Activations LeakyReLU / PReLU / Softplus / Mish (#109)

* Leaky_relu / prelu / softplus / mish

* added tests

* updated bench

* remove torch refs, add init to PReLU

* added arvix reference to mish

* added missing docs

python/mlx/nn/layers/__init__.py

@@ -7,6 +7,8 @@ from mlx.nn.layers.activations import (
     SELU,
     LeakyReLU,
     LogSigmoid,
+    Mish,
+    PReLU,
     ReLU,
     ReLU6,
     SiLU,
@@ -19,6 +21,8 @@ from mlx.nn.layers.activations import (
     gelu_fast_approx,
     leaky_relu,
     log_sigmoid,
+    mish,
+    prelu,
     relu,
     relu6,
     selu,

python/mlx/nn/layers/activations.py

@@ -176,6 +176,33 @@ def selu(x):
     See also :func:`elu`.
     """
     return elu(x, 1.67326) * 1.0507
+def prelu(x: mx.array, alpha: mx.array) -> mx.array:
+    r"""Applies the element-wise function:
+
+    .. math::
+        \text{PReLU}(x) = \max(0,x) + a * \min(0,x)
+
+    Here :math:`a` is an array.
+    """
+    return mx.maximum(0, x) + alpha * mx.minimum(0, x)
+
+
+def mish(x: mx.array) -> mx.array:
+    r"""Applies the Mish function, element-wise.
+    Mish: A Self Regularized Non-Monotonic Neural Activation Function.
+
+    Reference: https://arxiv.org/abs/1908.08681
+
+    .. math::
+        \text{Mish}(x) = x * \text{Tanh}(\text{Softplus}(x))
+
+    """
+    return x * mx.tanh(softplus(x))
+
+
+@_make_activation_module(mish)
+class Mish(Module):
+    pass
 
 
 @_make_activation_module(relu)
@@ -257,6 +284,15 @@ class LogSigmoid(Module):
     pass
 
 
+class PReLU(Module):
+    def __init__(self, num_parameters=1, init=0.25):
+        super().__init__()
+        self.weight = mx.full([num_parameters], init)
+
+    def __call__(self, x: mx.array):
+        return prelu(x, self.weight)
+
+
 class GELU(Module):
     r"""Applies the Gaussian Error Linear Units.
 

python/tests/mlx_tests.py

@@ -2,8 +2,10 @@
 
 import os
 import unittest
+from typing import Callable, List, Tuple
 
 import mlx.core as mx
+import numpy as np
 
 
 class MLXTestCase(unittest.TestCase):
@@ -16,3 +18,16 @@ class MLXTestCase(unittest.TestCase):
 
     def tearDown(self):
         mx.set_default_device(self.default)
+
+    def assertEqualArray(
+        self,
+        args: List[mx.array | float | int],
+        mlx_func: Callable[..., mx.array],
+        expected: mx.array,
+        atol=1e-2,
+        rtol=1e-2,
+    ):
+        mx_res = mlx_func(*args)
+        assert tuple(mx_res.shape) == tuple(expected.shape), "shape mismatch"
+        assert mx_res.dtype == expected.dtype, "dtype mismatch"
+        np.testing.assert_allclose(mx_res, expected, rtol=rtol, atol=atol)

python/tests/test_nn.py

@@ -449,31 +449,19 @@ class TestNN(mlx_tests.MLXTestCase):
         self.assertEqual(y.shape, [3])
         self.assertEqual(y.dtype, mx.float32)
 
-    def test_step_activation(self):
-        x = mx.arange(-3, 4)
-        expected = mx.array([0, 0, 0, 0, 0, 1, 1])
-        y = nn.Step()(x)
-        self.assertTrue(mx.array_equal(y, expected))
-
-        y = nn.Step(2)(x)
-        expected = mx.array([0, 0, 0, 0, 0, 0, 1])
-        self.assertTrue(mx.array_equal(y, expected))
-
-    def test_selu(self):
-        x = mx.arange(-3, 4)
-        expected = mx.array(
-            [
-                -1.670563817024231,
-                -1.5201621055603027,
-                -1.1113275289535522,
-                0.0,
-                1.0506999492645264,
-                2.1013998985290527,
-                3.152099847793579,
-            ]
+    def test_prelu(self):
+        self.assertEqualArray(
+            [mx.array([1.0, -1.0, 0.0, 0.5])],
+            nn.PReLU(),
+            mx.array([1.0, -0.25, 0.0, 0.5]),
+        )
+
+    def test_mish(self):
+        self.assertEqualArray(
+            [mx.array([1.0, -1.0, 0.0, 0.5])],
+            nn.Mish(),
+            mx.array([0.8651, -0.3034, 0.0000, 0.3752]),
         )
-        y = nn.SELU()(x)
-        self.assertTrue(mx.allclose(y, expected))
 
 
 if __name__ == "__main__":