Implement "RAdamW" optimizer (#107507)
Fixes #107282
## Overview
- basic design decision was followed as they made on #103881 (tensor operation, test cases, order & position of argument etc.)
- for the algorithm for decoupled weight decay, I referred to [1, 2]
## backwards-incompatible changes
- positional argument `decoupled_weight_decay` is added to:
- `torch.optim.radam`
The existing code which refers to these APIs can be affected.
Note: Positional argument `decoupled_weight_decay` is added to `torch.optim.RAdam`. However, since it was added to the last position and with default value, it is not affected.
## Reference
- [1] [Decoupled Weight Decay Regularization](https://arxiv.org/abs/1711.05101)
- [2] https://github.com/LiyuanLucasLiu/RAdam/blob/master/radam/radam.py#L5-L94
## TODO
- [x] implement tensor operation
- [x] implement test cases
- [x] modify doc-string
- [x] pass unit test code locally `python test/test_optim.py -k test_radam`
Pull Request resolved: https://github.com/pytorch/pytorch/pull/107507
Approved by: https://github.com/janeyx99
diff --git a/test/optim/test_optim.py b/test/optim/test_optim.py
index 812d52b..1677869 100644
--- a/test/optim/test_optim.py
+++ b/test/optim/test_optim.py
@@ -884,6 +884,8 @@
(optim.RAdam, dict(weight_decay=0)),
(optim.RAdam, dict(weight_decay=1, eps=1e-6)),
(optim.RAdam, dict(weight_decay=1)),
+ (optim.RAdam, dict(weight_decay=0, decoupled_weight_decay=True)),
+ (optim.RAdam, dict(weight_decay=1, decoupled_weight_decay=True)),
(optim.RMSprop, dict(weight_decay=1, momentum=1, centered=True)),
(optim.RMSprop, dict(weight_decay=1, momentum=0, centered=True)),
(optim.RMSprop, dict(weight_decay=1, momentum=1, centered=False)),
@@ -1500,6 +1502,23 @@
],
constructor_accepts_foreach=True,
)
+ # RAdamW tests
+ self._test_basic_cases(
+ lambda weight, bias, foreach: optim.RAdam(
+ [weight, bias], lr=1e-3, weight_decay=0.1, decoupled_weight_decay=True, foreach=foreach
+ ),
+ constructor_accepts_foreach=True,
+ )
+ self._test_basic_cases(
+ lambda weight, bias, foreach: optim.RAdam(
+ [weight, bias], lr=1e-3, weight_decay=0.1, decoupled_weight_decay=True, foreach=foreach
+ ),
+ [
+ lambda opt: ExponentialLR(opt, gamma=0.9),
+ lambda opt: ReduceLROnPlateau(opt),
+ ],
+ constructor_accepts_foreach=True,
+ )
with self.assertRaisesRegex(
ValueError, "Invalid beta parameter at index 0: 1.0"
):
@@ -2359,6 +2378,17 @@
*state.values(),
),
)
+ gradcheck(
+ _diff_fn,
+ (
+ p,
+ grad,
+ state,
+ torch.optim.RAdam,
+ {"lr": 0.9, "weight_decay": 0.1, "decoupled_weight_decay": True, "differentiable": True},
+ *state.values(),
+ ),
+ )
@unittest.skipIf(not TEST_CUDA, "test requires CUDA")
diff --git a/torch/optim/radam.py b/torch/optim/radam.py
index 4e57426..9753ab3 100644
--- a/torch/optim/radam.py
+++ b/torch/optim/radam.py
@@ -1,10 +1,19 @@
import math
+from typing import List, Optional
+
import torch
from torch import Tensor
-from .optimizer import (Optimizer, _use_grad_for_differentiable, _get_value, _dispatch_sqrt, _stack_if_compiling,
- _default_to_fused_or_foreach, _differentiable_doc, _foreach_doc)
-from typing import List, Optional
+from .optimizer import (
+ Optimizer,
+ _default_to_fused_or_foreach,
