qinzheng93 / easy-kpconv Goto Github PK

Which version of pykeops was used?
Were there some specific configs made?

Hi, I would like to ask about the input of the following method, is the shape of the input points (N,3) instead of (B,N,3)?
def grid_subsample_pack_mode(points, lengths, voxel_size):
"""Grid subsample in pack mode (fast version).
Args:
points (Tensor): the original points (N, 3).
lengths (LongTensor): the numbers of points in the batch (B,).
voxel_size (float): the voxel size.
Returns:
sampled_points (Tensor): the subsampled points (M, 3).
sampled_lengths (Tensor): the numbers of subsampled points in the batch (B,).
"""
batch_size = lengths.shape[0]

I've download, installed, and got the S3DIS example running. After confirming I can train the network I wanted to calibrate neighbors for my hardware (actually doing it to understand how it works, from what I can tell its hardware specific).

I've run into a type error originating on line 40 of examples/scene_segmentation/calibrate_neighbors.py where easy_kpconv.ops.calibrate_neighbors.calibrate_neighbors_pack_mode is called:

File "examples/scene_segmentation/calibrate_neighbors.py", line 40, in main
neighbor_limits = calibrate_neighbors_pack_mode(
File "/usr/local/lib/python3.8/dist-packages/torch/autograd/grad_mode.py", line 27, in decorate_context
return func(*args, **kwargs)
File "/workspace/projects/Easy-KPConv/easy_kpconv/ops/calibrate_neighbors.py", line 49, in calibrate_neighbors_pack_mode
data_dict = collate_fn([dataset[i]])
File "/usr/lib/python3.8/typing.py", line 875, in new
obj = super().new(cls, *args, **kwds)
TypeError: object.new() takes exactly one argument (the type to instantiate)

Specifically, from what I can tell after investigating for a bit, is this is an issue with the Python type hinting provided in the definition of calibrate_neighbors_pack_mode.

Any suggestions or ideas as to the issue?

Hi,
Thanks for your great work. I am an happy user of KpConv and would like to explore your repository since it offers more flexibility.
I would like to know how it performs (mIoU and Speed) compared to any of the two official implementations of KpConv .

Best,
kayode

Will it be expanded to the SemanticKITTI dataset?