Pytorch：为什么 print(model) 不显示激活函数？

Question

我需要从 pytorch 中经过训练的神经网络中提取权重、偏差和至少激活函数的类型。

我知道要提取权重和偏差，命令是：

model.parameters()

但我不知道如何提取层上使用的激活函数。这是我的网络

class NetWithODE(torch.nn.Module):
    def __init__(self, n_feature, n_hidden, n_output, sampling_interval, scaler_features):
        super(NetWithODE, self).__init__()
        self.hidden = torch.nn.Linear(n_feature, n_hidden)  # hidden layer
        self.predict = torch.nn.Linear(n_hidden, n_output)  # output layer
        self.sampling_interval = sampling_interval
        self.device = torch.device("cpu")
        self.dtype = torch.float
        self.scaler_features = scaler_features

    def forward(self, x):
        x0 = x.clone().requires_grad_(True)
        # activation function for hidden layer
        x = F.relu(self.hidden(x))
        # linear output, here r should be the output
        r = self.predict(x)
        # Now the r enters the integrator
        x = self.integrate(r, x0)

        return x

    def integrate(self, r, x0):
        # RK4 steps per interval
        M = 4
        DT = self.sampling_interval / M
        X = x0

        for j in range(M):
            k1 = self.ode(X, r)
            k2 = self.ode(X + DT / 2 * k1, r)
            k3 = self.ode(X + DT / 2 * k2, r)
            k4 = self.ode(X + DT * k3, r)
            X = X + DT / 6 * (k1 + 2 * k2 + 2 * k3 + k4)

        return X

    def ode(self, x0, r):
        qF = r[0, 0]
        qA = r[0, 1]
        qP = r[0, 2]
        mu = r[0, 3]

        FRU = x0[0, 0]
        AMC = x0[0, 1]
        PHB = x0[0, 2]
        TBM = x0[0, 3]

        fFRU = qF * TBM  
        fAMC = qA * TBM  
        fPHB = qP - mu * PHB
        fTBM = mu * TBM

        return torch.stack((fFRU, fAMC, fPHB, fTBM), 0)

如果我运行命令

print(model)

我明白了

NetWithODE(
  (hidden): Linear(in_features=4, out_features=10, bias=True)
  (predict): Linear(in_features=10, out_features=4, bias=True)
)

但是我在哪里可以获得激活函数（在本例中为 Relu）？

我有pytorch 1.4。

Answer 1

有两种向网络图添加操作的方法：低级函数方法和更高级的对象方法。您需要后者使您的结构可观察，在第一种情况下只是调用（不完全是，但是......）一个函数而不存储有关它的信息。所以，而不是

    def forward(self, x):
    ...
        x = F.relu(self.hidden(x))

它一定是类似的东西

def __init__(...):
    ...
    self.myFirstRelu= torch.nn.ReLU()

def forward(self, x):
    ...
    x1 = self.hidden(x)
    x2 = self.myFirstRelu(x1)

无论如何，混合使用这两种方法通常是个坏主意，尽管即使torchvision模型也存在这样的不一致：models.inception_v3例如不注册池>:-(（编辑：它已于 2020 年 6 月修复，谢谢，mitmul！）。

UPD: - Thanks, that works, now if I print I see ReLU(). But this seems to only print the function in the same order they are defined in the init. Is there a way to get the associations between layers and activation functions? For example I want to know which activation was applyed to layer 1, which to layer 2 end so on...

There is no uniform way, but here is some tricks: object way:

-just init them in order

-use torch.nn.Sequential

-hook callbacks on nodes like that -

def hook( m, i, o):
    print( m._get_name() )

for ( mo ) in model.modules():
    mo.register_forward_hook(hook)

functional and object way:

-make use of internal model graph, builded on forward pass, as torchviz do (https://github.com/szagoruyko/pytorchviz/blob/master/torchviz/dot.py), or just use plot generated by said torchviz.