+ _differentiable_doc,
+ _dispatch_sqrt,
+ _foreach_doc,
+ _get_value,
+ _stack_if_compiling,
+ _use_grad_for_differentiable,
+)
__all__ = ["RAdam", "radam"]
@@ -17,6 +26,7 @@
betas=(0.9, 0.999),
eps=1e-8,
weight_decay=0,
+ decoupled_weight_decay: bool = False,
*,
foreach: Optional[bool] = None,
differentiable: bool = False,
@@ -37,6 +47,7 @@
eps=eps,
weight_decay=weight_decay,
foreach=foreach,
+ decoupled_weight_decay=decoupled_weight_decay,
differentiable=differentiable,
)
super().__init__(params, defaults)
@@ -46,6 +57,7 @@
for group in self.param_groups:
group.setdefault("foreach", None)
group.setdefault("differentiable", False)
+ group.setdefault("decoupled_weight_decay", False)
state_values = list(self.state.values())
step_is_tensor = (len(state_values) != 0) and torch.is_tensor(
state_values[0]["step"]
@@ -115,6 +127,7 @@
eps=group["eps"],
foreach=group["foreach"],
differentiable=group["differentiable"],
+ decoupled_weight_decay=group["decoupled_weight_decay"],
)
return loss
@@ -128,15 +141,19 @@
&\textbf{input} : \gamma \text{ (lr)}, \: \beta_1, \beta_2
\text{ (betas)}, \: \theta_0 \text{ (params)}, \:f(\theta) \text{ (objective)}, \:
\lambda \text{ (weightdecay)}, \\
- &\hspace{13mm} \epsilon \text{ (epsilon)} \\
+ &\hspace{13mm} \epsilon \text{ (epsilon)}, \textit{decoupled\_weight\_decay} \\
&\textbf{initialize} : m_0 \leftarrow 0 \text{ ( first moment)},
v_0 \leftarrow 0 \text{ ( second moment)}, \\
&\hspace{18mm} \rho_{\infty} \leftarrow 2/(1-\beta_2) -1 \\[-1.ex]
&\rule{110mm}{0.4pt} \\
&\textbf{for} \: t=1 \: \textbf{to} \: \ldots \: \textbf{do} \\
- &\hspace{6mm}g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\
- &\hspace{5mm} \textbf{if} \: \lambda \neq 0 \\
- &\hspace{10mm} g_t \leftarrow g_t + \lambda \theta_{t-1} \\
+ &\hspace{6mm} g_t \leftarrow \nabla_{\theta} f_t (\theta_{t-1}) \\
+ &\hspace{6mm} \theta_t \leftarrow \theta_{t-1} \\
+ &\hspace{6mm} \textbf{if} \: \lambda \neq 0 \\
+ &\hspace{12mm}\textbf{if} \: \textit{decoupled\_weight\_decay} \\
+ &\hspace{18mm} \theta_t \leftarrow \theta_{t} - \gamma \lambda \theta_{t} \\
+ &\hspace{12mm}\textbf{else} \\
+ &\hspace{18mm} g_t \leftarrow g_t + \lambda \theta_{t} \\
&\hspace{6mm}m_t \leftarrow \beta_1 m_{t-1} + (1 - \beta_1) g_t \\
&\hspace{6mm}v_t \leftarrow \beta_2 v_{t-1} + (1-\beta_2) g^2_t \\
&\hspace{6mm}\widehat{m_t} \leftarrow m_t/\big(1-\beta_1^t \big) \\
@@ -146,9 +163,9 @@
&\hspace{12mm} l_t \leftarrow \frac{\sqrt{ (1-\beta^t_2) }}{ \sqrt{v_t} +\epsilon } \\
&\hspace{12mm} r_t \leftarrow
\sqrt{\frac{(\rho_t-4)(\rho_t-2)\rho_{\infty}}{(\rho_{\infty}-4)(\rho_{\infty}-2) \rho_t}} \\
- &\hspace{12mm}\theta_t \leftarrow \theta_{t-1} - \gamma \widehat{m_t} r_t l_t \\
+ &\hspace{12mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t} r_t l_t \\
&\hspace{6mm}\textbf{else} \\
- &\hspace{12mm}\theta_t \leftarrow \theta_{t-1} - \gamma \widehat{m_t} \\
+ &\hspace{12mm}\theta_t \leftarrow \theta_t - \gamma \widehat{m_t} \\
&\rule{110mm}{0.4pt} \\[-1.ex]
&\bf{return} \: \theta_t \\[-1.ex]
&\rule{110mm}{0.4pt} \\[-1.ex]
@@ -156,9 +173,13 @@
For further details regarding the algorithm we refer to `On the variance of the adaptive learning rate and beyond`_.
