Add support for Mobilenet v2 in Image Classification transfer learning

docs/samples/Microsoft.ML.Samples/Dynamic/ImageClassification/ResnetV2101TransferLearningEarlyStopping.cs

@@ -65,7 +65,7 @@ public static void Example()
                 var pipeline = mlContext.Transforms.LoadImages("Image", fullImagesetFolderPath, false, "ImagePath") // false indicates we want the image as a VBuffer<byte>
                     .Append(mlContext.Model.ImageClassification(
                         "Image", "Label",
-                        // Just by changing/selecting InceptionV3 here instead of 
+                        // Just by changing/selecting InceptionV3/MobilenetV2 here instead of 
                         // ResnetV2101 you can try a different architecture/pre-trained 
                         // model. 
                         arch: ImageClassificationEstimator.Architecture.ResnetV2101,

docs/samples/Microsoft.ML.Samples/Dynamic/ImageClassification/ResnetV2101TransferLearningTrainTestSplit.cs

@@ -62,7 +62,7 @@ public static void Example()
 
                 var pipeline = mlContext.Model.ImageClassification(
                         "Image", "Label",
-                        // Just by changing/selecting InceptionV3 here instead of 
+                        // Just by changing/selecting InceptionV3/MobilenetV2 here instead of 
                         // ResnetV2101 you can try a different architecture/
                         // pre-trained model. 
                         arch: ImageClassificationEstimator.Architecture.ResnetV2101,

src/Microsoft.ML.Dnn/DnnCatalog.cs

@@ -144,7 +144,7 @@ public static ImageClassificationEstimator ImageClassification(
         {
             var options = new ImageClassificationEstimator.Options()
             {
-                ModelLocation = arch == Architecture.ResnetV2101 ? @"resnet_v2_101_299.meta" : @"InceptionV3.meta",
+                ModelLocation = ModelLocation[arch],
                 InputColumns = new[] { featuresColumnName },
                 OutputColumns = new[] { scoreColumnName, predictedLabelColumnName },
                 LabelColumn = labelColumnName,
@@ -194,6 +194,14 @@ public static ImageClassificationEstimator ImageClassification(
                         client.DownloadFile(new Uri($"{baseGitPath}"), @"resnet_v2_101_299.meta");
                     }
                 }
+                else if(options.Arch == Architecture.MobilenetV2)
+                {
+                    var baseGitPath = @"https://tlcresources.blob.core.windows.net/image/MobileNetV2TensorFlow/mobilenet_v2.meta";
+                    using (WebClient client = new WebClient())
+                    {
+                        client.DownloadFile(new Uri($"{baseGitPath}"), @"mobilenet_v2.meta");
+                    }
+                }
             }
 
             var env = CatalogUtils.GetEnvironment(catalog);

src/Microsoft.ML.Dnn/ImageClassificationTransform.cs

@@ -588,7 +588,8 @@ private void TrainAndEvaluateClassificationLayer(string trainBottleneckFilePath,
 
                 tf.train.Saver().restore(evalSess, _checkpointPath);
                 (evaluationStep, prediction) = AddEvaluationStep(finalTensor, groundTruthInput);
-                (_jpegData, _resizedImage) = AddJpegDecoding(299, 299, 3);
+            var imageSize = ImageClassificationEstimator.ImagePreprocessingSize[options.Arch];
+            (_jpegData, _resizedImage) = AddJpegDecoding(imageSize.Item1, imageSize.Item2, 3);
             return (evalSess, _labelTensor, evaluationStep, prediction);
         }
 
@@ -827,12 +828,18 @@ internal ImageClassificationTransformer(IHostEnvironment env, Session session, s
                 _bottleneckOperationName = "module_apply_default/hub_output/feature_vector/SpatialSqueeze";
                 _inputTensorName = "Placeholder";
             }
+            else if(arch == ImageClassificationEstimator.Architecture.MobilenetV2)
+            {
+                _bottleneckOperationName = "import/MobilenetV2/Logits/Squeeze";
+                _inputTensorName = "import/input";
+            }
 
             _outputs = new[] { scoreColumnName, predictedLabelColumnName };
 
             if (loadModel == false)
             {
-                (_jpegData, _resizedImage) = AddJpegDecoding(299, 299, 3);
+                var imageSize = ImageClassificationEstimator.ImagePreprocessingSize[arch];
+                (_jpegData, _resizedImage) = AddJpegDecoding(imageSize.Item1, imageSize.Item2, 3);
                 _jpegDataTensorName = _jpegData.name;
                 _resizedImageTensorName = _resizedImage.name;
 
