search

mks0601 / 3DMPPE_POSENET_RELEASE

Official PyTorch implementation of "Camera Distance-aware Top-down Approach for 3D Multi-person Pose Estimation from a Single RGB Image", ICCV 2019
MIT License
807 stars 147 forks source link

There were different results in two different environments #89

Closed applech666 closed 3 years ago

applech666 commented 3 years ago

Hi, first of all, thank you for your great work. There were different results in two different environments.

  1. windows10, cuda9, torch=1.1, I can get a better result.
  2. ubuntu16.04 cuda10.1 torch1.4, I can get another result. Why do you think the same code gets inconsistent results in two environments?

thanks!

mks0601 commented 3 years ago

I haven't tested this code on window machines... could you use the same torch versions?

applech666 commented 3 years ago

windows10, cuda9, torch=1.1, In my opinion, I achieved a perfect result. But the same code doesn't work very well in either of the following environments:

  1. ubuntu16.04 cuda10 torch1.1,
  2. ubuntu cuda10.1 torch1.4, Is it the CUDA version that affects the reasoning effect of the model?
mks0601 commented 3 years ago

I guess the reason is a different version of torchvision? Several people said that different torchvision version make the batch normalization layer work different. Could you check the torchvision version?

applech666 commented 3 years ago

I just tried it, and on Linux, torchvision(0.4.1) stays the same as Windows, but it still doesn't improve

applech666 commented 3 years ago

supplement: ubuntu cuda=10.1 torch=1.3 torchvision=0.4.1

mks0601 commented 3 years ago

Sorry I don't get reasons :( You'd better update torch and torchvision

applech666 commented 3 years ago

I've solved the problem. On Linux, due to my image data set out of order. Thank you for your valuable time support

mks0601 commented 3 years ago

Good!