Adaptive Nesterov Momentum Algorithm (Adan) Optimizer.
This file contains code adapted from the Adan optimizer implementation in PyTorch:
https://github.com/sail-sg/Adan
The remainder of this file is unmodified from the original source, including the license:
Copyright 2022 Garena Online Private Limited
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.
Adan(params, lr=0.001, betas=(0.98, 0.92, 0.99), eps=1e-08, weight_decay=0.0, max_grad_norm=0.0, no_prox=False, foreach=True, fused=False)
Bases: Optimizer
Implements a pytorch variant of Adan
Adan was proposed in
Adan: Adaptive Nesterov Momentum Algorithm for
Faster Optimizing Deep Models[J].arXiv preprint arXiv:2208.06677, 2022.
https://arxiv.org/abs/2208.06677
Arguments:
params (iterable): iterable of parameters to optimize or
dicts defining parameter groups.
lr (float, optional): learning rate. (default: 1e-3)
betas (Tuple[float, float, flot], optional): coefficients used for
first- and second-order moments. (default: (0.98, 0.92, 0.99))
eps (float, optional): term added to the denominator to improve
numerical stability. (default: 1e-8)
weight_decay (float, optional): decoupled weight decay
(L2 penalty) (default: 0)
max_grad_norm (float, optional): value used to clip
global grad norm (default: 0.0 no clip)
no_prox (bool): how to perform the decoupled weight decay
(default: False)
foreach (bool): if True would use torch._foreach implementation.
It's faster but uses slightly more memory. (default: True)
fused (bool, optional): whether fused implementation is used.
(default: False)
Source code in manify/optimizers/_adan.py
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111 | def __init__(
self,
params,
lr=1e-3,
betas=(0.98, 0.92, 0.99),
eps=1e-8,
weight_decay=0.0,
max_grad_norm=0.0,
no_prox=False,
foreach: bool = True,
fused: bool = False,
):
if not 0.0 <= max_grad_norm:
raise ValueError("Invalid Max grad norm: {}".format(max_grad_norm))
if not 0.0 <= lr:
raise ValueError("Invalid learning rate: {}".format(lr))
if not 0.0 <= eps:
raise ValueError("Invalid epsilon value: {}".format(eps))
if not 0.0 <= betas[0] < 1.0:
raise ValueError("Invalid beta parameter at index 0: {}".format(betas[0]))
if not 0.0 <= betas[1] < 1.0:
raise ValueError("Invalid beta parameter at index 1: {}".format(betas[1]))
if not 0.0 <= betas[2] < 1.0:
raise ValueError("Invalid beta parameter at index 2: {}".format(betas[2]))
if fused:
_check_fused_available()
defaults = dict(
lr=lr,
betas=betas,
eps=eps,
weight_decay=weight_decay,
max_grad_norm=max_grad_norm,
no_prox=no_prox,
foreach=foreach,
fused=fused,
)
super().__init__(params, defaults)
|
step(closure=None)
Performs a single optimization step.
Source code in manify/optimizers/_adan.py
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241 | @torch.no_grad()
def step(self, closure=None):
"""Performs a single optimization step."""
loss = None
if closure is not None:
with torch.enable_grad():
loss = closure()
if self.defaults["max_grad_norm"] > 0:
device = self.param_groups[0]["params"][0].device
global_grad_norm = torch.zeros(1, device=device)
max_grad_norm = torch.tensor(self.defaults["max_grad_norm"], device=device)
for group in self.param_groups:
for p in group["params"]:
if p.grad is not None:
grad = p.grad
global_grad_norm.add_(grad.pow(2).sum())
global_grad_norm = torch.sqrt(global_grad_norm)
clip_global_grad_norm = torch.clamp(max_grad_norm / (global_grad_norm + group["eps"]), max=1.0).item()
else:
clip_global_grad_norm = 1.0
for group in self.param_groups:
params_with_grad = []
grads = []
exp_avgs = []
exp_avg_sqs = []
exp_avg_diffs = []
neg_pre_grads = []
beta1, beta2, beta3 = group["betas"]
# assume same step across group now to simplify things
# per parameter step can be easily support
# by making it tensor, or pass list into kernel
if "step" in group:
group["step"] += 1
else:
group["step"] = 1
bias_correction1 = 1.0 - beta1 ** group["step"]
bias_correction2 = 1.0 - beta2 ** group["step"]
bias_correction3 = 1.0 - beta3 ** group["step"]
for p in group["params"]:
if p.grad is None:
continue
params_with_grad.append(p)
grads.append(p.grad)
state = self.state[p]
if len(state) == 0:
state["exp_avg"] = torch.zeros_like(p)
state["exp_avg_sq"] = torch.zeros_like(p)
state["exp_avg_diff"] = torch.zeros_like(p)
if "neg_pre_grad" not in state or group["step"] == 1:
state["neg_pre_grad"] = p.grad.clone().mul_(-clip_global_grad_norm)
exp_avgs.append(state["exp_avg"])
exp_avg_sqs.append(state["exp_avg_sq"])
exp_avg_diffs.append(state["exp_avg_diff"])
neg_pre_grads.append(state["neg_pre_grad"])
if not params_with_grad:
continue
kwargs = dict(
params=params_with_grad,
grads=grads,
exp_avgs=exp_avgs,
exp_avg_sqs=exp_avg_sqs,
exp_avg_diffs=exp_avg_diffs,
neg_pre_grads=neg_pre_grads,
beta1=beta1,
beta2=beta2,
beta3=beta3,
bias_correction1=bias_correction1,
bias_correction2=bias_correction2,
bias_correction3_sqrt=math.sqrt(bias_correction3),
lr=group["lr"],
weight_decay=group["weight_decay"],
eps=group["eps"],
no_prox=group["no_prox"],
clip_global_grad_norm=clip_global_grad_norm,
)
if group["foreach"]:
if group["fused"]:
if torch.cuda.is_available():
_fused_adan_multi_tensor(**kwargs)
else:
raise ValueError("Fused Adan does not support CPU")
else:
_multi_tensor_adan(**kwargs)
elif group["fused"]:
if torch.cuda.is_available():
_fused_adan_single_tensor(**kwargs)
else:
raise ValueError("Fused Adan does not support CPU")
else:
_single_tensor_adan(**kwargs)
return loss
|