ccc-pytorch-LSTM（8）

最新推荐文章于 2024-04-29 16:12:40 发布

扔出去的回旋镖

最新推荐文章于 2024-04-29 16:12:40 发布

阅读量1.2k

点赞数

分类专栏： pytorch学习文章标签： pytorch lstm 深度学习

本文链接：https://blog.csdn.net/liubi32326/article/details/129431732

版权

pytorch学习专栏收录该内容

12 篇文章 1 订阅

订阅专栏

文章目录

一、LSTM简介

LSTM（long short-term memory）长短期记忆网络，RNN的改进，克服了RNN中“记忆低下”的问题。通过“门”结构实现信息的添加和移除，通过记忆元将序列处理过程中的相关信息一直传递下去，经典结构如下：

二、LSTM中的核心结构

记忆元（memory cell）-长期记忆：
在这里插入图片描述
就像一个cell一样，信息通过这条只有少量线性交互的线传递。传递过程中有3种“门”结构可以告诉它该学习或者保存哪些信息
三个门结构-短期记忆
遗忘门：用来决定当前状态哪些信息被移除

输入门：决定放入哪些信息到细胞状态

输出门：决定哪些信息用于输出

细节注意：

新的细胞状态只需要遗忘门和输入门就可以更新，公式为： $C_t=f_t*C_{t-1}+i_t* \tilde{C_t}$ （注意所有的 $*$ 都表示Hadamard 乘积）
只有隐状态h_t会传递到输出层，记忆元完全属于内部信息，不可手动修改

三、如何解决RNN中的梯度消失/爆炸问题

解决是指很大程度上缓解，不是让它彻底消失。先解释RNN为什么会有这些问题：
$\begin{aligned} &\frac{\partial L_t}{\partial U}= \sum_{k=0}^{t}\frac{\partial L_t}{\partial O_t}\frac{\partial O_t}{\partial S_t}(\prod_{j=k+1}^{t}\frac{\partial S_j}{\partial S_{j-1}})\frac{\partial S_k}{\partial U}\\&\frac{\partial L_t}{\partial W}= \sum_{k=0}^{t}\frac{\partial L_t}{\partial O_t}\frac{\partial O_t}{\partial S_t}(\prod_{j=k+1}^{t}\frac{\partial S_j}{\partial S_{j-1}})\frac{\partial S_k}{\partial W} \end{aligned}$ （具体过程可以看这里）

上面是训练过程任意时刻更新W、U需要用到的求偏导的结果。实际使用会加上激活函数，通常为tanh、sigmoid等
tanh和其导数图像如下

sigmoid和其导数如下

这些激活函数的导数都比1要小，又因为 $\prod_{j=k+1}^{t}\frac{\partial S_j}{\partial S_{j-1}}=\prod_{j=k+1}^{t}tanh'(W_s)$ ，所以当 $W_s$ 过小过大就会分别造成梯度消失和爆炸的问题，特别是过小。
LSTM如何缓解
由链式法则和三个门的公式可以得到：
$\begin{aligned} &\frac{\partial C_t}{\partial C_{t-1}}\\&=\frac{\partial C_t}{\partial f_t}\frac{\partial f_t}{\partial h_{t-1}}\frac{\partial h_{t-1}}{\partial C_{t-1}}+\frac{\partial C_t}{\partial i_t}\frac{\partial i_t}{\partial h_{t-1}}\frac{\partial h_{t-1}}{\partial C_{t-1}}+\frac{\partial C_t}{\partial \tilde{C_t}}\frac{\partial \tilde{C_t}}{\partial h_{t-1}}\frac{\partial h_{t-1}}{\partial C_{t-1}}+\frac{\partial C_t}{\partial C_{t-1}}\\ &=C_{t-1}\sigma '(\cdot)W_f*o_{t-1}tanh'(C_{t-1})+\tilde{C_t}\sigma '(\cdot)W_i*o_{t-1}tanh'(C_{t-1})\\&+i_ttanh'(\cdot)W_c*o_{t-1}tanh'(C_{t-1})+f_t \end{aligned}$

