search

threedle / tetGAN

29 stars 3 forks source link

predict code execution error (2) #3

Open shiyuze991202 opened 1 year ago

shiyuze991202 commented 1 year ago

I have trained the network on A40x1 with epoch60, batchsize 10, the dataset used is shapenet-bech, the subdivision grid used is 0.20 and depth is 3. Although tensorboard can read the logs successfully, it generates an error as shown in the figure, and the predict.py function reports an error when running. error, I wonder if you can provide some help: Snipaste_2023-09-30_16-15-25

ChocolateBiscuit commented 1 year ago

I believe the shape of weights should be two-dimensional, could you check the shape of weights where it throws an error?

shiyuze991202 commented 1 year ago

I believe the shape of weights should be two-dimensional, could you check the shape of weights where it throws an error?

sry i didn't get your point, but then I turn off the option 'laplacian smooth' then the error occured like this Snipaste_2023-10-02_08-23-31 then I was wondering if there's something going wrong with during traning, so I tried two ways to figure it out,

  1. I set 'drop_last' of 'dataloader' into 'True'
  2. i set the batchsize into a number which can be exact division by the size of dataset Neither of the two methods mentioned above solved the problem, and the same error still occurs. Is there something wrong with how I'm processing the data, or is there something wrong with my predict.py code? The parameters for my processing dataset are mentioned in the previous comment.
ChocolateBiscuit commented 1 year ago

It seems that the vertices list is empty for the predicted mesh. Is your model predicting all 0 occupancies? I would also double check the processed meshes to see if they are empty.