Skip to content
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

Jump to bottom

Upgrade TF.Net version from 0.10.10 to 0.11.3 #4205

Merged
merged 16 commits into from
Sep 17, 2019
Merged

Upgrade TF.Net version from 0.10.10 to 0.11.3 #4205

Show file tree
Hide file tree
Changes from 8 commits
Commits
Show all changes
16 commits
Select commit Hold shift + click to select a range
2186651
updated Tensorflow .Net from 0.10.10 to 0.11, changed decode_jpeg to …
ashbhandare Sep 7, 2019
df47f1a
Changed decode_image to decode_jpeg, modified few API changes for TF.…
ashbhandare Sep 9, 2019
efa7ea8
Fixed null graph error, reproducer for protobuff size error.
ashbhandare Sep 10, 2019
72b5715
Final changes with all ML Samples running with TF.NET version 0.11.3
ashbhandare Sep 11, 2019
9536843
removed local file diff.out
ashbhandare Sep 11, 2019
cd5be6d
Corrected for running all ML.Samples
ashbhandare Sep 11, 2019
3008242
cleanup
ashbhandare Sep 11, 2019
579bdb6
resolved common issurs with PR
ashbhandare Sep 12, 2019
e599ac9
removed extra tab
ashbhandare Sep 12, 2019
5eaa5d2
Added a check to set default graph only when current graph is not def…
ashbhandare Sep 13, 2019
42ec0d4
Changed to that tf.net version dependency can directly be changed fro…
ashbhandare Sep 13, 2019
90f7ffd
Removed unused variable, reverted ImageClassification.cs
ashbhandare Sep 13, 2019
6f3d33d
added previously missed package dependency variable for Tensorflow.NE…
ashbhandare Sep 13, 2019
ffc95e3
Created a sepatare function for string type so that MakeGetter can on…
ashbhandare Sep 16, 2019
e633904
Additional refactoring for MakeGetter for String type
ashbhandare Sep 16, 2019
9c2371c
Reverting the refactoring as the Microsoft.ML.Scenarios.TensorFlowSce…
ashbhandare Sep 16, 2019
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
2 changes: 1 addition & 1 deletion docs/samples/Microsoft.ML.Samples/Dynamic/TensorFlow/ImageClassification.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -38,7 +38,7 @@ public static void Example()
.ScoreTensorFlowModel(
new[] { nameof(OutputScores.output) },
new[] { nameof(TensorData.input) }, addBatchDimensionInput: true);

Copy link
Member

@safern safern Sep 12, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

Nit: tabs in empty line. #Resolved

// Run the pipeline and get the transformed values.
var estimator = pipeline.Fit(idv);
var transformedValues = estimator.Transform(idv);
Expand Down
66 changes: 35 additions & 31 deletions src/Microsoft.ML.Dnn/DnnRetrainTransform.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -235,7 +235,8 @@ private void CheckTrainingParameters(DnnRetrainEstimator.Options options)
inputTensor.InputType(index);
var tfInputShape = ((Tensor)inputTensor).TensorShape;

if (isInputVector && (tfInputShape == null || (tfInputShape.NDim == 0)))
var numInputDims = tfInputShape.ndim;
Copy link
Member

@codemzs codemzs Sep 16, 2019 •
edited
Loading

Choose a reason for hiding this comment

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

var numInputDims = tfInputShape.ndim; [](start = 12, length = 37)

It seems tfInputShape can be null. We should set numInputDims to -1 if tfInputShape is null otherwise assign numInputDims = tfInputShape.ndim; #Closed

if (isInputVector && (tfInputShape == null || (numInputDims == 0)))
{
var vecType = (VectorDataViewType)type;
var colTypeDims = new int[vecType.Dimensions.Length + 1];
Expand All @@ -245,13 +246,14 @@ private void CheckTrainingParameters(DnnRetrainEstimator.Options options)

tfInputShape = new TensorShape(colTypeDims);
}
if (tfInputShape.NDim != -1)
if (numInputDims != -1)
{
var newShape = new int[tfInputShape.NDim];
newShape[0] = tfInputShape[0] == 0 || tfInputShape[0] == -1 ? batchSize : tfInputShape[0];
var newShape = new int[numInputDims];
var dims = tfInputShape.dims;
newShape[0] = dims[0] == 0 || dims[0] == -1 ? batchSize : dims[0];

for (int j = 1; j < tfInputShape.NDim; j++)
newShape[j] = tfInputShape[j];
for (int j = 1; j < numInputDims; j++)
newShape[j] = dims[j];
tfInputShape = new TensorShape(newShape);
}

