Skip to content

Create API for extracting information about the nodes in a TensorFlow model #862

New issue

Have a question about this project? Sign up for a free GitHub account to open an issue and contact its maintainers and the community.

Sign up for GitHub

By clicking “Sign up for GitHub”, you agree to our terms of service and privacy statement. We’ll occasionally send you account related emails.

Already on GitHub? Sign in to your account

Merged
yaeldekel merged 12 commits into from
Sep 20, 2018
Merged

Create API for extracting information about the nodes in a TensorFlow model #862

Show file tree
Hide file tree
Changes from 3 commits
Commits
Show all changes
12 commits
Select commit Hold shift + click to select a range
0e1e2fb
Add a method that returns TensorFlow model outputs as an ISchema.
yaeldMS Sep 7, 2018
0476b47
Merge from master.
yaeldMS Sep 7, 2018
f632417
Update after merge with master
yaeldMS Sep 7, 2018
78b0ae4
Address PR comments.
yaeldMS Sep 11, 2018
9669fbe
Merge from master.
yaeldMS Sep 13, 2018
3702031
Add metadata with information about the operation type, and the input…
yaeldMS Sep 13, 2018
1f6a931
Add method that returns an enumerable of the information about graph …
yaeldMS Sep 14, 2018
478b020
Add the DnnAnalyzer project files.
yaeldMS Sep 14, 2018
8a9e2c4
Address code review comments
yaeldMS Sep 18, 2018
bc9fbc1
Merge with master
yaeldMS Sep 20, 2018
7ad8de9
Make needed changes after merge with master
yaeldMS Sep 20, 2018
b94987d
Fix bug when there is a node with 1 dimension that is unknown
yaeldMS Sep 20, 2018
File filter

Filter by extension

Filter by extension
Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
24 changes: 20 additions & 4 deletions src/Microsoft.ML.TensorFlow/TensorFlow/Tensorflow.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -4,8 +4,6 @@

using System;
using System.Runtime.InteropServices;
using System.Text;
using System.Globalization;
using System.Linq;

// We use this TF_Xxx as the native "TF_Xxx *" as those are opaque
Expand All @@ -24,9 +22,9 @@
using TF_DeviceList = System.IntPtr;

using size_t = System.UIntPtr;
using System.Numerics;
using System.Collections.Generic;
using System.Linq.Expressions;
using Microsoft.ML.Runtime.Internal.Utilities;
using Microsoft.ML.Runtime.Data;

@ericstj ericstj Sep 10, 2018 •
edited by yaeldekel
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Are these using statements actually necessary? I'm missing the additions that actually used them. #Resolved

@yaeldekel yaeldekel Sep 11, 2018

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for pointing it out. I added the new API in this file, but ended up moving it to TensorflowUtils.

In reply to: 216493797 [](ancestors = 216493797)


#pragma warning disable MSML_GeneralName
#pragma warning disable MSML_PrivateFieldName
Expand Down Expand Up @@ -700,6 +698,24 @@ public override string ToString()
IntPtr len;
return TF_GraphDebugString(Handle, out len);
}

[DllImport(NativeBinding.TensorFlowLibrary)]
internal static extern string TF_OperationName(TF_Operation oper);

@ericstj ericstj Sep 10, 2018 •
edited by yaeldekel
Loading

Copy link
Copy Markdown
Member

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Rather than directly exposing the TF C-API we should consider bringing back the TF# defined API (TFOperation class) on top of it. See zeahmed@b2a8016#diff-ec7ea5716f3c05f773d3e1507b4f486aL729, zeahmed@b2a8016#diff-ec7ea5716f3c05f773d3e1507b4f486aL748. #Resolved


[DllImport(NativeBinding.TensorFlowLibrary)]
internal static extern string TF_OperationOpType(TF_Operation oper);

[DllImport(NativeBinding.TensorFlowLibrary)]
internal static extern int TF_OperationNumOutputs(TF_Operation oper);

[DllImport(NativeBinding.TensorFlowLibrary)]
internal static extern TFDataType TF_OperationOutputType(TFOutput oper_out);

[DllImport(NativeBinding.TensorFlowLibrary)]
internal static extern int TF_OperationNumInputs(TF_Operation oper);

[DllImport(NativeBinding.TensorFlowLibrary)]
internal static unsafe extern TF_Operation TF_GraphNextOperation(TF_Graph graph, long* pos);
}

/// <summary>
Expand Down
79 changes: 78 additions & 1 deletion src/Microsoft.ML.TensorFlow/TensorFlow/TensorflowUtils.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -3,9 +3,16 @@
// See the LICENSE file in the project root for more information.

