Skip to content

4. Training a model

Training teaches a model to find your skeleton's nodes from the frames you labeled. With one labeled frame this first model won't be accurate, and that's fine. Early on, a fast model beats an accurate one: all it has to do is make correcting its predictions quicker than labeling from scratch. Accuracy improves quickly as you add labels.

Open the Training panel and check these settings:

Setting Value Why
Model Type Top-Down Finds each animal by its anchor, then the rest of its nodes1
Anchor Part torso Central and visible in most frames2
Max Epochs 5 (both models) Just enough to see training work3
Post-Training Inference Target Suggested frames Gives you predictions to correct in the next step

The Training panel set up for the tutorial

Then click Start Training (Start Remote Training on a remote worker).

Top-Down trains two models one after the other, and each gets a progress row. Click a row to see its loss curves: loss measures how far the model's predictions are from your labels, so it should go down as training goes on. When training finishes, the app runs the new model on your suggested frames.

On a GPU this run takes a few minutes. On a CPU it's much slower.

Note

Set Post-Training Inference Target before you start. It can't be changed while training runs.

For your own videos

This camera looks down on the mice, so any orientation is natural. For overhead videos like this, set Rotation Augmentation to ±180°: the model then sees each labeled frame at random angles.4 For side-view cameras, keep it small (±15°), since an upside-down animal would look unnatural.

See Training for every setting.

Next up: Correcting predictions

  1. The first model (centroid) finds each animal. The second (centered instance) looks at a crop around each one and finds the rest of its nodes. That's why Top-Down shows two config slots and trains twice. ↩

  2. The crop is centered on the anchor part on every frame, so the anchor should be central and almost always visible. Anchoring on a body part gives more consistent crops than the animal's center, which can land on different body parts from frame to frame. ↩

  3. An epoch is one pass over your labeled frames. Real projects train for many more, and training stops early on its own once it stops improving. ↩

  4. Augmentation makes random variations of your labeled frames, such as rotations and scaling, so the model doesn't memorize the exact images. ↩