Skip to main content

Adjusting and improving models

Make each experiment in a new draft version so the trained version and its metrics remain reproducible.

Create the next draft​

Use one of these actions in the model drawer:

  • Open the version picker and select New version.
  • From a trained version, select Clone version (new model).
  • In Create from this version, choose New version of this model or New model.

A cloned draft copies the model configuration. It does not automatically reuse trained weights. Choose how weights are initialized under Advanced Settings > Model initialization; see Reusing existing models.

Change one cause at a time​

Start with evidence from validation metrics and predictions:

  • Improve or correct labels when errors follow missing, ambiguous, or inconsistent ground truth.
  • Add representative data when errors cluster by scene, sensor, processing condition, or background.
  • Revisit the model type or class definitions when the desired output does not match the label semantics.
  • Adjust preprocessing only when it matches a known spectral-data problem.
  • Tune thresholds when the model type and output path apply them and the operating point produces the wrong precision-recall tradeoff. Current Classification test predictions and deployed class maps use argmax even when a non-argmax validation policy is selected.
  • Tune architecture, regularization, optimizer, or augmentation after checking data and label quality.

Record the hypothesis for each version. Comparing many simultaneous changes makes an improvement difficult to explain or reproduce.

Verify the change​

Train the draft and compare it with prior versions using the same validation metrics and spatial examples. Managed administrator accounts may also have a Compare tab; it is not available to every user.

Use validation to select preprocessing, hyperparameters, thresholds, checkpoints, and model versions. After choosing and locking the version, generate a test run once on untouched test data. If test results guide another iteration, that test set has become validation data; use a new independent holdout for the final estimate.