[ablaom]

This PR updates IterationControl to 0.6, making some new features available. This addresses some shortcomings summarized at JuliaAI/IterationControl.jl#64.

This PR also resolves the bug #83, but I will probably work on a non-breaking patch for this, minus a new test.

Needs:

• Merge and tag of Add detailed traces to the output of train! IterationControl.jl#72
• Locally test MLJ integration tests
• Locally test an updated version this MLJFlux example and the live training example works

Draft release notes

• (new feature) An update to IterationControl 0.6 means that a detailed trace is available for each control. Traces are accessed using report(mach).traces, where mach is the machine bound to the IteratedModel instance. This means explicit callbacks to record things like the out-of-sample loss are unnecessary, unless live tracking is required. On warm restarts, there is an option to extend the existing traces (set warm_restart_extends_traces = true). Refer to the new IteratedModel doc-string for details.

• (breaking) The form of report(mach) for a machine bound to a IteratedModel has
  changed. Whereas before, traces = report(mach).controls, now do traces = report(mach).traces. The form of traces has also changed; it is now a vector of
  IterationControl.ControlTrace objects; see IterationControl.jl version 0.6 for
  details.

• (breaking - new default) WithIterationsDo() is now the same as
  WithIterationsDo(identity)

• (breaking - new default) WithEvaluationDo() is now the same as
  WithEvaluationDo(identity)

• (breaking - new default) WithFittedParamsDo() is now the same as
  WithFittedParamsDO(identity)

• (breaking - new default) WithReportDo() is now the same as
  WithReportDo(identity)

• (breaking - new default) WithMachineDo() is now the same as
  WithMachineDo(identity)

• (breaking - new default) WithModelDo() is now the same as
  WithModelDo(identity)

• (mildly breaking) With the exception of Save, controls that had a <:Function
  type parameter (corresponding to a callback function) no longer do so.

• (bug fix) Resolve incorrect behavior when using scores instead of losses (Different behaviour for score versus loss. #83)

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB