LightGBM Save/Load round trip loses Softmax OneVersusAllModelParameters

[rauhs]

version: 0.11

Related: #1424

When training a softmax Multi classifier with LightGBM a save/load will lose the softmax it seems:

Also inspecting the model we can see it doesn't use ImplSoftmax but ImplRaw.

Reproduce:

 public class GenericSample
 {
 public string A { get; set; }
 public string Label { get; set; }
 }
 public static void ReproduceLightGbmPersistanceBug()
 {
 var data = Enumerable.Range(1, 100).Select(x => new GenericSample { A = $"{x % 20}", Label = $"{x % 10}" });
 var ctx = new MLContext();

 var options = new Options {
 UseSoftmax = true,
 };
 var pipe = ctx.Transforms.Categorical.OneHotEncoding("A")
 .Append(ctx.Transforms.Concatenate("Features", "A"))
 .Append(ctx.Transforms.Conversion.MapValueToKey("Label"))
 .Append(ctx.MulticlassClassification.Trainers.LightGbm(options));
 var dataView = ctx.Data.LoadFromEnumerable(data);
 ITransformer model = pipe.Fit(dataView);
 var scores = model.Transform(dataView).GetColumn<float[]>(ctx,"Score");

 Console.WriteLine($"Min: {scores.Select(x => x.Min()).Min()}");
 Console.WriteLine($"Max: {scores.Select(x => x.Max()).Max()}");

 var memoryStream = new MemoryStream();
 ctx.Model.Save(model, memoryStream);
 model = ctx.Model.Load(memoryStream);

 scores = model.Transform(dataView).GetColumn<float[]>(ctx,"Score");
 Console.WriteLine($"Min: {scores.Select(x => x.Min()).Min()}");
 Console.WriteLine($"Max: {scores.Select(x => x.Max()).Max()}");
 }

Output:

Min: 0.001027671
Max: 0.9907509
Min: -4.843706
Max: 2.027462

[najeeb-kazmi]

Reproduces with 1.0.0 as well. Classifying as bug.

This behavior is not present in other multiclass trainers SdcaMaximumEntropy, LbfgsMaximumEntropy, and OneVersusAll. Could be related to calibration not being set to true in the TrainerInfo for LightGbmTrainerBase.

[drakarah]

Is there a workaround for this? I'd like to combine multiple models together, all trained with LightGbm, but I currently can't combine the predictions because the min/max scores range of each model is different. If both models were to output the 0-1 range I could aggregate the probabilities more easily.