Add metrics of Accuracy and CosineSimilarity.

Author: Oceania2018

src/TensorFlowNET.Core/APIs/tf.linalg.cs

@@ -54,6 +54,12 @@ public Tensor inv(Tensor input, bool adjoint = false, string name = null)
             public Tensor global_norm(Tensor[] t_list, string name = null)
                 => clip_ops.global_norm(t_list, name: name);
 
+            public Tensor l2_normalize(Tensor x,
+                    int axis = 0,
+                    float epsilon = 1e-12f,
+                    string name = null)
+                => nn_impl.l2_normalize(x, axis: axis, epsilon: constant_op.constant(epsilon), name: name);
+
             public Tensor lstsq(Tensor matrix, Tensor rhs,
                 NDArray l2_regularizer = null, bool fast = true, string name = null)
                 => ops.matrix_solve_ls(matrix, rhs, l2_regularizer: l2_regularizer, fast: fast, name: name);

src/TensorFlowNET.Core/Keras/Metrics/IMetricsApi.cs

@@ -31,6 +31,13 @@ Tensor categorical_crossentropy(Tensor y_true, Tensor y_pred,
     /// <returns></returns>
     Tensor top_k_categorical_accuracy(Tensor y_true, Tensor y_pred, int k = 5);
 
+    /// <summary>
+    /// Calculates how often predictions equal labels.
+    /// </summary>
+    /// <returns></returns>
+    IMetricFunc Accuracy(string name = "accuracy",
+        TF_DataType dtype = TF_DataType.TF_FLOAT);
+
     /// <summary>
     /// Calculates how often predictions match binary labels.
     /// </summary>
@@ -56,6 +63,14 @@ IMetricFunc CategoricalCrossentropy(string name = "categorical_crossentropy",
     IMetricFunc CategoricalAccuracy(string name = "categorical_accuracy", 
         TF_DataType dtype = TF_DataType.TF_FLOAT);
 
+    /// <summary>
+    /// Computes the cosine similarity between the labels and predictions.
+    /// </summary>
+    /// <returns></returns>
+    IMetricFunc CosineSimilarity(string name = "cosine_similarity",
+        TF_DataType dtype = TF_DataType.TF_FLOAT,
+        Axis? axis = null);
+
     /// <summary>
     /// Computes how often targets are in the top K predictions.
     /// </summary>

src/TensorFlowNET.Keras/Metrics/Accuracy.cs

@@ -0,0 +1,11 @@
+namespace Tensorflow.Keras.Metrics;
+
+public class Accuracy : MeanMetricWrapper
+{
+    public Accuracy(string name = "accuracy", TF_DataType dtype = TF_DataType.TF_FLOAT) 
+        : base((yt, yp) => metrics_utils.accuracy(yt, yp), 
+            name: name, 
+            dtype: dtype)
+    {
+    }
+}

src/TensorFlowNET.Keras/Metrics/CosineSimilarity.cs

@@ -0,0 +1,11 @@
+namespace Tensorflow.Keras.Metrics;
+
+public class CosineSimilarity : MeanMetricWrapper
+{
+    public CosineSimilarity(string name = "cosine_similarity", TF_DataType dtype = TF_DataType.TF_FLOAT, Axis? axis = null) 
+        : base((yt, yp) => metrics_utils.cosine_similarity(yt, yp, axis: axis ?? -1), 
+            name: name, 
+            dtype: dtype)
+    {
+    }
+}

src/TensorFlowNET.Keras/Metrics/MetricsApi.cs

@@ -71,6 +71,9 @@ public Tensor top_k_categorical_accuracy(Tensor y_true, Tensor y_pred, int k = 5
             );
         }
 
+        public IMetricFunc Accuracy(string name = "accuracy", TF_DataType dtype = TF_DataType.TF_FLOAT)
+            => new Accuracy(name: name, dtype: dtype);
+
         public IMetricFunc BinaryAccuracy(string name = "binary_accuracy", TF_DataType dtype = TF_DataType.TF_FLOAT, float threshold = 5)
             => new BinaryAccuracy();
 
@@ -80,6 +83,9 @@ public IMetricFunc CategoricalAccuracy(string name = "categorical_accuracy", TF_
         public IMetricFunc CategoricalCrossentropy(string name = "categorical_crossentropy", TF_DataType dtype = TF_DataType.TF_FLOAT, bool from_logits = false, float label_smoothing = 0, Axis? axis = null)
             => new CategoricalCrossentropy();
 
