Skip to content

Commit

Permalink
Merge official main
Browse files Browse the repository at this point in the history 
* remotes/official/main:
  Update lgbm to v2.3.1 (dotnet#5851)
  Speed-up bitmap operations on images. Fixes dotnet#5856 (dotnet#5857)
  Onnx recursion limit (dotnet#5840)
  Speed up the inference of the saved_model(s). Fixes dotnet#5847 (dotnet#5848)

Signed-off-by: darth-vader-lg <luigi.generale@gmail.com>
  • Loading branch information
@darth-vader-lg
darth-vader-lg committed Jun 26, 2021
2 parents 6aa56a3 + 1b3cb77 commit f11c475
Show file tree
Hide file tree
Showing 7 changed files with 115 additions and 28 deletions.
2 changes: 1 addition & 1 deletion eng/Versions.props
View file
Original file line number Diff line number Diff line change
Expand Up @@ -21,7 +21,7 @@
<SystemThreadingChannelsPackageVersion>4.7.1</SystemThreadingChannelsPackageVersion>
<!-- Other/External dependencies -->
<GoogleProtobufPackageVersion>3.10.1</GoogleProtobufPackageVersion>
<LightGBMPackageVersion>2.2.3</LightGBMPackageVersion>
<LightGBMPackageVersion>2.3.1</LightGBMPackageVersion>
<MicrosoftExtensionsPackageVersion>2.1.0</MicrosoftExtensionsPackageVersion>
<MicrosoftMLOnnxRuntimePackageVersion>1.6.0</MicrosoftMLOnnxRuntimePackageVersion>
<MlNetMklDepsPackageVersion>0.0.0.9</MlNetMklDepsPackageVersion>
Expand Down
33 changes: 30 additions & 3 deletions src/Microsoft.ML.OnnxTransformer/OnnxCatalog.cs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -53,7 +53,7 @@ public static class OnnxCatalog
/// <param name="modelFile">The path of the file containing the ONNX model.</param>
/// <param name="shapeDictionary">ONNX shapes to be used over those loaded from <paramref name="modelFile"/>.
/// For keys use names as stated in the ONNX model, e.g. "input". Stating the shapes with this parameter
/// is particullarly useful for working with variable dimension inputs and outputs.
/// is particularly useful for working with variable dimension inputs and outputs.
/// </param>
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on, <see langword="null" /> to run on CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
Expand Down Expand Up @@ -110,7 +110,7 @@ public static class OnnxCatalog
/// <param name="modelFile">The path of the file containing the ONNX model.</param>
/// <param name="shapeDictionary">ONNX shapes to be used over those loaded from <paramref name="modelFile"/>.
/// For keys use names as stated in the ONNX model, e.g. "input". Stating the shapes with this parameter
/// is particullarly useful for working with variable dimension inputs and outputs.
/// is particularly useful for working with variable dimension inputs and outputs.
/// </param>
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on, <see langword="null" /> to run on CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
Expand Down Expand Up @@ -162,7 +162,7 @@ public static class OnnxCatalog
/// <param name="modelFile">The path of the file containing the ONNX model.</param>
/// <param name="shapeDictionary">ONNX shapes to be used over those loaded from <paramref name="modelFile"/>.
/// For keys use names as stated in the ONNX model, e.g. "input". Stating the shapes with this parameter
/// is particullarly useful for working with variable dimension inputs and outputs.
/// is particularly useful for working with variable dimension inputs and outputs.
/// </param>
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on, <see langword="null" /> to run on CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
Expand All @@ -176,6 +176,33 @@ public static class OnnxCatalog
=> new OnnxScoringEstimator(CatalogUtils.GetEnvironment(catalog), outputColumnNames, inputColumnNames,
modelFile, gpuDeviceId, fallbackToCpu, shapeDictionary: shapeDictionary);

/// <summary>
/// Create a <see cref="OnnxScoringEstimator"/>, which applies a pre-trained Onnx model to the <paramref name="inputColumnNames"/> columns.
/// Please refer to <see cref="OnnxScoringEstimator"/> to learn more about the necessary dependencies,
/// and how to run it on a GPU.
/// </summary>
/// <param name="catalog">The transform's catalog.</param>
/// <param name="outputColumnNames">The output columns resulting from the transformation.</param>
/// <param name="inputColumnNames">The input columns.</param>
/// <param name="modelFile">The path of the file containing the ONNX model.</param>
/// <param name="shapeDictionary">ONNX shapes to be used over those loaded from <paramref name="modelFile"/>.
/// For keys use names as stated in the ONNX model, e.g. "input". Stating the shapes with this parameter
/// is particularly useful for working with variable dimension inputs and outputs.
/// </param>
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on, <see langword="null" /> to run on CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
public static OnnxScoringEstimator ApplyOnnxModel(this TransformsCatalog catalog,
string[] outputColumnNames,
string[] inputColumnNames,
string modelFile,
IDictionary<string, int[]> shapeDictionary,
int? gpuDeviceId = null,
bool fallbackToCpu = false,
int recursionLimit = 100)
=> new OnnxScoringEstimator(CatalogUtils.GetEnvironment(catalog), outputColumnNames, inputColumnNames,
modelFile, gpuDeviceId, fallbackToCpu, shapeDictionary: shapeDictionary, recursionLimit);

