Configuration

[volumes.<name>]

training_mask, outputs

[train.<model>]

max_iterations, save_checkpoints_every, save_snapshots_every, devices, sampling

[predict.<model>]

num_gpus, num_workers

[segment]

method, num_workers

[segment.params]

bias, global_bias, filter_fragments, remove_debris

[refine.fill-holes]

xy

[evaluate]

thresholds

2d_labels

Set true when the labels are on single sections.

gt

The ground truth for the evaluate step.

labels

The path of the labels array, absolute or relative to the working directory. Volumes without labels give no training samples.

mask_dilation

With training_mask = "labels", training also uses this many voxels of background around each object. With 2d_labels = true, the mask grows in xy only.

outputs example "/scratch/data.zarr"

The container for predictions, segmentations and error maps. The default is a container in the round, such as round_01/data.zarr.

raw

The path of the raw Zarr array, such as data.zarr/raw/s0 or an OME-Zarr level. It cannot be a tif or an image folder, but bs prepare --raw converts those to a Zarr array.

target_mask

The segment and evaluate steps work only inside this mask. A [refine.mask] op zeros the labels outside it. bs predict covers the mask's ROI, cut to the raw ROI.

training_mask

Use "labels" for sparse labels, "everywhere" for dense labels, or the path of a mask for dense labels within that mask.

voxel_size

The voxel size in nm, z y x.

devices bs prepare writes 1

The training devices, as a count such as 1 or a list of GPU indices such as [0]. Lightning selects the accelerator automatically.

max_iterations bs prepare writes 30001

Training stops after this many iterations. With save_checkpoints_every = 5000, 30001 saves a last checkpoint at 30000.

sampling example "equal"

Use "equal" for equal batch shares, "labels" for shares by training-mask voxels (labelled voxels without a mask), or "inv" for inverse shares. Custom weights are a list in sample order, such as [3, 1], or a table by volume name.

save_checkpoints_every bs prepare writes 5000

The number of iterations between checkpoints.

save_snapshots_every bs prepare writes 1000

The number of iterations between snapshots, which are training batches you can view.

num_gpus bs prepare writes 1

The number of GPUs for prediction. With more than one, bs predict runs the blocks through daisy.

num_workers bs prepare writes 1

The number of prediction workers. With more than one, bs predict runs the blocks through daisy.

block_shape example [32, 512, 512]

The block shape in voxels, z y x. The default is the chunk shape of the affinities, deepened to at least 16 sections.

blockwise bs prepare writes true

With true, bs segment works in blocks with several workers, unless block_shape = "roi". With false, it segments the ROI as one block. Only ws and mws support blockwise = true. Use false for cc.

context

The block context that bs segment --block-context sets, in voxels, z y x.

method bs prepare writes "mws"

The segmentation method: mws (mutex watershed), ws (watershed, needs the waterz extra) or cc (connected components).

num_workers bs prepare writes 10

The number of block workers. With SQLite, the workers share one file, and each write waits up to 10 min for its lock. For more than 10 workers, use PostgreSQL in [segment.db].

roi_offset

The ROI offset in nm, z y x, as a TOML array such as [0, 0, 0].

roi_shape

The ROI shape in nm, z y x, as a TOML array such as [400, 4000, 4000].

seeds_dataset

A path the segment step reads.

bias ws unset; mws [-0.4, -0.4, -0.4, -0.7, -0.7, -0.7, -0.7, -0.7, -0.7]

The value to add to each affinity channel, one per offset. More negative values give more splits.

epsilon_agglomerate ws 0.0

Merges the watershed fragments of each block by mean-affinity waterz score up to this threshold. 0.0 merges none.

filter_fragments ws 0.1; mws 0.1

Removes fragments whose mean affinity over the first three channels is below this value. The filter keeps fragments that contain a seeds_dataset seed.

fragments_in_xy ws true

true makes watershed fragments section by section, and false makes them in 3D.

global_bias mws [1.0, -0.5]

The weight and bias of each edge score, which is weight x mean affinity + bias. A positive score attracts and a negative one repels. A repulsion can block a later merge.

merge_function ws "mean"

The waterz score: mean, hist_quant_10, hist_quant_25, hist_quant_50, hist_quant_75 or hist_quant_90. Each hist_quant score can also take the suffix _initmax.

min_seed_distance ws 10; mws unset

The width in voxels of the maximum filter that puts seeds at the peaks of the distance to the boundary. In mws, leaving it unset makes no seeds. mws stops with an error when you set it together with seeds_dataset.

noise_eps ws unset; mws 0.001; cc unset

Adds zero-mean Gaussian noise with this standard deviation to the affinities, to break ties.

randomized_strides mws true

true keeps each strided affinity with probability 1/prod(stride), and false keeps a regular grid.

remove_debris ws 64; mws 64; cc 64

Removes fragments smaller than this many voxels. The filter keeps fragments that contain a seeds_dataset seed.

seed_eps ws unset

Subtracts this value x the voxel distance to the nearest watershed seed from the affinities. Unset means no change.

sigma ws unset; mws unset; cc unset

The Gaussian smoothing of the affinities, in voxels, z y x. Unset means no smoothing.

strides mws [[1, 1, 1], [1, 1, 1], [1, 1, 1], [2, 9, 9], [2, 9, 9], [2, 9, 9], [3, 27, 27], [3, 27, 27], [3, 27, 27]]

The z y x stride of each affinity channel, one per offset. A larger stride keeps fewer edges. Fewer repulsive edges can give fewer splits, and fewer attractive edges can give more splits.

threshold cc 0.5

Joins neighboring voxels whose affinity is above this value, on a 0 to 1 scale. It uses the first three channels.

thresholds ws [0.2, 0.35, 0.5]

Writes one segmentation for each merge-score threshold.

db_file

The path of an SQLite file.

db_host example "localhost"

A PostgreSQL connection setting.

db_name example "bootstrapper"

A PostgreSQL connection setting.

db_password example "..."

