[Minor] Remove redundant computation.

Author: kentang-mit

torchsparse/nn/functional/sparseconv.py

@@ -150,8 +150,7 @@ def conv3d(inputs, kernel, bias=None, stride=1, dilation=1, transpose=False):
  idx_query = sphashquery(hash_query, hash_target)
  idx_query = list(convert_neighbor_map_gpu(idx_query))
  idx_query[1] = idx_query[1].to('cpu')
- sizes = ((idx_query[0][:, 0].max() + 1).item(),
- (idx_query[0][:, 1].max() + 1).item())
+ sizes = (features.shape[0], new_coords.shape[0])
  output_features = sparseconv_op(features, kernel, idx_query[0],
  idx_query[1], sizes, transpose)
  if bias is not None:
@@ -180,8 +179,7 @@ def conv3d(inputs, kernel, bias=None, stride=1, dilation=1, transpose=False):
  idx_query = sphashquery(hash_query, hash_target)
  idx_query = list(convert_neighbor_map_gpu(idx_query))
  idx_query[1] = idx_query[1].to('cpu')
- sizes = ((idx_query[0][:, 0].max() + 1).item(),
- (idx_query[0][:, 1].max() + 1).item())
+ sizes = (features.shape[0], features.shape[0])
  output_features = sparseconv_op(features, kernel, idx_query[0],
  idx_query[1], sizes, transpose)
  if bias is not None: