在固定某些层的多个 GPU 上训练单个 pytorch 模型？

Question

I met some problems when using pytorch DistributedDataParallel . The situation is:

1. My model is A , and it has been trained on a single GPU as usual. Suppose that there are three layers in A :

    class A(nn.module): def __init__(self): super(A,self).__init__() self.layer0 = layer0 self.layer1 = layer1 self.layer2 = layer2 def forward(self,x): x=self.layer0(x) x=self.layer1(x) x=self.layer2(x) return x

2. Now I have some new data. I want to fine-tune A with it on multiple GPUs. I need to wrap A as a multi-GPU model B .

3. But there are two training stages. In the 1st stage, I want to fix layer0 and layer1 of B . In the 2nd stage, only to fix layer0 . Then requires_grad of parameters in layer1 should be changed during training. However, DistributedDataParalleldoc says:

   You should never try to change your model's parameters after wrapping up your model with DistributedDataParallel.

In fact, I tried to use B.module to refer A wrapped in B . But the test results were abnormal compared to the single-GPU model. Maybe this way is disallowed.

What should I do? Is there any proper way to wrap my model? And what should be take care for when saving and loading the model?

Just run it on a single machine with multiple GPUs so you can ignore the distributed situation using multiple machines. Many thanks.

Update 2019.12.03

As suggested by @jodag, I tried DataParallel , but it didn't work. This time I didn't change anything in B (except training it) after wrapping it. For simplification, My code is like this (and I refered this ):

class B(nn.DataParallel):
     def __getattr__(self, name):
        try:
            return super().__getattr__(name)
        except AttributeError:
            return getattr(self.module, name)
a = A()
b = B(a,device_ids=[0,1])
b = b.cuda()
trained_param = b.layer2.parameters()
# trained_param = [{'params':b.layer2.parameters()},{'params':b.layer1.parameters()}]
optimizer = optim.Adam(trained_param)
b.train()
...
for x, label in data_loader:
    optimizer.zero_grad()
    x = x.to(0) # This line can be commented.
    y = b(x)
    l = loss(y, label)
    l.backword()
    optimizer.step()

Answer 1

If you only try to optimize part of the parameters, why not try controlling this via the optimizer , rather than the model?
You can leave your model as-is (wrapped in a DistributedDataParallel ) and pass only part of its parameters to the relevant optimizer.