Drag to reorder: steps run in the order shown. An edit takes effect when
you press Apply, which sends this list and the Postprocess tab's together.
Postprocessing
Drag to reorder: steps run in the order shown. An edit takes effect when
you press Apply, which sends this list and the Input tab's together.
Finetune: paint corrections, then train the model's last layer in this tab, or a 2D model's whole network.
Needs a model with full metadata.
Which model to start from
Ranks models by how well each one already matches your paint, to pick the best one to finetune.
Tick
/
(shown)
Match: the share of your painted voxels the model gets right, objects and background counting equally, which ranks them; then the share of painted objects it calls object, and of painted background it calls background. Threshold: the probability above which the model's prediction fits your paint best (0.5 is the model's own boundary), as Seed from Prediction takes it. Input: the raw's grey levels the model reads as 0 to 1, its normalizer's own, or the raw under your paint stretched when that matches clearly better. Use this model sets both, as Fit to Paint does for the model seeding the paint.
Little paint: scores a few points apart are effectively tied; paint a little in two or three places for a clearer ranking.
Training Configuration
The last layer trains in seconds; the whole network in minutes, on this machine's GPU.
Empty: taken from the model config.
Higher = more capacity. 0 = full finetune.
Usually 10–20. A 3D model's whole network trains one block a step: aim for 60 to 150 steps in all, epochs × blocks (the run's first line gives the blocks).
Tiles per step: each step's time and memory grow with it.
The head's batch: painted voxels per step.
1e-2 suits the last layer; lower it if the loss jumps about.
1e-4 suits the whole network; lower it if the loss jumps about or the model forgets what it knew.
Blank = manifest's, 0 = auto.
Share of patches from good regions.
BCE as before; MSE and margin as on the server.
Smaller = stricter.
Distance model: scribbles train on distance bounds, whatever the loss above.
Keeps the model near its own output where nothing is painted.
Of the painted chunks, up to 32,768 voxels. Good regions, once marked, decide instead.
0 for sharp 0/1 output.
Weights fg and bg equally.
Once a job waits for a restart, Shift+R over the viewer restarts it with these settings,
as Restart Training does, without asking first.
Run on a region: run the models served on the Models tab over a box of the volume, or all of it, on this machine's GPU, and save their predictions as OME-Zarr in a folder on this machine, to open here again or copy to the cluster.
Models
Each model is written to <folder>/<model>.zarr, through the Input and Postprocess steps it is served with.
Region
Draw a box: hold Ctrl and click two opposite corners in the viewer. Scroll between the two clicks to give the box depth; a flat box is one chunk deep. The latest box is used.
zyx
Output folder
No folder chosen
Review: Query the instance-review SQLite and record
bless/edit verdicts. Server navigates neuroglancer to each instance.
Session
Queue
Hover the cursor over a segment in NG and press t to pick
it. The current instance updates automatically — no Select click.
Current instance:
state
entry
rank
vox
em_mean
sphericity
fm_score
centroid (nm)
bbox (vox)
No more unreviewed instances matching the current filter.
Record edit verdict
Free-form JSON stored in the ledger; any valid JSON object is accepted.