using System;
using System.Collections.Generic;
using System.IO;
using System.Linq;
using System.Runtime.InteropServices;
using Microsoft.ML.Runtime;
using Microsoft.ML.Runtime.Data;
using Microsoft.ML.Runtime.ImageAnalytics.EntryPoints;
using Microsoft.ML.Runtime.Internal.Utilities;

using TF_Operation = System.IntPtr;

namespace Microsoft.ML.Transforms.TensorFlow
{
Expand All @@ -22,7 +29,54 @@ public static void Initialize()
ImageAnalytics.Initialize();
}

private static unsafe ISchema GetModelSchema(IExceptionContext ectx, TFGraph graph)

@Ivanidzo4ka Ivanidzo4ka Sep 17, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

unsafe [](start = 23, length = 6)

just curious, why it's unsafe? #Closed

@yaeldekel yaeldekel Sep 18, 2018

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Thanks for pointing it out, it doesn't need to be. In a previous iteration I was directly running the while loop over the graph that is now in the Graph class, so it was unsafe.

In reply to: 218204434 [](ancestors = 218204434)

{
long pos = 0;
TF_Operation oper;
var res = new List<KeyValuePair<string, ColumnType>>();
while ((oper = TFGraph.TF_GraphNextOperation(graph.handle, &pos)) != IntPtr.Zero)
{
var name = TFGraph.TF_OperationName(oper);
var type = TFGraph.TF_OperationOpType(oper);

@zeahmed zeahmed Sep 10, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

type [](start = 20, length = 4)

Its not being used anywhere. #Closed

var numOutputs = TFGraph.TF_OperationNumOutputs(oper);
if (numOutputs != 1)
continue;

var numInputs = TFGraph.TF_OperationNumInputs(oper);
if (numInputs == 0)

@zeahmed zeahmed Sep 10, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

numInputs [](start = 20, length = 9)

What does numInputs == 0 indicates? The input node does not have any input I believe??? #Closed

@yaeldekel yaeldekel Sep 11, 2018

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

The input has OpType "Placeholder". There are other nodes with numInputs==0, which have OpType "Const". I am not sure what they do but I think we don't want them in our output schema. What do you think?

In reply to: 216508337 [](ancestors = 216508337)

@zeahmed zeahmed Sep 11, 2018

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

What do you think if such nodes are not filter? Do you foresee adverse effect of doing this or not doing this?

In reply to: 216758114 [](ancestors = 216758114,216508337)

@yaeldekel yaeldekel Sep 13, 2018

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

I added the nodes with 0 inputs to the schema as well.

In reply to: 216817673 [](ancestors = 216817673,216758114,216508337)

continue;

var tfType = TFGraph.TF_OperationOutputType(new TFOutput(graph[name]));
var mlType = Tf2MlNetTypeOrNull(tfType);
if (mlType == null)
continue;

var shape = graph.GetTensorShape(new TFOutput(graph[name]));
var shapeArray = shape.ToIntArray();
var columnType = Utils.Size(shapeArray) > 0 && shapeArray.Skip(1).All(x => x > 0) ?
new VectorType(mlType, shapeArray[0] > 0 ? shapeArray : shapeArray.Skip(1).ToArray())
: new VectorType(mlType);
res.Add(new KeyValuePair<string, ColumnType>(name, columnType));
}
return new SimpleSchema(ectx, res.ToArray());
}

public static ISchema GetModelSchema(IExceptionContext ectx, string modelFile)
{
var bytes = File.ReadAllBytes(modelFile);
var session = LoadTFSession(ectx, bytes, modelFile);

@zeahmed zeahmed Sep 10, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

LoadTFSession [](start = 26, length = 13)

What about models that are not frozen? I assume it will be there once @abgoswam changes are there, right? #Closed

@yaeldekel yaeldekel Sep 11, 2018

Copy link
Copy Markdown
Author

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Yes. I assume that LoadTFSession will take care of the logic to decide which kind of model to load.

In reply to: 216509938 [](ancestors = 216509938)

return GetModelSchema(ectx, session.Graph);
}

internal static PrimitiveType Tf2MlNetType(TFDataType type)
{
var mlNetType = Tf2MlNetTypeOrNull(type);
if (mlNetType == null)
throw new NotSupportedException("TensorFlow type not supported.");
return mlNetType;
}

private static PrimitiveType Tf2MlNetTypeOrNull(TFDataType type)
{
switch (type)
{
Expand All @@ -35,8 +89,27 @@ internal static PrimitiveType Tf2MlNetType(TFDataType type)
case TFDataType.UInt64:
return NumberType.U8;
default:
throw new NotSupportedException("TensorFlow type not supported.");
return null;
}
}

internal static TFSession LoadTFSession(IExceptionContext ectx, byte[] modelBytes, string modelArg)
{
var graph = new TFGraph();
try
{
graph.Import(modelBytes, "");
}
catch (Exception ex)
{
if (!string.IsNullOrEmpty(modelArg))
throw ectx.Except($"TensorFlow exception triggered while loading model from '{modelArg}'");
#pragma warning disable MSML_NoMessagesForLoadContext
throw ectx.ExceptDecode(ex, "Tensorflow exception triggered while loading model.");
#pragma warning restore MSML_NoMessagesForLoadContext

