def addParameterUpdateOps(self, model):
if self.optimizer not in OPTIMIZER_DICT:
raise Exception(
"Optimizer {} unknown. Valid choices are {}".format(
self.optimizer, ", ".join(OPTIMIZER_DICT.keys())))
optimizer_rule = OPTIMIZER_DICT[self.optimizer]
if optimizer_rule == GRAD_OPTIMIZER.SGD:
build_sgd(
model,
self.learning_rate,
gamma=self.lr_decay,
policy=self.lr_policy,
stepsize=1,
)
elif optimizer_rule == GRAD_OPTIMIZER.ADAGRAD:
build_adagrad(model, self.learning_rate)
elif optimizer_rule == GRAD_OPTIMIZER.ADAM:
build_adam(model, self.learning_rate)
elif optimizer_rule == GRAD_OPTIMIZER.FTRL:
build_ftrl(model, self.learning_rate)
else:
print("Unrecognized in caffe2 setting, using default SGD",
optimizer_rule)
build_sgd(model, self.learning_rate)
def build_optimizer(self, model):
build_ftrl(model,
engine=None,
alpha=1.0,
beta=0.1,
lambda1=0.0,
lambda2=0.0)
def AddParameterUpdateOps(
model, optimizer_input="SGD", base_learning_rate=0.01, *args, **kwargs
):
if optimizer_input not in OPTIMIZER_DICT:
raise Exception(
"Optimizer {} unknown. Valid choices are {}"
.format(optimizer_input, ', '.join(OPTIMIZER_DICT.keys()))
)
optimizer_rule = OPTIMIZER_DICT[optimizer_input]
if optimizer_rule == GRAD_OPTIMIZER.SGD:
build_sgd(
model,
base_learning_rate,
gamma=kwargs['gamma'],
policy=kwargs['policy'],
stepsize=1
)
elif optimizer_rule == GRAD_OPTIMIZER.ADAGRAD:
build_adagrad(model, base_learning_rate)
elif optimizer_rule == GRAD_OPTIMIZER.ADAM:
build_adam(model, base_learning_rate)
elif optimizer_rule == GRAD_OPTIMIZER.FTRL:
build_ftrl(model, base_learning_rate)
else:
print(
"Unrecognized in caffe2 setting, using default SGD", optimizer_rule
)
build_sgd(model, base_learning_rate)
def build_optimizer(self, model):
self._skip_gpu = True
return build_ftrl(model,
engine=None,
alpha=1.0,
beta=0.1,
lambda1=0.0,
lambda2=0.0)
def build_optimizer(self, model, **kwargs):
self._skip_gpu = True
return build_ftrl(
model,
engine=None,
alpha=1.0,
beta=0.1,
lambda1=0.0,
lambda2=0.0,
**kwargs
)
