[Semantic Segmentation] - Problem when re-training Semantic segmentation, Accuracy decrease with parameter default?

[vanloctc]

Hi @charlesq34,

I'm really grateful when you public this source for me to have a chance to learn.
I'm trying to re-training PointNet Semantic Segmentation but seems to be the training process met overfitting/underfitting with the 10th epoch. I still try for researching the cause but I think you can know the more cleary result.
This is a summary result:

(log_train.txt) - almost file

EPOCH 009
mean loss: 271829.089959
accuracy: 0.549186

eval mean loss: 167083.881273
eval accuracy: 0.421601
eval avg class acc: 0.260372
**** EPOCH 010 ****

mean loss: 129570.972106
accuracy: 0.561843

eval mean loss: 103253.728045
eval accuracy: 0.484906
eval avg class acc: 0.297415
Model saved in file: log6/model.ckpt
**** EPOCH 011 ****

mean loss: 37190426.054835
accuracy: 0.354552

eval mean loss: 6327.641183
eval accuracy: 0.462198
eval avg class acc: 0.178751
**** EPOCH 012 ****

mean loss: 5618.020521
accuracy: 0.337940

eval mean loss: 4386.426989
eval accuracy: 0.203347
eval avg class acc: 0.133981

[vanloctc]

This image above is the plot of loss and accuracy of log_train.txt

[Swaraj-72]

Hello @vanloctc, have you found any feasible solution for your case? I have been facing similar issue with my custom dataset.