Skip to content

darts_postprocessing

darts_postprocessing

Postprocessing steps for the DARTS dataset.

__version__ module-attribute 

__version__ = importlib.metadata.version('darts-nextgen')

binarize 

binarize(
    probs: xarray.DataArray,
    threshold: float,
    min_object_size: int,
    mask: xarray.DataArray,
    device: typing.Literal["cuda", "cpu"] | int,
) -> xarray.DataArray

Binarize the probabilities based on a threshold and a mask.

Steps for binarization
  1. Dilate the mask. This will dilate the edges of holes in the mask as well as the edges of the tile.
  2. Binarize the probabilities based on the threshold.
  3. Remove objects at which overlap with either the edge of the tile or the noData mask.
  4. Remove small objects.

Parameters:

  • probs (xarray.DataArray) –

    Probabilities to binarize.

  • threshold (float) –

    Threshold to binarize the probabilities.

  • min_object_size (int) –

    Minimum object size to keep.

  • mask (xarray.DataArray) –

    Mask to apply to the binarized probabilities. Expects 0=negative, 1=postitive.

  • device (typing.Literal['cuda', 'cpu'] | int) –

    The device to use for removing small objects.

Returns:

Source code in darts-postprocessing/src/darts_postprocessing/postprocess.py 
 88
 89
 90
 91
 92
 93
 94
 95
 96
 97
 98
 99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
@stopwatch.f("Binarizing probabilities", printer=logger.debug, print_kwargs=["threshold", "min_object_size"])
def binarize(
    probs: xr.DataArray,
    threshold: float,
    min_object_size: int,
    mask: xr.DataArray,
    device: Literal["cuda", "cpu"] | int,
) -> xr.DataArray:
    """Binarize the probabilities based on a threshold and a mask.

    Steps for binarization:
        1. Dilate the mask. This will dilate the edges of holes in the mask as well as the edges of the tile.
        2. Binarize the probabilities based on the threshold.
        3. Remove objects at which overlap with either the edge of the tile or the noData mask.
        4. Remove small objects.

    Args:
        probs (xr.DataArray): Probabilities to binarize.
        threshold (float): Threshold to binarize the probabilities.
        min_object_size (int): Minimum object size to keep.
        mask (xr.DataArray): Mask to apply to the binarized probabilities. Expects 0=negative, 1=postitive.
        device (Literal["cuda", "cpu"] | int): The device to use for removing small objects.

    Returns:
        xr.DataArray: Binarized probabilities.

    """
    use_gpu = device == "cuda" or isinstance(device, int)

    # Warn user if use_gpu is set but no GPU is available
    if use_gpu and not CUCIM_AVAILABLE:
        logger.warning(
            f"Device was set to {device}, but GPU acceleration is not available. Calculating TPI and slope on CPU."
        )
        use_gpu = False

    # Where the output from the ensemble / segmentation is nan turn it into 0, else threshold it
    # Also, where there was no valid input data, turn it into 0
    # Using uint8 for further processing
    binarized = (probs.fillna(0) > threshold).astype("uint8")

    # Remove objects at which overlap with either the edge of the tile or the noData mask
    labels = binarized.copy(data=label(binarized, connectivity=2))
    edge_label_ids = np.unique(xr.where(~mask, labels, 0))
    binarized = ~labels.isin(edge_label_ids) & binarized

    # Remove small objects with GPU
    if use_gpu:
        device_nr = device if isinstance(device, int) else 0
        logger.debug(f"Moving binarized to GPU:{device}.")
        # Check if binarized is dask, if not persist it, since remove_small_objects_gpu can't be calculated from
        # cupy-dask arrays
        if binarized.chunks is not None:
            binarized = binarized.persist()
        with cp.cuda.Device(device_nr):
            binarized = binarized.cupy.as_cupy()
            binarized.values = remove_small_objects_gpu(
                binarized.astype(bool).expand_dims("batch", 0).data, min_size=min_object_size
            )[0]
            binarized = binarized.cupy.as_numpy()
            free_cupy()
    else:
        binarized.values = remove_small_objects(
            binarized.astype(bool).expand_dims("batch", 0).values, min_size=min_object_size
        )[0]

    # Convert back to bool
    binarized = binarized.astype("bool")

    return binarized

erode_mask 

erode_mask(
    mask: xarray.DataArray,
    size: int,
    device: typing.Literal["cuda", "cpu"] | int,
    edge_size: int | None = None,
) -> xarray.DataArray

Erode the mask, also set the edges to invalid.

