Merge pull request #1263 from ASolomatin/master

fix: Support for training a multi-input model using a dataset.
SciSharp · Jul 2, 2024 · 6a2d7e1 · 6a2d7e1
2 parents 7fb73cd + 93dda17
commit 6a2d7e1
Show file tree

Hide file tree

Showing 3 changed files with 107 additions and 2 deletions.
diff --git a/src/TensorFlowNET.Keras/Engine/Model.Evaluate.cs b/src/TensorFlowNET.Keras/Engine/Model.Evaluate.cs
@@ -112,7 +112,19 @@ public Dictionary<string, float> evaluate(IDatasetV2 x, int verbose = 1, bool is
                 Steps = data_handler.Inferredsteps
             });
 
-            return evaluate(data_handler, callbacks, is_val, test_function);
+            Func<DataHandler, OwnedIterator, Dictionary<string, float>> testFunction;
+
+            if (data_handler.DataAdapter.GetDataset().structure.Length > 2 ||
+                data_handler.DataAdapter.GetDataset().FirstInputTensorCount > 1)
+            {
+                testFunction = test_step_multi_inputs_function;
+            }
+            else
+            {
+                testFunction = test_function;
+            }
+
+            return evaluate(data_handler, callbacks, is_val, testFunction);
         }
 
         /// <summary>

diff --git a/src/TensorFlowNET.Keras/Engine/Model.Fit.cs b/src/TensorFlowNET.Keras/Engine/Model.Fit.cs
@@ -179,9 +179,20 @@ public ICallback fit(IDatasetV2 dataset,
                 StepsPerExecution = _steps_per_execution
             });
 
+            Func<DataHandler, OwnedIterator, Dictionary<string, float>> trainStepFunction;
+
+            if (data_handler.DataAdapter.GetDataset().structure.Length > 2 ||
+                data_handler.DataAdapter.GetDataset().FirstInputTensorCount > 1)
+            {
+                trainStepFunction = train_step_multi_inputs_function;
+            }
+            else
+            {
+                trainStepFunction = train_step_function;
+            }
 
             return FitInternal(data_handler, epochs, validation_step, verbose, callbacks, validation_data: validation_data,
-                    train_step_func: train_step_function);
+                    train_step_func: trainStepFunction);
         }
 
         History FitInternal(DataHandler data_handler, int epochs, int validation_step, int verbose, List<ICallback> callbackList, IDatasetV2 validation_data, 

diff --git a/test/TensorFlowNET.Keras.UnitTest/MultiInputModelTest.cs b/test/TensorFlowNET.Keras.UnitTest/MultiInputModelTest.cs
@@ -2,6 +2,7 @@
 using System;
 using Tensorflow.Keras.Optimizers;
 using Tensorflow.NumPy;
+using static Tensorflow.Binding;
 using static Tensorflow.KerasApi;
 
 namespace Tensorflow.Keras.UnitTest
@@ -54,10 +55,91 @@ public void LeNetModel()
             var x = new NDArray[] { x1, x2 };
             model.fit(x, dataset.Train.Labels, batch_size: 8, epochs: 3);
 
+            x1 = x1["0:8"];
+            x2 = x1;
+
+            x = new NDArray[] { x1, x2 };
+            var y = dataset.Train.Labels["0:8"];
+            (model as Engine.Model).evaluate(x, y);
+
             x1 = np.ones((1, 28, 28, 1), TF_DataType.TF_FLOAT);
             x2 = np.zeros((1, 28, 28, 1), TF_DataType.TF_FLOAT);
             var pred = model.predict((x1, x2));
             Console.WriteLine(pred);
         }
+
+        [TestMethod]
+        public void LeNetModelDataset()
+        {
+            var inputs = keras.Input((28, 28, 1));
+            var conv1 = keras.layers.Conv2D(16, (3, 3), activation: "relu", padding: "same").Apply(inputs);
+            var pool1 = keras.layers.MaxPooling2D((2, 2), 2).Apply(conv1);
+            var conv2 = keras.layers.Conv2D(32, (3, 3), activation: "relu", padding: "same").Apply(pool1);
+            var pool2 = keras.layers.MaxPooling2D((2, 2), 2).Apply(conv2);
+            var flat1 = keras.layers.Flatten().Apply(pool2);
+
+            var inputs_2 = keras.Input((28, 28, 1));
+            var conv1_2 = keras.layers.Conv2D(16, (3, 3), activation: "relu", padding: "same").Apply(inputs_2);
+            var pool1_2 = keras.layers.MaxPooling2D((4, 4), 4).Apply(conv1_2);
+            var conv2_2 = keras.layers.Conv2D(32, (1, 1), activation: "relu", padding: "same").Apply(pool1_2);
+            var pool2_2 = keras.layers.MaxPooling2D((2, 2), 2).Apply(conv2_2);
+            var flat1_2 = keras.layers.Flatten().Apply(pool2_2);
+
+            var concat = keras.layers.Concatenate().Apply((flat1, flat1_2));
+            var dense1 = keras.layers.Dense(512, activation: "relu").Apply(concat);
+            var dense2 = keras.layers.Dense(128, activation: "relu").Apply(dense1);
+            var dense3 = keras.layers.Dense(10, activation: "relu").Apply(dense2);
+            var output = keras.layers.Softmax(-1).Apply(dense3);
+
+            var model = keras.Model((inputs, inputs_2), output);
+            model.summary();
+
+            var data_loader = new MnistModelLoader();
+
+            var dataset = data_loader.LoadAsync(new ModelLoadSetting
+            {
+                TrainDir = "mnist",
+                OneHot = false,
+                ValidationSize = 59900,
+            }).Result;
+
+            var loss = keras.losses.SparseCategoricalCrossentropy();
+            var optimizer = new Adam(0.001f);
+            model.compile(optimizer, loss, new string[] { "accuracy" });
+
+            NDArray x1 = np.reshape(dataset.Train.Data, (dataset.Train.Data.shape[0], 28, 28, 1));
+
+            var multiInputDataset = tf.data.Dataset.zip(
+                tf.data.Dataset.from_tensor_slices(x1),
+                tf.data.Dataset.from_tensor_slices(x1),
+                tf.data.Dataset.from_tensor_slices(dataset.Train.Labels)
+            ).batch(8);
+            multiInputDataset.FirstInputTensorCount = 2;
+
+            model.fit(multiInputDataset, epochs: 3);
+
+            x1 = x1["0:8"];
+
+            multiInputDataset = tf.data.Dataset.zip(
+                tf.data.Dataset.from_tensor_slices(x1),
+                tf.data.Dataset.from_tensor_slices(x1),
+                tf.data.Dataset.from_tensor_slices(dataset.Train.Labels["0:8"])
+            ).batch(8);
+            multiInputDataset.FirstInputTensorCount = 2;
+
+            (model as Engine.Model).evaluate(multiInputDataset);
+
+            x1 = np.ones((1, 28, 28, 1), TF_DataType.TF_FLOAT);
+            var x2 = np.zeros((1, 28, 28, 1), TF_DataType.TF_FLOAT);
+
+            multiInputDataset = tf.data.Dataset.zip(
+                tf.data.Dataset.from_tensor_slices(x1),
+                tf.data.Dataset.from_tensor_slices(x2)
+            ).batch(8);
+            multiInputDataset.FirstInputTensorCount = 2;
+
+            var pred = model.predict(multiInputDataset);
+            Console.WriteLine(pred);
+        }
     }
 }