hc99's picture
Add files using upload-large-folder tool
fc0f7bd verified
# Copyright 2019 DeepMind Technologies Limited. All Rights Reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
# http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.
# ==============================================================================
"""Aliases for popular optimizers."""
import functools
from typing import Any, Callable, Optional, Union
import jax.numpy as jnp
from optax._src import base
from optax._src import clipping
from optax._src import combine
from optax._src import factorized
from optax._src import transform
from optax._src import wrappers
MaskOrFn = Optional[Union[Any, Callable[[base.Params], Any]]]
def adabelief(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-16,
eps_root: float = 1e-16) -> base.GradientTransformation:
r"""The AdaBelief optimizer.
AdaBelief is an adaptive learning rate optimizer that focuses on fast
convergence, generalization, and stability. It adapts the step size depending
on its "belief" in the gradient direction — the optimizer adaptively scales
the step size by the difference between the predicted and observed gradients.
AdaBelief is a modified version of :func:`optax.adam` and contains the same
number of parameters.
Let :math:`\alpha_t` represent the learning rate and :math:`\beta_1, \beta_2`,
:math:`\varepsilon`, :math:`\bar{\varepsilon}` represent the arguments
``b1``, ``b2``, ``eps`` and ``eps_root`` respectively. The learning rate is
indexed by :math:`t` since the learning rate may also be provided by a
schedule function.
The ``init`` function of this optimizer initializes an internal state
:math:`S_0 := (m_0, s_0) = (0, 0)`, representing initial estimates for the
first and second moments. In practice these values are stored as pytrees
containing all zeros, with the same shape as the model updates.
At step :math:`t`, the ``update`` function of this optimizer takes as
arguments the incoming gradients :math:`g_t` and optimizer state :math:`S_t`
and computes updates :math:`u_t` and new state :math:`S_{t+1}`. Thus, for
:math:`t > 0`, we have,
.. math::
\begin{align*}
m_t &\leftarrow \beta_1 \cdot m_{t-1} + (1-\beta_1) \cdot g_t \\
s_t &\leftarrow \beta_2 \cdot s_{t-1} + (1-\beta_2) \cdot (g_t - m_t)^2
+ \bar{\varepsilon} \\
\hat{m}_t &\leftarrow m_t / {(1-\beta_1^t)} \\
\hat{s}_t &\leftarrow s_t / {(1-\beta_2^t)} \\
u_t &\leftarrow -\alpha_t \cdot \hat{m}_t / \left(\sqrt{\hat{s}_t}
+ \varepsilon \right) \\
S_t &\leftarrow (m_t, s_t).
\end{align*}
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adabelief(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.38E+01
References:
Zhuang et al, 2020: https://arxiv.org/abs/2010.07468
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: Term added to the denominator to improve numerical stability.
eps_root: Term added to the second moment of the prediction error to
improve numerical stability. If backpropagating gradients through the
gradient transformation (e.g. for meta-learning), this must be non-zero.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_belief(b1=b1, b2=b2, eps=eps, eps_root=eps_root),
transform.scale_by_learning_rate(learning_rate),
)
def adadelta(
learning_rate: Optional[base.ScalarOrSchedule] = None,
rho: float = 0.9,
eps: float = 1e-6,
weight_decay: float = 0.0,
weight_decay_mask: MaskOrFn = None,
) -> base.GradientTransformation:
"""The Adadelta optimizer.
Adadelta is a stochastic gradient descent method that adapts learning rates
based on a moving window of gradient updates. Adadelta is a modification of
Adagrad.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> f = lambda x: jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adadelta(learning_rate=10.)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.36E+01
Objective function: 1.32E+01
Objective function: 1.29E+01
Objective function: 1.25E+01
Objective function: 1.21E+01
References:
[Matthew D. Zeiler, 2012](https://arxiv.org/pdf/1212.5701.pdf)
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
rho: A coefficient used for computing a running average of squared
gradients.
eps: Term added to the denominator to improve numerical stability.
weight_decay: Optional rate at which to decay weights.
weight_decay_mask: A tree with same structure as (or a prefix of) the params
PyTree, or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the transformation to, and `False` for those you want to skip.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.add_decayed_weights(weight_decay, mask=weight_decay_mask),
transform.scale_by_adadelta(rho=rho, eps=eps),
transform.scale_by_learning_rate(learning_rate),
)
def adafactor(
learning_rate: Optional[base.ScalarOrSchedule] = None,
min_dim_size_to_factor: int = 128,
decay_rate: float = 0.8,
decay_offset: int = 0,
multiply_by_parameter_scale: float = True,
clipping_threshold: Optional[float] = 1.0,
momentum: Optional[float] = None,
dtype_momentum: Any = jnp.float32,
weight_decay_rate: Optional[float] = None,
eps: float = 1e-30,
factored: bool = True,
weight_decay_mask: MaskOrFn = None,
) -> base.GradientTransformation:
"""The Adafactor optimizer.
Adafactor is an adaptive learning rate optimizer that focuses on fast
training of large scale neural networks. It saves memory by using a factored
estimate of the second order moments used to scale gradients.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adafactor(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.36E+01
References:
Shazeer and Stern, 2018: https://arxiv.org/abs/1804.04235
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
Note that the natural scale for Adafactor's LR is markedly different
from Adam, one doesn't use the 1/sqrt(hidden) correction for this optim
with attention-based models.
min_dim_size_to_factor: Only factor the statistics if two array dimensions
have at least this size.
decay_rate: Controls second-moment exponential decay schedule.
decay_offset: For fine-tuning, one may set this to the starting step
number of the fine-tuning phase.
multiply_by_parameter_scale: If True, then scale learning_rate by
parameter norm. If False, provided learning_rate is absolute step size.
clipping_threshold: Optional clipping threshold. Must be >= 1. If None,
clipping is disabled.
momentum: Optional value between 0 and 1, enables momentum and uses extra
memory if non-None! None by default.
dtype_momentum: Data type of momentum buffers.
weight_decay_rate: Optional rate at which to decay weights.
eps: Regularization constant for root mean squared gradient.
factored: Whether to use factored second-moment estimates.
weight_decay_mask: A tree with same structure as (or a prefix of)
the params PyTree, or a Callable that returns such a pytree given
the params/updates. The leaves should be booleans, `True`
for leaves/subtrees you want to apply the transformation to,
and `False` for those you want to skip.
