add Deep Learning digital recognizer

Author: asm-jaime

MLTests/MLTests.csproj

@@ -1,10 +1,15 @@
 <Project Sdk="Microsoft.NET.Sdk">
 
   <PropertyGroup>
-    <OutputType>Library</OutputType>
+    <OutputType>WinExe</OutputType>
     <TargetFramework>net6.0</TargetFramework>
     <ImplicitUsings>enable</ImplicitUsings>
     <Nullable>enable</Nullable>
+    <PlatformTarget>AnyCPU</PlatformTarget>
+  </PropertyGroup>
+
+  <PropertyGroup Condition="'$(Configuration)|$(Platform)'=='Release|AnyCPU'">
+    <Optimize>False</Optimize>
   </PropertyGroup>
 
   <ItemGroup>

MLTests/NDLDigitalRecognizer.cs

@@ -0,0 +1,79 @@
+using Keras.Datasets;
+using Numpy;
+
+namespace MLTests
+{
+    public class NDLDigitalRecognizer
+    {
+        private NDarray _weights01;
+        private NDarray _weights12;
+        private string _dataPath;
+
+        private static NDarray Relu(NDarray layer) => layer * (layer > 0);
+        private static NDarray ReluToDerivative(NDarray layer) => np.where(layer > 0, np.array(1), np.array(0));
+
+        public NDLDigitalRecognizer(string dataPath)
+        {
+            _weights01 = np.array(0f);
+            _weights12 = np.array(0f);
+            _dataPath = dataPath;
+        }
+
+        public void TrainOnMNISTData()
+        {
+            var ((x_train, y_train), (x_test, y_test)) = MNIST.LoadData(_dataPath);
+
+            var (images, labels) = (x_train["0:1000"].reshape(1000, 28 * 28) / 255, y_train["0:1000"]);
+            var one_hot_labels = np.zeros(labels.len, 10);
+            for (var id = 0; id < labels.len; id++)
+            {
+                var label = labels[id];
+                one_hot_labels[id][label] = np.array(1);
+            }
+            labels = one_hot_labels;
+
+            np.random.seed(1);
+
+            var (alpha, iterations, hidden_size, pixels_per_image, num_labels) = (0.005f, 30, 40, 784, 10);
+
+            _weights01 = 0.2f * np.random.rand(pixels_per_image, hidden_size) - 0.1f;
+            _weights12 = 0.2f * np.random.rand(hidden_size, num_labels) - 0.1f;
+
+            for (var i = 0; i < iterations; i++)
+            {
+                for (var imageId = 0; imageId < images.len; ++imageId)
+                {
+                    var layer_0 = images[$"{imageId}:{imageId + 1}"];
+                    var layer_1 = Relu(np.dot(layer_0, _weights01));
+                    var layer_2 = np.dot(layer_1, _weights12);
+
+                    var layer_2_delta = labels[$"{imageId}:{imageId + 1}"] - layer_2;
+                    var layer_1_delta = layer_2_delta.dot(_weights12.T) * ReluToDerivative(layer_1);
+
+                    _weights12 += alpha * layer_1.T.dot(layer_2_delta);
+                    _weights01 += alpha * layer_0.T.dot(layer_1_delta);
+                }
+            }
+        }
+
+        public int PredictNumberFromImageTest(int testId)
+        {
+            var ((x_train, y_train), (x_test, y_test)) = MNIST.LoadData(_dataPath);
+
+            var test_images = x_test.reshape(x_test.len, 28 * 28) / 255;
+            var test_labels = np.zeros(y_test.len, 10);
+            for (var id = 0; id < y_test.len; id++)
+            {
+                var test = y_test[id];
+                test_labels[id][test] = np.array(1);
+            }
+
+            var layer_0 = test_images[$"{testId}:{testId + 1}"];
+            var layer_1 = Relu(np.dot(layer_0, _weights01));
+            var layer_2 = np.dot(layer_1, _weights12);
+            var result = layer_2.GetData<double>();
+            var max = result.Max();
+            return result.Select((e, index) => new {e, index}).OrderBy(e => e.e).Last().index;
+        }
+    }
+}

MLTests/Program.cs

@@ -1,4 +1,8 @@
 using MLTests;
 
 //var vector = ImageToVector.GetVectorOnPath(@"./data/0/1.jpg");
-Console.WriteLine("ML");
+var recognizer = new NDLDigitalRecognizer(Path.GetFullPath(@"D:/projects.active/LearnML/data/mnist.npz"));
+recognizer.TrainOnMNISTData();
+var result = recognizer.PredictNumberFromImageTest(0);
+Console.WriteLine(result);
+Console.ReadLine();