65 lines
2.2 KiB
Python
65 lines
2.2 KiB
Python
import torch
|
||
from utils import *
|
||
|
||
|
||
# 手动实现torch.optim.SGD
|
||
class My_SGD:
|
||
def __init__(self, params: list[torch.Tensor], lr: float, weight_decay=0.0, momentum=0.0):
|
||
self.params = params
|
||
self.lr = lr
|
||
self.weight_decay = weight_decay
|
||
self.momentum = momentum
|
||
self.velocities = [torch.zeros_like(param.data) for param in params]
|
||
|
||
def step(self):
|
||
with torch.no_grad():
|
||
for index, param in enumerate(self.params):
|
||
if param.grad is not None:
|
||
if self.weight_decay > 0:
|
||
if len(param.data.shape) > 1:
|
||
param.grad.data = (param.grad.data + self.weight_decay * param.data)
|
||
self.velocities[index] = (self.momentum * self.velocities[index] - self.lr * param.grad)
|
||
param.data = param.data + self.velocities[index]
|
||
|
||
def zero_grad(self):
|
||
for param in self.params:
|
||
if param.grad is not None:
|
||
param.grad.data = torch.zeros_like(param.grad.data)
|
||
|
||
|
||
if __name__ == "__main__":
|
||
params1 = torch.tensor([[1.0, 2.0]], requires_grad=True)
|
||
params2 = torch.tensor([[1.0, 2.0]], requires_grad=True)
|
||
|
||
my_sgd = My_SGD(params=[params1], lr=0.5, momentum=1)
|
||
optim_sgd = torch.optim.SGD(params=[params2], lr=0.5, momentum=1)
|
||
my_sgd.zero_grad()
|
||
optim_sgd.zero_grad()
|
||
|
||
loss1 = 2 * params1.sum()
|
||
loss2 = 2 * params2.sum()
|
||
# 偏导为2
|
||
loss1.backward()
|
||
loss2.backward()
|
||
my_sgd.step()
|
||
optim_sgd.step()
|
||
# 结果为：w - lr * grad + momentum * velocity
|
||
# w[0] = 1 - 0.5 * 2 + 1 * 0 = 0
|
||
# w[1] = 2 - 0.5 * 2 + 1 * 0 = 1
|
||
print("My_SGD第1次反向传播结果：\n", params1.data)
|
||
print("torch.optim.SGD第1次反向传播结果：\n", params2.data)
|
||
|
||
my_sgd.zero_grad()
|
||
optim_sgd.zero_grad()
|
||
loss1 = -3 * params1.sum()
|
||
loss2 = -3 * params2.sum()
|
||
loss1.backward()
|
||
loss2.backward()
|
||
my_sgd.step()
|
||
optim_sgd.step()
|
||
# 结果为：w - lr * grad + momentum * velocity
|
||
# w[0] = 0 - 0.5 * -3 + 1 * (-0.5 * 2) = 0.5
|
||
# w[1] = 1 - 0.5 * -3 + 1 * (-0.5 * 2) = 1.5
|
||
print("My_SGD第2次反向传播结果：\n", params1.data)
|
||
print("torch.optim.SGD第2次反向传播结果：\n", params2.data)
|