Training wgangp always generate same gen_voxel.obj

[xiumingzhang]

Training wgangp always generate same gen_voxel.obj

[qiminchen]

Any thoughts about this or any direction to look at? Thank you!

[xiumingzhang]

Sorry for the delayed response. Can you try if
#25
solves this problem?

[qiminchen]

I tried different iso-surfaces threshold but none of them worked, even I tried training the wgangp using ShapeNet, it still visualized the triangle, any suggestions? Thanks in advance.

[qiminchen]

I might figure out what causes this issue,

1. Since VoxelGenerator has already added sigmoid layer
   

   GenRe-ShapeHD/networks/networks.py

   Line 100 in 62fc39b

   layers.append(nn.Sigmoid())

   
   use_sigmoid should be set to False when visualizing the gen_voxel, sigmoid(sigmoid('gen_voxel')) will lead to a minimum value of 0.5 so that iso-surfaces threshold always smaller than the min value of the generated voxel. This should solve the triangle problem, am I right?
   

   GenRe-ShapeHD/visualize/visualizer.py

   Lines 103 to 106 in 62fc39b

   	for k in cls.voxels:
   	prefix = '{:04d}_%02d_' % counter + k + '.obj'
   	cls._vis_voxel(pack.get(k), join(outdir, prefix), instance_cnt,
   	voxel_isosurf_th)

2. I'm wondering if we should add sigmoid or tanh to the last layer of the VoxelDiscriminator?

[qiminchen]

Update: issue solved by setting the use_sigmoid=False when visualizing the gen_voxel. Still wondering if we should add sigmoid or tanh to the last layer of the VoxelDiscriminator?

[xiumingzhang]

Glad that visualization issue has been solved.

It makes sense to add sigmoid or tanh to the last layer. I think adding either would work too.