fix stuff and add tests

Author: mawright

README.md

@@ -7,7 +7,7 @@
 Low-level utilities for PyTorch sparse tensors and operations.
 
 ## Introduction
-PyTorch's implementation of sparse tensors is lacking full support for many common operations. This repository contains a set of utilities for making PyTorch sparse tensors into more usable general-purpose sparse data structures. 
+PyTorch's implementation of sparse tensors is lacking full support for many common operations. This repository contains a set of utilities for making PyTorch sparse tensors into more usable general-purpose sparse data structures, particularly in the context of modern neural network architectures like Transformer-based models.
 
 For example, while the basic operation `index_select` has a sparse forward implementation, using it as part of an autograd graph alongside direct manipulation of the sparse tensor's values is not supported:
 ```python

pyproject.toml

@@ -54,5 +54,5 @@ exclude_lines = [
     "raise NotImplementedError",
     "if __name__ == .__main__.:",
     "if TYPE_CHECKING:",
-    "if torch.jit.is_scriping():",
+    "if torch.jit.is_scripting():",
 ]

pytorch_sparse_utils/batching/batch_utils.py

@@ -399,7 +399,7 @@ def concatenated_to_padded(
         >>> padded.shape
         torch.Size([3, 4, 3, 4])  # 3 batches, max length 4, 3x4 features
     """
-    validate_atleast_nd(tensor, 2)
+    validate_atleast_nd(tensor, 1)
     if not batch_offsets.ndim == 1:
         raise ValueError(f"Expected batch_offsets to be 1D, got {batch_offsets.ndim}")
 
@@ -418,9 +418,7 @@ def concatenated_to_padded(
 
     # Fast path: If all sequences are equal length can just return a view
     if torch.all(seq_lens == max_len):
-        if not tensor.is_contiguous():
-            tensor = tensor.contiguous()
-        out = tensor.view(out_shape)
+        out = tensor.reshape(out_shape)
         padding_mask = torch.zeros(
             batch_size, max_len, device=tensor.device, dtype=torch.bool
         )
@@ -511,7 +509,7 @@ def padded_to_concatenated(
         >>> offsets
         tensor([0, 0, 1])
     """
-    validate_atleast_nd(tensor, 3)
+    validate_atleast_nd(tensor, 2)
     batch_size, max_len = tensor.shape[:2]
     feature_dims = tensor.shape[2:]
 
@@ -656,7 +654,7 @@ def batch_offsets_from_sparse_tensor_indices(indices_tensor: Tensor) -> Tensor:
     """
     assert not torch.is_floating_point(indices_tensor)
 
-    if indices_tensor.shape[1] == 0:  # empty case
+    if indices_tensor.numel() == 0:  # empty case
         return torch.zeros(1, device=indices_tensor.device, dtype=indices_tensor.dtype)
 
     batch_indices = indices_tensor[0]

pytorch_sparse_utils/imports.py

@@ -8,7 +8,7 @@
     from MinkowskiEngine.MinkowskiNonlinearity import MinkowskiNonlinearityBase
 
     has_minkowskiengine = True
-except ImportError:
+except ImportError:  # pragma: no cover
     class DummyME:
         SparseTensor = None
     ME = DummyME()
@@ -19,7 +19,7 @@ class DummyME:
     from spconv.pytorch import SparseConvTensor
 
     has_spconv = True
-except ImportError:
+except ImportError:  # pragma: no cover
     spconv = None
     SparseConvTensor = None
     has_spconv = False