features_at, coords_at function error

[whitealex95]

Hi, I have some trouble using features_at function
It seems to not work when there is a skip in the batch index

>>> import torch
>>> from MinkowskiEngine import SparseTensor
>>> f = torch.Tensor([1,1,1,1])
>>> c = torch.Tensor([[0,0],[0,1],[2,0],[2,1]])
>>> x = SparseTensor(f,c)
>>> x.features_at(2)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/whitealex95/.anaconda3/envs/arrgen-min/lib/python3.7/site-packages/MinkowskiEngine/SparseTensor.py", line 366, in features_at
    batch_index)
  File "/home/whitealex95/.anaconda3/envs/arrgen-min/lib/python3.7/site-packages/MinkowskiEngine/MinkowskiCoords.py", line 224, in get_row_indices_at
    coords_key.CPPCoordsKey, out_coords_key.CPPCoordsKey, batch_index)
RuntimeError: src/coords_manager.cpp:876, assertion (batch_index < batch_indices.size()) faild. Invalid batch index:  2 , batch_index exceeds current maximum batch size:  2  of a provided batch indices:  [0, 2]

Thanks

[chrischoy]

Thanks for reporting the issue.
Yes, it currently assumes the index of the unique batch indices as the input, but I can see how this is a confusing name for the function. I'll update the function in the next version.

For now, to get unique batch indices, use

>>> x.coords_man.get_batch_indices()
{0, 2}

[whitealex95]

Hi, thank you for the answer.

In case we have the continuous batch indices:

>>> x.coords_man.get_batch_indices()
{0,1, 2, 3}

I think the function features_at works fine with code:

>>> x.features_at(2)

giving out the features with the batch index 2

Is there a built-n function to get the same result in case of skipping batch indices?
I'm currently using a custom function:

def feats_at_idx(x: SparseTensor, idx: int):
    data_idx = x.C[:, 0] == idx
    return x.F[data_idx]

I just want to know if there's a better way for this job

Thanks

[chrischoy]

So currently using the batch indices would be the best work around for now.

def feats_at_idx(x, idx):
  bs = list(x.coords_man.get_batch_indices())
  if idx not in bs:
    return torch.FloatTensor([[]])
  unique_idx = bs.index(idx)
  return x.features_at(unique_idx)

[whitealex95]

I'm sorry, but according to your function, the code

x.features_at(1)

should work in case of

>>> x.coords_man.get_batch_indices()
{0, 2}

However, I get the following error.

>>> import torch
>>> from MinkowskiEngine import SparseTensor
>>> f = torch.Tensor([1,1,1,1])
>>> c = torch.Tensor([[0,0],[0,1],[2,0],[2,1]])
>>> x = SparseTensor(f,c)
>>> x.features_at(1)
Traceback (most recent call last):
  File "<stdin>", line 1, in <module>
  File "/home/whitealex95/.anaconda3/envs/arrgen-min/lib/python3.7/site-packages/MinkowskiEngine/SparseTensor.py", line 366, in features_at
    batch_index)
  File "/home/whitealex95/.anaconda3/envs/arrgen-min/lib/python3.7/site-packages/MinkowskiEngine/MinkowskiCoords.py", line 224, in get_row_indices_at
    coords_key.CPPCoordsKey, out_coords_key.CPPCoordsKey, batch_index)
RuntimeError: src/coords_manager.cpp:876, assertion (batch_index < batch_indices.size()) faild. Invalid batch index:  2 , batch_index exceeds current maximum batch size:  2  of a provided batch indices:  [0, 2]

Did I misunderstand your function?

[chrischoy]

Right, I just pushed the latest commit that will now work as expected with the batch index. Please let me know if there's a problem.