public static void ToArray <T>(this Tensor tensor, ref T[] array) where T : unmanaged
{
Utils.EnsureSize(ref array, (int)tensor.size, (int)tensor.size, false);
var span = new Span <T>(array);
CopyTo(tensor, span);
}
protected override Delegate MakeGetter(DataViewRow input, int iinfo, Func <int, bool> activeOutput, out Action disposer)
{
disposer = null;
Host.AssertValue(input);
var outputCache = new OutputCache();
var activeOutputColNames = _parent.Outputs.Where((x, i) => activeOutput(i)).ToArray();
var type = Tf2MlNetType(_parent.TFOutputTypes[iinfo]).RawType;
Host.Assert(type == _parent.OutputTypes[iinfo].GetItemType().RawType);
var srcTensorGetters = GetTensorValueGetters(input, _inputColIndices, _isInputVector, _parent.TFInputTypes, _fullySpecifiedShapes);
return(Utils.MarshalInvoke(MakeGetter <int>, type, input, iinfo, srcTensorGetters, activeOutputColNames, outputCache));
}
internal static DataViewSchema GetModelSchema(IExceptionContext ectx, Graph graph, string opType = null)
{
var schemaBuilder = new DataViewSchema.Builder();
foreach (Operation op in graph)
{
if (opType != null && opType != op.OpType)
{
continue;
}
var tfType = op.OutputType(0);
// Determine element type in Tensorflow tensor. For example, a vector of floats may get NumberType.R4 here.
var mlType = Tf2MlNetTypeOrNull(tfType);
// If the type is not supported in ML.NET then we cannot represent it as a column in an Schema.
// We also cannot output it with a TensorFlowTransform, so we skip it.
// Furthermore, operators which have NumOutputs <= 0 needs to be filtered.
// The 'GetTensorShape' method crashes TensorFlow runtime
// (https://github.com/dotnet/machinelearning/issues/2156) when the operator has no outputs.
if (mlType == null || op.NumOutputs <= 0)
{
continue;
}
// There can be at most two metadata fields.
// 1. The first field always presents. Its value is this operator's type. For example,
// if an output is produced by an "Softmax" operator, the value of this field should be "Softmax".
// 2. The second field stores operators whose outputs are consumed by this operator. In other words,
// these values are names of some upstream operators which should be evaluated before executing
// the current operator. It's possible that one operator doesn't need any input, so this field
// can be missing.
var metadataBuilder = new DataViewSchema.Annotations.Builder();
// Create the first metadata field.
metadataBuilder.Add(TensorflowOperatorTypeKind, TextDataViewType.Instance, (ref ReadOnlyMemory <char> value) => value = op.OpType.AsMemory());
if (op.NumInputs > 0)
{
// Put upstream operators' names to an array (type: VBuffer) of string (type: ReadOnlyMemory<char>).
VBuffer <ReadOnlyMemory <char> > upstreamOperatorNames = default;
var bufferEditor = VBufferEditor.Create(ref upstreamOperatorNames, op.NumInputs);
for (int i = 0; i < op.NumInputs; ++i)
{
bufferEditor.Values[i] = op.inputs[i].op.name.AsMemory();
}
upstreamOperatorNames = bufferEditor.Commit(); // Used in metadata's getter.
// Create the second metadata field.
metadataBuilder.Add(TensorflowUpstreamOperatorsKind, new VectorDataViewType(TextDataViewType.Instance, op.NumInputs),
(ref VBuffer <ReadOnlyMemory <char> > value) => { upstreamOperatorNames.CopyTo(ref value); });
}
// Construct the final ML.NET type of a Tensorflow variable.
var tensorShape = op.output.TensorShape.dims;
if (tensorShape == null)
{
// primitive column type
schemaBuilder.AddColumn(op.name, mlType, metadataBuilder.ToAnnotations());
}
else
{
// vector column type
DataViewType columnType = new VectorDataViewType(mlType);
if (!(Utils.Size(tensorShape) == 1 && tensorShape[0] <= 0) &&
(Utils.Size(tensorShape) > 0 && tensorShape.Skip(1).All(x => x > 0)))
{
columnType = new VectorDataViewType(mlType, tensorShape[0] > 0 ? tensorShape : tensorShape.Skip(1).ToArray());
}
schemaBuilder.AddColumn(op.name, columnType, metadataBuilder.ToAnnotations());
}
}
return(schemaBuilder.ToSchema());
}