Returns:
The corresponding `GradientTransformation`.
"""
# The core of the algorithm is a procedure for rescaling gradients
# by a factored estimate of the root mean squared gradients.
# This reduces memory compared to algorithms such as Adam or RmsProp,
# by not having to hold a separate estimate for each weight.
tx = [
factorized.scale_by_factored_rms(
factored, decay_rate, decay_offset, min_dim_size_to_factor, eps)]
# This basic rescaling is typically combined with one or more of the following
# transformation (all can be disabled via adafactor's constructor args).
if clipping_threshold is not None:
tx.append(clipping.clip_by_block_rms(clipping_threshold))
if learning_rate is not None:
tx.append(transform.scale_by_learning_rate(learning_rate, flip_sign=False))
if multiply_by_parameter_scale:
tx.append(transform.scale_by_param_block_rms())
if momentum is not None:
tx.append(
transform.ema(momentum, debias=False, accumulator_dtype=dtype_momentum))
if weight_decay_rate is not None:
tx.append(transform.add_decayed_weights(
weight_decay_rate, mask=weight_decay_mask))
# In gradient "descent" we follow the negative gradient.
tx.append(transform.scale(-1))
return combine.chain(*tx)
def adagrad(
learning_rate: base.ScalarOrSchedule,
initial_accumulator_value: float = 0.1,
eps: float = 1e-7
) -> base.GradientTransformation:
"""The Adagrad optimizer.
Adagrad is an algorithm for gradient based optimization that anneals the
learning rate for each parameter during the course of training.
.. warning::
Adagrad's main limit is the monotonic accumulation of squared
gradients in the denominator: since all terms are >0, the sum keeps growing
during training and the learning rate eventually becomes vanishingly small.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adagrad(learning_rate=1.0)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 5.01E+00
Objective function: 2.40E+00
Objective function: 1.25E+00
Objective function: 6.86E-01
Objective function: 3.85E-01
References:
Duchi et al, 2011: https://jmlr.org/papers/v12/duchi11a.html
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
initial_accumulator_value: Initial value for the accumulator.
eps: A small constant applied to denominator inside of the square root
(as in RMSProp) to avoid dividing by zero when rescaling.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_rss(
initial_accumulator_value=initial_accumulator_value, eps=eps),
transform.scale_by_learning_rate(learning_rate),
)
def adam(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-8,
eps_root: float = 0.0,
mu_dtype: Optional[Any] = None,
*,
nesterov: bool = False
) -> base.GradientTransformation:
r"""The Adam optimizer.
Adam is an SGD variant with gradient scaling adaptation. The scaling
used for each parameter is computed from estimates of first and second-order
moments of the gradients (using suitable exponential moving averages).
Let :math:`\alpha_t` represent the learning rate and :math:`\beta_1, \beta_2`,
:math:`\varepsilon`, :math:`\bar{\varepsilon}` represent the arguments
``b1``, ``b2``, ``eps`` and ``eps_root`` respectively. The learning rate is
indexed by :math:`t` since the learning rate may also be provided by a
schedule function.
The ``init`` function of this optimizer initializes an internal state
:math:`S_0 := (m_0, v_0) = (0, 0)`, representing initial estimates for the
first and second moments. In practice these values are stored as pytrees
containing all zeros, with the same shape as the model updates.
At step :math:`t`, the ``update`` function of this optimizer takes as
arguments the incoming gradients :math:`g_t` and optimizer state :math:`S_t`
and computes updates :math:`u_t` and new state :math:`S_{t+1}`. Thus, for
:math:`t > 0`, we have,
.. math::
\begin{align*}
m_t &\leftarrow \beta_1 \cdot m_{t-1} + (1-\beta_1) \cdot g_t \\
v_t &\leftarrow \beta_2 \cdot v_{t-1} + (1-\beta_2) \cdot {g_t}^2 \\
\hat{m}_t &\leftarrow m_t / {(1-\beta_1^t)} \\
\hat{v}_t &\leftarrow v_t / {(1-\beta_2^t)} \\
u_t &\leftarrow -\alpha_t \cdot \hat{m}_t / \left({\sqrt{\hat{v}_t +
\bar{\varepsilon}} + \varepsilon} \right)\\
S_t &\leftarrow (m_t, v_t).
\end{align*}
With the keyword argument `nesterov=True`, the optimizer uses Nesterov
momentum, replacing the above :math:`\hat{m}_t` with
.. math::
\hat{m}_t \leftarrow
\beta_1 m_t / {(1-\beta_1^{t+1})} + (1 - \beta_1) g_t / {(1-\beta_1^t)}.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adam(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Kingma et al, `Adam: A Method for Stochastic Optimization
<https://arxiv.org/abs/1412.6980>`_, 2014
Dozat, `Incorporating Nesterov Momentum into Adam
<https://openreview.net/pdf?id=OM0jvwB8jIp57ZJjtNEZ>`_, 2016
.. warning::
PyTorch and optax's implementation follow Algorithm 1 of [Kingma et al.
2014]. Note that TensorFlow used instead the formulation just before Section
2.1 of the paper. See https://github.com/deepmind/optax/issues/571 for more
detail.
