前言

仅记录学习过程，有问题欢迎讨论

循环神经网络RNN（recurrent neural network）：

• 主要思想：将整个序列划分成多个时间步，将每一个时间步的信息依次输入模型，同时将模型输出的结果传给下一个时间步
• 自带了tanh的激活函数

代码

发现RNN效率高很多

import json
import randomimport numpy as np
import torch
import torch.nn as nn
import matplotlib.pyplot as plt
import torch.utils.data as Data"""
构建一个 用RNN实现的 判断某个字符的位置 的任务5 分类任务 判断 a出现的位置 返回index +1 or -1
"""class TorchModel(nn.Module):def __init__(self, sentence_length, hidden_size, vocab, input_dim, output_size):super(TorchModel, self).__init__()#self.emb = nn.Embedding(len(vocab) + 1, input_dim)self.rnn = nn.RNN(input_dim, hidden_size, batch_first=True)self.pool = nn.MaxPool1d(sentence_length)self.leaner = nn.Linear(hidden_size, output_size)self.loss = nn.functional.cross_entropydef forward(self, x, y=None):# x = 15 * 4x = self.emb(x)  # output = 15 * 4 * 10x, h = self.rnn(x)  # output = 15 * 4 * 20 h = 1*15*20x = self.pool(x.transpose(1, 2)).squeeze()  # output = 15 * 20 * (1,被去除)y_pred = self.leaner(x)  # output = 15 * 5if y is not None:return self.loss(y_pred, y)else:return y_pred# 创建字符集 只有6个 希望a出现的概率大点def build_vocab():chars = "abcdef"vocab = {}for index, char in enumerate(chars):vocab[char] = index + 1# vocab['unk'] = len(vocab) + 1return vocab# 构建样本集
def build_dataset(vocab, data_size, sentence_length):dataset_x = []dataset_y = []for i in range(data_size):x, y = build_simple(vocab, sentence_length)dataset_x.append(x)dataset_y.append(y)return torch.LongTensor(dataset_x), torch.LongTensor(dataset_y)# 构建样本
def build_simple(vocab, sentence_length):# 随机生成 长度为4的字符串x = [random.choice(list(vocab.keys())) for _ in range(sentence_length)]if x.count('a') != 0:y = x.index('a')else:y = 4# 转化为 数字x = [vocab[char] for char in list(x)]return x, ydef main():batch_size = 15simple_size = 500vocab = build_vocab()# 每个样本的长度为4sentence_length = 4# 样本的向量维度为10input_dim = 10# rnn的隐藏层 随便设置为20hidden_size = 20# 5 分类任务output_size = 5# 学习率lr = 0.02# 轮次epoch_size = 25model = TorchModel(sentence_length, hidden_size, vocab, input_dim, output_size)# 优化函数optim = torch.optim.Adam(model.parameters(), lr=lr)# 样本x, y = build_dataset(vocab, simple_size, sentence_length)dataset = Data.TensorDataset(x, y)dataiter = Data.DataLoader(dataset, batch_size)for epoch in range(epoch_size):epoch_loss = []model.train()for x, y_true in dataiter:loss = model(x, y_true)loss.backward()optim.step()optim.zero_grad()epoch_loss.append(loss.item())print("第%d轮 loss = %f" % (epoch + 1, np.mean(epoch_loss)))# evaluateacc = evaluate(model, vocab, sentence_length)  # 测试本轮模型结果return# 评估效果
def evaluate(model, vocab, sentence_length):model.eval()x, y = build_dataset(vocab, 200, sentence_length)correct, wrong = 0, 0with torch.no_grad():y_pred = model(x)for y_p, y_t in zip(y_pred, y):  # 与真实标签进行对比if int(torch.argmax(y_p)) == int(y_t):correct += 1  # 正样本判断正确else:wrong += 1print("正确预测个数：%d / %d, 正确率：%f" % (correct, correct + wrong, correct / (correct + wrong)))return correct / (correct + wrong)if __name__ == '__main__':main()

可以对model 优化一下

 def __init__(self, sentence_length, hidden_size, vocab, input_dim, output_size):super(TorchModel, self).__init__()# Embedding 层 变为稀疏self.emb = nn.Embedding(len(vocab) + 1, input_dim)self.rnn = nn.RNN(input_dim, input_dim, batch_first=True)self.pool = nn.AvgPool1d(sentence_length)self.leaner = nn.Linear(input_dim, sentence_length + 1)self.loss = nn.functional.cross_entropydef forward(self, x, y=None):# x = 15 * 4x = self.emb(x)  # output = 15 * 4 * 10x, h = self.rnn(x)  # output = 15 * 4 * 20 h = 1*15*20# x = self.pool(x.transpose# (1, 2)).squeeze()  # output = 15 * 20 * (1,被去除)# rnn 最后一维度包含之前所有信息h = h.squeeze()y_pred = self.leaner(h)  # output = 15 * 5if y is not None:return self.loss(y_pred, y)else:return y_pred