由相乘变成了相加，不容易叠加
sigmoid函数使单元间传递结果非常接近0或者1，使模型变成非线性，并且可以在学习过程中内部调整

四、情感分类实战（google colab）

环境和库：

!pip install torch
!pip install torchtext
!python -m spacy download en

# K80 gpu for 12 hours
import torch
from torch import nn, optim
from torchtext import data, datasets

print('GPU:', torch.cuda.is_available())

torch.manual_seed(123)

加载数据集：

TEXT = data.Field(tokenize='spacy')
LABEL = data.LabelField(dtype=torch.float)
train_data, test_data = datasets.IMDB.splits(TEXT, LABEL)

print(train_data.examples[15].text)
print(train_data.examples[15].label)

网络结构：

class RNN(nn.Module):
    
    def __init__(self, vocab_size, embedding_dim, hidden_dim):
        """
        """
        super(RNN, self).__init__()
        
        # [0-10001] => [100]
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        # [100] => [256]
        self.rnn = nn.LSTM(embedding_dim, hidden_dim, num_layers=2, 
                           bidirectional=True, dropout=0.5)
        # [256*2] => [1]
        self.fc = nn.Linear(hidden_dim*2, 1)
        self.dropout = nn.Dropout(0.5)
        
        
    def forward(self, x):
        """
        x: [seq_len, b] vs [b, 3, 28, 28]
        """
        # [seq, b, 1] => [seq, b, 100]
        embedding = self.dropout(self.embedding(x))
        
        # output: [seq, b, hid_dim*2]
        # hidden/h: [num_layers*2, b, hid_dim]
        # cell/c: [num_layers*2, b, hid_di]
        output, (hidden, cell) = self.rnn(embedding)
        
        # [num_layers*2, b, hid_dim] => 2 of [b, hid_dim] => [b, hid_dim*2]
        hidden = torch.cat([hidden[-2], hidden[-1]], dim=1)
        
        # [b, hid_dim*2] => [b, 1]
        hidden = self.dropout(hidden)
        out = self.fc(hidden)
        
        return out

Embedding

rnn = RNN(len(TEXT.vocab), 100, 256)

pretrained_embedding = TEXT.vocab.vectors
print('pretrained_embedding:', pretrained_embedding.shape)
rnn.embedding.weight.data.copy_(pretrained_embedding)
print('embedding layer inited.')

optimizer = optim.Adam(rnn.parameters(), lr=1e-3)
criteon = nn.BCEWithLogitsLoss().to(device)
rnn.to(device)

在这里插入图片描述
训练并测试

import numpy as np

def binary_acc(preds, y):
    """
    get accuracy
    """
    preds = torch.round(torch.sigmoid(preds))
    correct = torch.eq(preds, y).float()
    acc = correct.sum() / len(correct)
    return acc

def train(rnn, iterator, optimizer, criteon):
    
    avg_acc = []
    rnn.train()
    
    for i, batch in enumerate(iterator):
        
        # [seq, b] => [b, 1] => [b]
        pred = rnn(batch.text).squeeze(1)
        # 
        loss = criteon(pred, batch.label)
        acc = binary_acc(pred, batch.label).item()
        avg_acc.append(acc)
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if i%10 == 0:
            print(i, acc)
        
    avg_acc = np.array(avg_acc).mean()
    print('avg acc:', avg_acc)
    
    
def eval(rnn, iterator, criteon):
    
    avg_acc = []
    rnn.eval()
    
    with torch.no_grad():
        for batch in iterator:

            # [b, 1] => [b]
            pred = rnn(batch.text).squeeze(1)

            #
            loss = criteon(pred, batch.label)

            acc = binary_acc(pred, batch.label).item()
            avg_acc.append(acc)
        
    avg_acc = np.array(avg_acc).mean()
    
    print('>>test:', avg_acc)

for epoch in range(10):
    
    eval(rnn, test_iterator, criteon)
    train(rnn, train_iterator, optimizer, criteon)

最后得到的准确率结果如下：
在这里插入图片描述
完整colab链接：lstm
完整代码：

# -*- coding: utf-8 -*-
"""lstm

Automatically generated by Colaboratory.

