search

jsll / pytorch_6dof-graspnet

MIT License
146 stars 55 forks source link

Weird VAE results #14

Open bearpaw opened 3 years ago

bearpaw commented 3 years ago

Hi @jsll , I found the VAE results are pretty weird compared with the GAN results. Here are the examples:

python -m demo.main --generate_dense_grasps --num_grasp_samples 20 --grasp_sampler_folder checkpoints/vae_pretrained/

image image

python -m demo.main --grasp_sampler_folder checkpoints/vae_pretrained/

image image image Any ideas?

jsll commented 3 years ago

Interesting findings. As far as I know, this also happened to me sometimes with the VAE network but not this severely. I should have uploaded the correct VAE so the problem should not be there.

What I do know, is that I did not train the PyTorch VAE for nearly as long of a time as the TensorFlow one, I think I just trained it for a couple of days in contract to a week or so for the TensorFlow model. I would bet most of the problem stems from that fact.

bearpaw commented 3 years ago

Thanks Jens! That makes sense.

Just to clarify that the results I showed were generated from the pretrained model you provided. Would you be able to upload the correct VAE once you get a chance? Thanks!

jsll commented 3 years ago

I actually think that should be the pretrained model, at least I did not find anyone performing better. If I have time end of this week, I might setup a new training round and leave it training for a longer time.

bearpaw commented 3 years ago

Thanks! I was wondering there might be some potential bugs that generated results like these.