PyTorch 指定模型参数

Question

我正在尝试在 PyTorch 中创建一个卷积模型

• 固定一层（初始化为规定值）
• 学习了另一层（但初始猜测取自规定值）。

这是一个模型定义的示例代码：

import torch.nn as nn

class Net(nn.Module):
    def __init__(self, weights_fixed, weights_guess):
        super(Net, self).__init__()
        self.convL1 = nn.Conv1d(1, 3, 3, bias=False)
        self.convL1.weight = weights_fixed # I want to keep these weights fixed

        self.convL2 = nn.Conv1d(3, 1, 1, bias=False)
        self.convL1.weight = weights_guess # I want to learn these weights

    def forward(self, inp_batch):
        out1 = self.convL1(inp_batch)
        out2 = self.convL2(out1)

        return out2

及样品用途：

weights_fixed = ...
weights_guess = ...

model = Net(weights_fixed, weights_guess)

loss_fn = nn.CrossEntropyLoss()
optim = torch.optim.SGD(model.parameters(), lr=0.1, momentum=0.9)

train_dataset = ... #define training set here

for (X, y) in train_dataset:
    optim.zero_grad()
    out = model(X)
    loss = loss_fn(out, y)
    loss.backward()
    optim.step() 

我怎样才能使重量 weights_fixed - 固定和 weights_guess - 可学习？

我的猜测是 weights_fixed = nn.Parameter(W1,requires_grad=False) weights_guess = nn.Parameter(W2,requires_grad=True) 为了完整起见，在哪里 将 numpy 导入为 np 进口火炬

krnl = np.zeros((5,order+1))
krnl[:,0] = [ 0. , 1., 0. ]
krnl[:,1] = [-0.5, 0., 0.5]
krnl[:,2] = [ 1. ,-2., 1. ]
W1 = torch.tensor(krnl)

a = np.array((1.,2.,3.))
W2 = torch.tensor(a)

但我完全糊涂了。任何建议或参考将不胜感激。当然，我查看了 PyTorch 文档，但它并没有使我的理解更加清晰。

Answer 1

你可以这样做：

# this will be inside your class mostly
self.conv1.weight.requires_grad = False

这将是您定义优化器的地方：

optimizer = optim.SGD(filter(lambda p: p.requires_grad, net.parameters()), lr=0.1)

因此，优化器将只使用启用了渐变的参数。

Answer 2

只需将可学习的参数用nn.Parameter 包裹起来（requires_grad=True 是默认的，不需要指定这个），并且在没有nn.Parameter 包裹器的情况下将固定权重作为张量。

所有nn.Parameter的权重都会自动添加到net.parameters()，所以当你像optimizer = optim.SGD(net.parameters(), lr=0.01)那样进行训练时，固定的权重不会改变。

所以基本上是这样的：

weights_fixed = W1
weights_guess = nn.Parameter(W2)

Answer 3

您可以只将您想学习的参数传递给优化器：

optim = torch.optim.SGD(model.convL2.parameters(), lr=0.1, momentum=0.9)
# Now optimizer bypass parameters from convL1

如果您的模型有更多层，则必须将参数转换为列表：

params_to_update = list(model.convL2.parameters()) + list(model.convL3.parameters())
optim = torch.optim.SGD(params_to_update, lr=0.1, momentum=0.9)

如此处所述：https://discuss.pytorch.org/t/giving-multiple-parameters-in-optimizer/869

Answer 4

将您的模型定义修改为：

import torch.nn as nn

class Net(nn.Module):
    def __init__(self, weights_fixed, weights_guess):
        super(Net, self).__init__()
        self.convL1 = nn.Conv1d(1, 3, 3, bias=False)
        self.convL1.weight = weights_fixed # I want to keep these weights fixed
        self.convL1.requires_grad = False

        self.convL2 = nn.Conv1d(3, 1, 1, bias=False)
        self.convL1.weight = weights_guess # I want to learn these weights

    def forward(self, inp_batch):
        out1 = self.convL1(inp_batch)
        out2 = self.convL2(out1)

        return out2