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
example when computing (meta-)gradients through Adam.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
nesterov: Whether to use Nesterov momentum. The solver with
nesterov=True is equivalent to the :func:`optax.nadam` optimizer, and
described in [Dozat 2016].
Returns:
The corresponding `GradientTransformation`.
.. seealso:: :func:`optax.nadam`, :func:`optax.adamw`.
"""
return combine.chain(
transform.scale_by_adam(
b1=b1,
b2=b2,
eps=eps,
eps_root=eps_root,
mu_dtype=mu_dtype,
nesterov=nesterov,
),
transform.scale_by_learning_rate(learning_rate),
)
nadam = functools.partial(adam, nesterov=True)
nadam.__doc__ = (
r"""The NAdam optimizer.
Nadam is a variant of :func:`optax.adam` with Nesterov's momentum. The update
rule of this solver is as follows:
.. math::
\begin{align*}
m_t &\leftarrow \beta_1 \cdot m_{t-1} + (1-\beta_1) \cdot g_t \\
v_t &\leftarrow \beta_2 \cdot v_{t-1} + (1-\beta_2) \cdot {g_t}^2 \\
\hat{m}_t &\leftarrow
\beta_1 m_t / {(1-\beta_1^{t+1})} + (1 - \beta_1) g_t / {(1-\beta_1^t)}\\
\hat{v}_t &\leftarrow v_t / {(1-\beta_2^t)} \\
u_t &\leftarrow \alpha_t \cdot \hat{m}_t / \left({\sqrt{\hat{v}_t +
\bar{\varepsilon}} + \varepsilon} \right)\\
S_t &\leftarrow (m_t, v_t).
\end{align*}
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.nadam(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.38E+01
References:
Dozat, `Incorporating Nesterov Momentum into Adam
<https://openreview.net/pdf?id=OM0jvwB8jIp57ZJjtNEZ>`_, 2016
.. versionadded:: 0.1.9
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
example when computing (meta-)gradients through Adam.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
Returns:
The corresponding `GradientTransformation`.
.. seealso:: :func:`optax.adam`, :func:`optax.nadamw`.
"""
)
def adamw(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-8,
eps_root: float = 0.0,
mu_dtype: Optional[Any] = None,
weight_decay: float = 1e-4,
mask: Optional[Union[Any, Callable[[base.Params], Any]]] = None,
*,
nesterov: bool = False,
) -> base.GradientTransformation:
r"""Adam with weight decay regularization.
AdamW uses weight decay to regularize learning towards small weights, as
this leads to better generalization. In SGD you can also use L2 regularization
to implement this as an additive loss term, however L2 regularization
does not behave as intended for adaptive gradient algorithms such as Adam,
see [Loshchilov et al, 2019].
Let :math:`\alpha_t` represent the learning rate and :math:`\beta_1, \beta_2`,
:math:`\varepsilon`, :math:`\bar{\varepsilon}` represent the arguments
``b1``, ``b2``, ``eps`` and ``eps_root`` respectively. The learning rate is
indexed by :math:`t` since the learning rate may also be provided by a
schedule function. Let :math:`\lambda` be the weight decay and
:math:`\theta_t` the parameter vector at time :math:`t`.
The ``init`` function of this optimizer initializes an internal state
:math:`S_0 := (m_0, v_0) = (0, 0)`, representing initial estimates for the
first and second moments. In practice these values are stored as pytrees
containing all zeros, with the same shape as the model updates.
At step :math:`t`, the ``update`` function of this optimizer takes as
arguments the incoming gradients :math:`g_t`, the optimizer state :math:`S_t`
and the parameters :math:`\theta_t` and computes updates :math:`u_t` and
new state :math:`S_{t+1}`. Thus, for :math:`t > 0`, we have,
.. math::
\begin{align*}
m_t &\leftarrow \beta_1 \cdot m_{t-1} + (1-\beta_1) \cdot g_t \\
v_t &\leftarrow \beta_2 \cdot v_{t-1} + (1-\beta_2) \cdot {g_t}^2 \\
\hat{m}_t &\leftarrow m_t / {(1-\beta_1^t)} \\
\hat{v}_t &\leftarrow v_t / {(1-\beta_2^t)} \\
u_t &\leftarrow -\alpha_t \cdot \left( \hat{m}_t / \left({\sqrt{\hat{v}_t
+ \bar{\varepsilon}} + \varepsilon} \right) + \lambda \theta_{t} \right)\\
S_t &\leftarrow (m_t, v_t).
\end{align*}
This implementation can incorporate a momentum a la Nesterov introduced by
[Dozat 2016]. The resulting optimizer is then often referred as NAdamW.
With the keyword argument `nesterov=True`, the optimizer uses Nesterov
momentum, replacing the above :math:`\hat{m}_t` with
.. math::
\hat{m}_t \leftarrow
\beta_1 m_t / {(1-\beta_1^{t+1})} + (1 - \beta_1) g_t / {(1-\beta_1^t)}.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adamw(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Loshchilov et al, `Decoupled Weight Decay
Regularization <https://arxiv.org/abs/1711.05101>`_, 2019
Dozat, `Incorporating Nesterov Momentum into Adam
<https://openreview.net/pdf?id=OM0jvwB8jIp57ZJjtNEZ>`_, 2016
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
instance when computing (meta-)gradients through Adam.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
weight_decay: Strength of the weight decay regularization. Note that this
weight decay is multiplied with the learning rate. This is consistent
with other frameworks such as PyTorch, but different from
(Loshchilov et al, 2019) where the weight decay is only multiplied with
the "schedule multiplier", but not the base learning rate.
mask: A tree with same structure as (or a prefix of) the params PyTree,
or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the weight decay to, and `False` for those you want to skip. Note
that the Adam gradient transformations are applied to all parameters.
nesterov: Whether to use Nesterov momentum. The solver with
nesterov=True is equivalent to the :func:`optax.nadamw` optimizer. This
modification is described in [Dozat 2016].
Returns:
The corresponding `GradientTransformation`.
.. seealso:: :func:`optax.adam`, :func:`optax.nadamw`.
"""
return combine.chain(
transform.scale_by_adam(
b1=b1,
b2=b2,
eps=eps,
eps_root=eps_root,
mu_dtype=mu_dtype,
nesterov=nesterov,
),
transform.add_decayed_weights(weight_decay, mask),
transform.scale_by_learning_rate(learning_rate),
)
nadamw = functools.partial(adamw, nesterov=True)
nadamw.__doc__ = (
r"""NAdamW optimizer, implemented as part of the AdamW optimizer.
NadamW is variant of :func:`optax.adamw` with Nesterov's momentum. Compared
to AdamW, this optimizer replaces the assignment
.. math::
\hat{m}_t \leftarrow m_t / {(1-\beta_1^t)}
with
.. math::
\hat{m}_t \leftarrow
\beta_1 m_t / {(1-\beta_1^{t+1})} + (1 - \beta_1) g_t / {(1-\beta_1^t)}.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.nadamw(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.38E+01
References:
Loshchilov et al, `Decoupled Weight Decay
Regularization <https://arxiv.org/abs/1711.05101>`_, 2019
Dozat, `Incorporating Nesterov Momentum into Adam
<https://openreview.net/pdf?id=OM0jvwB8jIp57ZJjtNEZ>`_, 2016
.. versionadded:: 0.1.9
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
instance when computing (meta-)gradients through Adam.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
weight_decay: Strength of the weight decay regularization. Note that this
weight decay is multiplied with the learning rate. This is consistent
with other frameworks such as PyTorch, but different from
(Loshchilov et al, 2019) where the weight decay is only multiplied with
the "schedule multiplier", but not the base learning rate.
mask: A tree with same structure as (or a prefix of) the params PyTree,
or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the weight decay to, and `False` for those you want to skip. Note
that the Adam gradient transformations are applied to all parameters.
Returns:
The corresponding `GradientTransformation`.
.. seealso:: :func:`optax.adam`, :func:`optax.adamw`.
"""
)
def lion(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.99,
mu_dtype: Optional[Any] = None,
weight_decay: float = 1e-3,
mask: Optional[Union[Any, Callable[[base.Params], Any]]] = None,
) -> base.GradientTransformation:
"""The Lion optimizer.
Lion is discovered by symbolic program search. Unlike most adaptive optimizers
such as AdamW, Lion only tracks momentum, making it more memory-efficient.
The update of Lion is produced through the sign operation, resulting in a
larger norm compared to updates produced by other optimizers such as SGD and
AdamW. A suitable learning rate for Lion is typically 3-10x smaller than that
for AdamW, the weight decay for Lion should be in turn 3-10x larger than that
for AdamW to maintain a similar strength (lr * wd).
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.lion(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Chen et al, 2023: https://arxiv.org/abs/2302.06675
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Rate to combine the momentum and the current gradient.
b2: Exponential decay rate to track the momentum of past gradients.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
weight_decay: Strength of the weight decay regularization. Note that this
weight decay is multiplied with the learning rate. This is consistent
with other frameworks such as PyTorch, but different from
(Loshchilov et al, 2019) where the weight decay is only multiplied with
the "schedule multiplier", but not the base learning rate.
mask: A tree with same structure as (or a prefix of) the params PyTree,
or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the weight decay to, and `False` for those you want to skip. Note
that the Adam gradient transformations are applied to all parameters.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_lion(b1=b1, b2=b2, mu_dtype=mu_dtype),
transform.add_decayed_weights(weight_decay, mask),
transform.scale_by_learning_rate(learning_rate),
)
def amsgrad(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-8,
eps_root: float = 0.0,
mu_dtype: Optional[Any] = None,
) -> base.GradientTransformation:
"""The AMSGrad optimiser.
The original Adam can fail to converge to the optimal solution in some cases.
AMSGrad guarantees convergence by using a long-term memory of past gradients.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.amsgrad(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Reddi et al, 2018: https://openreview.net/forum?id=ryQu7f-RZ
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
instance when computing (meta-)gradients through Adam.
mu_dtype: Optional `dtype` to be used for the first order accumulator; if
`None` then the `dtype` is inferred from `params` and `updates`.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_amsgrad(
b1=b1, b2=b2, eps=eps, eps_root=eps_root, mu_dtype=mu_dtype),
transform.scale_by_learning_rate(learning_rate),
)
def fromage(
learning_rate: float,
min_norm: float = 1e-6
) -> base.GradientTransformation:
"""The Frobenius matched gradient descent (Fromage) optimizer.
Fromage is a learning algorithm that does not require learning rate tuning.
The optimizer is based on modeling neural network gradients via deep relative
trust (a distance function on deep neural networks). Fromage is similar to the
LARS optimizer and can work on a range of standard neural network benchmarks,
such as natural language Transformers and generative adversarial networks.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.fromage(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.37E+01
Objective function: 1.36E+01
References:
Bernstein et al, 2020: https://arxiv.org/abs/2002.03432
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
min_norm: A minimum value that the norm of the gradient updates and the norm
of the layer parameters can be clipped to to avoid dividing by zero when
computing the trust ratio (as in the LARS paper).
Returns:
The corresponding `GradientTransformation`.
"""
mult = 1 / jnp.sqrt(1 + learning_rate ** 2)
return combine.chain(
transform.scale_by_trust_ratio(min_norm),
transform.scale_by_learning_rate(learning_rate * mult),
transform.add_decayed_weights((mult - 1)),
)
def lars(
learning_rate: base.ScalarOrSchedule,
weight_decay: float = 0.,
weight_decay_mask: MaskOrFn = True,
trust_coefficient: float = 0.001,
eps: float = 0.,
trust_ratio_mask: MaskOrFn = True,
momentum: float = 0.9,
nesterov: bool = False,
) -> base.GradientTransformation:
"""The LARS optimizer.
