Question

我正在实施一个简单的RNN网络，它可以预测某个可变长度为1/0 时间序列数据。网络首先将训练数据馈送到LSTM单元，然后使用线性层进行分类。

通常，我们会使用迷你批次来训练网络。但是，问题是当我使用batch_size＆gt;时，这个简单的RNN网络没有训练。 1。

我设法创建一个可以重现问题的最小代码示例。如果你在第95行设置batch_size=1，网络会成功训练，但如果你设置batch_size=2，网络根本就没有训练，那么损失就会蹦蹦跳跳。（需要python3，pytorch＆gt; = 0.4.0）

import numpy as np
import random
import torch
import torch.nn as nn
import torch.optim as optim
from torch.nn.utils.rnn import pad_sequence, pack_padded_sequence


class ToyDataLoader(object):

    def __init__(self, batch_size):
        self.batch_size = batch_size
        self.index = 0
        self.dataset_size = 10

        # generate 10 random variable length training samples,
        # each time step has 1 feature dimension
        self.X = [
            [[1], [1], [1], [1], [0], [0], [1], [1], [1]],
            [[1], [1], [1], [1]],
            [[0], [0], [1], [1]],
            [[1], [1], [1], [1], [1], [1], [1]],
            [[1], [1]],
            [[0]],
            [[0], [0], [0], [0], [0], [0], [0]],
            [[1]],
            [[0], [1]],
            [[1], [0]]
        ]

        # assign labels for the toy traning set
        self.y = torch.LongTensor([0, 0, 0, 0, 0, 1, 1, 1, 1, 1])

    def __len__(self):
        return self.dataset_size // self.batch_size

    def __iter__(self):
        return self

    def __next__(self):
        if self.index + self.batch_size > self.dataset_size:
            self.index = 0
            raise StopIteration()
        if self.index == 0:  # shufle the dataset
            tmp = list(zip(self.X, self.y))
            random.shuffle(tmp)
            self.X, self.y = zip(*tmp)
            self.y = torch.LongTensor(self.y)
        X = self.X[self.index: self.index + self.batch_size]
        y = self.y[self.index: self.index + self.batch_size]
        self.index += self.batch_size
        return X, y


class NaiveRNN(nn.Module):
    def __init__(self):
        super(NaiveRNN, self).__init__()
        self.lstm = nn.LSTM(1, 128)
        self.linear = nn.Linear(128, 2)

    def forward(self, X):
        '''
        Parameter:
            X: list containing variable length training data
        '''

        # get the length of each seq in the batch
        seq_lengths = [len(x) for x in X]

        # convert to torch.Tensor
        seq_tensor = [torch.Tensor(seq) for seq in X]

        # sort seq_lengths and seq_tensor based on seq_lengths, required by torch.nn.utils.rnn.pad_sequence
        pairs = sorted(zip(seq_lengths, seq_tensor),
                       key=lambda pair: pair[0], reverse=True)
        seq_lengths = torch.LongTensor([pair[0] for pair in pairs])
        seq_tensor = [pair[1] for pair in pairs]

        # padded_seq shape: (seq_len, batch_size, feature_size)
        padded_seq = pad_sequence(seq_tensor)

        # pack them up
        packed_seq = pack_padded_sequence(padded_seq, seq_lengths.numpy())

        # feed to rnn
        packed_output, (ht, ct) = self.lstm(packed_seq)

        # linear classification layer
        y_pred = self.linear(ht[-1])

        return y_pred


def main():
    trainloader = ToyDataLoader(batch_size=2)  # not training at all! !!
    # trainloader = ToyDataLoader(batch_size=1) # it converges !!!

    model = NaiveRNN()
    criterion = nn.CrossEntropyLoss()
    optimizer = optim.Adadelta(model.parameters(), lr=1.0)

    for epoch in range(30):
        # switch to train mode
        model.train()

        for i, (X, labels) in enumerate(trainloader):

            # compute output
            outputs = model(X)
            loss = criterion(outputs, labels)

            # measure accuracy and record loss
            _, predicted = torch.max(outputs, 1)
            accu = (predicted == labels).sum().item() / labels.shape[0]

            # compute gradient and do SGD step
            optimizer.zero_grad()
            loss.backward()

            optimizer.step()

            print('Epoch: [{}][{}/{}]\tLoss {:.4f}\tAccu {:.3f}'.format(
                epoch, i, len(trainloader), loss, accu))


if __name__ == '__main__':
    main()

batch_size=1时的示例输出：

...
Epoch: [28][7/10]       Loss 0.1582     Accu 1.000
Epoch: [28][8/10]       Loss 0.2718     Accu 1.000
Epoch: [28][9/10]       Loss 0.0000     Accu 1.000
Epoch: [29][0/10]       Loss 0.2808     Accu 1.000
Epoch: [29][1/10]       Loss 0.0000     Accu 1.000
Epoch: [29][2/10]       Loss 0.0001     Accu 1.000
Epoch: [29][3/10]       Loss 0.0149     Accu 1.000
Epoch: [29][4/10]       Loss 0.1445     Accu 1.000
Epoch: [29][5/10]       Loss 0.2866     Accu 1.000
Epoch: [29][6/10]       Loss 0.0170     Accu 1.000
Epoch: [29][7/10]       Loss 0.0869     Accu 1.000
Epoch: [29][8/10]       Loss 0.0000     Accu 1.000
Epoch: [29][9/10]       Loss 0.0498     Accu 1.000

batch_size=2时的示例输出：

...
Epoch: [27][2/5]        Loss 0.8051     Accu 0.000
Epoch: [27][3/5]        Loss 1.2835     Accu 0.000
Epoch: [27][4/5]        Loss 1.0782     Accu 0.000
Epoch: [28][0/5]        Loss 0.5201     Accu 1.000
Epoch: [28][1/5]        Loss 0.6587     Accu 0.500
Epoch: [28][2/5]        Loss 0.3488     Accu 1.000
Epoch: [28][3/5]        Loss 0.5413     Accu 0.500
Epoch: [28][4/5]        Loss 0.6769     Accu 0.500
Epoch: [29][0/5]        Loss 1.0434     Accu 0.000
Epoch: [29][1/5]        Loss 0.4460     Accu 1.000
Epoch: [29][2/5]        Loss 0.9879     Accu 0.000
Epoch: [29][3/5]        Loss 1.0784     Accu 0.500
Epoch: [29][4/5]        Loss 0.6051     Accu 1.000

我搜索了很多材料但仍无法找出原因。

Answer 1

我认为一个主要问题是您将ht[-1]作为输入传递给线性图层 ht[-1]将包含上一个时间步的状态，该状态仅对具有最大长度的输入有效。

要解决此问题，您需要解压缩输出并获得与相应输入的最后长度相对应的输出以下是我们需要改变的方式：

# feed to rnn
packed_output, (ht, ct) = self.lstm(packed_seq)

# Unpack output
lstm_out, seq_len = pad_packed_sequence(packed_output)

# get vector containing last input indices
last_input = seq_len - 1

indices = torch.linspace(0, (seq_len.size(0)-1), steps=seq_len.size(0)).long()

# linear classification layer
y_pred = self.linear(lstm_out[last_input, indices, :])
return y_pred

我仍然无法使其与其余参数融合，但这应该会有所帮助。

当批量大小时，RNN不进行训练。 1，可变长度数据

1 个答案: