← All workshopsRead Deciding When a Training Run Has Done Enough

Optimisation and Training

Select the saved checkpoint

Two checkpoints share the best loss

Botie saved checkpoints while validation loss changed. More training does not automatically make the last saved state the best one to retain.

BOTIE

The first two validation losses tie. Which checkpoint satisfies this workshop's declared selection rule?
The inputs we are comparing

Validation losses in checkpoint order

Overfitting later

  1. 00.8
  2. 10.5
  3. 20.55
  4. 30.6

Tied best

  1. 00.4
  2. 10.4
  3. 20.5
Result for overfitting later
1
Choose an explanation to test
Validation recovers to an earlier best

This check uses the pieces from “Select the checkpoint by its validation record”.

BOTIE

Validation loss worsens and later returns to its earlier minimum. Return the zero-based checkpoint index chosen by the workshop's earliest-minimum rule.
Validation losses in checkpoint order
  1. 00.7
  2. 10.5
  3. 20.3
  4. 30.45
  5. 40.3

Use a number, text, or an array matching the requested result. For a matrix, put each row in brackets. Saving your prediction runs your current construction on these inputs and checks the earlier examples too.

Build a construction in the workspace above first. This case will test those pieces.

Why should the held-out test remain separate after choosing the checkpoint this way?

Selecting a model on validation data uses that data for a decision; a final test answers a different evaluation question.

This does not evaluate the final test or prove that later unobserved epochs would be worse.

Work through the Python challenges →
Continue the courseConvolution →
← Checking a Model Before a Long Training Run