def create_sparsetensor2d_file(file_name, rand_num_events, n_projections):
voxel_set_list = data_generator.build_sparse_tensor(
rand_num_events, n_projections=n_projections)
data_generator.write_sparse_tensors(file_name,
voxel_set_list,
dimension=2,
n_projections=n_projections)
def test_sparse_tensor_downsample(dimension, pooling):
# Create image Meta:
meta = image_meta_factory(dimension)
meta.set_projection_id(0)
for dim in range(dimension):
L = 10.
N = 128
meta.set_dimension(dim, L, N)
if dimension == 2:
st = larcv.SparseTensor2D()
if dimension == 3:
st = larcv.SparseTensor3D()
st.meta(meta)
# Get a set of voxels:
voxel_set_list = data_generator.build_sparse_tensor(1, n_projections=1)
indexes = voxel_set_list[0][0]['indexes']
values = voxel_set_list[0][0]['values']
n_voxels = voxel_set_list[0][0]['n_voxels']
for j in range(n_voxels):
if pooling == larcv.kPoolMax:
# Only use positive values for max pooling.
# Negative values have different behavior in sparse vs dense
# max pooling.
st.emplace(larcv.Voxel(indexes[j], numpy.abs(values[j])), False)
else:
st.emplace(larcv.Voxel(indexes[j], values[j]), False)
# Dense downsampling is tested against skimage
# Here, test against dense downsampling
compression = 2
st_dense = st.to_tensor()
st_dense_compressed = st_dense.compress(compression, pooling).as_array()
st_compressed = st.compress(compression, pooling)
st_compressed_dense = st_compressed.dense()
print(st_dense.as_array())
# Do some checks:
assert numpy.abs((st_compressed_dense.sum() - st_dense_compressed.sum()) /
st_dense_compressed.sum()) < 1e-6
max_index = numpy.prod(st_compressed_dense.shape)
for i in range(50):
index = numpy.random.randint(0, max_index)
assert numpy.abs(
st_compressed_dense.take(index) -
st_dense_compressed.take(index)) < 1e-4