RNN没有接受培训(PyTorch)

时间:2017-10-12 09:08:33

标签: machine-learning deep-learning pytorch rnn

训练 RNN 时,我无法理解我做错了什么。我正在尝试为序列上的 AND 操作训练 RNN (以了解它如何在简单任务上工作)。 但是我的网络不是学习,损失保持不变,并且它不能使事件超过模型。 你能帮我找到问题吗?

我正在使用的数据:

data = [
    [1, 1, 1, 1, 0, 0, 1, 1, 1],
    [1, 1, 1, 1],
    [0, 0, 1, 1],
    [0, 0, 0, 0, 0, 0, 0],
    [1, 1, 1, 1, 1, 1, 1],
    [1, 1],
    [0],
    [1],
    [1, 0]]
labels = [
    0,
    1, 
    0, 
    0,
    1,
    1,
    0,
    1,
    0
]

NN代码:

class AndRNN(nn.Module):
def __init__(self):
    super(AndRNN, self).__init__()
    self.rnn = nn.RNN(1, 10, 5)
    self.fc = nn.Sequential(
        nn.Linear(10, 30),
        nn.Linear(30, 2)
    )

def forward(self, input, hidden):
    x, hidden = self.rnn(input, hidden)
    x = self.fc(x[-1])        
    return x, hidden

def initHidden(self):
    return Variable(torch.zeros((5, 1, 10)))

训练循环:

criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(net.parameters(), lr=0.001, momentum=0.9)

correct = 0
for e in range(20):
for i in range(len(data)):
    tensor = torch.FloatTensor(data[i]).view(-1, 1, 1)
    label = torch.LongTensor([labels[i]])
    hidden = net.initHidden()
    optimizer.zero_grad()

    out, hidden = net(Variable(tensor), Variable(hidden.data))

    _, l = torch.topk(out, 1)
    if label[0] == l[0].data[0]:
        correct += 1

    loss = criterion(out, Variable(label))
    loss.backward()
    optimizer.step()

    print("Loss ", loss.data[0], "Accuracy ", (correct / (i + 1)))

张量的形状将是(sequence_len,1(批量大小),1),根据RNN的PyTorch文档是正确的

1 个答案:

答案 0 :(得分:1)

问题在于这一行:

out, hidden = net(Variable(tensor), Variable(hidden.data))

应该只是

out, hidden = net(Variable(tensor), hidden)

通过在此处设置Variable(hidden.data),您将在非常步骤创建一个新的hidden_​​state变量(全部为零),而不是从前一个状态传递隐藏状态。

我尝试了你的例子,并将优化器更改为Adam。有完整的代码。

class AndRNN(nn.Module):
    def __init__(self):
        super(AndRNN, self).__init__()
        self.rnn = nn.RNN(1, 10, 5)
        self.fc = nn.Sequential(
            nn.Linear(10, 30),
            nn.Linear(30, 2)
        )

    def forward(self, input, hidden):
        x, hidden = self.rnn(input, hidden)
        x = self.fc(x[-1])        
        return x, hidden

    def initHidden(self):
        return Variable(torch.zeros((5, 1, 10)))

net = AndRNN()    
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters())

correct = 0
for e in range(100):
    for i in range(len(data)):
        tensor = torch.FloatTensor(data[i]).view(-1, 1, 1)
        label = torch.LongTensor([labels[i]])
        hidden = net.initHidden()
        optimizer.zero_grad()

        out, hidden = net(Variable(tensor), hidden)

        loss = criterion(out, Variable(label))
        loss.backward()
        optimizer.step()
    if e % 25 == 0:
        print("Loss ", loss.data[0])

结果

Loss  0.6370733976364136
Loss  0.25336754322052
Loss  0.006924811284989119
Loss  0.002351854695007205