LARS is a layer-wise adaptive optimizer introduced to help scale SGD to
larger batch sizes. LARS later inspired the LAMB optimizer.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.lars(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.40E+01
References:
You et al, 2017: https://arxiv.org/abs/1708.03888
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
weight_decay: Strength of the weight decay regularization.
weight_decay_mask: A tree with same structure as (or a prefix of) the params
PyTree, or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the transformation to, and `False` for those you want to skip.
trust_coefficient: A multiplier for the trust ratio.
eps: Optional additive constant in the trust ratio denominator.
trust_ratio_mask: A tree with same structure as (or a prefix of) the params
PyTree, or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the transformation to, and `False` for those you want to skip.
momentum: Decay rate for momentum.
nesterov: Whether to use Nesterov momentum.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.add_decayed_weights(weight_decay, mask=weight_decay_mask),
wrappers.masked(
inner=transform.scale_by_trust_ratio(
trust_coefficient=trust_coefficient, eps=eps),
mask=trust_ratio_mask),
transform.scale_by_learning_rate(learning_rate),
transform.trace(decay=momentum, nesterov=nesterov),
)
def lamb(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-6,
eps_root: float = 0.0,
weight_decay: float = 0.,
mask: MaskOrFn = None,
) -> base.GradientTransformation:
"""The LAMB optimizer.
LAMB is a general purpose layer-wise adaptive large batch optimizer designed
to provide consistent training performance across a wide range of tasks,
including those that use attention-based models (such as Transformers) and
ResNet-50. The optimizer is able to work with small and large batch sizes.
LAMB was inspired by the LARS learning algorithm.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.lamb(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.36E+01
References:
You et al, 2019: https://arxiv.org/abs/1904.00962
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
instance when computing (meta-)gradients through Adam.
weight_decay: Strength of the weight decay regularization.
mask: A tree with same structure as (or a prefix of) the params PyTree,
or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the transformation to, and `False` for those you want to skip.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_adam(b1=b1, b2=b2, eps=eps, eps_root=eps_root),
transform.add_decayed_weights(weight_decay=weight_decay, mask=mask),
transform.scale_by_trust_ratio(),
transform.scale_by_learning_rate(learning_rate),
)
def noisy_sgd(
learning_rate: base.ScalarOrSchedule,
eta: float = 0.01,
gamma: float = 0.55,
seed: int = 0
) -> base.GradientTransformation:
r"""A variant of SGD with added noise.
Noisy SGD is a variant of :func:`optax.sgd` that incorporates Gaussian noise
into the updates. It has been found that adding noise to the gradients can
improve both the training error and the generalization error in very deep
networks.
The update :math:`u_t` is modified to include this noise as follows:
.. math::
u_t \leftarrow -\alpha_t (g_t + N(0, \sigma_t^2)),
where :math:`N(0, \sigma_t^2)` represents Gaussian noise with zero mean and a
variance of :math:`\sigma_t^2`.
The variance of this noise decays over time according to the formula
.. math::
\sigma_t^2 = \frac{\eta}{(1+t)^\gamma},
where :math:`\gamma` is the decay rate parameter ``gamma`` and :math:`\eta`
represents the initial variance ``eta``.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.noisy_sgd(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.35E+01
Objective function: 1.33E+01
Objective function: 1.32E+01
References:
Neelakantan et al, 2014: https://arxiv.org/abs/1511.06807
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
eta: Initial variance for the Gaussian noise added to gradients.
gamma: A parameter controlling the annealing of noise over time ``t``, the
variance decays according to ``(1+t)**(-gamma)``.
seed: Seed for the pseudo-random generation process.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.add_noise(eta, gamma, seed),
transform.scale_by_learning_rate(learning_rate),
)
def novograd(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.25,
eps: float = 1e-6,
eps_root: float = 0.0,
weight_decay: float = 0.,
) -> base.GradientTransformation:
"""NovoGrad optimizer.
NovoGrad is more robust to the initial learning rate and
weight initialization than other methods. For example,
NovoGrad works well without LR warm-up, while other methods require it.
NovoGrad performs exceptionally well for large batch training, e.g. it
outperforms other methods for ResNet-50 for all batches up to 32K.
In addition, NovoGrad requires half the memory compared to Adam.
It was introduced together with Jasper ASR model.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.novograd(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.37E+01
References:
Ginsburg et al, 2019: https://arxiv.org/abs/1905.11286
Li et al, 2019: https://arxiv.org/abs/1904.03288
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: An exponential decay rate to track the first moment of past gradients.
b2: An exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root (as
in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside
the square root (as in RMSProp), to avoid dividing by zero when rescaling.
This is needed for instance when computing (meta-)gradients through Adam.
weight_decay: Strength of the weight decay regularization.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_novograd(
b1=b1, b2=b2, eps=eps, eps_root=eps_root, weight_decay=weight_decay),
transform.scale_by_learning_rate(learning_rate),
)
def optimistic_gradient_descent(
learning_rate: base.ScalarOrSchedule,
alpha: base.ScalarOrSchedule = 1.0,
beta: base.ScalarOrSchedule = 1.0
) -> base.GradientTransformation:
"""An Optimistic Gradient Descent optimizer.