- This implementation uses the same weight_decay implementation as Adam (were the weight_decay is applied
- to the gradient) and not the one from AdamW (were weight_decay is applied to the update). This
- is different from the `author's implementation`_.
+ This implementation provides an option to use either the original weight_decay implementation as in Adam
+ (where the weight_decay is applied to the gradient) or the one from AdamW (where weight_decay is applied
+ to the weight) through the decoupled_weight_decay option. When decoupled_weight_decay is set to False
+ (default), it uses the original Adam style weight decay, otherwise, it uses the AdamW style which
+ corresponds more closely to the `author's implementation`_ in the RAdam paper. Further information
+ about decoupled weight decay can be found in `Decoupled Weight Decay Regularization`_.
+
""" + fr"""
Args:
params (iterable): iterable of parameters to optimize or dicts defining
@@ -169,6 +190,8 @@
eps (float, optional): term added to the denominator to improve
numerical stability (default: 1e-8)
weight_decay (float, optional): weight decay (L2 penalty) (default: 0)
+ decoupled_weight_decay (bool, optional): whether to use decoupled weight
+ decay as in AdamW to obtain RAdamW (default: False)
{_foreach_doc}
{_differentiable_doc}
@@ -176,6 +199,8 @@
https://arxiv.org/abs/1908.03265
.. _author's implementation:
https://github.com/LiyuanLucasLiu/RAdam
+ .. _Decoupled Weight Decay Regularization:
+ https://arxiv.org/abs/1711.05101
"""
@@ -188,6 +213,7 @@
state_steps: List[Tensor],
# kwonly args with defaults are not supported by functions compiled with torchscript issue #70627
# setting this as kwarg for now as functional API is compiled by torch/distributed/optim
+ decoupled_weight_decay: bool = False,
foreach: Optional[bool] = None,
differentiable: bool = False,
*,
@@ -229,6 +255,7 @@
lr=lr,
weight_decay=weight_decay,
eps=eps,
+ decoupled_weight_decay=decoupled_weight_decay,
differentiable=differentiable,
)
@@ -246,6 +273,7 @@
weight_decay: float,
eps: float,
differentiable: bool,
+ decoupled_weight_decay: bool,
):
for i, param in enumerate(params):
@@ -261,7 +289,10 @@
bias_correction2 = 1 - beta2 ** step
if weight_decay != 0:
- grad = grad.add(param, alpha=weight_decay)
+ if decoupled_weight_decay:
+ param.mul_(1 - lr * weight_decay)
+ else:
+ grad = grad.add(param, alpha=weight_decay)
# Decay the first and second moment running average coefficient
exp_avg.lerp_(grad, 1 - beta1)
@@ -306,6 +337,7 @@
lr: float,
weight_decay: float,
eps: float,
+ decoupled_weight_decay: bool,
differentiable: bool,
):
@@ -332,7 +364,10 @@
(1 - beta2 ** _get_value(step)) for step in grouped_state_steps]
if weight_decay != 0:
- grouped_grads = torch._foreach_add(grouped_grads, grouped_params, alpha=weight_decay)
+ if decoupled_weight_decay:
+ torch._foreach_mul_(grouped_params, 1 - lr * weight_decay)
+ else:
+ grouped_grads = torch._foreach_add(grouped_grads, grouped_params, alpha=weight_decay)
# Decay the first and second moment running average coefficient
torch._foreach_lerp_(grouped_exp_avgs, grouped_grads, 1 - beta1)
diff --git a/torch/optim/radam.pyi b/torch/optim/radam.pyi
index e0bbd53..19750ea 100644
--- a/torch/optim/radam.pyi
+++ b/torch/optim/radam.pyi
@@ -10,4 +10,5 @@
betas: Tuple[float, float] = ...,
eps: float = ...,
weight_decay: float = ...,
+ decoupled_weight_decay: bool = ...,
) -> None: ...