Original file is located at
    https://colab.research.google.com/drive/1GX0Rqur8T45MSYhLU9MYWAbycfLH4-Fu
"""

!pip install torch
!pip install torchtext
!python -m spacy download en


# K80 gpu for 12 hours
import torch
from torch import nn, optim
from torchtext import data, datasets

print('GPU:', torch.cuda.is_available())

torch.manual_seed(123)

TEXT = data.Field(tokenize='spacy')
LABEL = data.LabelField(dtype=torch.float)
train_data, test_data = datasets.IMDB.splits(TEXT, LABEL)

print('len of train data:', len(train_data))
print('len of test data:', len(test_data))

print(train_data.examples[15].text)
print(train_data.examples[15].label)

# word2vec, glove
TEXT.build_vocab(train_data, max_size=10000, vectors='glove.6B.100d')
LABEL.build_vocab(train_data)


batchsz = 30
device = torch.device('cuda')
train_iterator, test_iterator = data.BucketIterator.splits(
    (train_data, test_data),
    batch_size = batchsz,
    device=device
)

class RNN(nn.Module):
    
    def __init__(self, vocab_size, embedding_dim, hidden_dim):
        """
        """
        super(RNN, self).__init__()
        
        # [0-10001] => [100]
        self.embedding = nn.Embedding(vocab_size, embedding_dim)
        # [100] => [256]
        self.rnn = nn.LSTM(embedding_dim, hidden_dim, num_layers=2, 
                           bidirectional=True, dropout=0.5)
        # [256*2] => [1]
        self.fc = nn.Linear(hidden_dim*2, 1)
        self.dropout = nn.Dropout(0.5)
        
        
    def forward(self, x):
        """
        x: [seq_len, b] vs [b, 3, 28, 28]
        """
        # [seq, b, 1] => [seq, b, 100]
        embedding = self.dropout(self.embedding(x))
        
        # output: [seq, b, hid_dim*2]
        # hidden/h: [num_layers*2, b, hid_dim]
        # cell/c: [num_layers*2, b, hid_di]
        output, (hidden, cell) = self.rnn(embedding)
        
        # [num_layers*2, b, hid_dim] => 2 of [b, hid_dim] => [b, hid_dim*2]
        hidden = torch.cat([hidden[-2], hidden[-1]], dim=1)
        
        # [b, hid_dim*2] => [b, 1]
        hidden = self.dropout(hidden)
        out = self.fc(hidden)
        
        return out

rnn = RNN(len(TEXT.vocab), 100, 256)

pretrained_embedding = TEXT.vocab.vectors
print('pretrained_embedding:', pretrained_embedding.shape)
rnn.embedding.weight.data.copy_(pretrained_embedding)
print('embedding layer inited.')

optimizer = optim.Adam(rnn.parameters(), lr=1e-3)
criteon = nn.BCEWithLogitsLoss().to(device)
rnn.to(device)

import numpy as np

def binary_acc(preds, y):
    """
    get accuracy
    """
    preds = torch.round(torch.sigmoid(preds))
    correct = torch.eq(preds, y).float()
    acc = correct.sum() / len(correct)
    return acc

def train(rnn, iterator, optimizer, criteon):
    
    avg_acc = []
    rnn.train()
    
    for i, batch in enumerate(iterator):
        
        # [seq, b] => [b, 1] => [b]
        pred = rnn(batch.text).squeeze(1)
        # 
        loss = criteon(pred, batch.label)
        acc = binary_acc(pred, batch.label).item()
        avg_acc.append(acc)
        
        optimizer.zero_grad()
        loss.backward()
        optimizer.step()
        
        if i%10 == 0:
            print(i, acc)
        
    avg_acc = np.array(avg_acc).mean()
    print('avg acc:', avg_acc)
    
    
def eval(rnn, iterator, criteon):
    
    avg_acc = []
    
    rnn.eval()
    
    with torch.no_grad():
        for batch in iterator:

            # [b, 1] => [b]
            pred = rnn(batch.text).squeeze(1)

            #
            loss = criteon(pred, batch.label)

            acc = binary_acc(pred, batch.label).item()
            avg_acc.append(acc)
        
    avg_acc = np.array(avg_acc).mean()
    
    print('>>test:', avg_acc)

for epoch in range(10):
    
    eval(rnn, test_iterator, criteon)
    train(rnn, train_iterator, optimizer, criteon)