def convert_cube_job(
source_knossos_info: KnossosDatasetInfo, args: Tuple[View, int]
) -> None:
target_view, _ = args
time_start(f"Converting of {target_view.bounding_box}")
cube_size = cast(Tuple[int, int, int], (CUBE_EDGE_LEN,) * 3)
offset = target_view.bounding_box.in_mag(target_view.mag).topleft
size = target_view.bounding_box.in_mag(target_view.mag).size
buffer = np.zeros(size.to_tuple(), dtype=target_view.get_dtype())
with open_knossos(source_knossos_info) as source_knossos:
for x in range(0, size.x, CUBE_EDGE_LEN):
for y in range(0, size.y, CUBE_EDGE_LEN):
for z in range(0, size.z, CUBE_EDGE_LEN):
cube_data = source_knossos.read(
(offset + Vec3Int(x, y, z)).to_tuple(), cube_size
)
buffer[
x : (x + CUBE_EDGE_LEN),
y : (y + CUBE_EDGE_LEN),
z : (z + CUBE_EDGE_LEN),
] = cube_data
target_view.write(buffer)
time_stop(f"Converting of {target_view.bounding_box}")
def convert_knossos(
source_path: Path,
target_path: Path,
layer_name: str,
dtype: str,
voxel_size: Tuple[float, float, float],
data_format: DataFormat,
chunk_size: Vec3Int,
chunks_per_shard: Vec3Int,
mag: int = 1,
args: Optional[Namespace] = None,
) -> None:
source_knossos_info = KnossosDatasetInfo(source_path, dtype)
target_dataset = Dataset(target_path, voxel_size, exist_ok=True)
target_layer = target_dataset.get_or_add_layer(
layer_name,
COLOR_CATEGORY,
data_format=data_format,
dtype_per_channel=dtype,
)
with open_knossos(source_knossos_info) as source_knossos:
knossos_cubes = np.array(list(source_knossos.list_cubes()))
if len(knossos_cubes) == 0:
logging.error(
"No input KNOSSOS cubes found. Make sure to pass the path which points to a KNOSSOS magnification (e.g., testdata/knossos/color/1)."
)
exit(1)
min_xyz = knossos_cubes.min(axis=0) * CUBE_EDGE_LEN
max_xyz = (knossos_cubes.max(axis=0) + 1) * CUBE_EDGE_LEN
target_layer.bounding_box = BoundingBox(
Vec3Int(min_xyz), Vec3Int(max_xyz - min_xyz)
)
target_mag = target_layer.get_or_add_mag(
mag, chunk_size=chunk_size, chunks_per_shard=chunks_per_shard
)
with get_executor_for_args(args) as executor:
target_mag.for_each_chunk(
partial(convert_cube_job, source_knossos_info),
chunk_size=chunk_size * chunks_per_shard,
executor=executor,
progress_desc=f"Converting knossos layer {layer_name}",
)
def _parse_vec3_int(value: str) -> Vec3Int:
parts = [int(part.strip()) for part in value.split(",")]
if len(parts) == 1:
return Vec3Int.full(parts[0])
elif len(parts) == 3:
return Vec3Int(*parts)
else:
raise TypeError(f"Cannot convert `{value}` to Vec3Int.")
def downsample_test_helper(WT1_path: Path, tmp_path: Path, use_compress: bool,
chunk_size: Vec3Int) -> None:
source_path = WT1_path
target_path = tmp_path / "WT1_wkw"
source_ds = Dataset.open(source_path)
target_ds = source_ds.copy_dataset(target_path,
chunk_size=chunk_size,
chunks_per_shard=16)
target_layer = target_ds.get_layer("color")
mag1 = target_layer.get_mag("1")
target_layer.delete_mag("2-2-1") # This is not needed for this test
# The bounding box has to be set here explicitly because the downsampled data is written to a different dataset.
target_layer.bounding_box = source_ds.get_layer("color").bounding_box
mag2 = target_layer._initialize_mag_from_other_mag("2", mag1, use_compress)
