《PyTorch深度学习实践10》——循环神经网络-基础篇（Basic-Recurrent Neural Network）

一、RNN简介

       RNN网络最大的特点就是可以处理序列特征，就是我们的一组动态特征。比如，我们可以通过将前三天每天的特征（是否下雨，是否有太阳等）输入到网络，从而来预测第四天的天气。
       我们可以看RNN的网络结构如下：

二、RNN Cell用法

import torch

batch_size = 1 # 批处理大小
seq_len = 3 # 序列长度
input_size = 4 # 输入维度
hidden_size = 2 # 隐藏层维度

cell = torch.nn.RNNCell(input_size=input_size, hidden_size=hidden_size)

# (seq, batch, features)
dataset = torch.randn(seq_len, batch_size, input_size)
print(dataset)
hidden = torch.zeros(batch_size, hidden_size)
print(hidden)

for idx, input in enumerate(dataset):
    print( '=' * 20, idx, '=' * 20)
    print( 'Input size: ', input.shape)
    hidden = cell(input, hidden)
    print( 'outputs size: ', hidden.shape)
    print(hidden)

三、RNN用法

import torch

batch_size = 1 # 批处理大小
seq_len = 3 # 序列长度
input_size = 4 # 输入维度
hidden_size = 2 # 隐藏层维度
num_layers = 4  # 隐藏层数量

cell = torch.nn.RNN(input_size=input_size, hidden_size=hidden_size, num_layers=num_layers)

# (seqLen, batchSize, inputSize)
inputs = torch.randn(seq_len, batch_size, input_size)
hidden = torch.zeros(num_layers, batch_size, hidden_size)
out, hidden = cell(inputs, hidden)

print( 'Output size:', out.shape)
print( 'Output:', out)
print( 'Hidden size: ', hidden.shape)
print( 'Hidden: ', hidden)

batch_first参数：

三、实例：hello换序

       任务描述：我们需要训练一个模型，输入是“hello”，使输出是“ohlol”。如下图所示：

       方法描述：首先我们可以将“hello”中每个字母对应一个索引，之后得到输入“hello”和输出“ohlol”的编码分别为10223和31232。对编码中的每一个数字，都可以转换成一个四维张量（通过在对应张量对应索引填充为1，其余填充为0），如下图所示。这样我们的输入序列有5个元素，每个元素的维度为4。

1.RNN Cell

import torch

input_size = 4 # 输入维度，每个字母对应张量维度
hidden_size = 4
batch_size = 1

# 准备数据集
idx2char = ['e', 'h', 'l', 'o']
x_data = [1, 0, 2, 3, 3] # hello对应编码
y_data = [3, 1, 2, 3, 2] # ohlol对应编码

one_hot_lookup = [[1, 0, 0, 0],
                  [0, 1, 0, 0],
                  [0, 0, 1, 0],
                  [0, 0, 0, 1]]

x_one_hot = [one_hot_lookup[x] for x in x_data]

inputs = torch.Tensor(x_one_hot).view(-1, batch_size, input_size)
labels = torch.LongTensor(y_data).view(-1, 1)
print(inputs.shape, labels.shape)

# 构建模型
class Model(torch.nn.Module):
    def __init__(self, input_size, hidden_size, batch_size):
        super(Model, self).__init__()
        self.batch_size = batch_size
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.rnncell = torch.nn.RNNCell(input_size=self.input_size, hidden_size=self.hidden_size)

    def forward(self, inputs, hidden):
        hidden = self.rnncell(inputs, hidden)
        return hidden

    def init_hidden(self):
        return torch.zeros(self.batch_size, self.hidden_size)

net = Model(input_size, hidden_size, batch_size)

# 损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.1)

# 训练
for epoch in range(15):
    loss = 0
    optimizer.zero_grad()
    hidden = net.init_hidden()
    print( 'Predicted string: ', end= '')
    for input, label in zip(inputs, labels):
        hidden = net(input, hidden)
        loss += criterion(hidden, label)
        _, idx = hidden.max(dim=1)
        print(idx2char[idx.item()], end= '')
    loss.backward()
    optimizer.step()
    print( ', Epoch [%d/15] loss=%.4f' % (epoch+1, loss.item()))

2.RNN

import torch

input_size = 4
hidden_size = 4
batch_size = 1
seq_len = 5
num_layers = 1

# 准备数据集
idx2char = ['e', 'h', 'l', 'o']
x_data = [1, 0, 2, 3, 3] # hello对应编码
y_data = [3, 1, 2, 3, 2] # ohlol对应编码

one_hot_lookup = [[1, 0, 0, 0],
                  [0, 1, 0, 0],
                  [0, 0, 1, 0],
                  [0, 0, 0, 1]]
x_one_hot = [one_hot_lookup[x] for x in x_data]

inputs = torch.Tensor(x_one_hot).view(seq_len, batch_size, input_size)
labels = torch.LongTensor(y_data)
print(inputs.shape, labels.shape)

