Enter the URL of an OME-Zarr volume below to open it in the viewer. iNo volume is open. The URL must be that of a public OME-Zarr volume, one served with CORS, such as an OpenOrganelle dataset; or pick a folder on this machine. The models run in this browser.
or drop an OME-Zarr, zarr or N5 folder on this panel
Drag to reorder; edits take effect on Apply.
iSteps run in the order shown, the input normalization
before the model and the postprocessing after it. An edit in either list
takes effect when you press Apply.
The raw channel every model reads.
Finetune: paint corrections, then train the model in this tab. iIts last layer, or a 2D model's whole network.
Needs a model with full metadata.
GPU Queues for Rank, Use this model and training
Which model to start from
Ranks models by how well each one already matches your paint.
iUse it to pick the best one to finetune. Listed as the Models tab lists them: the ones served now, cellmap's, the Hugging Face repos you added, your finetunes; the ticked ones are scored.
Tick
/
Cellpose models are not ranked. iThey predict flows, not objects and background.
Match: how much of your paint the model already gets right. iObjects and background count equally (50% is guessing, 100% is all of it). Use this model makes it the model to finetune, with the row's threshold, input normalization and channel.
Little paint: paint in two or three places for a clearer ranking. iWith little paint, scores a few points apart are effectively tied.
Training Configuration
The channel picked learns your paint. iThe others are held to what the model predicted.
The last layer trains in seconds; the whole network in minutes. iThe whole network trains on this machine's GPU.
Empty: taken from the model config.
Higher = more capacity. 0 = full finetune.
Usually 10–20. iA 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. iLower it if the loss jumps about.
1e-4 suits the whole network. iLower 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.
Where distillation applies. iOf 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.
Shift+R over the viewer restarts a waiting job with these settings.
iAs Restart Training does, but without asking first.
Make one at huggingface.co/settings/tokens (a Write token). iIt is kept in this page only and sent only to Hugging Face.
iIts links will not open it: the app reads public repositories only.
Hugging Face
On a server, for Add a model
In the browser or on a server, for Add a model
Try it in the browser
Open on this dataset
The smaller copy, for Add a model
Try the smaller copy in the browser
Training Logs
Loss vs Epoch
No loss data yet
Save predictions: run the models over a region and save them to a folder. iRuns the models served on the Models tab over a box of the volume, or all of it, on this machine's GPU, and saves 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. iIt is written through the Processing tab's steps it is served with.
Region
Draw a box: hold Ctrl and click two opposite corners in the viewer. iScroll 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: record a verdict on each instance of a review db.
iQuery the instance-review SQLite and record
bless/edit verdicts. Server navigates neuroglancer to each instance.
Session
Queue
Press t over a segment in NG to pick it.
iHover the cursor over the segment first. 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.