# The actual size of mag1 is (4600, 4600, 512).
# To keep this test case fast, we are only downsampling a small part
offset = (4096, 4096, 0)
size = (504, 504, 512)
source_buffer = mag1.read(
absolute_offset=offset,
size=size,
)[0]
assert np.any(source_buffer != 0)
downsample_cube_job(
(
mag1.get_view(absolute_offset=offset, size=size),
mag2.get_view(
absolute_offset=offset,
size=size,
),
0,
),
Vec3Int(2, 2, 2),
InterpolationModes.MAX,
Vec3Int.full(128),
)
assert np.any(source_buffer != 0)
target_buffer = mag2.read(absolute_offset=offset, size=size)[0]
assert np.any(target_buffer != 0)
assert np.all(target_buffer == downsample_cube(source_buffer, [2, 2, 2],
InterpolationModes.MAX))
def test_downsample_multi_channel(tmp_path: Path) -> None:
num_channels = 3
size = (32, 32, 10)
source_data = (128 * np.random.randn(num_channels, size[0], size[1],
size[2])).astype("uint8")
ds = Dataset(tmp_path / "multi-channel-test", (1, 1, 1))
l = ds.add_layer(
"color",
COLOR_CATEGORY,
dtype_per_channel="uint8",
num_channels=num_channels,
)
mag1 = l.add_mag("1", chunks_per_shard=32)
print("writing source_data shape", source_data.shape)
mag1.write(source_data)
assert np.any(source_data != 0)
mag2 = l._initialize_mag_from_other_mag("2", mag1, False)
downsample_cube_job(
(l.get_mag("1").get_view(), l.get_mag("2").get_view(), 0),
Vec3Int(2, 2, 2),
InterpolationModes.MAX,
BUFFER_SHAPE,
)
channels = []
for channel_index in range(num_channels):
channels.append(
downsample_cube(source_data[channel_index], [2, 2, 2],
InterpolationModes.MAX))
joined_buffer = np.stack(channels)
target_buffer = mag2.read()
assert np.any(target_buffer != 0)
assert np.all(target_buffer == joined_buffer)
def upsample_test_helper(tmp_path: Path, use_compress: bool) -> None:
ds = Dataset(tmp_path, voxel_size=(10.5, 10.5, 5))
layer = ds.add_layer("color", COLOR_CATEGORY)
mag2 = layer.add_mag([2, 2, 2])
offset = Vec3Int(WKW_CUBE_SIZE, 2 * WKW_CUBE_SIZE, 0)
mag2.write(
absolute_offset=offset,
data=(np.random.rand(*BUFFER_SHAPE) * 255).astype(np.uint8),
)
mag1 = layer._initialize_mag_from_other_mag("1-1-2", mag2, use_compress)
source_buffer = mag2.read(
absolute_offset=offset,
size=BUFFER_SHAPE,
)[0]
assert np.any(source_buffer != 0)
upsample_cube_job(
(
mag2.get_view(absolute_offset=offset, size=BUFFER_SHAPE),
mag1.get_view(
absolute_offset=offset,
size=BUFFER_SHAPE,
),
0,
),
[0.5, 0.5, 1.0],
BUFFER_SHAPE,
)
assert np.any(source_buffer != 0)
target_buffer = mag1.read(absolute_offset=offset, size=BUFFER_SHAPE)[0]
assert np.any(target_buffer != 0)
assert np.all(target_buffer == upsample_cube(source_buffer, [2, 2, 1]))
def tile_cubing(
target_path: Path,
layer_name: str,
batch_size: int,
input_path_pattern: str,
voxel_size: Tuple[int, int, int],
args: Optional[Namespace] = None,
) -> None:
decimal_lengths = get_digit_counts_for_dimensions(input_path_pattern)
(
min_dimensions,
max_dimensions,
arbitrary_file,
file_count,
) = detect_interval_for_dimensions(input_path_pattern, decimal_lengths)
if not arbitrary_file:
logging.error(
f"No source files found. Maybe the input_path_pattern was wrong. You provided: {input_path_pattern}"
)
return
# Determine tile size from first matching file
tile_width, tile_height = image_reader.read_dimensions(arbitrary_file)
num_z = max_dimensions["z"] - min_dimensions["z"] + 1
num_x = (max_dimensions["x"] - min_dimensions["x"] + 1) * tile_width
num_y = (max_dimensions["y"] - min_dimensions["y"] + 1) * tile_height
x_offset = min_dimensions["x"] * tile_width
y_offset = min_dimensions["y"] * tile_height
