65 lines
2.3 KiB
Python
65 lines
2.3 KiB
Python
import torch
|
||
from utils import *
|
||
|
||
|
||
class My_Adam:
|
||
def __init__(self, params: list[torch.Tensor], lr=1e-3, betas=(0.9, 0.999), eps=1e-8):
|
||
self.params = params
|
||
self.lr = lr
|
||
self.beta1 = betas[0]
|
||
self.beta2 = betas[1]
|
||
self.eps = eps
|
||
self.t = 0
|
||
self.momentums = [torch.zeros_like(param.data) for param in params]
|
||
self.velocities = [torch.zeros_like(param.data) for param in params]
|
||
|
||
def step(self):
|
||
self.t += 1
|
||
with torch.no_grad():
|
||
for index, param in enumerate(self.params):
|
||
if param.grad is not None:
|
||
self.momentums[index] = (self.beta1 * self.momentums[index] + (1 - self.beta1) * param.grad)
|
||
self.velocities[index] = (self.beta2 * self.velocities[index] + (1 - self.beta2) * param.grad ** 2)
|
||
|
||
momentums_hat = self.momentums[index] / (1 - self.beta1 ** self.t)
|
||
velocities_hat = self.velocities[index] / (1 - self.beta2 ** self.t)
|
||
|
||
param.data = param.data - self.lr * momentums_hat / (torch.sqrt(velocities_hat) + self.eps)
|
||
|
||
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_Adam(params=[params1], lr=1, betas=(0.5, 0.5), eps=1e-8)
|
||
optim_sgd = torch.optim.Adam(params=[params2], lr=1, betas=(0.5, 0.5), eps=1e-8)
|
||
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()
|
||
print("My_Adam第1次反向传播结果：\n", params1.data)
|
||
print("torch.optim.Adam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()
|
||
# 偏导为-3
|
||
loss1.backward()
|
||
loss2.backward()
|
||
my_sgd.step()
|
||
optim_sgd.step()
|
||
print("My_Adam第2次反向传播结果：\n", params1.data)
|
||
print("torch.optim.Adam第2次反向传播结果：\n", params2.data)
|