How to calculate the loss during training.

[gabrielpeixoto-cvai]

I am not sure this question has been asked before. I searched the issues by keyword and could not find anything. I want to use this chamfer distance as the loss to train a network (more specifically, a pointnet-like autoencoder).
Currently, I am using it like this (based on the python version):

import dist_chamfer_2D
loss_chamfer = dist_chamfer_2D.chamfer_2DDist()
dist1, dist2, idx1, idx2 = loss_chamfer(x.permute(0, 2, 1), dec_y.permute(0, 2, 1))
loss = (dist1.min(dim=0)[0].mean()) + (dist2.min(dim=0)[0].mean())

Is this the correct way of using it?

However, the reconstructed result does not look good. I also tried to define the loss as:

loss = torch.sum(dist1) + torch.sum(dist2)

Which had a better overall qualitative result but still not as I expected.
The problem should not be hard, and I am trying to learn a representation for a simple 2D/3D point cloud (composed of squares and circles ).

Example below (blue is the original, red is the decoder output):

[ThibaultGROUEIX]

• v1 will not work well most likely : you are backpropagating gradients only to two points because of min()
• v2 is reasonnable. People use a variant usually

loss = torch.mean(dist1**2) + torch.mean(dist2**2)