如何有两个优化器，以便一个优化器训练整个参数，另一个训练部分参数?

Question

我有一个模型：

class MyModel(nn.Module):
def __init__(self):
super(MyModel, self).__init__()
self.conv1 = nn.Conv2d(128, 128, (3,3))
self.conv2 = nn.Conv2d(128, 256, (3,3))
self.conv3 = nn.Conv2d(256, 256, (3,3))
def forward(self,):
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
x = F.relu(self.conv3(x))
return x
model = MyModel()

我想以这样一种方式训练模型，即在每个训练步骤中DATA_X1都应该训练['conv1', 'conv2', 'conv3']层和DATA_X2应仅训练['conv3']层。

我尝试制作两个优化器：

# Full parameters train
all_params = model.parameters()
all_optimizer = optim.Adam(all_params, lr=0.01)
# Partial parameters train
partial_params = model.parameters()
for p, (name, param) in zip(list(partial_params), model.named_parameters()):
if name in ['conv3']:
p.requires_grad = True
else:
p.requires_grad = False
partial_optimizer = optim.Adam(partial_params, lr=0.01)

但这会影响优化器，required_grad = False

有什么办法可以做到这一点吗？

Answer 1

为什么不将此功能构建到模型中？

class MyModel(nn.Module):
def __init__(self):
super(MyModel, self).__init__()
self.conv1 = nn.Conv2d(128, 128, (3,3))
self.conv2 = nn.Conv2d(128, 256, (3,3))
self.conv3 = nn.Conv2d(256, 256, (3,3))
self.partial_grad = False  # a flag
def forward(self, x):
if self.partial_grad:
with torch.no_grad():
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))
else:
x = F.relu(self.conv1(x))
x = F.relu(self.conv2(x))     
x = F.relu(self.conv3(x))
return x

现在，您可以使用一个包含所有参数的优化器，并且可以根据您的训练数据打开和关闭model.partial_grad：

optimizer.zero_grad()
model.partial_grad = False  # prep for DATA_X1 training
x1, y1 = DATA_X1.item()  # this is not really a code, but you get the point
out = model(x1)
loss = criterion(out, y1)
loss.backward()
optimizer.step()  
# do a partial opt for DATA_X2
optimizer.zero_grad()
model.partial_grad = True  # prep for DATA_X2 training
x2, y2 = DATA_X2.item()  # this is not really a code, but you get the point
out = model(x2)
loss = criterion(out, y2)
loss.backward()
optimizer.step()  

拥有单个优化器应该更有益，因为您可以跟踪两个数据集中的动量和参数变化。