pointGAN

pointGAN copied to clipboard

Hi,

I was wondering if you had encountered convergence issues when training an point autoencoder with EMD as loss function. I find that using a strictly convolutional decoder (instead of dense layers) leads to degenerate solutions. Did you encounter similar issues ? I found that very weird.

Also, I was getting seg faults with the pytorch wrapper you provided in Pytorch 0.4, so I rewrote one. I checked it against this. and it gave the same results, however something weird might be happening. You can find my wrapper [here] (https://gist.github.com/pclucas14/826c129aea315b02cbde0ec23ceb1068).

Thanks for your open source code btw, It's been a real help for my research.

-Lucas

Have you try the emd loss when training autoencoder? This can‘t converge. Do you have the same problem?

I have not tried in an autoencoding setup yet, will get back to you

Do you know, why in emd folder test.py file the point has the dimension(1500, 6)？？ This means the point has 6 columns. When i try to change the 6 to 3, appear a error. @pclucas14

and how to use your emd.py

What line are you referring to ?

dist = EMDModule()

p1 = torch.rand(10,1000,6) p2 = torch.rand(10,1500,6) points1 = Variable(p1,requires_grad = True) points2 = Variable(p2, requires_grad = True)

I don't understand why this is (10, 1000, 6 ). Why it is 6? A point is N * 3.

@pclucas14

The code you are referring to is not mine unfortunately. My guess is that the EMD loss is still defined for n >=1. However, I'm not sure if the CUDA implementation has support for n != 3

@pclucas14 have you implement a new code about EMD?

I have posted my wrapper for Hao Su's EMD implementation. I have not tested it however.

@pclucas14 can you give me the link?

https://gist.github.com/pclucas14/826c129aea315b02cbde0ec23ceb1068. Please let me know if it works for you.

ok thanks.

@gaohongfein Did you figure out why 6 is here p1 = torch.rand(10,1000,6)? I also have segmentation error when I use 3 here p1 = torch.rand(10,1000,3) @fxia22 thanks.

@RyanCV the original EMD code was developed for 6D point cloud (xyz + normal), so there is 6. You can set the last 3 dimensions to 0 for xyz only point cloud.

@fxia22 For the version of pytorch, in the readme it said pip install http://download.pytorch.org/whl/cu80/torch-0.1.11.post5-cp27-none-linux_x86_64.whl

but it cannot compile emd in pytorch-0.4, any solution? Because I found those two versions of pytorch are very different in many functions.