}
return new TFSession(graph);
}

internal static unsafe void FetchData<T>(IntPtr data, T[] result)
Expand All @@ -57,6 +130,10 @@ internal static bool IsTypeSupported(TFDataType tfoutput)
{
case TFDataType.Float:
case TFDataType.Double:
case TFDataType.UInt8:
case TFDataType.UInt16:
case TFDataType.UInt32:
case TFDataType.UInt64:
return true;
default:
return false;
Expand Down
22 changes: 3 additions & 19 deletions src/Microsoft.ML.TensorFlow/TensorflowTransform.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -158,22 +158,6 @@ public static IDataTransform Create(IHostEnvironment env, ModelLoadContext ctx,
public static IRowMapper Create(IHostEnvironment env, ModelLoadContext ctx, ISchema inputSchema)
=> Create(env, ctx).MakeRowMapper(inputSchema);

private TFSession LoadTFSession(byte[] modelBytes)
{
var graph = new TFGraph();
try
{
graph.Import(modelBytes, "");
}
catch (Exception ex)
{
#pragma warning disable MSML_NoMessagesForLoadContext
throw _host.ExceptDecode(ex, "Tensorflow exception triggered while loading model.");
#pragma warning restore MSML_NoMessagesForLoadContext
}
return new TFSession(graph);
}

private static byte[] CheckFileAndRead(IHostEnvironment env, string modelFile)
{
env.CheckNonWhiteSpace(modelFile, nameof(modelFile));
Expand All @@ -182,16 +166,16 @@ private static byte[] CheckFileAndRead(IHostEnvironment env, string modelFile)
}

public TensorFlowTransform(IHostEnvironment env, string modelFile, string[] inputs, string[] outputs) :
this(env, CheckFileAndRead(env, modelFile), inputs, outputs)
this(env, CheckFileAndRead(env, modelFile), inputs, outputs, modelFile)

@Ivanidzo4ka Ivanidzo4ka Sep 11, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

modelFile [](start = 73, length = 9)

is it model file or model args? #Closed

{
}

private TensorFlowTransform(IHostEnvironment env, byte[] modelBytes, string[] inputs, string[] outputs)
private TensorFlowTransform(IHostEnvironment env, byte[] modelBytes, string[] inputs, string[] outputs, string modelFile = null)
{
Contracts.CheckValue(env, nameof(env));
_host = env.Register(nameof(RegistrationName));
_host.CheckValue(modelBytes, nameof(modelBytes));
Session = LoadTFSession(modelBytes);
Session = TensorFlowUtils.LoadTFSession(_host, modelBytes, modelFile);

@Ivanidzo4ka Ivanidzo4ka Sep 11, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

Abhishek is working on brining another way to load session, so I think it's better to extend CheckFileAndRead function and force it to return you Session instead of byte array. So private constructor would just accept session. #Closed

foreach (var input in inputs)
{
_host.CheckNonWhiteSpace(input, nameof(inputs));
Expand Down
11 changes: 11 additions & 0 deletions test/Microsoft.ML.Tests/ScenariosWithDirectInstantiation/TensorflowTests.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -8,6 +8,7 @@
using Microsoft.ML.Runtime.ImageAnalytics;
using Microsoft.ML.Runtime.LightGBM;
using Microsoft.ML.Transforms;
using Microsoft.ML.Transforms.TensorFlow;
using System.Collections.Generic;
using System.IO;
using Xunit;
Expand Down Expand Up @@ -70,6 +71,16 @@ public void TensorFlowTransformMatrixMultiplicationTest()
}
}

[Fact]
public void TensorFlowListLayersMnistConv()
{
var model_location = "mnist_model/frozen_saved_model.pb";
using (var env = new TlcEnvironment(seed: 1, conc: 1))
{
var schema = TensorFlowUtils.GetModelSchema(env, model_location);

@zeahmed zeahmed Sep 10, 2018 •
edited
Loading

Copy link
Copy Markdown
Contributor

Choose a reason for hiding this comment

The reason will be displayed to describe this comment to others. Learn more.

schema [](start = 20, length = 6)

It would be nice to have schema checked against the actual model information. It seems like mnist_model/frozen_saved_model.pb is a big model. Matrix multiplication model used in above test may be a good for this test.
It can be implemented as a new test in addition to this one. #Closed

}
}

[Fact]
public void TensorFlowTransformMNISTConvTest()
{
Expand Down