search

HuguesTHOMAS / KPConv-PyTorch

Kernel Point Convolution implemented in PyTorch
MIT License
760 stars 153 forks source link

Training sensaturban does not converge #135

Open zhu-xiao-di opened 2 years ago

zhu-xiao-di commented 2 years ago

Hello, I use the dataset sensaturban according to train S3DIS. Py training kpconv, without changing parameters, using two Birmingham block 0.ply to Birmingham block_ 4. Ply (all split into 2x2 sizes), but the model does not converge. What parameters do I need to change, or can you provide code for training sensaturban, just like the code used for training sensaturban results in your paper? Thank you!

zhu-xiao-di commented 2 years ago

Please help me

HuguesTHOMAS commented 2 years ago

I am sorry I never tried this dataset myself, you can maybe ask the author of sensaturban dataset what parameters they used to get their results. Look at the in_radius and first_subsampling_dl parameters which are the ones that should be adapted to new datasets. Other issues already mention that.

However, given what you are saying, that the network does not converge, there could a problem with the way you load the data. Even with bad parameters, the network should converge, so maybe have a look at the batch points to make sure they are good.

zhu-xiao-di commented 2 years ago

Thank you very much!

Chang-007 commented 2 years ago

hello, I want to run this network on the SensatUrban dataset. Do you know how to convert the SensatUrban dataset to S3DIS format, thanks!