public static SuggestedPipeline GetNextInferredPipeline(MLContext context,
IEnumerable <SuggestedPipelineRunDetail> history,
DatasetColumnInfo[] columns,
TaskKind task,
bool isMaximizingMetric,
CacheBeforeTrainer cacheBeforeTrainer,
IChannel logger,
IEnumerable <TrainerName> trainerAllowList = null)
{
var availableTrainers = RecipeInference.AllowedTrainers(context, task,
ColumnInformationUtil.BuildColumnInfo(columns), trainerAllowList);
var transforms = TransformInferenceApi.InferTransforms(context, task, columns).ToList();
var transformsPostTrainer = TransformInferenceApi.InferTransformsPostTrainer(context, task, columns).ToList();
// if we haven't run all pipelines once
if (history.Count() < availableTrainers.Count())
{
return(GetNextFirstStagePipeline(context, history, availableTrainers, transforms, transformsPostTrainer, cacheBeforeTrainer));
}
// get top trainers from stage 1 runs
var topTrainers = GetTopTrainers(history, availableTrainers, isMaximizingMetric);
// sort top trainers by # of times they've been run, from lowest to highest
var orderedTopTrainers = OrderTrainersByNumTrials(history, topTrainers);
// keep as hash set of previously visited pipelines
var visitedPipelines = new HashSet <SuggestedPipeline>(history.Select(h => h.Pipeline));
// iterate over top trainers (from least run to most run),
// to find next pipeline
foreach (var trainer in orderedTopTrainers)
{
var newTrainer = trainer.Clone();
// repeat until passes or runs out of chances
const int maxNumberAttempts = 10;
var count = 0;
do
{
// sample new hyperparameters for the learner
if (!SampleHyperparameters(context, newTrainer, history, isMaximizingMetric, logger))
{
// if unable to sample new hyperparameters for the learner
// (ie SMAC returned 0 suggestions), break
break;
}
var suggestedPipeline = SuggestedPipelineBuilder.Build(context, transforms, transformsPostTrainer, newTrainer, cacheBeforeTrainer);
// make sure we have not seen pipeline before
if (!visitedPipelines.Contains(suggestedPipeline))
{
return(suggestedPipeline);
}
} while (++count <= maxNumberAttempts);
}
return(null);
}
/// <summary>
/// Initializes a new instance of <see cref="ExperimentSettings"/>.
/// </summary>
public ExperimentSettings()
{
MaxExperimentTimeInSeconds = 24 * 60 * 60;
CancellationToken = default;
CacheDirectoryName = "Microsoft.ML.AutoML";
CacheBeforeTrainer = CacheBeforeTrainer.Auto;
MaxModels = int.MaxValue;
}
/// <summary>
/// Initializes a new instance of <see cref="ExperimentSettings"/>.
/// </summary>
public ExperimentSettings()
{
MaxExperimentTimeInSeconds = 24 * 60 * 60;
CancellationToken = default;
CacheDirectory = new DirectoryInfo(Path.Combine(Path.GetTempPath(), "Microsoft.ML.AutoML"));
CacheBeforeTrainer = CacheBeforeTrainer.Auto;
MaxModels = int.MaxValue;
}
private static SuggestedPipeline BuildSuggestedPipeline(SuggestedTrainer trainer,
CacheBeforeTrainer cacheBeforeTrainer = CacheBeforeTrainer.Auto)
{
return(SuggestedPipelineBuilder.Build(_context,
new List <SuggestedTransform>(),
new List <SuggestedTransform>(),
trainer, cacheBeforeTrainer));
}
private static SuggestedPipeline GetNextFirstStagePipeline(MLContext context,
IEnumerable <SuggestedPipelineRunDetail> history,
IEnumerable <SuggestedTrainer> availableTrainers,
ICollection <SuggestedTransform> transforms,
ICollection <SuggestedTransform> transformsPostTrainer,
CacheBeforeTrainer cacheBeforeTrainer)
{
var trainer = availableTrainers.ElementAt(history.Count());
return(SuggestedPipelineBuilder.Build(context, transforms, transformsPostTrainer, trainer, cacheBeforeTrainer));
}
public static SuggestedPipeline Build(MLContext context,
ICollection <SuggestedTransform> transforms,
ICollection <SuggestedTransform> transformsPostTrainer,
SuggestedTrainer trainer,
CacheBeforeTrainer cacheBeforeTrainerSettings)
{
var trainerInfo = trainer.BuildTrainer().Info;
AddNormalizationTransforms(context, trainerInfo, transforms);
var cacheBeforeTrainer = ShouldCacheBeforeTrainer(trainerInfo, cacheBeforeTrainerSettings);
return(new SuggestedPipeline(transforms, transformsPostTrainer, trainer, context, cacheBeforeTrainer));
}
private static bool ShouldCacheBeforeTrainer(TrainerInfo trainerInfo, CacheBeforeTrainer cacheBeforeTrainerSettings)
{
return(cacheBeforeTrainerSettings == CacheBeforeTrainer.On || (cacheBeforeTrainerSettings == CacheBeforeTrainer.Auto && trainerInfo.WantCaching));
}
