CellMap Flow Dashboard

No Viewer Connected

No volume open. Enter the URL of an OME-Zarr volume below to open it in the viewer.

or drop an OME-Zarr, zarr or N5 folder on this panel

Open Pipeline Builder →

Add a model

Local Catalog

Hugging Face

Click to load Hugging Face models...

BioImage Model Zoo

Click to load BioImage Model Zoo models...

Cellpose

Each z slice is segmented in 2D. The DINO models (cpdino, cpdino-vitb) are not listed: they need the dinov3 package, which the cellpose4 environment does not have. Cellpose-SAM's weights are for non-commercial use.

Server Config

No cluster here (no bsub): models and finetuning run on this machine, and these settings are not used.

GPU Queue

Charge Group

Time Limit

LSF run limit, HH:MM or minutes. Jobs are killed at this point. Leave blank for the queue default, which is 2 hours.

Try other GPU queues

If the queue above is full or closed, submit to whichever GPU queue has capacity instead of waiting. Uncheck to always use the queue above, however long it takes to start.

Cores per Worker

Number of Workers

Models run in this browser, on its GPU through WebGPU, or else on its CPU.

Resample if no scale matches the model

Drag to reorder. Steps apply in the order shown.

⋮⋮ MinMaxNormalizer

min_value

max_value

invert

⋮⋮ LambdaNormalizer

expression

⋮⋮ ChannelSelector

⋮⋮ Dilate

⋮⋮ EuclideanDistance

⋮⋮ ZScoreNormalizer

Finetune: paint corrections, then train on the GPU cluster. Needs a model with full metadata.

Select Model

Crops are sized to the model's output.

Output Path for Zarr Files Where crops are saved; MinIO must be able to read it.

Start or resume an annotation volume

Paint 1 = background, 2 and up = foreground, an id per object.

Draws a box in the viewer around each region annotated so far. Only when you click, and not while painting: it drops the selected draw tool.

Patch on screen

Label it

Prediction from

Seed, split and box settings

Method

Box (voxels) Threshold Min object (voxels) Connectivity Per z slice

AI-assisted annotation

Good regions

Copies the annotations from MinIO to disk.

Log

Training Configuration

Checkpoint Path (Optional) Advanced Empty: taken from the model config.

LoRA Rank Higher = more capacity. 0 = full finetune.

Number of Epochs Usually 10–20.

Batch Size Higher = faster, more GPU memory.

Learning Rate 1e-4 suits LoRA.

Patches per Epoch YAML override Blank = manifest's, 0 = auto.

Good-Region Rehearsal Advanced Share of patches from good regions.

Loss Function Margin suits scribbles.

Margin margin loss only Smaller = stricter.

Distillation Weight Keeps the model near its original output. Keep it small.

Distillation Scope Advanced Where distillation applies.

Label Smoothing 0 for sharp 0/1 output.

Balance fg/bg classes

Weights fg and bg equally.

Augmentation

GPU Queue

Auto-load model after training (serves it and adds the layer)

Training Logs

Auto-scroll

Loss vs Epoch No loss data yet

No training job running...

Review: Query the instance-review SQLite and record bless/edit verdicts. Server navigates neuroglancer to each instance.

Session

review.sqlite path

reviewer

segmentation layer (optional)

Queue

order

min_vox

Hover the cursor over a segment in NG and press t to pick it. The current instance updates automatically — no Select click.

Auto-advance

Go to ID

Progress

(open a review db to see progress)

blessed / edited / erased / unreviewed:

Local Catalog

Hugging Face

BioImage Model Zoo

Cellpose

Server Config

Model Configuration

MinIO Status

Training Configuration

Training Status

Inference Server

Training Logs

Local Catalog

Hugging Face

cellmap

BioImage Model Zoo

bioimage.io

Cellpose

Cellpose 4

Server Config

Server Configuration

Model Configuration

MinIO Status

Training Configuration

Training Status

Inference Server

Training Logs