Bases: OptiModule
Adaptive Moment Estimation (Adam) module.
This module implements the following algorithm:
Init(beta_1, beta_2, eps):
state_m = 0
state_v = 0
Step(grads, step):
state_m = beta_1*state_m + (1-beta_1)*grads
state_v = beta_2*state_v + (1-beta_2)*(grads**2)
m_hat = state_m / (1 - beta_1**step)
v_hat = state_v / (1 - beta_2**step)
grads = state_m / (sqrt(v_hat) + eps)
In other words, gradients are first momentum-corrected, as
is the accumulated sum of squared past gradients. Both are
bias-corrected, then the former are scaled down based upon
the latter AdaGrad-style (indirectly adapting the learning
rate) and returned. This is the Adam [1] algorithm.
Optionally, the AMSGrad [2] algorithm may be implemented,
with a similar formula but using the element-wise maximum
of present-and-past v_hat values as a scaling factor. This
guarantees that the learning rate is shrinked across time,
at least from the point of view of this module (a warm-up
schedule might for example counteract this).
References
- [1]
Kingma and Ba, 2014.
Adam: A Method for Stochastic Optimization.
https://arxiv.org/abs/1412.6980
- [2]
Reddi et al., 2018.
On the Convergence of Adam and Beyond.
https://arxiv.org/abs/1904.09237
Source code in declearn/optimizer/modules/_adaptive.py
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
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293 | class AdamModule(OptiModule):
"""Adaptive Moment Estimation (Adam) module.
This module implements the following algorithm:
Init(beta_1, beta_2, eps):
state_m = 0
state_v = 0
Step(grads, step):
state_m = beta_1*state_m + (1-beta_1)*grads
state_v = beta_2*state_v + (1-beta_2)*(grads**2)
m_hat = state_m / (1 - beta_1**step)
v_hat = state_v / (1 - beta_2**step)
grads = state_m / (sqrt(v_hat) + eps)
In other words, gradients are first momentum-corrected, as
is the accumulated sum of squared past gradients. Both are
bias-corrected, then the former are scaled down based upon
the latter AdaGrad-style (indirectly adapting the learning
rate) and returned. This is the Adam [1] algorithm.
Optionally, the AMSGrad [2] algorithm may be implemented,
with a similar formula but using the element-wise maximum
of present-and-past v_hat values as a scaling factor. This
guarantees that the learning rate is shrinked across time,
at least from the point of view of this module (a warm-up
schedule might for example counteract this).
References
----------
- [1]
Kingma and Ba, 2014.
Adam: A Method for Stochastic Optimization.
https://arxiv.org/abs/1412.6980
- [2]
Reddi et al., 2018.
On the Convergence of Adam and Beyond.
https://arxiv.org/abs/1904.09237
"""
name: ClassVar[str] = "adam"
def __init__(
self,
beta_1: float = 0.9,
beta_2: float = 0.99,
amsgrad: bool = False,
eps: float = 1e-7,
) -> None:
"""Instantiate the Adam gradients-adaptation module.
Parameters
----------
beta_1: float
Beta parameter for the momentum correction
applied to the input gradients.
beta_2: float
Beta parameter for the momentum correction
applied to the adaptive scaling term.
amsgrad: bool, default=False
Whether to implement the AMSGrad algorithm
rather than the base Adam one.
eps: float, default=1e-7
Numerical-stability improvement term, added
to the (divisor) adapative scaling term.
"""
self.ewma_1 = EWMAModule(beta=beta_1)
self.ewma_2 = EWMAModule(beta=beta_2)
self.steps = 0
self.eps = eps
self.amsgrad = amsgrad
self.vmax = None # type: Optional[Vector]
def get_config(
self,
) -> Dict[str, Any]:
return {
"beta_1": self.ewma_1.beta,
"beta_2": self.ewma_2.beta,
"amsgrad": self.amsgrad,
"eps": self.eps,
}
def run(
self,
gradients: Vector,
) -> Vector:
# Compute momentum-corrected state variables.
m_t = self.ewma_1.run(gradients)
v_t = self.ewma_2.run(gradients**2)
# Apply bias correction to the previous terms.
m_h = m_t / (1 - (self.ewma_1.beta ** (self.steps + 1)))
v_h = v_t / (1 - (self.ewma_2.beta ** (self.steps + 1)))
# Optionally implement the AMSGrad algorithm.
if self.amsgrad:
if self.vmax is not None:
v_h = v_h.maximum(self.vmax)
self.vmax = v_h
# Compute and return the adapted gradients.
gradients = m_h / ((v_h**0.5) + self.eps)
self.steps += 1
return gradients
def get_state(
self,
) -> Dict[str, Any]:
state = {
"steps": self.steps,
"vmax": self.vmax,
} # type: Dict[str, Any]
state["momentum"] = self.ewma_1.get_state()
state["velocity"] = self.ewma_2.get_state()
return state
def set_state(
self,
state: Dict[str, Any],
) -> None:
for key in ("momentum", "velocity", "steps", "vmax"):
if key not in state:
raise KeyError(f"Missing required state variable '{key}'.")
self.ewma_1.set_state(state["momentum"])
self.ewma_2.set_state(state["velocity"])
self.steps = state["steps"]
self.vmax = state["vmax"]
|
__init__(beta_1=0.9, beta_2=0.99, amsgrad=False, eps=1e-07)
Instantiate the Adam gradients-adaptation module.
Parameters:
Name |
Type |
Description |
Default |
beta_1 |
float
|
Beta parameter for the momentum correction
applied to the input gradients. |
0.9
|
beta_2 |
float
|
Beta parameter for the momentum correction
applied to the adaptive scaling term. |
0.99
|
amsgrad |
bool
|
Whether to implement the AMSGrad algorithm
rather than the base Adam one. |
False
|
eps |
float
|
Numerical-stability improvement term, added
to the (divisor) adapative scaling term. |
1e-07
|
Source code in declearn/optimizer/modules/_adaptive.py
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 | def __init__(
self,
beta_1: float = 0.9,
beta_2: float = 0.99,
amsgrad: bool = False,
eps: float = 1e-7,
) -> None:
"""Instantiate the Adam gradients-adaptation module.
Parameters
----------
beta_1: float
Beta parameter for the momentum correction
applied to the input gradients.
beta_2: float
Beta parameter for the momentum correction
applied to the adaptive scaling term.
amsgrad: bool, default=False
Whether to implement the AMSGrad algorithm
rather than the base Adam one.
eps: float, default=1e-7
Numerical-stability improvement term, added
to the (divisor) adapative scaling term.
"""
self.ewma_1 = EWMAModule(beta=beta_1)
self.ewma_2 = EWMAModule(beta=beta_2)
self.steps = 0
self.eps = eps
self.amsgrad = amsgrad
self.vmax = None # type: Optional[Vector]
|