@@ -1080,7 +1087,28 @@ public sealed class ImageClassificationEstimator : IEstimator<ImageClassificatio
         public enum Architecture
         {
             ResnetV2101,
-            InceptionV3
+            InceptionV3,
+            MobilenetV2
+        };
+
+        /// <summary>
+        /// Dictionary mapping model architecture to model location.
+        /// </summary>
+        internal static IReadOnlyDictionary<Architecture, string> ModelLocation = new Dictionary<Architecture, string>
+        {
+            { Architecture.ResnetV2101, @"resnet_v2_101_299.meta" },
+            { Architecture.InceptionV3, @"InceptionV3.meta" },
+            { Architecture.MobilenetV2, @"mobilenet_v2.meta" }
+        };
+
+        /// <summary>
+        /// Dictionary mapping model architecture to image input size supported.
+        /// </summary>
+        internal static IReadOnlyDictionary<Architecture, Tuple<int,int>> ImagePreprocessingSize = new Dictionary<Architecture, Tuple<int,int>>
+        {
+            { Architecture.ResnetV2101, new Tuple<int, int>(299,299) },
+            { Architecture.InceptionV3, new Tuple<int, int>(299,299) },
+            { Architecture.MobilenetV2, new Tuple<int, int>(224,224) }
         };
 
         /// <summary>

test/Microsoft.ML.TestFramework/Attributes/TensorflowTheoryAttribute.cs

@@ -0,0 +1,23 @@
+// Licensed to the .NET Foundation under one or more agreements.
+// The .NET Foundation licenses this file to you under the MIT license.
+// See the LICENSE file in the project root for more information.
+
+using System;
+namespace Microsoft.ML.TestFramework.Attributes
+{
+    /// <summary>
+    /// A theory for tests requiring TensorFlow.
+    /// </summary>
+    public sealed class TensorFlowTheoryAttribute : EnvironmentSpecificTheoryAttribute
+    {
+        public TensorFlowTheoryAttribute() : base("TensorFlow is 64-bit only")
+        {
+        }
+
+        /// <inheritdoc />
+        protected override bool IsEnvironmentSupported()
+        {
+            return Environment.Is64BitProcess;
+        }
+    }
+}

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

@@ -1207,8 +1207,10 @@ public void TensorFlowStringTest()
             Assert.Equal(string.Join(" ", input.B).Replace("/", " "), textOutput.BOut[0]);
         }
 
-        [TensorFlowFact]
-        public void TensorFlowImageClassification()
+        [TensorFlowTheory]
+        [InlineData(ImageClassificationEstimator.Architecture.ResnetV2101)]
+        [InlineData(ImageClassificationEstimator.Architecture.MobilenetV2)]
+        public void TensorFlowImageClassification(ImageClassificationEstimator.Architecture arch)
         {
             string assetsRelativePath = @"assets";
             string assetsPath = GetAbsolutePath(assetsRelativePath);
@@ -1249,10 +1251,10 @@ public void TensorFlowImageClassification()
             var pipeline = mlContext.Transforms.LoadImages("Image", fullImagesetFolderPath, false, "ImagePath") // false indicates we want the image as a VBuffer<byte>
                 .Append(mlContext.Model.ImageClassification(
                     "Image", "Label",
-                    // Just by changing/selecting InceptionV3 here instead of 
+                    // Just by changing/selecting InceptionV3/MobilenetV2 here instead of 
                     // ResnetV2101 you can try a different architecture/pre-trained 
                     // model. 
-                    arch: ImageClassificationEstimator.Architecture.ResnetV2101,
+                    arch: arch,
                     epoch: 50,
                     batchSize: 10,
                     learningRate: 0.01f,
@@ -1384,7 +1386,7 @@ public void TensorFlowImageClassificationEarlyStoppingIncreasing()
             var pipeline = mlContext.Transforms.LoadImages("Image", fullImagesetFolderPath, false, "ImagePath") // false indicates we want the image as a VBuffer<byte>
                 .Append(mlContext.Model.ImageClassification(
                     "Image", "Label",
-                    // Just by changing/selecting InceptionV3 here instead of 
+                    // Just by changing/selecting InceptionV3/MobilenetV2 here instead of 
                     // ResnetV2101 you can try a different architecture/pre-trained 
                     // model. 
                     arch: ImageClassificationEstimator.Architecture.ResnetV2101,
@@ -1473,7 +1475,7 @@ public void TensorFlowImageClassificationEarlyStoppingDecreasing()
             var pipeline = mlContext.Transforms.LoadImages("Image", fullImagesetFolderPath, false, "ImagePath") // false indicates we want the image as a VBuffer<byte>
                 .Append(mlContext.Model.ImageClassification(
                     "Image", "Label",
-                    // Just by changing/selecting InceptionV3 here instead of 
+                    // Just by changing/selecting InceptionV3/MobilenetV2 here instead of 
                     // ResnetV2101 you can try a different architecture/pre-trained 
                     // model. 
                     arch: ImageClassificationEstimator.Architecture.ResnetV2101,