Optimistic gradient descent is an approximation of extra-gradient methods
which require multiple gradient calls to compute the next update. It has
strong formal guarantees for last-iterate convergence in min-max games, for
which standard gradient descent can oscillate or even diverge.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.optimistic_gradient_descent(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.37E+01
Objective function: 1.35E+01
Objective function: 1.33E+01
Objective function: 1.32E+01
Objective function: 1.30E+01
References:
Mokhtari et al, 2019: https://arxiv.org/abs/1901.08511v2
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
alpha: Coefficient for generalized OGD.
beta: Coefficient for generalized OGD negative momentum.
Returns:
A `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_optimistic_gradient(alpha=alpha, beta=beta),
transform.scale_by_learning_rate(learning_rate)
)
def radam(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-8,
eps_root: float = 0.0,
threshold: float = 5.0,
*,
nesterov: bool = False,
) -> base.GradientTransformation:
"""The Rectified Adam optimizer.
The adaptive learning rate in Adam has undesirably large variance in early
stages of training, due to the limited number of training samples used to
estimate the optimizer's statistics. Rectified Adam addresses this issue
by analytically reducing the large variance.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.radam(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.35E+01
Objective function: 1.33E+01
Objective function: 1.32E+01
References:
Liu et al, 2020: https://arxiv.org/abs/1908.03265
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
eps_root: A small constant applied to denominator inside the square root (as
in RMSProp), to avoid dividing by zero when rescaling. This is needed for
instance when computing (meta-)gradients through Adam.
threshold: Threshold for variance tractability.
nesterov: Whether to use Nesterov momentum.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_radam(
b1=b1,
b2=b2,
eps=eps,
eps_root=eps_root,
threshold=threshold,
nesterov=nesterov,
),
transform.scale_by_learning_rate(learning_rate),
)
def rmsprop(
learning_rate: base.ScalarOrSchedule,
decay: float = 0.9,
eps: float = 1e-8,
initial_scale: float = 0.,
centered: bool = False,
momentum: Optional[float] = None,
nesterov: bool = False
) -> base.GradientTransformation:
# pylint: disable=line-too-long
r"""A flexible RMSProp optimizer.
RMSProp is an SGD variant with learning rate adaptation. The `learning_rate`
used for each weight is scaled by a suitable estimate of the magnitude of the
gradients on previous steps. Several variants of RMSProp can be found
in the literature. This alias provides an easy to configure RMSProp
optimizer that can be used to switch between several of these variants.
..warning::
PyTorch and optax's RMSprop implementations differ and could impact
performance. In the denominator, optax uses :math:`$\sqrt{v + \epsilon}$`
whereas PyTorch uses :math:`$\sqrt{v} + \epsilon$`. See
https://github.com/google-deepmind/optax/issues/532 for more detail.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.rmsprop(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.39E+01
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.37E+01
Objective function: 1.36E+01
References:
Tieleman and Hinton, 2012: http://www.cs.toronto.edu/~hinton/coursera/lecture6/lec6.pdf
Graves, 2013: https://arxiv.org/abs/1308.0850
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
decay: Decay used to track the magnitude of previous gradients.
eps: A small numerical constant to avoid dividing by zero when rescaling.
initial_scale: Initial value of accumulators tracking the magnitude of
previous updates. PyTorch uses `0`, TF1 uses `1`. When reproducing results
from a paper, verify the value used by the authors.
centered: Whether the second moment or the variance of the past gradients is
used to rescale the latest gradients.
momentum: Decay rate used by the momentum term, when it is set to `None`,
then momentum is not used at all.
nesterov: Whether Nesterov momentum is used.
Returns:
The corresponding `GradientTransformation`.
"""
# pylint: enable=line-too-long
if centered:
return combine.chain(
transform.scale_by_stddev(
decay=decay, eps=eps, initial_scale=initial_scale),
transform.scale_by_learning_rate(learning_rate),
(transform.trace(decay=momentum, nesterov=nesterov)
if momentum is not None else base.identity())
)
return combine.chain(
transform.scale_by_rms(
decay=decay, eps=eps, initial_scale=initial_scale),
transform.scale_by_learning_rate(learning_rate),
(transform.trace(decay=momentum, nesterov=nesterov)
if momentum is not None else base.identity())
)
def sgd(
learning_rate: base.ScalarOrSchedule,
momentum: Optional[float] = None,
nesterov: bool = False,
accumulator_dtype: Optional[Any] = None,
) -> base.GradientTransformation:
r"""A canonical Stochastic Gradient Descent optimizer.
This implements stochastic gradient descent. It also includes support for
momentum, and Nesterov acceleration, as these are standard practice when
using stochastic gradient descent to train deep neural networks.
The canonical stochastic gradient descent returns an update
:math:`u_t` of the form
.. math::
u_t \leftarrow -\alpha_t g_t,
where :math:`g_t` is the gradient of the objective (potentially preprocessed
by other transformations) and :math:`\alpha_t` is the ``learning_rate`` at
time :math:`t` (constant or selected by an :class:`optax.Schedule`).
Stochastic gradient descent with momentum takes two possible forms.
.. math::
\begin{align*}
m_t &\leftarrow g_t + \mu m_{t-1} \\
u_t &\leftarrow \begin{cases}
-\alpha_t m_t & \text{ if } \texttt{nesterov = False} \\
-\alpha_t (g_t + \mu m_t) & \text{ if } \texttt{nesterov = True}
\end{cases} \\
S_t &\leftarrow m_t,
\end{align*}
where :math:`\mu` is the ``momentum`` parameter and :math:`S_t` is the state
of the optimizer.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.sgd(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.38E+01
Objective function: 1.37E+01
Objective function: 1.35E+01
Objective function: 1.33E+01
Objective function: 1.32E+01
References:
Sutskever et al, `On the importance of initialization and momentum in deep
learning <http://proceedings.mlr.press/v28/sutskever13.pdf>`_, 2013
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
momentum: Decay rate used by the momentum term, when it is set to ``None``,
then momentum is not used at all.
nesterov: Whether Nesterov momentum is used.
accumulator_dtype: Optional ``dtype`` to be used for the accumulator; if
``None`` then the ``dtype`` is inferred from ``params`` and ``updates``.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
(transform.trace(decay=momentum, nesterov=nesterov,
accumulator_dtype=accumulator_dtype)
if momentum is not None else base.identity()),
transform.scale_by_learning_rate(learning_rate)
)
def sm3(
learning_rate: float,
momentum: float = 0.9
) -> base.GradientTransformation:
"""The SM3 optimizer.
