Add two-variable scenario in Tensor shape inference for TensorflowTransform

build/Dependencies.props

@@ -55,7 +55,7 @@
     <MicrosoftExtensionsTestPackageVersion>3.0.1</MicrosoftExtensionsTestPackageVersion>
     <MicrosoftMLTestDatabasesPackageVersion>0.0.6-test</MicrosoftMLTestDatabasesPackageVersion>
     <MicrosoftMLTestModelsPackageVersion>0.0.6-test</MicrosoftMLTestModelsPackageVersion>
-    <MicrosoftMLTensorFlowTestModelsVersion>0.0.11-test</MicrosoftMLTensorFlowTestModelsVersion>
+    <MicrosoftMLTensorFlowTestModelsVersion>0.0.12-test</MicrosoftMLTensorFlowTestModelsVersion>
     <MicrosoftMLOnnxTestModelsVersion>0.0.6-test</MicrosoftMLOnnxTestModelsVersion>
     <SystemDataSqlClientVersion>4.6.1</SystemDataSqlClientVersion>
     <XunitCombinatorialVersion>1.2.7</XunitCombinatorialVersion>

src/Microsoft.ML.TensorFlow/TensorflowTransform.cs

@@ -529,6 +529,18 @@ public Mapper(TensorFlowTransformer parent, DataViewSchema inputSchema) :
                         if (typeValueCount % valCount != 0)
                             throw Contracts.Except($"Input shape mismatch: Input '{_parent.Inputs[i]}' has shape {originalShape.ToString()}, but input data is of length {typeValueCount}.");
 
+                        // This cover the 2-variable senario e.g. [?, ?, ?, C] where we can assume typeDims provides the information of [W, H, C]
+                        // The shape will become [?, W, H, C]
+                        var originalShapeDims = originalShape.dims;
+                        var originalShapeNdim = originalShape.ndim;
+                        if (numOfUnkDim == 3 && colTypeDims.Length == 3 && originalShapeNdim == numOfUnkDim + 1 && originalShapeDims[1] == -1)
+                        {
+                            originalShapeDims[1] = colTypeDims[0];
+                            originalShapeDims[2] = colTypeDims[1];
+                            valCount *= originalShapeDims[1] * originalShapeDims[2];
+                            numOfUnkDim -= 2;
+                        }
+
                         // If the shape is multi-dimensional, we should be able to create the length of the vector by plugging
                         // in a single value for the unknown shapes. For example, if the shape is [?,?,3], then there should exist a value
                         // d such that d*d*3 is equal to the length of the input column.
@@ -537,8 +549,6 @@ public Mapper(TensorFlowTransformer 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 originalShapeNdim = originalShape.ndim;
-                        var originalShapeDims = originalShape.dims;
                         var l = new int[originalShapeNdim];
                         for (int ishape = 0; ishape < originalShapeNdim; ishape++)
                             l[ishape] = originalShapeDims[ishape] == -1 ? (int)d : originalShapeDims[ishape];

test/Microsoft.ML.Tests/ScenariosWithDirectInstantiation/TensorflowTests.cs

@@ -1900,5 +1900,40 @@ private static string GetTemporaryDirectory()
             Directory.CreateDirectory(tempDirectory);
             return tempDirectory;
         }
+
+        [TensorFlowFact]
+        public void TensorflowPlaceholderShapeInferenceTest()
+        {
+            //frozen_model_variadic_input_shape.pb is modified by frozen_model.pb 
+            //the shape of placeholder is changed from [?, w, h, c] to [?, ?, ?, c]
+            string modelLocation = "cifar_model/frozen_model_variadic_input_shape.pb";
+
+            int imageHeight = 32;
+            int imageWidth = 32;
+            string dataFile = GetDataPath("images/images.tsv");
+            string imageFolder = Path.GetDirectoryName(dataFile);
+
+            IDataView data = _mlContext.Data.LoadFromTextFile(dataFile, new[] {
+                new TextLoader.Column("imagePath", DataKind.String, 0),
+                new TextLoader.Column("name", DataKind.String, 1)
+            });
+
+            Tensorflow.TensorShape[] tfInputShape;
+
+            using (var tfModel = _mlContext.Model.LoadTensorFlowModel(modelLocation))
+            {
+                var pipeline = _mlContext.Transforms.LoadImages("Input", imageFolder, "imagePath")
+                    .Append(_mlContext.Transforms.ResizeImages("Input", imageHeight, imageWidth))
+                    .Append(_mlContext.Transforms.ExtractPixels("Input", interleavePixelColors: true))
+                    .Append(tfModel.ScoreTensorFlowModel("Output", "Input"));
+
+                var transformer = pipeline.Fit(data);
+
+                tfInputShape = transformer.LastTransformer.TFInputShapes;
+            }
+
+            Assert.Equal(imageHeight, tfInputShape.ElementAt(0)[1].dims[0]);
+            Assert.Equal(imageWidth, tfInputShape.ElementAt(0)[2].dims[0]);
+        }
     }
 }