RNN未训练(PyTorch)
问题描述:
训练时我无法弄错我做错了RNN。我试图训练RNN对于和序列操作(了解它如何在简单的任务上工作)。 但是我的网络没有学习,损失保持不变,并且它不能模拟事件。 你能帮我找到问题吗?RNN未训练(PyTorch)
数据我使用:
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
]
代码为NN:
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)))
形状张量将是(sequence_len,1(这是一批大小),1),这是正确的根据PyTorch文档RNN
答
问题是这一行:
out, hidden = net(Variable(tensor), Variable(hidden.data))
应该是简单
out, hidden = net(Variable(tensor), hidden)
通过具有Variable(hidden.data)
在这里,你是在非常创建一个新的hidden_state变量(全零)步骤,而不是通过隐藏状态从以前的状态。
我试过你的例子,并将优化器改为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
谢谢,看起来合理! –