Some issues (errors) about this work

[072jiajia]

Thanks for bringing the idea of attention to wsss in point cloud!
I met two issues in this work

In https://github.com/plusmultiply/mprm/blob/master/datasets/Scannet_subcloud.py#L798
You make the center of the sub-cloud different every time, does it mean that we still need point-level labels to get the sub-cloud-level labels?
because the sub-clouds generated by the same seeds might be slightly different.

I think your attention submodules are implemented incorrectly
In short, it should be stacked_length = inputs['stacked_length_out'] in
https://github.com/plusmultiply/mprm/blob/master/models/network_blocks_mprm.py#L968
https://github.com/plusmultiply/mprm/blob/master/models/network_blocks_mprm.py#L1026

[072jiajia]

For the second problem,
The error can be checkout by modifying
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
def body(start_ind, batch_inds_0, x, xn):
batch_length = tf.gather(stacked_length, batch_inds_0)
end_ind = start_ind + batch_length
# end_ind = start_ind + stacked_length[batch_inds_0]
query = x1[start_ind:end_ind]
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^
to
vvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvvv
def body(start_ind, batch_inds_0, x, xn):
batch_length = tf.gather(stacked_length, batch_inds_0)
end_ind = start_ind + batch_length
debug_dummy_feature = x1[end_ind-1]
zero = tf.reduce_mean(debug_dummy_feature) * 0.
query = x1[start_ind:end_ind] + zero
^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^^

It will come up with an Out of Range Error.
This error didn't occur because when "slicing" a tensor, if the index is out of range, it will return tensors with size 0
So the dimension still matches the original one

But in fact, it becomes an attention module crossing some parts of other batches, which is different from what is written in the paper