import torch as t
 
 
class LeNet(t.nn.Module):
 def __init__(self):
  super(LeNet, self).__init__()
  self.features = t.nn.Sequential(
   t.nn.Conv2d(3, 6, 5),
   t.nn.ReLU(),
   t.nn.MaxPool2d(2, 2),
   t.nn.Conv2d(6, 16, 5),
   t.nn.ReLU(),
   t.nn.MaxPool2d(2, 2)
  )
  # 由于調整shape并不是一個class層，
  # 所以在涉及這種操作（非nn.Module操作）需要拆分為多個模型
  self.classifiter = t.nn.Sequential(
   t.nn.Linear(16*5*5, 120),
   t.nn.ReLU(),
   t.nn.Linear(120, 84),
   t.nn.ReLU(),
   t.nn.Linear(84, 10)
  )
 
 def forward(self, x):
  x = self.features(x)
  x = x.view(-1, 16*5*5)
  x = self.classifiter(x)
  return x
 
net = LeNet()

from torch import optim
 
# 通常的step優化過程
optimizer = optim.SGD(params=net.parameters(), lr=1)
optimizer.zero_grad() # net.zero_grad()
 
input_ = t.autograd.Variable(t.randn(1, 3, 32, 32))
output = net(input_)
output.backward(output)
 
optimizer.step()

# # 直接對不同的網絡模塊制定不同學習率
optimizer = optim.SGD([{'params': net.features.parameters()}, # 默認lr是1e-5
      {'params': net.classifiter.parameters(), 'lr': 1e-2}], lr=1e-5)

# # 以層為單位，為不同層指定不同的學習率
# ## 提取指定層對象
special_layers = t.nn.ModuleList([net.classifiter[0], net.classifiter[3]])
# ## 獲取指定層參數id
special_layers_params = list(map(id, special_layers.parameters()))
print(special_layers_params)
# ## 獲取非指定層的參數id
base_params = filter(lambda p: id(p) not in special_layers_params, net.parameters())
optimizer = t.optim.SGD([{'params': base_params},
       {'params': special_layers.parameters(), 'lr': 0.01}], lr=0.001)

'''調整學習率'''
# 新建optimizer或者修改optimizer.params_groups對應的學習率
# # 新建optimizer更簡單也更推薦，optimizer十分輕量級，所以開銷很小
# # 但是新的優化器會初始化動量等狀態信息，這對于使用動量的優化器（momentum參數的sgd）可能會造成收斂中的震蕩
# ## optimizer.param_groups:長度2的list，optimizer.param_groups[0]：長度6的字典
print(optimizer.param_groups[0]['lr'])
old_lr = 0.1
optimizer = optim.SGD([{'params': net.features.parameters()},
      {'params': net.classifiter.parameters(), 'lr': old_lr*0.1}], lr=1e-5)

import torch
from torch.optim.optimizer import Optimizer, required

class SGD(Optimizer):
 def __init__(self, params, lr=required, momentum=0, dampening=0, weight_decay1=0, weight_decay2=0, nesterov=False):
  defaults = dict(lr=lr, momentum=momentum, dampening=dampening,
      weight_decay1=weight_decay1, weight_decay2=weight_decay2, nesterov=nesterov)
  if nesterov and (momentum <= 0 or dampening != 0):
   raise ValueError("Nesterov momentum requires a momentum and zero dampening")
  super(SGD, self).__init__(params, defaults)

 def __setstate__(self, state):
  super(SGD, self).__setstate__(state)
  for group in self.param_groups:
   group.setdefault('nesterov', False)

 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_decay1 = group['weight_decay1']
   weight_decay2 = group['weight_decay2']
   momentum = group['momentum']
   dampening = group['dampening']
   nesterov = group['nesterov']

   for p in group['params']:
    if p.grad is None:
     continue
    d_p = p.grad.data
    if weight_decay1 != 0:
     d_p.add_(weight_decay1, torch.sign(p.data))
    if weight_decay2 != 0:
     d_p.add_(weight_decay2, p.data)
    if momentum != 0:
     param_state = self.state[p]
     if 'momentum_buffer' not in param_state:
      buf = param_state['momentum_buffer'] = torch.zeros_like(p.data)
      buf.mul_(momentum).add_(d_p)
     else:
      buf = param_state['momentum_buffer']
      buf.mul_(momentum).add_(1 - dampening, d_p)
     if nesterov:
      d_p = d_p.add(momentum, buf)
     else:
      d_p = buf

    p.data.add_(-group['lr'], d_p)

  return loss