Help with Training Inputs

[avichapman]

Hi.

I am attempting to replicate the training of the SFPNet in order to investigate other training improvements.

Can you please give me some insight on how to do the training? I don't see any example training code in this repo.

Specifically, in the Semantic model's forward method, there are three inputs - input_data, xyz and batch.

I assume that input_data is a float tensor of shape BxNxF, where the first three in the F dimension are also x, y and z.

Is the xyz the voxel coordinates as a long tensor or the original lidar coordinates as a float tensor? Is the shape BxNx3?

Is the batch coordinate just torch.arange(batch_size)? Is the shape B?

Thank you for your help.

Avi

[Cavendish518]

Hi,

input is SparseConvTensor (check spconv lib). batch record the ID (due to spconv lib) xyz is the lidar coordinates (check the code in dataloader and the function related to xyz in backbone.py)

For training details, you can refer to issue #2.

[avichapman]

Ah, ok.

So in training mode for a single entry in a minibatch, this would be torch.arange([number of samples])?
And in evaluation mode, this would be set in collation_fn_voxelmean_tta using the inds_recons value returned from the loader?