What's the best hyper parameter of this porject?

[sumai-1998]

so many opts, and in training.md, just two default training setting command. I want to train with the best parameter setting

[ThibaultGROUEIX]

Hi @4ssass1n0r, Thanks for using the code. Do you want to do single-image reconstruction or auto-encoders? Best, Thibault

[ThibaultGROUEIX]

In any case, you can use the command line i provided for those two cases, there should be close to optimal. At least, they correspond to what we do in the paper. https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default

Have a nice day, Thibault

[sumai-1998]

In any case, you can use the command line i provided for those two cases, there should be close to optimal. At least, they correspond to what we do in the paper. https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default

Have a nice day, Thibault

every thanks! I was running ```python train.py --shapenet13 --visdom_port 8890''', it seems this is no better than [https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default] right?

[sumai-1998]

In any case, you can use the command line i provided for those two cases, there should be close to optimal. At least, they correspond to what we do in the paper. https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default Have a nice day, Thibault

every thanks! I was running ```python train.py --shapenet13 --visdom_port 8890''', it seems this is no better than [https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default] right?

I almost done with this command, I will compare the three cmd and their performance, today I just download the dataset, so I asked this question, thanks a lost

[sumai-1998]

but if anyone have a better parameter setting, I hope you can share here thanks

[ThibaultGROUEIX]

In any case, you can use the command line i provided for those two cases, there should be close to optimal. At least, they correspond to what we do in the paper. https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default Have a nice day, Thibault

every thanks! I was running ```python train.py --shapenet13 --visdom_port 8890''', it seems this is no better than [https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default] right?

No python train.py --shapenet13 --dir_name log/atlasnet_autoencoder_25_squares --nb_primitives 25 --template_type "SQUARE" is better. Otherwize you train with a single sphere primitive. You can check the defaults in this webpage : https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md

[sumai-1998]

In any case, you can use the command line i provided for those two cases, there should be close to optimal. At least, they correspond to what we do in the paper. https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default Have a nice day, Thibault

every thanks! I was running ```python train.py --shapenet13 --visdom_port 8890''', it seems this is no better than [https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md#autoencoder-default] right?

No python train.py --shapenet13 --dir_name log/atlasnet_autoencoder_25_squares --nb_primitives 25 --template_type "SQUARE" is better. Otherwize you train with a single sphere primitive. You can check the defaults in this webpage : https://github.com/ThibaultGROUEIX/AtlasNet/blob/master/doc/training.md

I got it now, thanks bro