Expand Down Expand Up @@ -382,10 +384,9 @@ private void TrainCore(DnnRetrainEstimator.Options options, IDataView input, IDa
runner.AddInput(inputs[i], srcTensorGetters[i].GetBufferedBatchTensor());

Tensor[] tensor = runner.Run();
var buffer = tensor[0].Data();
loss = tensor.Length > 0 && tensor[0] != IntPtr.Zero ? (float)tensor[0].Data<float>()[0] : 0.0f;
metric = tensor.Length > 1 && tensor[1] != IntPtr.Zero ? (float)tensor[1].Data<float>()[0] : 0.0f;
var b = tensor.Length > 2 && tensor[2] != IntPtr.Zero ? (float[])tensor[2].Data<float>() : null;
loss = tensor.Length > 0 && tensor[0] != IntPtr.Zero ? (float)tensor[0].ToArray<float>()[0] : 0.0f;
eerhardt marked this conversation as resolved.
Show resolved Hide resolved
metric = tensor.Length > 1 && tensor[1] != IntPtr.Zero ? (float)tensor[1].ToArray<float>()[0] : 0.0f;
var b = tensor.Length > 2 && tensor[2] != IntPtr.Zero ? (float[])tensor[2].ToArray<float>() : null;
Copy link
Member

@eerhardt eerhardt Sep 13, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

I don't see b being used anywhere. Can it be removed? #Resolved

return (loss, metric);
}

Expand Down Expand Up @@ -639,7 +640,7 @@ internal static (TF_DataType[] tfOutputTypes, DataViewType[] outputTypes, (Opera
// i.e. the first dimension (if unknown) is assumed to be batch dimension.
// If there are other dimension that are unknown the transformer will return a variable length vector.
// This is the work around in absence of reshape transformer.
int[] dims = shape.NDim > 0 ? shape.Dimensions.Skip(shape[0] == -1 ? 1 : 0).ToArray() : new[] { 0 };
int[] dims = shape.ndim > 0 ? shape.dims.Skip(shape.dims[0] == -1 ? 1 : 0).ToArray() : new[] { 0 };
for (int j = 0; j < dims.Length; j++)
dims[j] = dims[j] == -1 ? 0 : dims[j];
if (dims == null || dims.Length == 0)
Expand Down Expand Up @@ -780,7 +781,7 @@ public Mapper(DnnRetrainTransformer parent, DataViewSchema inputSchema) :
if (type.GetItemType() != expectedType)
throw Host.ExceptSchemaMismatch(nameof(inputSchema), "input", _parent._inputs[i], expectedType.ToString(), type.ToString());
var originalShape = _parent._tfInputShapes[i];
var shape = originalShape.Dimensions;
var shape = originalShape.dims;

var colTypeDims = vecType.Dimensions.Select(dim => (int)dim).ToArray();
if (shape == null || (shape.Length == 0))
Expand Down Expand Up @@ -811,18 +812,20 @@ public Mapper(DnnRetrainTransformer parent, DataViewSchema inputSchema) :
throw Contracts.Except($"Input shape mismatch: Input '{_parent._inputs[i]}' has shape {originalShape.ToString()}, but input data is of length {typeValueCount}.");

// Fill in the unknown dimensions.
var l = new int[originalShape.NDim];
for (int ishape = 0; ishape < originalShape.NDim; ishape++)
l[ishape] = originalShape[ishape] == -1 ? (int)d : originalShape[ishape];
var originalShapeDims = originalShape.dims;
var originalShapeNdim = originalShape.ndim;
var l = new int[originalShapeNdim];
for (int ishape = 0; ishape < originalShapeNdim; ishape++)
l[ishape] = originalShapeDims[ishape] == -1 ? (int)d : originalShapeDims[ishape];
Copy link

@yaeldekel yaeldekel Sep 15, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

(int)d [](start = 74, length = 6)

Should this cast also be done outside the loop? #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

This is not an addition in this PR. If required, I can do this change

In reply to: 324450404 [](ancestors = 324450404)

