即使所有变量的 requires_grad = False，PyTorch 损失也会减少

Question

当我用 PyTorch 创建神经网络时，使用torch.nn.Sequential 方法定义层时，参数似乎默认有requires_grad = False。但是，当我训练这个网络时，损失会减少。如果没有通过渐变更新图层，这怎么可能？

例如，这是定义我的网络的代码：

class Network(torch.nn.Module):

def __init__(self):
    super(Network, self).__init__()
    self.layers = torch.nn.Sequential(
        torch.nn.Linear(10, 5),
        torch.nn.Linear(5, 2)
    )
    print('Network Parameters:')
    model_dict = self.state_dict()
    for param_name in model_dict:
        param = model_dict[param_name]
        print('Name: ' + str(param_name))
        print('\tRequires Grad: ' + str(param.requires_grad))

def forward(self, input):
    prediction = self.layers(input)
    return prediction

然后打印出来：

Network Parameters:
Name: layers.0.weight
    Requires Grad: False
Name: layers.0.bias
    Requires Grad: False
Name: layers.1.weight
    Requires Grad: False
Name: layers.1.bias
    Requires Grad: False

那么这是训练我的网络的代码：

network = Network()
network.train()
optimiser = torch.optim.SGD(network.parameters(), lr=0.001)
criterion = torch.nn.MSELoss()
inputs = np.random.random([100, 10]).astype(np.float32)
inputs = torch.from_numpy(inputs)
labels = np.random.random([100, 2]).astype(np.float32)
labels = torch.from_numpy(labels)


while True:
    prediction = network.forward(inputs)
    loss = criterion(prediction, labels)
    print('loss = ' + str(loss.item()))
    optimiser.zero_grad()
    loss.backward()
    optimiser.step()

然后打印出来：

loss = 0.284633219242
loss = 0.278225809336
loss = 0.271959483624
loss = 0.265835255384
loss = 0.259853869677
loss = 0.254015892744
loss = 0.248321473598
loss = 0.242770522833
loss = 0.237362638116
loss = 0.232097044587
loss = 0.226972639561
loss = 0.221987977624
loss = 0.217141270638
loss = 0.212430402637
loss = 0.207852959633
loss = 0.203406244516
loss = 0.199087426066
loss = 0.19489350915
loss = 0.190821439028
loss = 0.186868071556
loss = 0.183030322194
loss = 0.179305106401
loss = 0.175689414144
loss = 0.172180294991
loss = 0.168774917722
loss = 0.165470585227
loss = 0.162264674902
loss = 0.159154698253

如果所有参数都有requires_grad = False，为什么损失会减少？

Answer 1

这很有趣——state_dict() 和 parameters() 之间似乎有区别：

class Network(torch.nn.Module):
    def __init__(self):
        super(Network, self).__init__()
        self.layers = torch.nn.Sequential(
            torch.nn.Linear(10, 5),
            torch.nn.Linear(5, 2)
        )
        print(self.layers[0].weight.requires_grad) # True
        print(self.state_dict()['layers.0.weight'].requires_grad) # False
        print(list(self.parameters())[0].requires_grad) # True

    def forward(self, input):
        prediction = self.layers(input)
        return prediction

所以看起来你的损失正在减少，因为网络实际上正在学习，因为requires_grad 是 True。 （一般来说，为了调试，我更喜欢查询实际对象（self.layers[0]...）。

[EDIT] Ahah - 发现问题：有一个 keep_vars 布尔选项，您可以将其传递给 state_dict，它执行以下操作（除其他外）：(https://github.com/pytorch/pytorch/blob/master/torch/nn/modules/module.py#L665)

for name, param in self._parameters.items():
    if param is not None:
        destination[prefix + name] = param if keep_vars else param.data

所以，如果您想要实际的param，请使用keep_vars=True——如果您只想要数据，请使用默认的keep_vars=False。

所以：

print(self.layers[0].weight.requires_grad) # True
print(self.state_dict(keep_vars=True)['layers.0.weight'].requires_grad) # True
print(list(self.parameters())[0].requires_grad) # True