5. Correcting predictions¶

Expect rough predictions: the model learned from one frame. Correcting them is how it improves. This loop of predict, correct and retrain is the core of SLEAP: each model makes better predictions, so each round needs fewer corrections.

First, save a new version with File ▸ Save As… (Cmd+Shift+S / Ctrl+Shift+S). It suggests the next name, labels.v002.slp, so you keep the earlier version.

Correct¶

In the Suggestions panel, frames with a Mean Score have predictions. On a few of them:

1. Accept the predictions: double-click a predicted instance, or press Cmd+Shift+A / Ctrl+Shift+A for every one on the frame.
2. Drag any nodes that are off.

Predicted nodes are yellow. An accepted node is red until you touch it, then green, so you can see which nodes you've checked. Leaving a correct node red is fine: an accepted instance counts as your label either way.

While you correct:

• Label every mouse on the frame, even one the model missed: add it with Cmd+I / Ctrl+I. A frame with a missing animal teaches the model that the animal isn't there.
• Don't guess hidden parts. If you can't see a body part, mark it non-visible (hover over the node and press V). A guessed point teaches the model to be confident where it shouldn't be.
• Check nodes predicted as non-visible. The model sometimes gets this wrong. Press V over the node to mark it visible again, then place it.

Correct at least two frames, then save. The more you correct, the better the next model.

Retrain¶

1. In the Training panel, click Train Again.
2. Keep the same settings, but raise Max Epochs to 50.
3. Click Start Training.

You don't have to wait for all 50 epochs: Stop Early keeps what's been trained so far (click it once per model).

Repeat this correct-and-retrain loop until predictions rarely need fixing.

Next up: Tracking a new video