search

PRBonn / lidar-bonnetal

Semantic and Instance Segmentation of LiDAR point clouds for autonomous driving
http://semantic-kitti.org
MIT License
961 stars 206 forks source link

Failed to train SqueezeSeg from scratch #63

Closed lunw1024 closed 2 years ago

lunw1024 commented 4 years ago

@tano297 Hi, I used the exact config as your pretrained squeezeseg model, and I didn't do any modifications to the code. Why did I got an increasing training loss? BTW, lowering the learning rate won't improve its peak performance. Any help would be appreciated. image image image image

lunw1024 commented 4 years ago

I also tested SqueezeSegv3, and it seemed to work fine, getting to 0.26 iou after 1 epoch.

NagarajDesai1 commented 4 years ago

Can you share your valid_loss curve?

lunw1024 commented 4 years ago

@NagarajDesai1 valid_loss for squeezesegv1: image

jbehley commented 3 years ago

Sorry for the late reply: Is this your own data? If yes, the number of training data examples should be increased as it seems that your model is overfitting to the training data. Or you should use more data augmentation, like adding noise to the point cloud.

lunw1024 commented 3 years ago

Nope, I used the SemanticKITTI. The problem is that the training loss increases as well. So I am getting confused.