package ffn

Get desktop application:
View/edit binary Protocol Buffers messages

Settings related to online alignment.

Used in: InferenceRequest

• optional AlignmentOptions.AlignType type = 1

  required

• optional bool save_raw = 6

  Other options. -------------- Set true to write out the subvolumes of aligned raw imagery that are used for inference.alignmentoptions

Used in: AlignmentOptions

• UNKNOWN_ALIGNMENT = 0

• NO_ALIGNMENT = 1

Describes a 3-dimensional axis-aligned hyperrectangular region of voxels. The region contains whole numbers of rows, columns and layers of voxels (i.e., not fractional ones).

Used in: BoundingBoxes, InferenceRequest

• optional proto.Vector3j start = 1

  The minimum X, Y and Z coordinates of any voxel in the region.

  required

• optional proto.Vector3j size = 2

  The size of the region in the X, Y and Z dimensions, in units of voxels. The maximum X, Y and Z coordinates of any voxel in the region are start + size - 1. INVARIANT: size.x, y and z are all >0.

  required

• optional string description = 3

  An optional string associated with this bounding box.

• optional uint64 object_label = 4

  An optional object label associated with this bounding box. This is uint64 to match usage of uint64 for segmentations represented by NDArrays.

A sequence of bounding boxes.

• repeated BoundingBox box = 1

• optional SegmentationSource segmentation1 = 1

• optional SegmentationSource segmentation2 = 2

• optional string segmentation_output_dir = 3

• optional ConsensusRequest.ConsensusType type = 4

• optional int32 split_min_size = 7

  Minimal size (in voxels) of a component created in split consensus.

Used in: ConsensusRequest

• CONSENSUS_SPLIT = 2

  Computes the intersection of two segmentations.

Used in: MaskConfig

• optional string expression = 1

  Valid numpy expression, where 'x', 'y', and 'z' are dense index arrays defining the coordinates of every voxel in the current subvolume, in the global coordinate system of the volume. SECURITY WARNING: This gets passed to Python's eval, which will allow execution of arbitrary code. This option is for internal use only.

Used in: TaskCounters

• optional string name = 1

• optional int64 value = 2

Used in: InferenceRequest, VolumeMaskOptions

• oneof volume_path

  • string volinfo = 1

    Path to the TextFormat VolumeInfo proto for the volume.

  • string hdf5 = 3

    file_path:hdf5_internal_dataset_path for the volume.

• optional string decorator_specs = 2

  For volinfo Volumes only. JSON list of specs to decorate the volume via volume_decorator.from_specs. E.g.: '[{"decorator": "ZSub", "args": [{"2198": 2197}]}]' If left unspecified, the undecorated volume is used.

• optional uint64 id = 1

  ID of the original segment from which resegmentation started.

• optional proto.Vector3j start = 2

  Center point of the subvolume within which the segmentation was generated.

• optional int32 num_voxels = 3

  Number of voxels created in the resegmentation run.

• map<uint64, EndpointResegmentationResult.OverlapInfo> overlaps = 4

  Key is a segment ID in the original segmentation.

• optional EndpointResegmentationResult.OverlapInfo source = 5

  Information about the overlap with the seeding segment. Stored separately from the map for easier querying.

• optional proto.Vector3j segmentation_radius = 6

  Size of the area around the decision point used for segmentation.

• optional string tag = 7

  Free-form description of the resegmentation experiment.

Stores statistics about the overlap between an original segment and one created in the current resegmentation run.

Used in: EndpointResegmentationResult

• optional int32 num_overlapping = 1

  Number of overlapping voxels.

• optional int32 num_original = 2

  Number of voxels in the original segment.

Used in: MaskConfig

• repeated MaskChannelConfig channels = 1

Configures the details of the FFN inference process. Passed to the Canvas. FoV stands for 'Field of View' of the network.

Used in: InferenceRequest

• optional float init_activation = 1

  Settings affecting how the FFN is run. -------------------------------------- Threshold and padding values referring to soft object mask voxels are always specified as probabilities, even if the model operates in logit space. Value with which the starting seed is populated.

• optional float pad_value = 2

  Filler value to which unexplored areas are set.

• optional float move_threshold = 3

  Threshold that has to be matched or exceeded at the center of a candidate FoV position in order for the FoV to be moved to that position.

• optional float consistency_threshold = 4

  Average (over the FoV) magnitude of per voxel probability change between subsequent FFN iterations, below which to stop iterating the FFN further without moving the FoV. Note that setting this to a positive value will cause at least two FFN inference calls to be made for every FoV position.

• optional float disco_seed_threshold = 5

  Negative values disable the disconnected seed bias. If >= 0, specifies the fraction of voxels within the prediction FoV that need to be active in order for the disconnected voxel freezing to be applied. Settings this to a small nonzero value can reduce split errors for thin processes.

• optional proto.Vector3j min_boundary_dist = 6

  Settings affecting how FFN predictions are converted into a segmentation. ------------------------------------------------------------------------- Minimal separation of the seed voxel and any previously segmented voxel.

• optional float segment_threshold = 7

  Probability threshold determining how the soft object mask is converted into a binary segmentation. The lower the value, the more spatially extended the segments will be (which can also lead to more merge errors).

• optional int32 min_segment_size = 8

  Minimum number of voxels that a segment needs to have in order to be retained in the segmentation.

Passed to the Runner.

Used in: ResegmentationRequest

• optional DecoratedVolume image = 24

  Input image and normalization parameters.

• optional float image_mean = 2

• optional float image_stddev = 3

• optional string reference_histogram = 4

• repeated MaskConfig histogram_masks = 26

  Exclusion mask for the purpose of histogram matching. Normally used to mask out areas of invalid data (empty space, resin, etc).

• repeated MaskConfig masks = 5

  Exclusion mask. The final mask is formed by the logical sum of the individual masks.

• repeated MaskConfig seed_masks = 30

  Same as above, but for seed placement.

• optional DecoratedVolume shift_mask = 6

  2-channel volume specifying a local (x,y) offset between slice 'z' and 'z+1'.

• optional BoundingBox shift_mask_fov = 22

  Field of view around the current position to consider. A position will be invalid for movement if the shift_mask volume contains shift channel values at or above shift_mask_threshold within this field of view. If this is left blank, the default behavior of inference.Runner is to use the input field of view of the network. Specified relative to current position, so legal for start to be negative. Note: since shift_mask volumes are w.r.t. the preceding slice, it makes sense to bias the shift_mask_fov 1 slice positive. For anisotropic data, a value of start=(r, r, 0), size=(2r + 1, 2r + 1, 1) can give good results.

• optional int32 shift_mask_scale = 7

  Resolution scaling factor specifying how much smaller the pixel size of the shift mask is compared to the image volume.

• optional int32 shift_mask_threshold = 8

  If the magnitude of either component of the shift vector matches or exceeds this value within the FoV of the FFN, the object mask will not be extended within the FoV.

• optional string movement_policy_name = 9

  Movement policy. ---------------- Name of the movement policy as <module_name>.<model_class>.

• optional string movement_policy_args = 10

  JSON string with arguments to be passed to the model constructor.

• optional string model_name = 11

  FFN model. --------- Name of the FFN model as <module_name>.<model_class>.

• optional string model_args = 12

  JSON string with arguments to be passed to the model constructor.

• optional string model_checkpoint_path = 13

• optional int32 batch_size = 27

  Batch size to use during inference.

• optional int32 concurrent_requests = 28

  Number of additional subvolumes that the runner should handle. This should be >= batch_size.

• optional InferenceOptions inference_options = 14

• optional string segmentation_output_dir = 15

  Directory where the segmentation results should be saved. This directory will also contain the checkpoints (if any).

• optional int32 checkpoint_interval = 16

  How often to save checkpoints, in seconds.

• optional string seed_policy = 17

  Name of the seed policy. The code expects a function called policy_<seed_policy> in the `seed` module.

• optional string seed_policy_args = 19

  JSON string with arguments to be passed to the seed policy.

• optional SelfPredictionOptions self_prediction = 18

• optional AlignmentOptions alignment_options = 20

  Options for the online aligner.

• optional DecoratedVolume init_segmentation = 25

  Initial segmentation with which to prepopulate the canvas.

Specifies how to convert a channel from a VolumeStore volume into a Boolean exclusion mask. The mask will be formed by: min_value <= [channel_value] <= max_value or, if `values` is set, to all voxels containing one of the values specified therein. If `invert` is true, the complement of the mask computed according to the above rules is used.

Used in: ImageMaskOptions, VolumeMaskOptions

• optional int32 channel = 1

• optional float min_value = 2

• optional float max_value = 3

• repeated uint64 values = 5

• optional bool invert = 4

Specifies how to convert a VolumeStore volume into a Boolean mask. The per-channel masks built as described above are combined as a logical sum to form the final mask, which can then be optionally inverted.

Used in: InferenceRequest, MaskConfigs

• oneof source

  • VolumeMaskOptions volume = 1

  • ImageMaskOptions image = 2

  • CoordinateExpressionOptions coordinate_expression = 3

• optional bool invert = 4

Used in: SegmentationSource

• repeated MaskConfig masks = 1

• optional proto.Vector3j point = 1

  Decision point.

