Added Adafactor (#415)

* Added adafactor

* Added Adafactor and ran pre-commit

* modified operations

* Added docstrings

* Switched two ops to fix a bug

* added underscore for internal functions and removed the plus sign in the last return statment

* Removed parameter rms from the optimizer state because its not needed

* Added simple MNIST test for Adafactor and temporary training log

* remove test files

* nits in docs

* comment nit

---------

Co-authored-by: Awni Hannun <awni@apple.com>

docs/src/python/optimizers.rst

@@ -40,6 +40,7 @@ model's parameters and the **optimizer state**.
    SGD
    RMSprop
    Adagrad
+   Adafactor
    AdaDelta
    Adam
    AdamW

python/mlx/optimizers.py

@@ -1,7 +1,7 @@
 # Copyright © 2023 Apple Inc.
 
 import math
-from typing import List
+from typing import List, Optional, Tuple
 
 import mlx.core as mx
 from mlx.utils import tree_map
@@ -76,7 +76,7 @@ class Optimizer:
 
 
 class SGD(Optimizer):
-    r"""Stochastic gradient descent optimizer.
+    r"""The stochastic gradient descent optimizer.
 
     Updates a parameter :math:`w` with a gradient :math:`g` as follows
 
@@ -141,7 +141,7 @@ class SGD(Optimizer):
 
 
 class RMSprop(Optimizer):
-    r"""Implementation of the RMSprop optimizer [1].
+    r"""The RMSprop optimizer [1].
 
     [1]: Tieleman, T. and Hinton, G. 2012. Lecture 6.5-rmsprop, coursera: Neural networks for machine learning
 
@@ -190,7 +190,7 @@ class RMSprop(Optimizer):
 
 
 class Adagrad(Optimizer):
-    r"""Implementation of the Adagrad optimizer [1].
+    r"""The Adagrad optimizer [1].
 
     Our Adagrad implementation follows the original paper. In detail,
 
@@ -235,7 +235,7 @@ class Adagrad(Optimizer):
 
 
 class AdaDelta(Optimizer):
-    r"""Implementation of the AdaDelta optimizer with learning rate[1].
+    r"""The AdaDelta optimizer with a learning rate [1].
 
     Our AdaDelta implementation follows the original paper. In detail,
 
@@ -294,7 +294,7 @@ class AdaDelta(Optimizer):
 
 
 class Adam(Optimizer):
-    r"""Implementation of the Adam optimizer [1].
+    r"""The Adam optimizer [1].
 
     Our Adam implementation follows the original paper and omits the bias
     correction in the first and second moment estimates. In detail,
@@ -346,7 +346,7 @@ class Adam(Optimizer):
 
 
 class AdamW(Adam):
-    r"""Implementation of the AdamW optimizer [1].
+    r"""The AdamW optimizer [1].
 
     Following the above convention, in contrast with [1], we do not use bias
     correction in the first and second moments for AdamW. We update the weights
@@ -395,8 +395,7 @@ class AdamW(Adam):
 
 
 class Adamax(Adam):
-    r"""Implementation of the Adamax optimizer. It is a variant of Adam based
+    r"""The Adamax optimizer, a variant of Adam based on the infinity norm [1].
-    on the infinity norm [1].
 
     Our Adam implementation follows the original paper and omits the bias
     correction in the first and second moment estimates. In detail,
@@ -449,7 +448,7 @@ class Adamax(Adam):
 
 
 class Lion(Optimizer):
-    r"""Implementation of the Lion optimizer [1].
+    r"""The Lion optimizer [1].
 
