v0.2.0

FluxTraining v0.2.0

Diff since v0.1.3

Added

• New training loop API that is easier to extend. Defining a Phase and step! is all you need. See the new tutorial and the new reference.
  • Relevant functions: epoch!, step!, runstep, runepoch
• Added CHANGELOG.md (this file)
• AbstractValidationPhase as supertype for validation phases
• Documentation for callback helpers on reference page

Changed

• Batch* renamed to Step*:
  • events: BatchBegin now StepBegin, BatchEnd now StepEnd
  • CancelBatchException now CancelStepException.
  • field Learner.batch now Learner.step
• Learner.step/batch is no longer a special struct but now a PropDict, allowing you to set arbitrary fields.
• Learner.model can now be a NamedTuple/Tuple of models for use in custom training loops. Likewise, learner.params now resembles the structure of learner.model, allowing separate access to parameters of different models.
• Callbacks
  • Added init! method for callback initilization, replacing the Init event which required a Phase to implement.
  • Scheduler now has internal step counter and no longer relies on Recorder's history. This makes it easier to replace the scheduler without needing to offset the new schedules.
  • EarlyStopping callback now uses criteria from EarlyStopping.jl

Removed

• Removed old training API. Methods fitbatch!, fitbatchphase!, fitepoch!, fitepochphase! have all been removed.

Closed issues:

• Scheduler applies schedules per batch by default (#68)
• Recorder does not work with models with non-Array inputs. (#80)

Merged pull requests:

• CompatHelper: bump compat for "BSON" to "0.3" (#69) (@github-actions[bot])
• use EarlyStopping.jl for stopping criteria (#72) (@lorenzoh)
• CompatHelper: bump compat for "PrettyTables" to "0.12" (#73) (@github-actions[bot])
• Move documentation to Pollen.jl (#77) (@lorenzoh)
• Revert onecycle (#78) (@lorenzoh)
• Remove samplesfield from History (#81) (@lorenzoh)
• New training API and QoL improvements (v0.2.0) (#83) (@lorenzoh)