• optional uint64 id_a = 2

  IDs of the two original segments.

• optional uint64 id_b = 3

• optional proto.Vector3j segmentation_radius = 4

  Size of the area around the decision point used for segmentation.

• optional string tag = 5

  Free-form description of the resegmentation experiment.

• optional PairResegmentationResult.EvalResult eval = 6

Used in: PairResegmentationResult

• optional proto.Vector3j radius = 1

  Size of the area around the decision point used for evaluation of metrics stored in this proto.

• optional float iou = 2

  The Jaccard index between the two segments.

• optional SegmentResult from_a = 3

• optional SegmentResult from_b = 4

• optional float max_edt_a = 5

  Max distance from background in the original segments.

• optional float max_edt_b = 6

• optional int32 num_voxels_a = 7

  Number of voxels of the original segment visible within the resegmentation subvolume.

• optional int32 num_voxels_b = 8

Metrics from a single FFN inference run.

Used in: EvalResult

• optional proto.Vector3j origin = 1

  Location of the seed for this run in the global coordinate system.

• optional int32 num_voxels = 2

  Number of voxels created in this run.

• optional int32 deleted_voxels = 3

  Fraction of voxels flagged as deleted during inference.

• optional float segment_a_consistency = 4

  Fraction of voxels of the original segment reconstructed in the current segmentation run.

• optional float segment_b_consistency = 5

• optional float max_edt = 6

  Max distance from border in the reconstructed segment.

Used in: ResegmentationRequest

• optional uint64 id_a = 1

  Objects for which segmentation is to be performed. `id_b` can be omitted, in which case this is treated as describing an endpoint extension request.

• optional uint64 id_b = 2

• optional proto.Vector3j point = 3

  Decision point associated with the object pair. Normally this is (one of) the points of maximum proximity of the two objects.

• optional InferenceRequest inference = 1

  General FFN inference settings. Note that resegmentation requires that the initial segmentation volume is specified via `init_segmentation`.

• repeated ResegmentationPoint points = 2

  Defines the work to be done. A single request normally covers multiple points so that configuration options do not have to be specified separately per point.

• optional proto.Vector3j radius = 5

  Number of voxels around the decision point. Defines the segmentation subvolume size.

• optional string output_directory = 6

  Output options.

• optional int32 subdir_digits = 7

  If >0, shard results by the first N digits of the hash of the ID pair.

• optional int32 max_retry_iters = 8

  Maximum number of segmentation attempts for a given starting segment. Segmentation is attempted again in case the resulting objects fail to recover the specified fraction of the original segment.

• optional proto.Vector3j exclusion_radius = 9

  Radius of a subvolume around the seed point to exclude from future seeding, in case of a failed segmentation run.

• optional proto.Vector3j init_exclusion_radius = 11

  Defines a cuboid area around the decision point which to exclude from seeding.

• optional float segment_recovery_fraction = 10

  Fraction of voxels of the original segment that have to be recreated in order for the segmentation attempt to be considered successful.

• optional bool terminate_early = 12

  If true, when the first segment is not successfully recovered, the segmentation of the second segment is not attempted.

• optional proto.Vector3j analysis_radius = 13

  If not specified, same as `radius`. Defines a radius within which to compute `segment_recovery_fraction`. A typical recommended setting is: radius - ffn_fov_radius to make the analysis less sensitive to edge effects.

Describes a segmentation stored as subvolumes & basic transformations to be applied to it when it is read.

Used in: ConsensusRequest

• optional string directory = 1

  Directory where subvolume files are stored.

• optional float threshold = 2

  If specified, restrict segments to areas which at least this predicted object probability.

• optional bool split_cc = 3

  Whether to recompute connected components.

• optional int32 min_size = 4

  Minimum number of voxels a segment needs to have in order to be retained.

• optional MaskConfigs mask = 5

  Exclusion mask to apply to the read data.

Settings related to self-prediction halting. TODO(phli): Add more detailed explanations of the parameters and the halting mechanism.

Used in: InferenceRequest

• optional float threshold = 1

• optional float orig_threshold = 2

• repeated CounterValue counters = 1

• optional proto.Vector3j point = 2

  Representative point for the task for which counters are stored (e.g. subvolume corner).

• optional string filename = 3

  Path to the file from which the counters were retrieved.

Used in: MaskConfig

• optional DecoratedVolume mask = 1

• repeated MaskChannelConfig channels = 2