SM3 (Square-root of Minima of Sums of Maxima of Squared-gradients Method) is a
memory-efficient adaptive optimizer designed to decrease memory overhead when
training very large models, such as the Transformer for machine translation,
BERT for language modeling, and AmoebaNet-D for image classification. SM3: 1)
applies to tensors of arbitrary dimensions and any predefined cover of the
parameters; 2) adapts the learning rates in an adaptive and data-driven manner
(like Adagrad and unlike Adafactor); and 3) comes with rigorous convergence
guarantees in stochastic convex optimization settings.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.sm3(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.40E+01
References:
Anil et al, 2019: https://arxiv.org/abs/1901.11150
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
momentum: Decay rate used by the momentum term (when it is not set to
`None`, then momentum is not used at all).
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_sm3(momentum),
transform.scale(-learning_rate),
)
def yogi(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-3,
) -> base.GradientTransformation:
# pylint: disable=line-too-long
"""The Yogi optimizer.
Yogi is an adaptive optimizer, which provides control in tuning the effective
learning rate to prevent it from increasing. By doing so, it focuses on
addressing the issues of convergence and generalization in exponential moving
average-based adaptive methods (such as Adam and RMSprop). Yogi is a
modification of Adam and uses the same parameters.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.yogi(learning_rate=0.002)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
References:
Zaheer et al, 2018: https://proceedings.neurips.cc/paper/2018/file/90365351ccc7437a1309dc64e4db32a3-Paper.pdf
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the second moment of past gradients.
eps: A small constant applied to denominator outside of the square root
(as in the Adam paper) to avoid dividing by zero when rescaling.
Returns:
The corresponding `GradientTransformation`.
"""
# pylint: enable=line-too-long
return combine.chain(
transform.scale_by_yogi(b1=b1, b2=b2, eps=eps),
transform.scale_by_learning_rate(learning_rate),
)
def adamax(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-8,
) -> base.GradientTransformation:
r"""A variant of the Adam optimizer that uses the infinity norm.
AdaMax is a variant of the :func:`optax.adam` optimizer. By generalizing
Adam's :math:`L^2` norm to an :math:`L^p` norm and taking the limit as
:math:`p \rightarrow \infty`, we obtain a simple and stable update rule.
Let :math:`\alpha_t` represent the learning rate and :math:`\beta_1, \beta_2`,
:math:`\varepsilon` represent the arguments
``b1``, ``b2`` and ``eps`` respectively. The learning rate is
indexed by :math:`t` since the learning rate may also be provided by a
schedule function.
The ``init`` function of this optimizer initializes an internal state
:math:`S_0 := (m_0, v_0) = (0, 0)`, representing initial estimates for the
first and second moments. In practice these values are stored as pytrees
containing all zeros, with the same shape as the model updates.
At step :math:`t`, the ``update`` function of this optimizer takes as
arguments the incoming gradients :math:`g_t` and optimizer state :math:`S_t`
and computes updates :math:`u_t` and new state :math:`S_{t+1}`. Thus, for
:math:`t > 0`, we have,
.. math::
\begin{align*}
m_t &\leftarrow \beta_1 \cdot m_{t-1} + (1-\beta_1) \cdot g_t \\
v_t &\leftarrow \max(\left| g_t \right| + \varepsilon, \beta_2 \cdot
v_{t-1}) \\
\hat{m}_t &\leftarrow m_t / (1-\beta_1^t) \\
u_t &\leftarrow -\alpha_t \cdot \hat{m}_t / v_t \\
S_t &\leftarrow (m_t, v_t).
\end{align*}
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adamax(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Kingma et al, 2014: https://arxiv.org/abs/1412.6980
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the maximum of past gradients.
eps: A small constant applied to denominator to avoid dividing by zero when
rescaling.
Returns:
The corresponding `GradientTransformation`.
.. seealso:: :func:`optax.adam`, :func:`optax.adamaxw`.
"""
return combine.chain(
transform.scale_by_adamax(b1=b1, b2=b2, eps=eps,),
transform.scale_by_learning_rate(learning_rate),
)
def adamaxw(
learning_rate: base.ScalarOrSchedule,
b1: float = 0.9,
b2: float = 0.999,
eps: float = 1e-8,
weight_decay: float = 1e-4,
mask: Optional[Union[Any, Callable[[base.Params], Any]]] = None,
) -> base.GradientTransformation:
"""Adamax with weight decay regularization.
AdamaxW uses weight decay to regularize learning towards small weights, as
this leads to better generalization. In SGD you can also use L2 regularization
to implement this as an additive loss term, however L2 regularization
does not behave as intended for adaptive gradient algorithms such as Adam.
WARNING: Sometimes you may want to skip weight decay for BatchNorm scale or
for the bias parameters. You can use `optax.masked` to make your own AdamaxW
variant where `additive_weight_decay` is applied only to a subset of `params`.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.adamaxw(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Loshchilov et al, 2019: https://arxiv.org/abs/1711.05101
Args:
learning_rate: A global scaling factor, either fixed or evolving along
iterations with a scheduler, see :func:`optax.scale_by_learning_rate`.
b1: Exponential decay rate to track the first moment of past gradients.
b2: Exponential decay rate to track the maximum of past gradients.
eps: A small constant applied to denominator to avoid dividing by zero when
rescaling.
weight_decay: Strength of the weight decay regularization. Note that this
weight decay is multiplied with the learning rate. This is consistent
with other frameworks such as PyTorch, but different from
(Loshchilov et al, 2019) where the weight decay is only multiplied with
the "schedule multiplier", but not the base learning rate.
mask: A tree with same structure as (or a prefix of) the params PyTree,
or a Callable that returns such a pytree given the params/updates.
