Could not reproduce paper results

[LounesDev]

Hi,

I tried to train your network on the Ape object. So I ran the command :
python train_net.py --cfg_file configs/linemod.yaml model ape cls_type ape but after 200 epochs I cannot find the same results than you. I followed every steps of your tutorial.

Regards

[pengsida]

Could you provide the tensorboard curves?
https://github.com/zju3dv/clean-pvnet#tensorboard

[LounesDev]

[pengsida]

It seems that the training totally failed.

[LounesDev]

Are there parameters that I have to set to make it works ?

[pengsida]

I am not sure. I will check it.

[LounesDev]

I also see that loss value become NaN after a large number of epochs. Somewhere between 150 & 200 epochs (not always the same epoch).

[pengsida]

I retrain the network on ape and get similar training curves.
Maybe there are some problems in the data you use for training.
I also found a bug in python run.py --type linemod cls_type ape, which had been fixed.