_fullySpecifiedShapes[i] = new TensorShape(l);
}

if (_parent._addBatchDimensionInput)
{
var l = new int[_fullySpecifiedShapes[i].NDim + 1];
var l = new int[_fullySpecifiedShapes[i].ndim + 1];
l[0] = 1;
for (int ishape = 1; ishape < l.Length; ishape++)
l[ishape] = _fullySpecifiedShapes[i][ishape - 1];
l[ishape] = _fullySpecifiedShapes[i].dims[ishape - 1];
_fullySpecifiedShapes[i] = new TensorShape(l);
}
}
Expand Down Expand Up @@ -857,7 +860,7 @@ protected override Delegate MakeGetter(DataViewRow input, int iinfo, Func<int, b
return Utils.MarshalInvoke(MakeGetter<int>, type, input, iinfo, srcTensorGetters, activeOutputColNames, outputCache);
}

private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[] srcTensorGetters, string[] activeOutputColNames, OutputCache outputCache)
private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[] srcTensorGetters, string[] activeOutputColNames, OutputCache outputCache) where T: unmanaged
Copy link
Member

@eerhardt eerhardt Sep 13, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

where T: unmanaged

Can this ever be used with a string type? If so, this is going to fail at runtime since we use reflection to call this method. #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

I have made a separate function for when T is string to avoid this failure.

In reply to: 324317321 [](ancestors = 324317321)

{
Host.AssertValue(input);

Expand All @@ -868,7 +871,7 @@ private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[
UpdateCacheIfNeeded(input.Position, srcTensorGetters, activeOutputColNames, outputCache);

var tensor = outputCache.Outputs[_parent._outputs[iinfo]];
dst = tensor.Data<T>()[0];
dst = tensor.ToArray<T>()[0];
};
return valuegetter;
}
Expand All @@ -881,7 +884,7 @@ private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[
UpdateCacheIfNeeded(input.Position, srcTensorGetters, activeOutputColNames, outputCache);

var tensor = outputCache.Outputs[_parent._outputs[iinfo]];
var tensorSize = tensor.TensorShape.Dimensions.Where(x => x > 0).Aggregate((x, y) => x * y);
var tensorSize = tensor.TensorShape.dims.Where(x => x > 0).Aggregate((x, y) => x * y);

var editor = VBufferEditor.Create(ref dst, (int)tensorSize);
DnnUtils.FetchStringData(tensor, editor.Values);
Expand All @@ -896,11 +899,11 @@ private Delegate MakeGetter<T>(DataViewRow input, int iinfo, ITensorValueGetter[
UpdateCacheIfNeeded(input.Position, srcTensorGetters, activeOutputColNames, outputCache);

var tensor = outputCache.Outputs[_parent._outputs[iinfo]];
var tensorSize = tensor.TensorShape.Dimensions.Where(x => x > 0).Aggregate((x, y) => x * y);
var tensorSize = tensor.TensorShape.dims.Where(x => x > 0).Aggregate((x, y) => x * y);

var editor = VBufferEditor.Create(ref dst, (int)tensorSize);

DnnUtils.FetchData<T>(tensor.Data<T>(), editor.Values);
DnnUtils.FetchData<T>(tensor.ToArray<T>(), editor.Values);
dst = editor.Commit();
};
return valuegetter;
Expand All @@ -912,6 +915,7 @@ private void UpdateCacheIfNeeded(long position, ITensorValueGetter[] srcTensorGe
{
if (outputCache.Position != position)
{
_parent._session.graph.as_default();
Runner runner = new Runner(_parent._session);

// Feed the inputs.
Expand Down Expand Up @@ -972,12 +976,12 @@ public TensorValueGetter(DataViewRow input, int colIndex, TensorShape tfShape, b
_tfShape = tfShape;
long size = 0;
_position = 0;
if (tfShape.Dimensions.Length != 0)
if (tfShape.dims.Length != 0)
{
size = 1;
foreach (var dim in tfShape.Dimensions)
foreach (var dim in tfShape.dims)
size *= dim;
_dims = _tfShape.Dimensions.Select(x => (long)x).ToArray();
_dims = _tfShape.dims.Select(x => (long)x).ToArray();
}
if (keyType)
_bufferedDataLong = new long[size];
Expand All @@ -993,13 +997,13 @@ public Tensor GetTensor()
if (_keyType)
{
var tensor = new Tensor(new[] { Convert.ToInt64(scalar) - 1 });
tensor.SetShape(_tfShape);
tensor.set_shape(_tfShape);
return tensor;
}
else
{
var tensor = new Tensor(new[] { scalar });
tensor.SetShape(_tfShape);
tensor.set_shape(_tfShape);
return tensor;
}
}
Expand Down Expand Up @@ -1095,16 +1099,16 @@ public TensorValueGetterVec(DataViewRow input, int colIndex, TensorShape tfShape