+        public IMetricFunc CosineSimilarity(string name = "cosine_similarity", TF_DataType dtype = TF_DataType.TF_FLOAT, Axis? axis = null)
+            => new CosineSimilarity(name: name, dtype: dtype, axis: axis ?? -1);
+
         public IMetricFunc TopKCategoricalAccuracy(int k = 5, string name = "top_k_categorical_accuracy", TF_DataType dtype = TF_DataType.TF_FLOAT)
             => new TopKCategoricalAccuracy(k: k, name: name, dtype: dtype);
 

src/TensorFlowNET.Keras/Metrics/metrics_utils.cs

@@ -4,12 +4,26 @@ namespace Tensorflow.Keras.Metrics;
 
 public class metrics_utils
 {
+    public static Tensor accuracy(Tensor y_true, Tensor y_pred)
+    {
+        if (y_true.dtype != y_pred.dtype)
+            y_pred = tf.cast(y_pred, y_true.dtype);
+        return tf.cast(tf.equal(y_true, y_pred), keras.backend.floatx());
+    }
+
     public static Tensor binary_matches(Tensor y_true, Tensor y_pred, float threshold = 0.5f)
     {
         y_pred = tf.cast(y_pred > threshold, y_pred.dtype);
         return tf.cast(tf.equal(y_true, y_pred), keras.backend.floatx());
     }
 
+    public static Tensor cosine_similarity(Tensor y_true, Tensor y_pred, Axis? axis = null)
+    {
+        y_true = tf.linalg.l2_normalize(y_true, axis: axis ?? -1);
+        y_pred = tf.linalg.l2_normalize(y_pred, axis: axis ?? -1);
+        return tf.reduce_sum(y_true * y_pred, axis: axis ?? -1);
+    }
+
     /// <summary>
     /// Creates float Tensor, 1.0 for label-prediction match, 0.0 for mismatch.
     /// </summary>

test/TensorFlowNET.Keras.UnitTest/Metrics/MetricsTest.cs

@@ -14,6 +14,26 @@ namespace TensorFlowNET.Keras.UnitTest;
 [TestClass]
 public class MetricsTest : EagerModeTestBase
 {
+    /// <summary>
+    /// https://www.tensorflow.org/api_docs/python/tf/keras/metrics/Accuracy
+    /// </summary>
+    [TestMethod]
+    public void Accuracy()
+    {
+        var y_true = np.array(new[,] { { 1 }, { 2 }, { 3 }, { 4 } });
+        var y_pred = np.array(new[,] { { 0f }, { 2f }, { 3f }, { 4f } });
+        var m = tf.keras.metrics.Accuracy();
+        m.update_state(y_true, y_pred);
+        var r = m.result().numpy();
+        Assert.AreEqual(r, 0.75f);
+
+        m.reset_states();
+        var weights = np.array(new[] { 1f, 1f, 0f, 0f });
+        m.update_state(y_true, y_pred, sample_weight: weights);
+        r = m.result().numpy();
+        Assert.AreEqual(r, 0.5f);
+    }
+
     /// <summary>
     /// https://www.tensorflow.org/api_docs/python/tf/keras/metrics/BinaryAccuracy
     /// </summary>
@@ -23,14 +43,14 @@ public void BinaryAccuracy()
         var y_true = np.array(new[,] { { 1 }, { 1 },{ 0 }, { 0 } });
         var y_pred = np.array(new[,] { { 0.98f }, { 1f }, { 0f }, { 0.6f } });
         var m = tf.keras.metrics.BinaryAccuracy();
-        /*m.update_state(y_true, y_pred);
+        m.update_state(y_true, y_pred);
         var r = m.result().numpy();
         Assert.AreEqual(r, 0.75f);
 
-        m.reset_states();*/
+        m.reset_states();
         var weights = np.array(new[] { 1f, 0f, 0f, 1f });
         m.update_state(y_true, y_pred, sample_weight: weights);
-        var r = m.result().numpy();
+        r = m.result().numpy();
         Assert.AreEqual(r, 0.5f);
     }
 
@@ -74,6 +94,26 @@ public void CategoricalCrossentropy()
         Assert.AreEqual(r, 1.6271976f);
     }
 
+    /// <summary>
+    /// https://www.tensorflow.org/api_docs/python/tf/keras/metrics/CosineSimilarity
+    /// </summary>
+    [TestMethod]
+    public void CosineSimilarity()
+    {
+        var y_true = np.array(new[,] { { 0, 1 }, { 1, 1 } });
+        var y_pred = np.array(new[,] { { 1f, 0f }, { 1f, 1f } });
+        var m = tf.keras.metrics.CosineSimilarity(axis: 1);
+        m.update_state(y_true, y_pred);
+        var r = m.result().numpy();
+        Assert.AreEqual(r, 0.49999997f);
+
+        m.reset_states();
+        var weights = np.array(new[] { 0.3f, 0.7f });
+        m.update_state(y_true, y_pred, sample_weight: weights);
+        r = m.result().numpy();
+        Assert.AreEqual(r, 0.6999999f);
+    }
+
     /// <summary>
     /// https://www.tensorflow.org/api_docs/python/tf/keras/metrics/TopKCategoricalAccuracy
     /// </summary>