search

Juyong / CaricatureFace

The source code for paper "Landmark Detection and 3D Face Reconstruction for Caricature using a Nonlinear Parametric Model".
578 stars 109 forks source link

problem of " Error detected in AddmmBackward" in training #36

Closed aiaini66 closed 2 years ago

aiaini66 commented 2 years ago

When training to 500 epochs, the following error is reported because the vertices used to calculate the keypoint loss use the target:

Screenshot from 2022-02-17 10-25-18

What is the reason for this and how should I deal it?

aiaini66 commented 2 years ago

@Juyong I‘m looking forward to your reply!

aiaini66 commented 2 years ago

@RainbowRui I‘m looking forward to your reply!

RainbowRui commented 2 years ago

Do you use our recommended version of PyTorch?

aiaini66 commented 2 years ago

@RainbowRui i used pytorch1.4, but cuda is 11.2, i guess it is a version problem of cuda？

aiaini66 commented 2 years ago

@RainbowRui @Juyong I want to render the vertex values predicted by the network to the 2d plane, but I found that the magnitude of the trans matrix from model_1 is very different from the magnitude of the vertices obtained from model_2. How should the vertex values from model_2 be processed to match the trans matrix? The order of magnitude is consistent, or how do you normalize the vertex values of each cartoon graph? Screenshot from 2022-02-21 14-36-17 hope you reply. thanks!!!

RainbowRui commented 2 years ago

We use weak perspective projection in our method, and you can refer to Formula 9 in our paper.

aiaini66 commented 2 years ago

Thank you very much !!!