long size = 0;
_position = 0;
if (tfShape.Dimensions.Length != 0)
if (tfShape.dims.Length != 0)
{
size = 1;
foreach (var dim in tfShape.Dimensions)
foreach (var dim in tfShape.dims)
size *= dim;
}
_bufferedData = new T[size];
_bufferedDataSize = size;
if (_tfShape.Dimensions != null)
_dims = _tfShape.Dimensions.Select(x => (long)x).ToArray();
if (_tfShape.dims != null)
_dims = _tfShape.dims.Select(x => (long)x).ToArray();
}

public Tensor GetTensor()
Expand Down
11 changes: 5 additions & 6 deletions src/Microsoft.ML.Dnn/DnnUtils.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -11,7 +11,7 @@
using Microsoft.ML.Data;
using Microsoft.ML.Runtime;
using Tensorflow;
using static Tensorflow.Python;
using static Tensorflow.Binding;

namespace Microsoft.ML.Transforms.Dnn
{
Expand Down Expand Up @@ -92,11 +92,10 @@ internal static Session LoadTFSession(IExceptionContext ectx, byte[] modelBytes,

internal static Graph LoadMetaGraph(string path)
{
return tf_with(tf.Graph().as_default(), graph =>
{
tf.train.import_meta_graph(path);
return graph;
});
var graph = new Graph();
graph = graph.as_default();
tf.train.import_meta_graph(path);
Copy link

@yaeldekel yaeldekel Sep 15, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

tf.train.import_meta_graph(path); [](start = 12, length = 33)

It seems like there is no connection between graph and tf, does this line do anything to graph? #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

This creates a new graph and the 'as_default()' sets this as the tf default graph https://github.com/SciSharp/TensorFlow.NET/blob/41432600c8263cf972b13930c31479ffb412fa65/src/TensorFlowNET.Core/Graphs/Graph.cs#L134. The 'graph' is where the meta graph is loaded and returned.

In reply to: 324450631 [](ancestors = 324450631)

return graph;
Copy link
Member

@codemzs codemzs Sep 11, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

remove new lines. #Resolved

}

internal static Session LoadTFSessionByModelFilePath(IExceptionContext ectx, string modelFile, bool metaGraph = false)
Expand Down
41 changes: 19 additions & 22 deletions src/Microsoft.ML.Dnn/ImageClassificationTransform.cs
View file
Open in desktop
Original file line number Diff line number Diff line change
Expand Up @@ -20,7 +20,7 @@
using static Microsoft.ML.Data.TextLoader;
using static Microsoft.ML.Transforms.Dnn.DnnUtils;
using static Microsoft.ML.Transforms.ImageClassificationEstimator;
using static Tensorflow.Python;
using static Tensorflow.Binding;
using Architecture = Microsoft.ML.Transforms.ImageClassificationEstimator.Architecture;

[assembly: LoadableClass(ImageClassificationTransformer.Summary, typeof(IDataTransform), typeof(ImageClassificationTransformer),
Expand Down Expand Up @@ -170,7 +170,7 @@ private void CheckTrainingParameters(ImageClassificationEstimator.Options option
Host.CheckNonWhiteSpace(options.LabelColumn, nameof(options.LabelColumn));
Host.CheckNonWhiteSpace(options.TensorFlowLabel, nameof(options.TensorFlowLabel));

if (_session.graph.OperationByName(options.TensorFlowLabel) == null)
if (_session.graph.OperationByName(_labelTensor.name.Split(':')[0]) == null)
Copy link

@yaeldekel yaeldekel Sep 15, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

_labelTensor.name.Split(':')[0] [](start = 47, length = 31)

If this is the name of the operation you are searching for in the graph, then the error message should mention the correct name. In this case, the options argument is not needed here. #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

the options argument is not used anymore. The _labelTensor has the correct name that the error message would need

