调用方法:
torch.optim.SGD(params, lr=<required parameter>, momentum=0, dampening=0, weight_decay=0, nesterov=False)
momentum: 动量参数
dampen ing:梯度抑制参数
weight_cay:L2的参数
nesterov:是否使用neterov动量
def step(self, closure=None):
"""Performs a single optimization step.
Arguments:
closure (callable, optional): A closure that reevaluates the model
and returns the loss.
"""
loss = None
if closure is not None:
loss = closure()
for group in self.param_groups:
weight_decay = group['weight_decay'] # 权重衰减系数
momentum = group['momentum'] # 动量因子,0.9或0.8
dampening = group['dampening'] # 梯度抑制因子
nesterov = group['nesterov'] # 是否使用nesterov动量
for p in group['params']:
if p.grad is None:
continue
d_p = p.grad.data
if weight_decay != 0: # 进行正则化
# add_表示原处改变,d_p = d_p + weight_decay*p.data
d_p.add_(weight_decay, p.data)
if momentum != 0:
param_state = self.state[p] # 之前的累计的数据,v(t-1)
# 进行动量累计计算
if 'momentum_buffer' not in param_state:
buf = param_state['momentum_buffer'] = torch.clone(d_p).detach()
else:
# 之前的动量
buf = param_state['momentum_buffer']
# buf= buf*momentum + (1-dampening)*d_p
buf.mul_(momentum).add_(1 - dampening, d_p)
if nesterov: # 使用neterov动量
# d_p= d_p + momentum*buf
d_p = d_p.add(momentum, buf)
else:
d_p = buf
# p = p - lr*d_p
p.data.add_(-group['lr'], d_p)
return loss
Pytorch中的SGD采用下述方式
也即
代码中,
d_p = d_p + weight_decayp.data # 权重衰减,这里实际上是做的 L2正则
buf = bufmomentum + (1-dampening)d_p # 计算动量,即v
若采用nesterov动量
d_p= d_p + momentumbuf
否则 d_p = buf
最后更新
p = p - lr*d_p
说明一下,L2正则的地方
l
o
s
s
=
∣
∣
y
−
y
∗
∣
∣
+
λ
2
∣
∣
w
∣
∣
2
loss=||y-y^{*}||+\frac{\lambda}{2}||w||^2
loss=∣∣y−y∗∣∣+2λ∣∣w∣∣2
求完导数后
∂
l
o
s
s
∂
w
=
∂
∣
∣
y
−
y
∗
∣
∣
∂
w
+
λ
∗
w
\frac{\partial loss}{\partial w}=\frac{\partial ||y-y^{*}||}{\partial w}+\lambda*w
∂w∂loss=∂w∂∣∣y−y∗∣∣+λ∗w
上式右边的前半部分在loss.backward()已经求出,后半部分既是代码中进行正则化的部分,计算完毕再进行参数更新。weight_decay即是
λ
\lambda
λ.
普通不带动量的SGD中权重衰减和L2正则是等价的。
在这里,还需要注意,采用d_p = p.grad.data修改时,会直接修改p.grad.data的数据