Vectorized/VoxNet-Tensorflow

Problem with training

Opened this issue · 1 comments

Thank you for the implementation.

I followed the code voxnet_train.py. However, after several saved checkpoints 100, the percentage of training and test are still low around 0.14. It is not like you described " After 20 minutes of training from scratch on a Nvidia Titan X, it is able to get 86%+ test accuracy."

Could you verify again the voxnet_train.py?

Thank you very much

Sorry for the late reply.

Which version of Tensorflow are you using?
I'm using 1.9.0 GPU version and have tested it on another machine.
Not sure if a newer Tensorflow version will break things.

If you have modified or re-implemented it, do check the learning rates and the layers are similar.

Check that the cross-entropy loss is decreasing as you train. It should converge to +/- 0.2.
The learning rate begins at 0.001 is slowly lowered as training goes by using exponential decay.