Parameters:

  • mask (xarray.DataArray) –

    The mask to erode.

  • size (int) –

    The size of the disk to use for erosion and the edge-cropping.

  • device (typing.Literal['cuda', 'cpu'] | int) –

    The device to use for erosion.

  • edge_size (int, default: None ) –

    Define a different edge erosion width, will use size parameter if None.

Returns:

Source code in darts-postprocessing/src/darts_postprocessing/postprocess.py 
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
@stopwatch.f("Eroding mask", printer=logger.debug, print_kwargs=["size"])
def erode_mask(
    mask: xr.DataArray, size: int, device: Literal["cuda", "cpu"] | int, edge_size: int | None = None
) -> xr.DataArray:
    """Erode the mask, also set the edges to invalid.

    Args:
        mask (xr.DataArray): The mask to erode.
        size (int): The size of the disk to use for erosion and the edge-cropping.
        device (Literal["cuda", "cpu"] | int): The device to use for erosion.
        edge_size (int, optional): Define a different edge erosion width, will use size parameter if None.

    Returns:
        xr.DataArray: The dilated and inverted mask.

    """
    # Clone mask to avoid in-place operations
    mask = mask.copy()

    # Change to dtype uint8 for faster skimage operations
    mask = mask.astype("uint8")

    use_gpu = device == "cuda" or isinstance(device, int)

    # Warn user if use_gpu is set but no GPU is available
    if use_gpu and not CUCIM_AVAILABLE:
        logger.warning(
            f"Device was set to {device}, but GPU acceleration is not available. Calculating TPI and slope on CPU."
        )
        use_gpu = False

    # Dilate the mask with GPU
    if use_gpu:
        device_nr = device if isinstance(device, int) else 0
        logger.debug(f"Moving mask to GPU:{device}.")
        # Check if mask is dask, if not persist it, since dilation can't be calculated from cupy-dask arrays
        if mask.chunks is not None:
            mask = mask.persist()
        with cp.cuda.Device(device_nr):
            mask = mask.cupy.as_cupy()
            mask.values = binary_erosion_gpu(mask.data, disk_gpu(size))
            mask = mask.cupy.as_numpy()
            free_cupy()
    else:
        mask.values = binary_erosion(mask.values, disk(size))

    if edge_size is None:
        edge_size = size

    # Mask edges
    mask[:edge_size, :] = 0
    mask[-edge_size:, :] = 0
    mask[:, :edge_size] = 0
    mask[:, -edge_size:] = 0

    return mask

prepare_export 

prepare_export(
    tile: xarray.Dataset,
    bin_threshold: float = 0.5,
    mask_erosion_size: int = 10,
    min_object_size: int = 32,
    quality_level: int
    | typing.Literal[
        "high_quality", "low_quality", "none"
    ] = 0,
    ensemble_subsets: list[str] = [],
    device: typing.Literal["cuda", "cpu"]
    | int = darts_postprocessing.postprocess.DEFAULT_DEVICE,
    edge_erosion_size: int | None = None,
) -> xarray.Dataset

Prepare the export, e.g. binarizes the data and convert the float probabilities to uint8.

Parameters:

  • tile (xarray.Dataset) –

    Input tile from inference and / or an ensemble.

  • bin_threshold (float, default: 0.5 ) –

    The threshold to binarize the probabilities. Defaults to 0.5.

  • mask_erosion_size (int, default: 10 ) –

    The size of the disk to use for mask erosion and the edge-cropping. Defaults to 10.

  • min_object_size (int, default: 32 ) –

    The minimum object size to keep in pixel. Defaults to 32.

