FAQ¶
Do I need to install anything?¶
No. app.sleap.ai is the full labeling app. Install the desktop build if you want native file access, local GPU training, or offline use.
Do I need Python?¶
Not for labeling. Training and inference need sleap-nn — either installed locally through the Environment panel (desktop, no conda required) or running on a remote worker you connect to.
Is my data uploaded anywhere?¶
No. The browser app runs entirely on your machine — files are read and written locally through the browser's file-system APIs, and nothing is sent to a server.
The exception is deliberate: if you connect to a
remote worker, the files that job needs go to that
worker — there is no SLEAP server or account in between. The browser connects
over WebSocket, normally a plain, unencrypted ws:// address, so keep the worker
on a network you trust, such as your lab network or a VPN. The desktop app can
also connect over iroh, which is encrypted. Either way, the app and the worker
check each other's keys from pairing before any job data is sent.
Can I open files from the legacy SLEAP GUI?¶
Yes. .slp is the same format. You can label here, analyze in Python with
sleap-io, and open the same file in the Qt GUI, without
converting.
Are the keyboard shortcuts the same as SLEAP?¶
Yes — they follow SLEAP's defaults. See the shortcuts reference.
My video won't play¶
Most likely a codec WebCodecs can't decode. The desktop app transcodes these for you; in the browser, convert first. See Troubleshooting.
How many frames should I label before training?¶
Fewer than you'd think for the first run — label one frame completely, train for 5 epochs, and see what happens end to end. Then correct predictions instead of labeling from scratch. See the Tutorial.
For a finished model, a few hundred labeled frames (roughly 100–500) is typically enough to predict accurately across thousands of frames.
How good the predictions get depends on three things:
- Label quality — consistent, accurate labels. Label Quality Check catches the common mistakes.
- Label quantity and variety — enough frames, from across your videos. Suggestions pick them for you.
- Training settings — the model type, crop size and augmentation suited to your data. See Training.
Which model type should I use?¶
Top-Down for multiple animals, Single Animal for one. Bottom-up needs a
fully connected skeleton and suits crowded scenes. The + ID variants add
learned identity. See Training.
What crop size should I use?¶
Run Analyze ▸ Instance Size Distribution, set the rotation preset to match your planned augmentation, and take the 95th or 99th percentile. See the guide.
Can I train without a GPU?¶
Technically yes, practically no for anything real. Use a remote worker instead — it works from the browser.
Will I lose work if the app crashes?¶
The app keeps a background draft of unsaved changes and offers to restore it on next launch. It's a safety net, not a substitute for Cmd+S. See Saving & Recovery.
How do I know which version I'm running?¶
Help ▸ About SLEAP Label, or the window title. See Release channels.
Can I cite a specific version?¶
Yes — every release gets a permanent URL like app.sleap.ai/v0.1.2/ that is
never modified afterwards. Cite that rather than the bare domain.
Where do I report bugs?¶
Help ▸ Report Issue, or the issue tracker. Attach the output of Help ▸ Collect Diagnostics….