search

TiagoCortinhal / SalsaNext

Uncertainty-aware Semantic Segmentation of LiDAR Point Clouds for Autonomous Driving
MIT License
417 stars 102 forks source link

Bug: `proj_argmax` undefined. #12

Closed EthanZhangYi closed 4 years ago

EthanZhangYi commented 4 years ago

https://github.com/TiagoCortinhal/SalsaNext/blob/1ee74804219635e660a50c6ed2f53dd09096db83/train/tasks/semantic/modules/user.py#L160-L169

TiagoCortinhal commented 4 years ago

I just pushed it =)

zeng-hello-world commented 3 years ago

@TiagoCortinhal You didn't fix it with uncertainty network.

TiagoCortinhal commented 3 years ago

Hello @zeyu-hello!

If you check the if condition without uncertainty you will notice the only thing missing is proj_argmax = proj_output[0].argmax(dim=0). I will push it later today (and check if anything else is missing) but for now you can try adding this!

Best

RaphaelLorenzo commented 3 years ago

Hello @TiagoCortinhal and thank you for the great work ! It seems that proj_argmax it is still undefined for the case of self.uncertainty==True. I fixed it by using : proj_output2,log_var2 = self.model(proj_in) proj_output = proj_output_r.var(dim=0, keepdim=True).mean(dim=1) proj_argmax = proj_output2[0].argmax(dim=0)

I Hope it is consistent ?

Using proj_argmax = proj_output[0].argmax(dim=0) does not work (gives a 1 dimensional tensor of size 2048)