Support SigmoidFocalCrossEntropy, better for imbalanced multi-class task.

Author: Oceania2018

src/TensorFlowNET.Core/Keras/Losses/ILossesApi.cs

@@ -38,4 +38,19 @@ ILossFunc Huber(string reduction = null,
 
     ILossFunc LogCosh(string reduction = null, 
         string name = null);
+
+    /// <summary>
+    /// Implements the focal loss function.
+    /// </summary>
+    /// <param name="from_logits"></param>
+    /// <param name="alpha"></param>
+    /// <param name="gamma"></param>
+    /// <param name="reduction"></param>
+    /// <param name="name"></param>
+    /// <returns></returns>
+    ILossFunc SigmoidFocalCrossEntropy(bool from_logits = false,
+        float alpha = 0.25f,
+        float gamma = 2.0f,
+        string reduction = "none",
+        string name = "sigmoid_focal_crossentropy");
 }

src/TensorFlowNET.Keras/Losses/LossesApi.cs

@@ -37,5 +37,16 @@ public ILossFunc Huber(string reduction = null, string name = null, Tensor delta
 
         public ILossFunc LogCosh(string reduction = null, string name = null)
             => new LogCosh(reduction: reduction, name: name);
+
+        public ILossFunc SigmoidFocalCrossEntropy(bool from_logits = false, 
+                float alpha = 0.25F, 
+                float gamma = 2, 
+                string reduction = "none", 
+                string name = "sigmoid_focal_crossentropy")
+            => new SigmoidFocalCrossEntropy(from_logits: from_logits, 
+                alpha: alpha, 
+                gamma: gamma, 
+                reduction: reduction, 
+                name: name);
     }
 }

src/TensorFlowNET.Keras/Losses/SigmoidFocalCrossEntropy.cs

@@ -0,0 +1,48 @@
+using static HDF.PInvoke.H5L.info_t;
+
+namespace Tensorflow.Keras.Losses;
+
+public class SigmoidFocalCrossEntropy : LossFunctionWrapper, ILossFunc
+{
+    float _alpha;
+    float _gamma;
+
+    public SigmoidFocalCrossEntropy(bool from_logits = false,
+        float alpha = 0.25f,
+        float gamma = 2.0f,
+        string reduction = "none",
+        string name = "sigmoid_focal_crossentropy") :
+        base(reduction: reduction,
+             name: name, 
+             from_logits: from_logits)
+    {
+        _alpha = alpha;
+        _gamma = gamma;
+    }
+
+
+    public override Tensor Apply(Tensor y_true, Tensor y_pred, bool from_logits = false, int axis = -1)
+    {
+        y_true = tf.cast(y_true, dtype: y_pred.dtype);
+        var ce = keras.backend.binary_crossentropy(y_true, y_pred, from_logits: from_logits);
+        var pred_prob = from_logits ? tf.sigmoid(y_pred) : y_pred;
+
+        var p_t = (y_true * pred_prob) + ((1f - y_true) * (1f - pred_prob));
+        Tensor alpha_factor = constant_op.constant(1.0f);
+        Tensor modulating_factor = constant_op.constant(1.0f);
+
+        if(_alpha > 0)
+        {
+            var alpha = tf.cast(constant_op.constant(_alpha), dtype: y_true.dtype);
+            alpha_factor = y_true * alpha + (1f - y_true) * (1f - alpha);
+        }
+
+        if (_gamma > 0)
+        {
+            var gamma = tf.cast(constant_op.constant(_gamma), dtype: y_true.dtype);
+            modulating_factor = tf.pow(1f - p_t, gamma);
+        }
+
+        return tf.reduce_sum(alpha_factor * modulating_factor * ce, axis = -1);
+    }
+}

test/TensorFlowNET.Keras.UnitTest/Losses/LossesTest.cs

@@ -5,6 +5,7 @@
 using System.Text;
 using System.Threading.Tasks;
 using Tensorflow;
+using Tensorflow.NumPy;
 using TensorFlowNET.Keras.UnitTest;
 using static Tensorflow.Binding;
 using static Tensorflow.KerasApi;
@@ -48,4 +49,17 @@ public void BinaryCrossentropy()
         loss = bce.Call(y_true, y_pred);
         Assert.AreEqual(new float[] { 0.23515666f, 1.4957594f}, loss.numpy());
     }
+
+    /// <summary>
+    /// https://www.tensorflow.org/addons/api_docs/python/tfa/losses/SigmoidFocalCrossEntropy
+    /// </summary>
+    [TestMethod]
+    public void SigmoidFocalCrossEntropy()
+    {
+        var y_true = np.expand_dims(np.array(new[] { 1.0f, 1.0f, 0 }));
+        var y_pred = np.expand_dims(np.array(new[] { 0.97f, 0.91f, 0.03f }));
+        var bce = tf.keras.losses.SigmoidFocalCrossEntropy();
+        var loss = bce.Call(y_true, y_pred);
+        Assert.AreEqual(new[] { 6.8532745e-06f, 1.909787e-04f, 2.0559824e-05f }, loss.numpy());
+    }
 }