Add log_sigmoid activation

keras/api/_tf_keras/keras/activations/__init__.py

@@ -18,6 +18,7 @@
 from keras.src.activations.activations import hard_tanh
 from keras.src.activations.activations import leaky_relu
 from keras.src.activations.activations import linear
+from keras.src.activations.activations import log_sigmoid
 from keras.src.activations.activations import log_softmax
 from keras.src.activations.activations import mish
 from keras.src.activations.activations import relu

keras/api/activations/__init__.py

@@ -18,6 +18,7 @@
 from keras.src.activations.activations import hard_tanh
 from keras.src.activations.activations import leaky_relu
 from keras.src.activations.activations import linear
+from keras.src.activations.activations import log_sigmoid
 from keras.src.activations.activations import log_softmax
 from keras.src.activations.activations import mish
 from keras.src.activations.activations import relu

keras/src/activations/__init__.py

@@ -10,6 +10,7 @@
 from keras.src.activations.activations import hard_tanh
 from keras.src.activations.activations import leaky_relu
 from keras.src.activations.activations import linear
+from keras.src.activations.activations import log_sigmoid
 from keras.src.activations.activations import log_softmax
 from keras.src.activations.activations import mish
 from keras.src.activations.activations import relu
@@ -47,6 +48,7 @@
     linear,
     mish,
     log_softmax,
+    log_sigmoid,
 }
 
 ALL_OBJECTS_DICT = {fn.__name__: fn for fn in ALL_OBJECTS}

keras/src/activations/activations.py

@@ -433,6 +433,19 @@ def hard_sigmoid(x):
     return ops.hard_sigmoid(x)
 
 
+@keras_export("keras.activations.log_sigmoid")
+def log_sigmoid(x):
+    """Logarithm of the sigmoid activation function.
+
+    It is defined as `f(x) = log(1 / (1 + exp(-x)))`.
+
+    Args:
+        x: Input tensor.
+
+    """
+    return ops.log_sigmoid(x)
+
+
 @keras_export(["keras.activations.hard_silu", "keras.activations.hard_swish"])
 def hard_silu(x):
     """Hard SiLU activation function, also known as Hard Swish.

keras/src/activations/activations_test.py

@@ -40,6 +40,10 @@ def _ref_hard_sigmoid(x):
     return z
 
 
+def _ref_log_sigmoid(x):
+    return -1 * _ref_softplus(-x)
+
+
 def _ref_hard_silu(x):
     return x * np.minimum(np.maximum(0.0, x + 3.0), 6.0) * (1.0 / 6.0)
 
@@ -337,6 +341,45 @@ def test_hard_sigmoid(self):
             result_positive_above_1, expected_positive_above_1, rtol=1e-05
         )
 
+    def test_log_sigmoid(self):
+        # Basic test for random values between 0 and 1
+        x = np.random.uniform(0, 1, (2, 5))
+        result = activations.log_sigmoid(x[np.newaxis, :])[0]
+        expected = np.vectorize(_ref_log_sigmoid)(x)
+        self.assertAllClose(result, expected, rtol=1e-05)
+
+        # Test with 1D array
+        x_1d = np.random.uniform(-10, 10, 5)
+        result_1d = activations.log_sigmoid(x_1d)
+        expected_1d = np.vectorize(_ref_log_sigmoid)(x_1d)
+        self.assertAllClose(result_1d, expected_1d, rtol=1e-05)
+
+        # Test with 3D array
+        x_3d = np.random.uniform(-10, 10, (3, 3, 3))
+        result_3d = activations.log_sigmoid(x_3d)
+        expected_3d = np.vectorize(_ref_log_sigmoid)(x_3d)
+        self.assertAllClose(result_3d, expected_3d, rtol=1e-05)
+
+        # Test large positive values
+        x_large_positive = np.random.uniform(10, 100, (2, 5))
+        result_large_positive = activations.log_sigmoid(x_large_positive)
+        expected_large_positive = np.vectorize(_ref_log_sigmoid)(
+            x_large_positive
+        )
+        self.assertAllClose(
+            result_large_positive, expected_large_positive, rtol=1e-05
+        )
+
+        # Test large negative values
+        x_large_negative = np.random.uniform(-100, -10, (2, 5))
+        result_large_negative = activations.log_sigmoid(x_large_negative)
+        expected_large_negative = np.vectorize(_ref_log_sigmoid)(
+            x_large_negative
+        )
+        self.assertAllClose(
+            result_large_negative, expected_large_negative, rtol=1e-05
+        )
+
     def test_hard_silu(self):
         # Basic test for random values between -3 and 3
         x = np.random.uniform(-3, 3, (2, 5)).astype("float32")