how to align the order of indices

[repoww]

pooling = spconv.SparseAvgPool3d(kernel_size=2, stride=2)
conv1= spconv.SparseConv3d(
        3,
        3,
        2,
        stride=2,
        padding=0,
        dilation=1,
        bias=False
    )
data = spconv.SparseConvTensor(
        features=features.cuda(),
        indices=batch_indices.cuda(),
        spatial_shape=batch_indices.max(0)[0][1:] + 1,
        batch_size=1
    )
coordinates = spconv.SparseConvTensor(
        features=raw_coordinates.cuda(),
        indices=batch_indices.cuda(),
        spatial_shape=batch_indices.max(0)[0][1:] + 1,
        batch_size=1
    )

coords = pooling(coordinates)
embeddings = conv1(data)

How can I align the indices of coords and embbeddings to be the same to use coords as input of positional encoding.
The two inputs share the same indices, with conv(stride=2) and pooling(stride=2), the output indices should be the same?

Looks like the set of the conv output and pooling output indices are the same but in different orders.

Thanks

[repoww]

Probably, I should do pooling of raw_coordinates based on the indice_dict from embeddings and just replace the features of embeddings

[RobotDogCyberslacking]

Probably, I should do pooling of raw_coordinates based on the indice_dict from embeddings and just replace the features of embeddings

I don't think it will work

test_pooling = spconv.SparseAvgPool3d(kernel_size=2,stride=2,padding=0,indice_key='test')

sp_conv = spconv.SparseConv3d(3,3,kernel_size=2,stride = 2,bias= False,indice_key='test').cuda()

test_pooled_coords = test_pooling(coordinates)

test_conv_coords = sp_conv(coordinates)

assert(test_pooled_coords.indices == test_conv_coords.indices).all() # False

and replace_feature function does not take indices of new features as input, it will simply replace the features(with incurract indices)
https://github.com/traveller59/spconv/blob/125a194d895b1bc3ad6ff907bc72641548397b32/spconv/pytorch/core.py#L198

The best way I can think of indices alignment is to reorder the features, you can reference the code like sparse_add
https://github.com/traveller59/spconv/pull/395