非序列模型的 Pytorch 梯度问题

Question

我无法重现此Pytorch tutorial。

他们介绍的模型是：

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(RNN, self).__init__()

        self.hidden_size = hidden_size
        
        self.i2h = nn.Linear(input_size + hidden_size, hidden_size)
        self.i2o = nn.Linear(input_size + hidden_size, output_size)
        self.softmax = nn.LogSoftmax(dim=1)

    def forward(self, input, hidden):
        combined = torch.cat((input, hidden), 1)
        hidden = self.i2h(combined)
        output = self.i2o(combined)
        output = self.softmax(output)
        return output, hidden

    def initHidden(self, batch_size):
        return torch.zeros(batch_size, self.hidden_size, dtype=torch.float32, requires_grad=True)

此模型再现了 RNN 单元内发生的事情。

在编码时，我遇到了模型内部的渐变问题。

重现问题的代码如下：

import torch
import torch.nn as nn

# Toy data to reproduce the issue
toy_data_batch = torch.tensor([[0, 1], [1, 0], [1, 0]], dtype=torch.float32)
toy_label_batch = torch.tensor([2, 0, 3], dtype=torch.long)

class RNN(nn.Module):
    def __init__(self, input_size, hidden_size, output_size):
        super(RNN, self).__init__()

        self.hidden_size = hidden_size
        
        self.i2h = nn.Linear(input_size + hidden_size, hidden_size)
        self.i2o = nn.Linear(input_size + hidden_size, output_size)
        self.softmax = nn.LogSoftmax(dim=1)

    def forward(self, input, hidden):
        combined = torch.cat((input, hidden), 1)
        hidden = self.i2h(combined)
        output = self.i2o(combined)
        output = self.softmax(output)
        return output, hidden

    def initHidden(self, batch_size):
        return torch.zeros(batch_size, self.hidden_size, dtype=torch.float32, requires_grad=True)

# Model initialization
input_size = 2
hidden_size = 2
output_size = 4  # Targets in {0, 1, 2, 3}
batch_size = 3  # 3 data points in the batch
learning_rate = 5e-3

rnn = RNN(input_size, hidden_size, output_size)
hidden = rnn.initHidden(batch_size)  # init hidden layer with zeros

# Negative log likelihood as it is classification
criterion = nn.NLLLoss()

# Forward pass
output, hidden = rnn(toy_data_batch, hidden)
#output, hidden = rnn(toy_data_batch, hidden) ### BUG: if I remove the comment here, It works

# Loss computation
loss = criterion(output, toy_label_batch)

# Backward pass
loss.backward()

print(rnn.i2o.weight.grad)  # This one is fine
print(rnn.i2h.weight.grad)  # This one isn't (has type None)

# This will fail, because of the None gradient
for weight in rnn.parameters():
    weight.data.add_(weight.grad.data, alpha=-learning_rate)

输出是：

tensor([[-0.1892,  0.0462,  0.0000,  0.0000],
        [ 0.1274,  0.1133,  0.0000,  0.0000],
        [ 0.1455, -0.2525,  0.0000,  0.0000],
        [-0.0837,  0.0930,  0.0000,  0.0000]])
None

---------------------------------------------------------------------------
AttributeError                            Traceback (most recent call last)
<ipython-input-3-5f02113fddf6> in <module>
     54 # This will fail, because of the None gradient
     55 for weight in rnn.parameters():
---> 56     weight.data.add_(weight.grad.data, alpha=-learning_rate)

AttributeError: 'NoneType' object has no attribute 'data'

我注意到如果我取消注释该行

output, hidden = rnn(toy_data_batch, hidden)
#output, hidden = rnn(toy_data_batch, hidden) ### BUG: if I remove the comment here, It works

一切正常。在我看来，隐藏变量的初始化存在问题。由于我打开了参数“requires_grad”，我不知道该怎么做。

提前感谢您，任何帮助将不胜感激

Answer 1

self.i2h 没有渐变，因为它没有在模型的第一步中使用。当您反向传播时，您的模型仅在第一阶段使用self.i2o，因此self.i2h 对输出没有影响。但是，当您进入第二步时，它使用了 hidden，该 self.i2h 是使用 self.i2h 计算的，因此通过该层存在可追踪的梯度。