# 构建模型
class Model(torch.nn.Module):
    def __init__(self, input_size, hidden_size, batch_size, num_layers=1):
        super(Model, self).__init__()
        self.num_layers = num_layers
        self.batch_size = batch_size
        self.input_size = input_size
        self.hidden_size = hidden_size
        self.rnn = torch.nn.RNN(input_size=self.input_size, hidden_size=self.hidden_size, )

    def forward(self, inputs):
        hidden = torch.zeros(self.num_layers, self.batch_size, self.hidden_size)
        out, _ = self.rnn(inputs, hidden)    # 注意维度是(seqLen, batch_size, hidden_size)
        return out.view(-1, self.hidden_size) # 为了容易计算交叉熵这里调整维度为(seqLen * batch_size, hidden_size)

net = Model(input_size, hidden_size, batch_size)

# 损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.1)

# 训练
for epoch in range(15):
    optimizer.zero_grad()
    outputs = net(inputs)
    # print(outputs.shape, labels.shape)
    # 这里的outputs维度是([seqLen * batch_size, hidden]), labels维度是([seqLen])
    loss = criterion(outputs, labels)
    loss.backward()
    optimizer.step()

    _, idx = outputs.max(dim=1)
    idx = idx.data.numpy()
    print('Predicted: ', ''.join([idx2char[x] for x in idx]), end='')
    print(', Epoch [%d/15] loss = %.3f' % (epoch + 1, loss.item()))

四、Embedding

• One-hot encoding of words and characters:
  (1) The one-hot vectors are high-dimension.
  (2) The one-hot vectors are sparse.
  (3) The one-hot vectors are hardcoded.

• Do we have a way to associate a vector with a word/character with following specification:
  (1) Lower-dimension
  (2) Dense
  (3) Learned from data

• A popular and powerful way is called EMBEDDING.

我们使用embedding将模型变成如下所示：

import torch

# parameters
num_class = 4
input_size = 4
hidden_size = 8
embedding_size = 10
num_layers = 2
batch_size = 1
seq_len = 5

# 准备数据集
idx2char = ['e', 'h', 'l', 'o']
x_data = [[1, 0, 2, 2, 3]]  # (batch, seq_len)
y_data = [3, 1, 2, 3, 2]    # (batch * seq_len)

inputs = torch.LongTensor(x_data)   # Input should be LongTensor: (batchSize, seqLen)
labels = torch.LongTensor(y_data)   # Target should be LongTensor: (batchSize * seqLen)

# 构建模型
class Model(torch.nn.Module):
    def __init__(self):
        super(Model, self).__init__()
        self.emb = torch.nn.Embedding(input_size, embedding_size)
        self.rnn = torch.nn.RNN(input_size=embedding_size, hidden_size=hidden_size, num_layers=num_layers, batch_first=True)
        self.fc = torch.nn.Linear(hidden_size, num_class)

    def forward(self, x):
        hidden = torch.zeros(num_layers, x.size(0), hidden_size)
        x = self.emb(x)  # (batch, seqLen, embeddingSize)
        x, _ = self.rnn(x, hidden)  # 输出(𝒃𝒂𝒕𝒄𝒉𝑺𝒊𝒛𝒆, 𝒔𝒆𝒒𝑳𝒆𝒏, hidden_size)
        x = self.fc(x)  # 输出(𝒃𝒂𝒕𝒄𝒉𝑺𝒊𝒛𝒆, 𝒔𝒆𝒒𝑳𝒆𝒏, 𝒏𝒖𝒎𝑪𝒍𝒂𝒔𝒔)
        return x.view(-1, num_class)  # reshape to use Cross Entropy: (𝒃𝒂𝒕𝒄𝒉𝑺𝒊𝒛𝒆×𝒔𝒆𝒒𝑳𝒆𝒏, 𝒏𝒖𝒎𝑪𝒍𝒂𝒔𝒔)

net = Model()

# 损失函数和优化器
criterion = torch.nn.CrossEntropyLoss()
optimizer = torch.optim.Adam(net.parameters(), lr=0.05)

# 训练模型
for epoch in range(15):
    optimizer.zero_grad()
    outputs = net(inputs)
    loss = criterion(outputs, labels)
    loss.backward()
    optimizer.step()

    _, idx = outputs.max(dim=1)
    idx = idx.data.numpy()
    print('Predicted: ', ''.join([idx2char[x] for x in idx]), end='')
    print(', Epoch [%d/15] loss = %.3f' % (epoch + 1, loss.item()))

本文为系列文章：

上一篇	《Pytorch深度学习实践》目录	下一篇
卷积神经网络-高级篇（Advanced-Convolution Neural Network）	资料	循环神经网络-高级篇（Advanced-Recurrent Neural Network）【未更新】