     Since updates are computed through the sign operation, they tend to
     have larger norm than for other optimizers such as SGD and Adam.
@@ -502,3 +501,139 @@ class Lion(Optimizer):
         if weight_decay > 0:
             parameter = (1 - lr * weight_decay) * parameter
         return parameter - lr * mx.sign(c)
+
+
+class Adafactor(Optimizer):
+    r"""The Adafactor optimizer.
+
+    Our Adafactor implementation follows the original paper: `Adafactor:
+    Adaptive Learning Rates with Sublinear Memory Cost
+    <https://arxiv.org/abs/1804.04235>`_
+
+    Args:
+        learning_rate (float, optional): The learning rate. Default: ``None``.
+        eps (tuple(float, float), optional): The first term :math:`\epsilon_1`
+            added to the square of the gradients to improve numerical
+            stability and the second term :math:`\epsilon_2` is used for
+            parameter scaling if ``parameter_scale`` is set to ``True``.
+            Default: ``(1e-30, 1e-3)``.
+        clip_threshold (float, optional): Clips the unscaled update at
+            ``clip_threshold``. Default: ``1.0``.
+        decay_rate (float, optional): Coefficient for the running average
+            of the squared gradient. Default: ``-0.8``.
+        beta_1 (float, optional): If set to a value bigger than zero
+            then first moment will be used. Default: ``None``.
+        weight_decay (float, optional): The weight decay :math:`\lambda`.
+            Default: ``0.0``.
+        scale_parameter (bool, optional): If set to ``True`` the learning rate
+            will be scaled by :math:`\max(\epsilon_1, \text{RMS}(w_{t-1}))`.
+            Default: ``True``.
+        relative_step (bool, optional): If set to ``True`` the ``learning_rate``
+            will be ignored and relative step size will be computed.
+            Default: ``True``.
+        warmup_init (bool, optional): If set to ``True`` then the relative
+            step size will be calculated by the current step. Default:
+            ``False``.
+    """
+
+    def __init__(
+        self,
+        learning_rate: Optional[float] = None,
+        eps: Tuple[float, float] = (1e-30, 1e-3),
+        clip_threshold: float = 1.0,
+        decay_rate: float = -0.8,
+        beta_1: Optional[float] = None,
+        weight_decay: float = 0.0,
+        scale_parameter: bool = True,
+        relative_step: bool = True,
+        warmup_init: bool = False,
+    ):
+        super().__init__()
+        self.learning_rate = learning_rate
+        self.eps = eps
+        self.clip_threshold = clip_threshold
+        self.decay_rate = decay_rate
+        self.beta_1 = beta_1
+        self.weight_decay = weight_decay
+        self.scale_parameter = scale_parameter
+        self.relative_step = relative_step
+        self.warmup_init = warmup_init
+
+    def _compute_rms(self, inputs):
+        return mx.sqrt(mx.mean(mx.square(inputs)))
+
+    def _compute_learning_rate(self, step, parameter_rms):
+        relative_step_size = self.learning_rate
+        if self.relative_step:
+            min_step = 1e-6 * step if self.warmup_init else 1e-2
+            relative_step_size = min(min_step, 1 / math.sqrt(step))
+
+        parameter_scale = 1.0
+        if self.scale_parameter:
+            parameter_scale = mx.maximum(self.eps[1], parameter_rms)
+        return parameter_scale * relative_step_size
+
+    def _approximate_exp_moving_avg(self, exp_avg_sq_row, exp_avg_sq_col):
+        r_factor = mx.rsqrt(
+            exp_avg_sq_row / mx.mean(exp_avg_sq_row, axis=-1, keepdims=True)
+        )
+        c_factor = mx.rsqrt(exp_avg_sq_col)
+        return mx.matmul(
+            mx.expand_dims(r_factor, axis=-1), mx.expand_dims(c_factor, axis=0)
+        )
+
+    def apply_single(
+        self, gradient: mx.array, parameter: mx.array, state: OptimizerState
+    ):
+        """Performs the Adafactor parameter and state update."""
+        gradient_shape = gradient.shape
+        factored = len(gradient_shape) >= 2
+        step = state.get("step", 0) + 1
+        state["step"] = step
+        use_first_moment = self.beta_1 is not None
+
+        parameter_rms = self._compute_rms(parameter)
+        learning_rate = self._compute_learning_rate(step, parameter_rms)
+        beta_2 = 1.0 - math.pow(step, self.decay_rate)
+        update = mx.square(gradient) + self.eps[0]
+
+        if factored:
+            exp_avg_sq_row = state.get(
+                "exp_avg_sq_row", mx.zeros(gradient_shape[:-1], dtype=gradient.dtype)
+            )
+            exp_avg_sq_col = state.get(
+                "exp_avg_sq_col",
+                mx.zeros(
+                    gradient_shape[:-2] + gradient_shape[-1:], dtype=gradient.dtype
+                ),
+            )
+            exp_avg_sq_row = (beta_2 * exp_avg_sq_row) + (
+                (1 - beta_2) * mx.mean(update, axis=-1)
+            )
+            exp_avg_sq_col = (beta_2 * exp_avg_sq_col) + (
+                (1 - beta_2) * mx.mean(update, axis=-2)
+            )
+            state["exp_avg_sq_row"] = exp_avg_sq_row
+            state["exp_avg_sq_col"] = exp_avg_sq_col
+            update = self._approximate_exp_moving_avg(exp_avg_sq_row, exp_avg_sq_col)
+            update = update * gradient
+        else:
+            exp_avg_sq = state.get("exp_avg_sq", mx.zeros_like(gradient))
+            exp_avg_sq = (beta_2 * exp_avg_sq) + ((1 - beta_2) * update)
+            state["exp_avg_sq"] = exp_avg_sq
+            update = mx.rsqrt(exp_avg_sq) * gradient
+
+        update = update / mx.maximum(
+            1.0, self._compute_rms(update) / self.clip_threshold
+        )
+        update = learning_rate * update
+
+        if use_first_moment:
+            exp_avg = state.get("exp_avg", mx.zeros_like(gradient))
+            exp_avg = (self.beta_1 * exp_avg) + ((1 - self.beta_1) * update)
+            state["exp_avg"] = exp_avg
+            update = exp_avg
+
+        if self.weight_decay != 0:
+            parameter += parameter * (-self.weight_decay * learning_rate)
+        return parameter - update

python/tests/test_optimizers.py

@@ -39,6 +39,24 @@ class TestOptimizers(mlx_tests.MLXTestCase):
             all_equal = all(v for _, v in mlx.utils.tree_flatten(equal_shape))
             self.assertTrue(all_equal)
 
+    def test_adafactor(self):
+        x = mx.zeros((5, 5))
+        grad = mx.ones_like(x)
+        optimizer = opt.Adafactor()
+        for _ in range(2):
+            xp = optimizer.apply_single(grad, x, optimizer.state)
+            self.assertEqual(xp.dtype, x.dtype)
+            self.assertEqual(xp.shape, x.shape)
+
+        x = mx.zeros((5, 5), mx.float16)
+        grad = mx.ones_like(x)
+        optimizer = opt.Adafactor()
+        for _ in range(2):
+            xp = optimizer.apply_single(grad, x, optimizer.state)
+            self.assertEqual(xp.dtype, x.dtype)
+            self.assertEqual(xp.shape, x.shape)
+        self.assertEqual(optimizer.state["step"], 2)
+
 
 if __name__ == "__main__":
     unittest.main()