A PostgreSQL connection setting.

db_user example "postgres"

A PostgreSQL connection setting.

num_workers

Number of workers

max_size

Remove objects larger than this many voxels. With 0 or no value, the command keeps large objects.

min_size default 0

Remove objects smaller than this many voxels (dust)

num_workers default 20

The number of daisy workers that process the blocks.

out_array

The label array to write. Default: the input path with the suffix _size_filtered.

min_size default 0

Compute the mean and standard deviation from the objects with at least this many voxels.

num_std default 3.0

Remove objects whose size differs from the mean by more than this many standard deviations.

num_workers default 20

The number of daisy workers that process the blocks.

out_array

The label array to write. Default: the input path with the suffix _outlier_filtered.

min_z default 1

Remove objects whose z-extent is this many slices or fewer

num_workers default 20

The number of daisy workers that process the blocks.

out_array

The label array to write. Default: the input path with the suffix _z_filtered.

merge_ids

Merge groups as a TOML array of comma-separated strings, such as ["12,34", "56,78"]. Each group merges into its first id.

num_workers default 20

The number of daisy workers that process the blocks.

out_array

The label array to write. Default: the input path with the suffix _remapped.

remove_ids

Comma-separated label ids to remove (set to 0)

block_size default 2048

The write-block size in xy, in voxels, rounded to a multiple of the chunk size. In z, each block is one chunk deep.

context default 64

The halo in voxels, in xy with --xy and otherwise in xy and z. The minimum is the iteration count for dilate and erode, twice the iteration count for opening and closing, and the width of the largest hole plus its wall for fill-holes.

exclude_ids

Label ids that dilate never grows and fill-holes never absorbs, such as 1,4,8. In config files, use exclude_ids = [1, 4, 8].

iterations default 1

The number of iterations for dilate, erode, opening and closing.

max_area default 0

The largest enclosed region in which fill-holes absorbs objects, in voxels (pixels with --xy), such as 84000. A value of 0 keeps every object.

num_workers default 20

The number of daisy workers that process the blocks.

out_array

The label array to write. Default: append the operation to the input path, such as labels_fill_holes.

xy bs prepare writes true; default false

Fills holes section by section; false fills them in 3D.

num_workers default 20

The number of daisy workers that process the blocks.

out_array

The label array to write. Default: the input path with the suffix _masked.

out_dir

A path the evaluate step writes.

out_result

Result JSON path. Each volume gets its own file, and with several checks each file name also gets the check name.

roi_offset

The ROI offset in nm, z y x, as a TOML array such as [0, 0, 0].

roi_shape

The ROI shape in nm, z y x, as a TOML array such as [400, 4000, 4000].

seg_datasets

The segmentations the evaluate step reads, as a list of paths such as ["volume.zarr/labels"].

thresholds bs prepare writes [0.1, 1.0]

The lower and upper bounds for the mean squared affinity or LSD error. The error mask includes values strictly between these bounds.

skeletons_file

The path of a ground-truth skeletons file.

params

The table [evaluate.pred.params].

thresholds

Two values that override [evaluate] thresholds.

lsd_sigma

For 3d_lsds predictions, the Gaussian standard deviation in nm that training used. Use one value or a z y x array.

roi_offset

The ROI offset in nm, z y x, as a TOML array such as [0, 0, 0].

roi_shape

The ROI shape in nm, z y x, as a TOML array such as [400, 4000, 4000].

samples

An array of tables, [[train.<model>.samples]], one per labelled volume. Not taken by 3d_affs_from_2d_affs, 3d_affs_from_2d_lsd, 3d_affs_from_2d_mtlsd, 3d_affs_from_3d_lsd.

setup_dir

The directory that contains the model scripts and checkpoints.

voxel_size

The voxel size in nm, z y x, as a TOML array such as [40, 4, 4].

labels

The path of the volume's labels.

mask

The path of the volume's training mask.

raw

The path of the volume's raw array.

volume

The volume's name.

chain_str

The models before this one, each as <model>_<iteration>, joined by --from--.

checkpoint

The last checkpoint of the train step, or the shipped one.

input_datasets

A list of paths to raw arrays or the preceding predict step's output arrays.

output_datasets_prefix

A path the predict step writes.

roi_offset

The ROI offset in nm, z y x, as a TOML array such as [0, 0, 0].

roi_shape

The ROI shape in nm, z y x, as a TOML array such as [400, 4000, 4000].

setup_dir

The directory that contains the model scripts and checkpoints.

affs_dataset

A path the segment step reads.

fragments_dataset

A path the segment step writes.

lut_dir

A path the segment step writes.

mask_dataset

The target mask.

seg_dataset_prefix

A path the segment step writes.

aff_neighborhood mws [[-1, 0, 0], [0, -1, 0], [0, 0, -1], [-2, 0, 0], [0, -9, 0], [0, 0, -9], [-3, 0, 0], [0, -27, 0], [0, 0, -27]]

The affinity offsets in the model's net_config.json.

in_array

The mws segmentation that the segment step writes with its params. For ws or cc, set in_array in [refine].

mask_array

The mask array, at the labels' voxel size. Labels become 0 where the mask is 0.

mask_dataset

The target mask.

seg_datasets_prefix

A path the evaluate step reads.

labels_dataset

A path the evaluate step reads.

pred_dataset

A path the evaluate step reads.

aff_neighborhood

The affinity offsets in the model's net_config.json.