In reply to: 324450724 [](ancestors = 324450724)

throw Host.ExceptParam(nameof(options.TensorFlowLabel), $"'{options.TensorFlowLabel}' does not exist in the model");
}

Expand Down Expand Up @@ -243,7 +243,7 @@ private void CacheFeaturizedImagesToDisk(IDataView input, string labelColumnName
var imageTensor = imageProcessor.ProcessImage(imagePathStr);
runner.AddInput(imageTensor, 0);
var featurizedImage = runner.Run()[0]; // Reuse memory?
writer.WriteLine(label - 1 + "," + string.Join(",", featurizedImage.Data<float>()));
writer.WriteLine(label - 1 + "," + string.Join(",", featurizedImage.ToArray<float>()));
eerhardt marked this conversation as resolved.
Show resolved Hide resolved
featurizedImage.Dispose();
imageTensor.Dispose();
metrics.Bottleneck.Index++;
Expand Down Expand Up @@ -378,8 +378,8 @@ private void TrainAndEvaluateClassificationLayer(string trainBottleneckFilePath,
.AddInput(new Tensor(labelBatchPtr, labelTensorShape, TF_DataType.TF_INT64, labelBatchSizeInBytes), 1)
.Run();

metrics.Train.Accuracy += outputTensors[0].Data<float>()[0];
metrics.Train.CrossEntropy += outputTensors[1].Data<float>()[0];
metrics.Train.Accuracy += outputTensors[0].ToArray<float>()[0];
metrics.Train.CrossEntropy += outputTensors[1].ToArray<float>()[0];

outputTensors[0].Dispose();
outputTensors[1].Dispose();
Expand Down Expand Up @@ -429,7 +429,7 @@ private void TrainAndEvaluateClassificationLayer(string trainBottleneckFilePath,
.AddInput(new Tensor(labelBatchPtr, labelTensorShape, TF_DataType.TF_INT64, labelBatchSizeInBytes), 1)
.Run();

metrics.Train.Accuracy += outputTensors[0].Data<float>()[0];
metrics.Train.Accuracy += outputTensors[0].ToArray<float>()[0];
metrics.Train.BatchProcessedCount += 1;
batchIndex = 0;

Expand Down Expand Up @@ -458,17 +458,13 @@ private void TrainAndEvaluateClassificationLayer(string trainBottleneckFilePath,
Tensor evaluationStep = null;
Tensor prediction = null;
Tensor bottleneckTensor = evalGraph.OperationByName(_bottleneckOperationName);

tf_with(evalGraph.as_default(), graph =>
{
var (_, _, groundTruthInput, finalTensor) = AddFinalRetrainOps(classCount, options.LabelColumn,
evalGraph.as_default();
Copy link

@yaeldekel yaeldekel Sep 15, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

evalGraph.as_default(); [](start = 12, length = 23)

Does this mutate evalGraph? In DnnUtils.cs line 96 you wrote graph = graph.as_default();. #Resolved

Copy link

Choose a reason for hiding this comment

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

In line 842 in this file there is also a call to as_default() without assignment.

In reply to: 324450796 [](ancestors = 324450796)

Copy link
Contributor Author

Choose a reason for hiding this comment

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

This sets the tensorflow default graph to the evalGraph. This does not mutate the evalGraph.

In reply to: 324450796 [](ancestors = 324450796)

var (_, _, groundTruthInput, finalTensor) = AddFinalRetrainOps(classCount, options.LabelColumn,
options.ScoreColumnName, options.LearningRate, bottleneckTensor, false);

tf.train.Saver().restore(evalSess, _checkpointPath);
(evaluationStep, prediction) = AddEvaluationStep(finalTensor, groundTruthInput);
(_jpegData, _resizedImage) = AddJpegDecoding(299, 299, 3);
});

return (evalSess, _labelTensor, evaluationStep, prediction);
}

Expand Down Expand Up @@ -530,14 +526,15 @@ private void VariableSummaries(RefVariable var)
private (Operation, Tensor, Tensor, Tensor) AddFinalRetrainOps(int classCount, string labelColumn,
string scoreColumnName, float learningRate, Tensor bottleneckTensor, bool isTraining)
{
var (batch_size, bottleneck_tensor_size) = (bottleneckTensor.TensorShape.Dimensions[0], bottleneckTensor.TensorShape.Dimensions[1]);
var bottleneckTensorDims = bottleneckTensor.TensorShape.dims;
var (batch_size, bottleneck_tensor_size) = (bottleneckTensorDims[0], bottleneckTensorDims[1]);
tf_with(tf.name_scope("input"), scope =>
{
if (isTraining)
{
_bottleneckInput = tf.placeholder_with_default(
bottleneckTensor,
shape: bottleneckTensor.TensorShape.Dimensions,
shape: bottleneckTensorDims ,
name: "BottleneckInputPlaceholder");
}

