New features request: Upsampling layer (after pooled features); sparse tensor concatenation.

[ZXP-S-works]

Hi traveller,

Thanks for the great work! I think the following features could nice to have:

1. Upsampling layer. Although there is spconv.SparseInverseConv3d(), it seems limited to upsampling a stride-convoluted (spconv.SpaseConv3d(stride>1)) tensor.

Empirically I find for standard CNNs, maxpooling>=2 performs better than stride>=2 and

Upsampling a tensor then perform convolution is better than transpose conv.

Thus based on these two experience I am suggesting to implement either i) Interpolation function (this could be non-trivial for sparse tensors) or ii) SparseInverseConv3d that supports maxpooling.

2. Sparse tensor concatenation. This function should be very similar to Sparse_add, except it's concatenation, which could maintain more feature information. This actually could be done by:

def sparse_cat(tensor1, tensor2):
    tensor1 = replace_feature(tensor1, torch.cat((tensor1.features, torch.zeros_like(tensor1.features)), -1))
    tensor2 = replace_feature(tensor2, torch.cat((torch.zeros_like(tensor2.features), tensor2.features), -1))
    return Fsp.sparse_add(tensor1, tensor2)

Lastly it would be nice to show some example code for a Unet. All of these requests are the missing part for a standard U-net. Though there are some workarounds, they sacrifice performance.