public void TestOnnxTransformWithCustomShapes()
{
// The loaded model has input shape [-1, 3] and output shape [-1].
var modelFile = Path.Combine(Directory.GetCurrentDirectory(), "unknowndimensions", "test_unknowndimensions_float.onnx");
var dataPoints = new InputWithCustomShape[]
{
// It's a flattened 3-by-3 tensor.
// [1.1, 1.3, 1.2]
// |1.9, 1.3, 1.2|
// [1.1, 1.3, 1.8]
new InputWithCustomShape()
{
input = new float[] { 1.1f, 1.3f, 1.2f, 1.9f, 1.3f, 1.2f, 1.1f, 1.3f, 1.8f }
},
// It's a flattened 3-by-3 tensor.
// [0, 0, 1]
// |1, 0, 0|
// [1, 0, 0]
new InputWithCustomShape()
{
input = new float[] { 0f, 0f, 1f, 1f, 0f, 0f, 1f, 0f, 0f }
}
};
var shapeDictionary = new Dictionary <string, int[]>()
{
{ nameof(InputWithCustomShape.input), new int[] { 3, 3 } }
};
var dataView = ML.Data.LoadFromEnumerable(dataPoints);
var pipeline = new OnnxScoringEstimator[3];
var onnxTransformer = new OnnxTransformer[3];
var transformedDataViews = new IDataView[3];
// Test three public ONNX APIs with the custom shape.
// Test 1.
pipeline[0] = ML.Transforms.ApplyOnnxModel(
new[] { nameof(PredictionWithCustomShape.argmax) }, new[] { nameof(InputWithCustomShape.input) },
modelFile, shapeDictionary);
onnxTransformer[0] = pipeline[0].Fit(dataView);
transformedDataViews[0] = onnxTransformer[0].Transform(dataView);
// Test 2.
pipeline[1] = ML.Transforms.ApplyOnnxModel(
nameof(PredictionWithCustomShape.argmax), nameof(InputWithCustomShape.input),
modelFile, shapeDictionary);
onnxTransformer[1] = pipeline[1].Fit(dataView);
transformedDataViews[1] = onnxTransformer[1].Transform(dataView);
// Test 3.
pipeline[2] = ML.Transforms.ApplyOnnxModel(modelFile, shapeDictionary);
onnxTransformer[2] = pipeline[2].Fit(dataView);
transformedDataViews[2] = onnxTransformer[2].Transform(dataView);
// Conduct the same check for all the 3 called public APIs.
foreach (var transformedDataView in transformedDataViews)
{
var transformedDataPoints = ML.Data.CreateEnumerable <PredictionWithCustomShape>(transformedDataView, false).ToList();
// One data point generates one transformed data point.
Assert.Equal(dataPoints.Count(), transformedDataPoints.Count);
// Check result numbers. They are results of applying ONNX argmax along the second axis; for example
// [1.1, 1.3, 1.2] ---> [1] because 1.3 (indexed by 1) is the largest element.
// |1.9, 1.3, 1.2| ---> |0| 1.9 0
// [1.1, 1.3, 1.8] ---> [2] 1.8 2
var expectedResults = new long[][]
{
new long[] { 1, 0, 2 },
new long[] { 2, 0, 0 }
};
for (int i = 0; i < transformedDataPoints.Count; ++i)
{
Assert.Equal(transformedDataPoints[i].argmax, expectedResults[i]);
}
}
for (int i = 0; i < 3; i++)
{
(onnxTransformer[i] as IDisposable)?.Dispose();
}
}
public void TestOnnxTransformSaveAndLoadWithCustomShapes()
{
// The loaded model has input shape [-1, 3] and output shape [-1].
var modelFile = Path.Combine(Directory.GetCurrentDirectory(), "unknowndimensions", "test_unknowndimensions_float.onnx");
var dataPoints = new InputWithCustomShape[]
{
// It's a flattened 3-by-3 tensor.
// [1.1, 1.3, 1.2]
// |1.9, 1.3, 1.2|
// [1.1, 1.3, 1.8]
new InputWithCustomShape()
{
input = new float[] { 1.1f, 1.3f, 1.2f, 1.9f, 1.3f, 1.2f, 1.1f, 1.3f, 1.8f }
},
// It's a flattened 3-by-3 tensor.
// [0, 0, 1]
// |1, 0, 0|
// [1, 0, 0]
new InputWithCustomShape()
{
input = new float[] { 0f, 0f, 1f, 1f, 0f, 0f, 1f, 0f, 0f }
}
};
var shapeDictionary = new Dictionary <string, int[]>()
{
{ nameof(InputWithCustomShape.input), new int[] { 3, 3 } }
};
var dataView = ML.Data.LoadFromEnumerable(dataPoints);
var pipeline = ML.Transforms.ApplyOnnxModel(nameof(PredictionWithCustomShape.argmax),
nameof(InputWithCustomShape.input), modelFile, shapeDictionary);
var model = pipeline.Fit(dataView);
// Save the trained ONNX transformer into file and then load it back.
ITransformer loadedModel = null;
var tempPath = Path.GetTempFileName();
using (var file = new SimpleFileHandle(Env, tempPath, true, true))
{
// Save.
using (var fs = file.CreateWriteStream())
ML.Model.Save(model, null, fs);
// Load.
using (var fs = file.OpenReadStream())
loadedModel = ML.Model.Load(fs, out var schema);
}
var transformedDataView = loadedModel.Transform(dataView);
// Conduct the same check for all the 3 called public APIs.
var transformedDataPoints = ML.Data.CreateEnumerable <PredictionWithCustomShape>(transformedDataView, false).ToList();
// One data point generates one transformed data point.
Assert.Equal(dataPoints.Count(), transformedDataPoints.Count);
// Check result numbers. They are results of applying ONNX argmax along the second axis; for example
// [1.1, 1.3, 1.2] ---> [1] because 1.3 (indexed by 1) is the largest element.
// |1.9, 1.3, 1.2| ---> |0| 1.9 0
// [1.1, 1.3, 1.8] ---> [2] 1.8 2
var expectedResults = new long[][]
{
new long[] { 1, 0, 2 },
new long[] { 2, 0, 0 }
};
for (int i = 0; i < transformedDataPoints.Count; ++i)
{
Assert.Equal(transformedDataPoints[i].argmax, expectedResults[i]);
}
(model as IDisposable)?.Dispose();
(loadedModel as IDisposable)?.Dispose();
}
