def process_tile(self, tile):
''
flat_data = tile.reshape((tile.shape[0], -1))
if self.task_data.use_torch:
import torch
masks = self.task_data.masks.get(self.meta.slice, transpose=True)
# CuPy back-end disables torch in get_task_data
# FIXME use GPU torch with CuPy array?
result = torch.mm(
torch.from_numpy(flat_data),
torch.from_numpy(masks),
).numpy()
# Required due to https://github.com/cupy/cupy/issues/4072
elif self.backend == 'cupy' and self.task_data.masks.use_sparse:
masks = self.task_data.masks.get(self.meta.slice, transpose=False)
result = masks.dot(flat_data.T).T
# Required due to https://github.com/scipy/scipy/issues/13211
elif (self.backend == 'numpy' and self.task_data.masks.use_sparse
and 'scipy.sparse' in self.task_data.masks.use_sparse):
masks = self.task_data.masks.get(self.meta.slice, transpose=True)
result = rmatmul(flat_data, masks)
else:
masks = self.task_data.masks.get(self.meta.slice, transpose=True)
result = flat_data @ masks
# '+' is the correct merge for dot product
self.results.intensity[:] += result
def test_rmatmul_4():
le = np.zeros((2, 3))
ri = np.zeros((3, 2))
# Not a csc or csr matrix
with pytest.raises(ValueError):
rmatmul(le, ri)
def test_rmatmul_3():
le = np.zeros((2, 3))
ri = scipy.sparse.csr_matrix(np.zeros((5, 6)))
# Shape mismatch
with pytest.raises(ValueError):
rmatmul(le, ri)
def test_rmatmul_2():
le = np.zeros((2, 3))
ri = np.zeros((4, 5, 6))
# 3D shape right
with pytest.raises(ValueError, ):
rmatmul(le, ri)
def test_rmatmul_1():
le = np.zeros((1, 2, 3))
ri = scipy.sparse.csr_matrix(np.zeros((5, 6)))
# 3D shape left
with pytest.raises(ValueError):
rmatmul(le, ri)
def test_rmatmul_csc():
le = np.random.random((2, 3))
ri = scipy.sparse.csr_matrix(np.random.random((3, 2)))
assert np.allclose(rmatmul(le, ri), le @ ri)