/// <summary>
/// Create <see cref="DnnImageFeaturizerEstimator"/>, which applies one of the pre-trained DNN models in
/// <see cref="DnnImageModelSelector"/> to featurize an image.
Expand Down
57 changes: 44 additions & 13 deletions src/Microsoft.ML.OnnxTransformer/OnnxTransform.cs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -87,6 +87,9 @@ internal sealed class Options : TransformInputBase

[Argument(ArgumentType.Multiple, HelpText = "Shapes used to overwrite shapes loaded from ONNX file.", SortOrder = 5)]
public CustomShapeInfo[] CustomShapeInfos;

[Argument(ArgumentType.AtMostOnce, HelpText = "Protobuf CodedInputStream recursion limit.", SortOrder = 6)]
public int RecursionLimit = 100;
}

/// <summary>
Expand Down Expand Up @@ -126,8 +129,9 @@ private static VersionInfo GetVersionInfo()
modelSignature: "ONNXSCOR",
// version 10001 is single input & output.
// version 10002 = multiple inputs & outputs
verWrittenCur: 0x00010002,
verReadableCur: 0x00010002,
// version 10003 = custom protobuf recursion limit
verWrittenCur: 0x00010003,
verReadableCur: 0x00010003,
verWeCanReadBack: 0x00010001,
loaderSignature: LoaderSignature,
loaderAssemblyName: typeof(OnnxTransformer).Assembly.FullName);
Expand Down Expand Up @@ -184,7 +188,26 @@ private static OnnxTransformer Create(IHostEnvironment env, ModelLoadContext ctx
}
}

var options = new Options() { InputColumns = inputs, OutputColumns = outputs, CustomShapeInfos = loadedCustomShapeInfos };
int recursionLimit;

// Recursion limit change
if (ctx.Header.ModelVerWritten >= 0x00010003)
{
recursionLimit = ctx.Reader.ReadInt32();
}
else
{
// Default if not written inside ONNX model
recursionLimit = 100;
}

var options = new Options()
{
InputColumns = inputs,
OutputColumns = outputs,
CustomShapeInfos = loadedCustomShapeInfos,
RecursionLimit = recursionLimit
};