  • quality_level (int | str, default: 0 ) –

    The quality level to use for the mask. If a string maps to int. high_quality -> 2, low_quality=1, none=0 (apply no masking). Defaults to 0.

  • ensemble_subsets (list[str], default: [] ) –

    The ensemble subsets to use for the binarization. Defaults to [].

  • device (typing.Literal['cuda', 'cpu'] | int, default: darts_postprocessing.postprocess.DEFAULT_DEVICE ) –

    The device to use for dilation. Defaults to "cuda" if cuda for cucim is available, else "cpu".

  • edge_erosion_size (int, default: None ) –

    If the edge-cropping should have a different witdth, than the (inner) mask erosion, set it here. Defaults to mask_erosion_size.

Returns:

Source code in darts-postprocessing/src/darts_postprocessing/postprocess.py 
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
@stopwatch.f(
    "Preparing export",
    printer=logger.debug,
    print_kwargs=[
        "bin_threshold",
        "mask_erosion_size",
        "edge_erosion_size",
        "min_object_size",
        "quality_level",
        "ensemble_subsets",
    ],
)
def prepare_export(
    tile: xr.Dataset,
    bin_threshold: float = 0.5,
    mask_erosion_size: int = 10,
    min_object_size: int = 32,
    quality_level: int | Literal["high_quality", "low_quality", "none"] = 0,
    ensemble_subsets: list[str] = [],
    device: Literal["cuda", "cpu"] | int = DEFAULT_DEVICE,
    edge_erosion_size: int | None = None,
) -> xr.Dataset:
    """Prepare the export, e.g. binarizes the data and convert the float probabilities to uint8.

    Args:
        tile (xr.Dataset): Input tile from inference and / or an ensemble.
        bin_threshold (float, optional): The threshold to binarize the probabilities. Defaults to 0.5.
        mask_erosion_size (int, optional): The size of the disk to use for mask erosion and the edge-cropping.
            Defaults to 10.
        min_object_size (int, optional): The minimum object size to keep in pixel. Defaults to 32.
        quality_level (int | str, optional): The quality level to use for the mask. If a string maps to int.
            high_quality -> 2, low_quality=1, none=0 (apply no masking). Defaults to 0.
        ensemble_subsets (list[str], optional): The ensemble subsets to use for the binarization.
            Defaults to [].
        device (Literal["cuda", "cpu"] | int, optional): The device to use for dilation.
            Defaults to "cuda" if cuda for cucim is available, else "cpu".
        edge_erosion_size (int, optional): If the edge-cropping should have a different witdth, than the (inner) mask
            erosion, set it here. Defaults to `mask_erosion_size`.

    Returns:
        xr.Dataset: Output tile.

    """
    quality_level = (
        quality_level
        if isinstance(quality_level, int)
        else {"high_quality": 2, "low_quality": 1, "none": 0}[quality_level]
    )
    mask = tile["quality_data_mask"] >= quality_level
    if quality_level > 0:
        mask = erode_mask(mask, mask_erosion_size, device, edge_size=edge_erosion_size)  # 0=positive, 1=negative
    tile["extent"] = mask.copy()
    tile["extent"].attrs = {"long_name": "Extent of the segmentation"}

    def _prep_layer(tile: xr.Dataset, subset: str | None = None):
        layername = "probabilities" if subset is None else f"probabilities-{subset}"
        binarized_layername = "binarized_segmentation" if subset is None else f"binarized_segmentation-{subset}"

        # Mask the segmentation
        tile[layername] = tile[layername].where(mask)

        # Binarize the segmentation
        tile[binarized_layername] = binarize(tile[layername], bin_threshold, min_object_size, mask, device)
        tile[binarized_layername].attrs = {"long_name": "Binarized Segmentation"}
        return tile

    tile = _prep_layer(tile)

    # get the names of the model probabilities if available
    # for example 'tcvis' from 'probabilities-tcvis'
    for ensemble_subset in ensemble_subsets:
        tile = _prep_layer(tile, ensemble_subset)

    return tile