Expand All @@ -559,7 +556,8 @@ private void VariableSummaries(RefVariable var)
RefVariable layerBiases = null;
tf_with(tf.name_scope("biases"), delegate
{
layerBiases = tf.Variable(tf.zeros(classCount), name: "final_biases");
TensorShape shape = new TensorShape(classCount);
layerBiases = tf.Variable(tf.zeros(shape), name: "final_biases");
VariableSummaries(layerBiases);
});

Expand Down Expand Up @@ -599,11 +597,9 @@ private void AddTransferLearningLayer(string labelColumn,
string scoreColumnName, float learningRate, int classCount)
{
_bottleneckTensor = Graph.OperationByName(_bottleneckOperationName);
tf_with(Graph.as_default(), delegate
{
(_trainStep, _crossEntropy, _labelTensor, _softMaxTensor) =
(_trainStep, _crossEntropy, _labelTensor, _softMaxTensor) =
AddFinalRetrainOps(classCount, labelColumn, scoreColumnName, learningRate, _bottleneckTensor, true);
});

}

// Factory method for SignatureLoadDataTransform.
Expand Down Expand Up @@ -757,7 +753,7 @@ private protected override void SaveModel(ModelSaveContext ctx)
var buffer = _session.graph.ToGraphDef(status);
ctx.SaveBinaryStream("TFModel", w =>
{
w.WriteByteArray(buffer.Data);
w.WriteByteArray(buffer.MemoryBlock.ToArray());
Copy link
Member

@codemzs codemzs Sep 11, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

w.WriteByteArray(buffer.MemoryBlock.ToArray()); [](start = 16, length = 47)

can you please do a diff between the graph with your change and master and see if there are any changes? I know the graph size increased and I want to know why and by how much. #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

The size of the files are exactly the same between the master and my change, but doing a diff shows that the files differ. Are the trained values expected to be exactly the same?

In reply to: 323478331 [](ancestors = 323478331)

});
status.Check(true);
}
Expand Down Expand Up @@ -830,8 +826,8 @@ public void UpdateCacheIfNeeded()
_imagePathGetter(ref _imagePath);
var processedTensor = _imageProcessor.ProcessImage(_imagePath.ToString());
var outputTensor = _runner.AddInput(processedTensor, 0).Run();
ClassProbabilities = outputTensor[0].Data<float>();
PredictedLabel = (UInt32)outputTensor[1].Data<long>()[0];
ClassProbabilities = outputTensor[0].ToArray<float>();
PredictedLabel = (UInt32)outputTensor[1].ToArray<long>()[0];
outputTensor[0].Dispose();
outputTensor[1].Dispose();
processedTensor.Dispose();
Expand All @@ -843,6 +839,7 @@ public void UpdateCacheIfNeeded()
protected override Delegate MakeGetter(DataViewRow input, int iinfo, Func<int, bool> activeOutput, out Action disposer)
{
disposer = null;
_parent._session.graph.as_default();
Copy link
Member

@codemzs codemzs Sep 16, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

_parent._session.graph.as_default(); [](start = 16, length = 36)

not needed. #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Required for the Microsoft.ML.Scenarios.TensorFlowScenariosTests.TensorFlowImageClassification

In reply to: 324890773 [](ancestors = 324890773)

Copy link
Member

@codemzs codemzs Sep 17, 2019 •
edited by ashbhandare
Loading

Choose a reason for hiding this comment

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

_parent._session.graph.as_default(); [](start = 16, length = 36)

is this needed? #Resolved

Copy link
Contributor Author

Choose a reason for hiding this comment

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

Yes, test Microsoft.ML.Scenarios.TensorFlowScenariosTests.TensorFlowImageClassification is failing without this

In reply to: 324938993 [](ancestors = 324938993)

Host.AssertValue(input);
var cache = new OutputCache(input, _parent);

Expand Down
Loading