public IList <TRunDetail> Execute()
{
var stopwatch = Stopwatch.StartNew();
var iterationResults = new List <TRunDetail>();
do
{
var iterationStopwatch = Stopwatch.StartNew();
// get next pipeline
var getPipelineStopwatch = Stopwatch.StartNew();
var pipeline = PipelineSuggester.GetNextInferredPipeline(_context, _history, _datasetColumnInfo, _task,
_optimizingMetricInfo.IsMaximizing, _experimentSettings.CacheBeforeTrainer, _trainerAllowList);
var pipelineInferenceTimeInSeconds = getPipelineStopwatch.Elapsed.TotalSeconds;
// break if no candidates returned, means no valid pipeline available
if (pipeline == null)
{
break;
}
// evaluate pipeline
_logger.Trace($"Evaluating pipeline {pipeline.ToString()}");
(SuggestedPipelineRunDetail suggestedPipelineRunDetail, TRunDetail runDetail)
= _runner.Run(pipeline, _modelDirectory, _history.Count + 1);
_history.Add(suggestedPipelineRunDetail);
WriteIterationLog(pipeline, suggestedPipelineRunDetail, iterationStopwatch);
runDetail.RuntimeInSeconds = iterationStopwatch.Elapsed.TotalSeconds;
runDetail.PipelineInferenceTimeInSeconds = getPipelineStopwatch.Elapsed.TotalSeconds;
ReportProgress(runDetail);
iterationResults.Add(runDetail);
// if model is perfect, break
if (_metricsAgent.IsModelPerfect(suggestedPipelineRunDetail.Score))
{
break;
}
// If after third run, all runs have failed so far, throw exception
if (_history.Count() == 3 && _history.All(r => !r.RunSucceeded))
{
throw new InvalidOperationException($"Training failed with the exception: {_history.Last().Exception}");
}
} while (_history.Count < _experimentSettings.MaxModels &&
!_experimentSettings.CancellationToken.IsCancellationRequested &&
stopwatch.Elapsed.TotalSeconds < _experimentSettings.MaxExperimentTimeInSeconds);
return(iterationResults);
}
public IList <TRunDetail> Execute()
{
var iterationResults = new List <TRunDetail>();
// Create a timer for the max duration of experiment. When given time has
// elapsed, MaxExperimentTimeExpiredEvent is called to interrupt training
// of current model. Timer is not used if no experiment time is given, or
// is not a positive number.
if (_experimentSettings.MaxExperimentTimeInSeconds > 0)
{
_maxExperimentTimeTimer = new Timer(
new TimerCallback(MaxExperimentTimeExpiredEvent), null,
_experimentSettings.MaxExperimentTimeInSeconds * 1000, Timeout.Infinite
);
}
// If given max duration of experiment is 0, only 1 model will be trained.
// _experimentSettings.MaxExperimentTimeInSeconds is of type uint, it is
// either 0 or >0.
else
{
_experimentTimerExpired = true;
}
// Add second timer to check for the cancelation signal from the main MLContext
// to the active child MLContext. This timer will propagate the cancelation
// signal from the main to the child MLContexs if the main MLContext is
// canceled.
_mainContextCanceledTimer = new Timer(new TimerCallback(MainContextCanceledEvent), null, 1000, 1000);
// Pseudo random number generator to result in deterministic runs with the provided main MLContext's seed and to
// maintain variability between training iterations.
int?mainContextSeed = ((IHostEnvironmentInternal)_context.Model.GetEnvironment()).Seed;
_newContextSeedGenerator = (mainContextSeed.HasValue) ? RandomUtils.Create(mainContextSeed.Value) : null;
do
{
try
{
var iterationStopwatch = Stopwatch.StartNew();
// get next pipeline
var getPipelineStopwatch = Stopwatch.StartNew();
// A new MLContext is needed per model run. When max experiment time is reached, each used
// context is canceled to stop further model training. The cancellation of the main MLContext
// a user has instantiated is not desirable, thus additional MLContexts are used.
_currentModelMLContext = _newContextSeedGenerator == null ? new MLContext() : new MLContext(_newContextSeedGenerator.Next());
_currentModelMLContext.Log += RelayCurrentContextLogsToLogger;
var pipeline = PipelineSuggester.GetNextInferredPipeline(_currentModelMLContext, _history, _datasetColumnInfo, _task,
_optimizingMetricInfo.IsMaximizing, _experimentSettings.CacheBeforeTrainer, _logger, _trainerAllowList);
// break if no candidates returned, means no valid pipeline available
if (pipeline == null)
{
break;
}
// evaluate pipeline
_logger.Trace($"Evaluating pipeline {pipeline.ToString()}");
(SuggestedPipelineRunDetail suggestedPipelineRunDetail, TRunDetail runDetail)
= _runner.Run(pipeline, _modelDirectory, _history.Count + 1);
_history.Add(suggestedPipelineRunDetail);
WriteIterationLog(pipeline, suggestedPipelineRunDetail, iterationStopwatch);
runDetail.RuntimeInSeconds = iterationStopwatch.Elapsed.TotalSeconds;
runDetail.PipelineInferenceTimeInSeconds = getPipelineStopwatch.Elapsed.TotalSeconds;
ReportProgress(runDetail);
iterationResults.Add(runDetail);
// if model is perfect, break
if (_metricsAgent.IsModelPerfect(suggestedPipelineRunDetail.Score))
{
break;
}
// If after third run, all runs have failed so far, throw exception
if (_history.Count() == 3 && _history.All(r => !r.RunSucceeded))
{
throw new InvalidOperationException($"Training failed with the exception: {_history.Last().Exception}");
}
}
catch (OperationCanceledException e)
{
// This exception is thrown when the IHost/MLContext of the trainer is canceled due to
// reaching maximum experiment time. Simply catch this exception and return finished
// iteration results.
_logger.Warning(_operationCancelledMessage, e.Message);
return(iterationResults);
}
catch (AggregateException e)
{
// This exception is thrown when the IHost/MLContext of the trainer is canceled due to
// reaching maximum experiment time. Simply catch this exception and return finished
// iteration results. For some trainers, like FastTree, because training is done in parallel
// in can throw multiple OperationCancelledExceptions. This causes them to be returned as an
// AggregateException and misses the first catch block. This is to handle that case.
if (e.InnerExceptions.All(exception => exception is OperationCanceledException))
{
_logger.Warning(_operationCancelledMessage, e.Message);
return(iterationResults);
}
throw;
}
} while (_history.Count < _experimentSettings.MaxModels &&
!_experimentSettings.CancellationToken.IsCancellationRequested &&
!_experimentTimerExpired);
return(iterationResults);
}