num_channels = image_reader.read_channel_count(arbitrary_file)
logging.info("Found source files: count={} with tile_size={}x{}".format(
file_count, tile_width, tile_height))
if args is None or not hasattr(args, "dtype") or args.dtype is None:
dtype = image_reader.read_dtype(arbitrary_file)
else:
dtype = args.dtype
target_ds = Dataset(target_path, voxel_size=voxel_size, exist_ok=True)
is_segmentation_layer = layer_name == "segmentation"
if is_segmentation_layer:
target_layer = target_ds.get_or_add_layer(
layer_name,
SEGMENTATION_CATEGORY,
dtype_per_channel=dtype,
num_channels=num_channels,
largest_segment_id=0,
)
else:
target_layer = target_ds.get_or_add_layer(
layer_name,
COLOR_CATEGORY,
dtype_per_channel=dtype,
num_channels=num_channels,
)
bbox = BoundingBox(
Vec3Int(x_offset, y_offset, min_dimensions["z"]),
Vec3Int(num_x, num_y, num_z),
)
if target_layer.bounding_box.volume() == 0:
# If the layer is empty, we want to set the bbox directly because extending it
# would mean that the bbox would always start at (0, 0, 0)
target_layer.bounding_box = bbox
else:
target_layer.bounding_box = target_layer.bounding_box.extended_by(bbox)
target_mag_view = target_layer.get_or_add_mag(
Mag(1), block_len=DEFAULT_CHUNK_SIZE.z)
with get_executor_for_args(args) as executor:
job_args = []
# Iterate over all z batches
for z_batch in get_regular_chunks(min_dimensions["z"],
max_dimensions["z"],
DEFAULT_CHUNK_SIZE.z):
# The z_batch always starts and ends at a multiple of DEFAULT_CHUNK_SIZE.z.
# However, we only want the part that is inside the bounding box
z_batch = range(
max(list(z_batch)[0], target_layer.bounding_box.topleft.z),
min(
list(z_batch)[-1] + 1,
target_layer.bounding_box.bottomright.z),
)
z_values = list(z_batch)
job_args.append((
target_mag_view.get_view(
(x_offset, y_offset, z_values[0]),
(num_x, num_y, len(z_values)),
),
z_values,
input_path_pattern,
batch_size,
(tile_width, tile_height, num_channels),
min_dimensions,
max_dimensions,
decimal_lengths,
dtype,
num_channels,
))
largest_segment_id_per_chunk = wait_and_ensure_success(
executor.map_to_futures(tile_cubing_job, job_args),
f"Tile cubing layer {layer_name}",
)
if is_segmentation_layer:
largest_segment_id = max(largest_segment_id_per_chunk)
cast(SegmentationLayer,
target_layer).largest_segment_id = largest_segment_id
def cubing(
source_path: Path,
target_path: Path,
layer_name: str,
batch_size: Optional[int],
channel_index: Optional[int],
sample_index: Optional[int],
dtype: Optional[str],
target_mag_str: str,
data_format: DataFormat,
chunk_size: Vec3Int,
chunks_per_shard: Vec3Int,
interpolation_mode_str: str,
start_z: int,
skip_first_z_slices: int,
pad: bool,
voxel_size: Tuple[float, float, float],
executor_args: Namespace,
) -> Layer:
source_files = find_source_filenames(source_path)
all_num_x, all_num_y = zip(*[
image_reader.read_dimensions(source_files[i])
for i in range(len(source_files))
])
num_x = max(all_num_x)
num_y = max(all_num_y)
# All images are assumed to have equal channels and samples
num_channels = image_reader.read_channel_count(source_files[0])
num_samples = image_reader.read_sample_count(source_files[0])
num_output_channels = num_channels * num_samples
if channel_index is not None:
# if there is no c axis, but someone meant to only use one channel/sample, set the sample index instead
if sample_index is None and num_channels == 1 and channel_index > 0:
sample_index = channel_index
channel_index = 0
assert (
0 <= channel_index < num_channels
), "Selected channel is invalid. Please check the number of channels in the source file."
num_output_channels = num_samples
if sample_index is not None:
# if no channel axis exists, it is valid to only set the sample index. Set channel index to 0 to avoid confusion
if channel_index is None and num_channels == 1:
channel_index = 0
assert (channel_index is not None
), "Sample index is only valid if a channel index is also set."
assert (
0 <= sample_index < num_samples
), "Selected sample is invalid. Please check the number of samples in the source file."
num_output_channels = 1
num_z_slices_per_file = image_reader.read_z_slices_per_file(
source_files[0])
assert (num_z_slices_per_file == 1 or len(source_files)
== 1), "Multi page TIFF support only for single files"
if num_z_slices_per_file > 1:
num_z = num_z_slices_per_file
else:
num_z = len(source_files)
if dtype is None:
dtype = image_reader.read_dtype(source_files[0])
if batch_size is None:
batch_size = DEFAULT_CHUNK_SIZE.z
target_mag = Mag(target_mag_str)
target_ds = Dataset(target_path, voxel_size=voxel_size, exist_ok=True)
is_segmentation_layer = layer_name == "segmentation"
if is_segmentation_layer:
target_layer = target_ds.get_or_add_layer(
layer_name,
SEGMENTATION_CATEGORY,
dtype_per_channel=dtype,
num_channels=num_output_channels,
largest_segment_id=0,
)
else:
target_layer = target_ds.get_or_add_layer(
layer_name,
COLOR_CATEGORY,
dtype_per_channel=dtype,
num_channels=num_output_channels,
data_format=data_format,
)
target_layer.bounding_box = target_layer.bounding_box.extended_by(
BoundingBox(
Vec3Int(0, 0, start_z + skip_first_z_slices) * target_mag,
Vec3Int(num_x, num_y, num_z - skip_first_z_slices) * target_mag,
))
target_mag_view = target_layer.get_or_add_mag(
target_mag,
chunks_per_shard=chunks_per_shard,
chunk_size=chunk_size,
)
interpolation_mode = parse_interpolation_mode(interpolation_mode_str,
target_layer.category)
if target_mag != Mag(1):
logging.info(
f"Downsampling the cubed image to {target_mag} in memory with interpolation mode {interpolation_mode}."
)
logging.info("Found source files: count={} size={}x{}".format(
num_z, num_x, num_y))
with get_executor_for_args(executor_args) as executor:
job_args = []
# We iterate over all z sections
for z in range(skip_first_z_slices, num_z, DEFAULT_CHUNK_SIZE.z):
# The z is used to access the source files. However, when writing the data, the `start_z` has to be considered.
max_z = min(num_z, z + DEFAULT_CHUNK_SIZE.z)
# Prepare source files array
if len(source_files) > 1:
source_files_array = source_files[z:max_z]
else:
source_files_array = source_files * (max_z - z)
# Prepare job
job_args.append((
target_mag_view.get_view(
(0, 0, z + start_z),
(num_x, num_y, max_z - z),
),
target_mag,
interpolation_mode,
source_files_array,
batch_size,
pad,
channel_index,
sample_index,
dtype,
target_layer.num_channels,
))
largest_segment_id_per_chunk = wait_and_ensure_success(
executor.map_to_futures(cubing_job, job_args),
progress_desc=f"Cubing from {skip_first_z_slices} to {num_z}",
)
if is_segmentation_layer:
largest_segment_id = max(largest_segment_id_per_chunk)
cast(SegmentationLayer,
target_layer).largest_segment_id = largest_segment_id
# Return layer
return target_layer