The leaves should be booleans, `True` for leaves/subtrees you want to
apply the weight decay to, and `False` for those you want to skip. Note
that the Adamax gradient transformations are applied to all parameters.
Returns:
The corresponding `GradientTransformation`.
.. seealso:: :func:`optax.adam`, :func:`optax.adamax`.
"""
return combine.chain(
transform.scale_by_adamax(b1=b1, b2=b2, eps=eps),
transform.add_decayed_weights(weight_decay, mask),
transform.scale_by_learning_rate(learning_rate),
)
def rprop(
learning_rate: float,
eta_minus: float = 0.5,
eta_plus: float = 1.2,
min_step_size: float = 1e-6,
max_step_size: float = 50.0,
) -> base.GradientTransformation:
"""The Rprop optimizer.
Rprop, short for resillient backpropogation, is a first order variant of
gradient descent. It responds only to the sign of the gradient by increasing
or decreasing the step size selected per parameter exponentially to speed up
convergence and avoid oscillations.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.rprop(learning_rate=0.003)
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... grad = jax.grad(f)(params)
... updates, opt_state = solver.update(grad, opt_state, params)
... params = optax.apply_updates(params, updates)
... print('Objective function: {:.2E}'.format(f(params)))
Objective function: 1.40E+01
Objective function: 1.40E+01
Objective function: 1.39E+01
Objective function: 1.39E+01
Objective function: 1.38E+01
References:
Riedmiller and Braun. `A direct adaptive method for faster backpropagation
learning: the RPROP algorithm
<https://ieeexplore.ieee.org/document/298623>`_, 1993
Igel and Hüsken. `Empirical evaluation of the improved Rprop learning
algorithms
<https://www.sciencedirect.com/science/article/abs/pii/S0925231201007007>`_,
2003
Args:
learning_rate: The initial step size.
eta_minus: Multiplicative factor for decreasing step size. This is applied
when the gradient changes sign from one step to the next.
eta_plus: Multiplicative factor for increasing step size. This is applied
when the gradient has the same sign from one step to the next.
min_step_size: Minimum allowed step size. Smaller steps will be clipped to
this value.
max_step_size: Maximum allowed step size. Larger steps will be clipped to
this value.
Returns:
The corresponding `GradientTransformation`.
"""
return combine.chain(
transform.scale_by_rprop(
learning_rate=learning_rate,
eta_minus=eta_minus,
eta_plus=eta_plus,
min_step_size=min_step_size,
max_step_size=max_step_size,
),
transform.scale(-1.0),
)
def polyak_sgd(
max_learning_rate: float = 1.,
scaling: base.ScalarOrSchedule = 1.,
f_min: float = 0.0,
eps: float = 0.0,
) -> base.GradientTransformationExtraArgs:
r"""SGD with Polyak step-size.
This solver implements the SGD with Polyak step size of (Loizou et al. 2021).
It sets the step-size as
.. math::
s \min\left\{\frac{f(x) - f^\star}{\|\nabla f(x)\|^2 + \epsilon},
\gamma_{\max}\right\}\,,
where :math:`f` is the function from which a gradient is computed,
:math:`\gamma_{\max}` is a maximal acceptable learning rate set by
``max_learning_rate``, :math:`\epsilon` is a constant preventing division by
zero set with ``eps``, :math:`s` scales the formula by ``scaling``, and
:math:`f^\star` is a guess of the minimum value of the function set with
``f_min``.
Examples:
>>> import optax
>>> import jax
>>> import jax.numpy as jnp
>>> def f(x): return jnp.sum(x ** 2) # simple quadratic function
>>> solver = optax.polyak_sgd()
>>> params = jnp.array([1., 2., 3.])
>>> print('Objective function: ', f(params))
Objective function: 14.0
>>> opt_state = solver.init(params)
>>> for _ in range(5):
... value, grad = jax.value_and_grad(f)(params)
... params, opt_state = solver.update(grad, opt_state, params, value=value)
... print('Objective function: ', f(params))
Objective function: 3.5
Objective function: 0.875
Objective function: 0.21875
Objective function: 0.0546875
Objective function: 0.013671875
.. warning::
This method requires knowledge of an approximate value of the of the
objective function minimum, passed through the ``f_min`` argument.
For models that interpolate the data, this can be set to 0 (default
value).
Failing to set an appropriate value for ``f_min`` can lead to
divergence or convergence to a suboptimal solution.
References:
Loizou et al. `Stochastic polyak step-size for SGD: An adaptive learning
rate for fast convergence <https://arxiv.org/abs/2002.10542>`_, 2021
Berrada et al., `Training neural networks for and by interpolation
<https://arxiv.org/pdf/1906.05661.pdf>`_, 2020
Args:
max_learning_rate: a maximum step size to use (defaults to 1).
scaling: A global scaling factor, either fixed or evolving along
iterations with a scheduler (defaults to 1).
f_min: a lower bound on the objective function (defaults to 0). Corresponds
to :math:`f^\star` in the formula above.
eps: a value to add in the denominator of the update (defaults to 0).
Returns:
A :class:`GradientTransformationExtraArgs`, where the ``update`` function
takes an additional keyword argument ``value`` containing the current
value of the objective function.
"""
return combine.chain(
sgd(learning_rate=scaling),
transform.scale_by_polyak(
max_learning_rate=max_learning_rate, f_min=f_min, eps=eps
),
)