return new OnnxTransformer(env, options, modelBytes);
}
Expand Down Expand Up @@ -221,13 +244,13 @@ private static IRowMapper Create(IHostEnvironment env, ModelLoadContext ctx, Dat
Host.CheckNonWhiteSpace(options.ModelFile, nameof(options.ModelFile));
Host.CheckIO(File.Exists(options.ModelFile), "Model file {0} does not exists.", options.ModelFile);
// Because we cannot delete the user file, ownModelFile should be false.
Model = new OnnxModel(options.ModelFile, options.GpuDeviceId, options.FallbackToCpu, ownModelFile: false, shapeDictionary: shapeDictionary);
Model = new OnnxModel(options.ModelFile, options.GpuDeviceId, options.FallbackToCpu, ownModelFile: false, shapeDictionary: shapeDictionary, options.RecursionLimit);
}
else
{
// Entering this region means that the byte[] is passed as the model. To feed that byte[] to ONNXRuntime, we need
// to create a temporal file to store it and then call ONNXRuntime's API to load that file.
Model = OnnxModel.CreateFromBytes(modelBytes, env, options.GpuDeviceId, options.FallbackToCpu, shapeDictionary: shapeDictionary);
Model = OnnxModel.CreateFromBytes(modelBytes, env, options.GpuDeviceId, options.FallbackToCpu, shapeDictionary: shapeDictionary, options.RecursionLimit);
}
}
catch (OnnxRuntimeException e)
Expand Down Expand Up @@ -258,16 +281,18 @@ private static IRowMapper Create(IHostEnvironment env, ModelLoadContext ctx, Dat
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on. Null for CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
/// <param name="shapeDictionary"></param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
internal OnnxTransformer(IHostEnvironment env, string modelFile, int? gpuDeviceId = null,
bool fallbackToCpu = false, IDictionary<string, int[]> shapeDictionary = null)
bool fallbackToCpu = false, IDictionary<string, int[]> shapeDictionary = null, int recursionLimit = 100)
: this(env, new Options()
{
ModelFile = modelFile,
InputColumns = new string[] { },
OutputColumns = new string[] { },
GpuDeviceId = gpuDeviceId,
FallbackToCpu = fallbackToCpu,
CustomShapeInfos = shapeDictionary?.Select(pair => new CustomShapeInfo(pair.Key, pair.Value)).ToArray()
CustomShapeInfos = shapeDictionary?.Select(pair => new CustomShapeInfo(pair.Key, pair.Value)).ToArray(),
RecursionLimit = recursionLimit
})
{
}
Expand All @@ -283,16 +308,18 @@ private static IRowMapper Create(IHostEnvironment env, ModelLoadContext ctx, Dat
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on. Null for CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
/// <param name="shapeDictionary"></param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
internal OnnxTransformer(IHostEnvironment env, string[] outputColumnNames, string[] inputColumnNames, string modelFile, int? gpuDeviceId = null, bool fallbackToCpu = false,
IDictionary<string, int[]> shapeDictionary = null)
IDictionary<string, int[]> shapeDictionary = null, int recursionLimit = 100)
: this(env, new Options()
{
ModelFile = modelFile,
InputColumns = inputColumnNames,
OutputColumns = outputColumnNames,
GpuDeviceId = gpuDeviceId,
FallbackToCpu = fallbackToCpu,
CustomShapeInfos = shapeDictionary?.Select(pair => new CustomShapeInfo(pair.Key, pair.Value)).ToArray()
CustomShapeInfos = shapeDictionary?.Select(pair => new CustomShapeInfo(pair.Key, pair.Value)).ToArray(),
RecursionLimit = recursionLimit
})
{
}
Expand Down Expand Up @@ -325,6 +352,8 @@ private protected override void SaveModel(ModelSaveContext ctx)
ctx.SaveNonEmptyString(info.Name);
ctx.Writer.WriteIntArray(info.Shape);
}

ctx.Writer.Write(_options.RecursionLimit);
}

private protected override IRowMapper MakeRowMapper(DataViewSchema inputSchema) => new Mapper(this, inputSchema);
Expand Down Expand Up @@ -807,10 +836,11 @@ public sealed class OnnxScoringEstimator : TrivialEstimator<OnnxTransformer>
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on. Null for CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
/// <param name="shapeDictionary"></param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
[BestFriend]
internal OnnxScoringEstimator(IHostEnvironment env, string modelFile, int? gpuDeviceId = null, bool fallbackToCpu = false,
IDictionary<string, int[]> shapeDictionary = null)
: this(env, new OnnxTransformer(env, new string[] { }, new string[] { }, modelFile, gpuDeviceId, fallbackToCpu, shapeDictionary))
IDictionary<string, int[]> shapeDictionary = null, int recursionLimit = 100)
: this(env, new OnnxTransformer(env, new string[] { }, new string[] { }, modelFile, gpuDeviceId, fallbackToCpu, shapeDictionary, recursionLimit))
{
}

Expand All @@ -825,9 +855,10 @@ public sealed class OnnxScoringEstimator : TrivialEstimator<OnnxTransformer>
/// <param name="gpuDeviceId">Optional GPU device ID to run execution on. Null for CPU.</param>
/// <param name="fallbackToCpu">If GPU error, raise exception or fallback to CPU.</param>
/// <param name="shapeDictionary"></param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
internal OnnxScoringEstimator(IHostEnvironment env, string[] outputColumnNames, string[] inputColumnNames, string modelFile,
int? gpuDeviceId = null, bool fallbackToCpu = false, IDictionary<string, int[]> shapeDictionary = null)
: this(env, new OnnxTransformer(env, outputColumnNames, inputColumnNames, modelFile, gpuDeviceId, fallbackToCpu, shapeDictionary))
int? gpuDeviceId = null, bool fallbackToCpu = false, IDictionary<string, int[]> shapeDictionary = null, int recursionLimit = 100)
: this(env, new OnnxTransformer(env, outputColumnNames, inputColumnNames, modelFile, gpuDeviceId, fallbackToCpu, shapeDictionary, recursionLimit))
{
}

