Stride for layer4 in resnet

Question

Stride for layer4 in resnet

ruotianluo opened this issue 7 years ago · 5 comments

ruotianluo commented 7 years ago

@philokey, I will create a new branch for this so that the current pretrained won't be messed up.

This issue was originally opened in ruotianluo/pytorch-resnet#2

Answer 1 · 2017-09-08T05:40:59.000Z

You can check the convert_from_tensorflow branch now.

Answer 2 · 2017-09-09T05:46:01.000Z

Thanks, I will try it.

Answer 3 · 2017-09-09T07:02:27.000Z

It's now significantly slower (because the stride 1 needs more computation) than tf-faster-rcnn, I will investigate what's the bottleneck. (0.31 -> 0.37 s per iter)

Answer 4 · 2017-10-03T14:59:22.000Z

@ruotianluo I'm trying to train a resnet101+faster rcnn model starting from image net weights. Changing the stride in layer4 from 2->1 does indeed increase the training time. Did you ever resolve the bottleneck?

Answer 5 · 2017-10-03T15:24:21.000Z

@Feynman27 I haven't profiled the code in detail. But it should be faster after commit df3cc05