlePaddle/models/tree/develop/fluid/PaddleCV/image_classification/#using-mixed-precision-training .

But, we haven't experimented on PyramidBox. We have plans to do F16 for detection in the future.