Expand Down
14 changes: 8 additions & 6 deletions src/Microsoft.ML.OnnxTransformer/OnnxUtils.cs
View file
Original file line number Diff line number Diff line change
Expand Up @@ -164,8 +164,9 @@ public OnnxVariableInfo(string name, OnnxShape shape, Type typeInOnnxRuntime, Da
/// <param name="ownModelFile">If true, the <paramref name="modelFile"/> will be deleted when <see cref="OnnxModel"/> is
/// no longer needed.</param>
/// <param name="shapeDictionary"></param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
public OnnxModel(string modelFile, int? gpuDeviceId = null, bool fallbackToCpu = false,
bool ownModelFile=false, IDictionary<string, int[]> shapeDictionary = null)
bool ownModelFile=false, IDictionary<string, int[]> shapeDictionary = null, int recursionLimit = 100)
{
// If we don't own the model file, _disposed should be false to prevent deleting user's file.
_disposed = false;
Expand Down Expand Up @@ -204,7 +205,7 @@ public OnnxVariableInfo(string name, OnnxShape shape, Type typeInOnnxRuntime, Da

// The CodedInputStream auto closes the stream, and we need to make sure that our main stream stays open, so creating a new one here.
using (var modelStream = new FileStream(modelFile, FileMode.Open, FileAccess.Read, FileShare.Delete | FileShare.Read))
using (var codedStream = Google.Protobuf.CodedInputStream.CreateWithLimits(modelStream, Int32.MaxValue, 100))
using (var codedStream = Google.Protobuf.CodedInputStream.CreateWithLimits(modelStream, Int32.MaxValue, recursionLimit))
model = OnnxCSharpToProtoWrapper.ModelProto.Parser.ParseFrom(codedStream);

// Parse actual input and output types stored in the loaded ONNX model to get their DataViewType's.
Expand Down Expand Up @@ -321,7 +322,7 @@ private static bool CheckOnnxShapeCompatibility(IEnumerable<int> left, IEnumerab

/// <summary>
/// Create an OnnxModel from a byte[]. Usually, a ONNX model is consumed by <see cref="OnnxModel"/> as a file.
/// With <see cref="CreateFromBytes(byte[], IHostEnvironment)"/> and <see cref="CreateFromBytes(byte[], IHostEnvironment, int?, bool, IDictionary{string, int[]})"/>,
/// With <see cref="CreateFromBytes(byte[], IHostEnvironment)"/> and <see cref="CreateFromBytes(byte[], IHostEnvironment, int?, bool, IDictionary{string, int[]}, int)"/>,
/// it's possible to use in-memory model (type: byte[]) to create <see cref="OnnxModel"/>.
/// </summary>
/// <param name="modelBytes">Bytes of the serialized model</param>
Expand All @@ -335,7 +336,7 @@ public static OnnxModel CreateFromBytes(byte[] modelBytes, IHostEnvironment env)
/// Create an OnnxModel from a byte[]. Set execution to GPU if required.
/// Usually, a ONNX model is consumed by <see cref="OnnxModel"/> as a file.
/// With <see cref="CreateFromBytes(byte[], IHostEnvironment)"/> and
/// <see cref="CreateFromBytes(byte[], IHostEnvironment, int?, bool, IDictionary{string, int[]})"/>,
/// <see cref="CreateFromBytes(byte[], IHostEnvironment, int?, bool, IDictionary{string, int[]}, int)"/>,
/// it's possible to use in-memory model (type: byte[]) to create <see cref="OnnxModel"/>.
/// </summary>
/// <param name="modelBytes">Bytes of the serialized model.</param>
Expand All @@ -345,14 +346,15 @@ public static OnnxModel CreateFromBytes(byte[] modelBytes, IHostEnvironment env)
/// <param name="shapeDictionary">User-provided shapes. If the key "myTensorName" is associated
/// with the value [1, 3, 5], the shape of "myTensorName" will be set to [1, 3, 5].
/// The shape loaded from <paramref name="modelBytes"/> would be overwritten.</param>
/// <param name="recursionLimit">Optional, specifies the Protobuf CodedInputStream recursion limit. Default value is 100.</param>
/// <returns>An <see cref="OnnxModel"/></returns>
public static OnnxModel CreateFromBytes(byte[] modelBytes, IHostEnvironment env, int? gpuDeviceId = null, bool fallbackToCpu = false,
IDictionary<string, int[]> shapeDictionary = null)
IDictionary<string, int[]> shapeDictionary = null, int recursionLimit = 100)
{
var tempModelFile = Path.Combine(((IHostEnvironmentInternal)env).TempFilePath, Path.GetRandomFileName());
File.WriteAllBytes(tempModelFile, modelBytes);
return new OnnxModel(tempModelFile, gpuDeviceId, fallbackToCpu,
ownModelFile: true, shapeDictionary: shapeDictionary);
ownModelFile: true, shapeDictionary: shapeDictionary, recursionLimit);
}

/// <summary>
Expand Down

0 comments on commit f11c475

Please sign in to comment.