- 递归神经网络 （RNN）：
定义：RNN 是一类专为顺序数据处理而设计的人工神经网络。
顺序处理：RNN 保持一个隐藏状态，该状态捕获有关序列中先前输入的信息，使其适用于涉及顺序依赖关系的任务。 - 词嵌入：
定义：词嵌入是捕获语义关系的词的密集向量表示。
重要性：它们允许神经网络学习上下文信息和单词之间的关系。
实现：使用预先训练的词嵌入（Word2Vec、GloVe）或在模型中包含嵌入层。 - 文本标记化和填充：
代币化：将文本分解为单个单词或子单词。
填充：通过添加零或截断来确保所有序列具有相同的长度。 - Keras 中的顺序模型：
实现：利用 Keras 库中的 Sequential 模型创建线性层堆栈。 - 嵌入层：
实现：向模型添加嵌入层，将单词转换为密集向量。
配置：指定输入维度、输出维度（嵌入大小）和输入长度。 - 循环层（LSTM 或 GRU）：
LSTM 和 GRU：长短期记忆 （LSTM） 和门控循环单元 （GRU） 层有助于捕获长期依赖关系。
实现：将一个或多个 LSTM 或 GRU 层添加到模型中。 - 致密层：
目的：密集层用于最终分类输出。
实现：添加一个或多个具有适当激活函数的密集层。 - 激活功能：
选择：ReLU（整流线性单元）或tanh是隐藏层中激活函数的常见选择。 - 损失函数和优化器：
损失函数：稀疏分类交叉熵通常用于文本分类任务。
优化：Adam 或 RMSprop 是常用的优化器。 - 批处理和排序：
批处理：在批量输入序列上训练模型。
处理不同长度的物料：使用填充来处理不同长度的序列。 - 培训流程：
汇编：使用所选的损失函数、优化器和指标编译模型。
训练：将模型拟合到训练数据，在单独的集合上进行验证。 - 防止过拟合：
技术：实现 dropout 或 recurrent dropout 层以防止过拟合。
正规化：如果需要，请考虑 L1 或 L2 正则化。 - 超参数调优：
参数：根据验证性能调整超参数，例如学习率、批量大小和循环单元数。 - 评估指标：
指标：选择适当的指标，如准确率、精确率、召回率或 F1 分数进行评估。
import os
import warnings
warnings.filterwarnings('ignore')
import tensorflow as tf
import numpy as np
import pprint
import logging
import time
from collections import Counterfrom pathlib import Path
from tqdm import tqdm
(x_train, y_train), (x_test, y_test) = tf.keras.datasets.imdb.load_data()
_word2idx = tf.keras.datasets.imdb.get_word_index()
word2idx = {w: i+3 for w, i in _word2idx.items()}
word2idx['<pad>'] = 0
word2idx['<start>'] = 1
word2idx['<unk>'] = 2
idx2word = {i: w for w, i in word2idx.items()}def sort_by_len(x, y):x, y = np.asarray(x), np.asarray(y)idx = sorted(range(len(x)), key=lambda i: len(x[i]))return x[idx], y[idx]x_train, y_train = sort_by_len(x_train, y_train)
x_test, y_test = sort_by_len(x_test, y_test)def write_file(f_path, xs, ys):with open(f_path, 'w',encoding='utf-8') as f:for x, y in zip(xs, ys):f.write(str(y)+'\t'+' '.join([idx2word[i] for i in x][1:])+'\n')write_file('./data/train.txt', x_train, y_train)
write_file('./data/test.txt', x_test, y_test)counter = Counter()
with open('./data/train.txt',encoding='utf-8') as f:for line in f:line = line.rstrip()label, words = line.split('\t')words = words.split(' ')counter.update(words)words = ['<pad>'] + [w for w, freq in counter.most_common() if freq >= 10]
print('Vocab Size:', len(words))Path('./vocab').mkdir(exist_ok=True)with open('./vocab/word.txt', 'w',encoding='utf-8') as f:for w in words:f.write(w+'\n')word2idx = {}
with open('./vocab/word.txt',encoding='utf-8') as f:for i, line in enumerate(f):line = line.rstrip()word2idx[line] = i
embedding = np.zeros((len(word2idx)+1, 50)) with open('./data/glove.6B.50d.txt',encoding='utf-8') as f: count = 0for i, line in enumerate(f):if i % 100000 == 0:print('- At line {}'.format(i)) line = line.rstrip()0sp = line.split(' ')word, vec = sp[0], sp[1:]if word in word2idx:count += 1embedding[word2idx[word]] = np.asarray(vec, dtype='float32') print("[%d / %d] words have found pre-trained values"%(count, len(word2idx)))
np.save('./vocab/word.npy', embedding)
print('Saved ./vocab/word.npy')
def data_generator(f_path, params):with open(f_path,encoding='utf-8') as f:print('Reading', f_path)for line in f:line = line.rstrip()label, text = line.split('\t')text = text.split(' ')x = [params['word2idx'].get(w, len(word2idx)) for w in text]if len(x) >= params['max_len']:x = x[:params['max_len']]else:x += [0] * (params['max_len'] - len(x))y = int(label)yield x, ydef dataset(is_training, params):_shapes = ([params['max_len']], ())_types = (tf.int32, tf.int32)if is_training:ds = tf.data.Dataset.from_generator(lambda: data_generator(params['train_path'], params),output_shapes=_shapes,output_types=_types, )ds = ds.shuffle(params['num_samples'])ds = ds.batch(params['batch_size'])ds = ds.prefetch(tf.data.experimental.AUTOTUNE) else:ds = tf.data.Dataset.from_generator(lambda: data_generator(params['test_path'], params),output_shapes=_shapes,output_types=_types, )ds = ds.batch(params['batch_size'])ds = ds.prefetch(tf.data.experimental.AUTOTUNE)return dsclass Model(tf.keras.Model):def __init__(self, params):super().__init__()self.embedding = tf.Variable(np.load('./vocab/word.npy'),dtype=tf.float32,name='pretrained_embedding',trainable=False, )self.drop1 = tf.keras.layers.Dropout(params['dropout_rate'])self.drop2 = tf.keras.layers.Dropout(params['dropout_rate'])self.drop3 = tf.keras.layers.Dropout(params['dropout_rate'])self.rnn1 = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(params['rnn_units'], return_sequences=True))self.rnn2 = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(params['rnn_units'], return_sequences=True))self.rnn3 = tf.keras.layers.Bidirectional(tf.keras.layers.LSTM(params['rnn_units'], return_sequences=True))self.drop_fc = tf.keras.layers.Dropout(params['dropout_rate'])self.fc = tf.keras.layers.Dense(2 * params['rnn_units'], tf.nn.elu)self.out_linear = tf.keras.layers.Dense(2)def call(self, inputs, training=False):if inputs.dtype != tf.int32:inputs = tf.cast(inputs, tf.int32)batch_sz = tf.shape(inputs)[0]rnn_units = 2 * params['rnn_units']x = tf.nn.embedding_lookup(self.embedding, inputs)x = tf.reshape(x, (batch_sz * 10 * 10, 10, 50))x = self.drop1(x, training=training)x = self.rnn1(x)x = tf.reduce_max(x, 1)x = tf.reshape(x, (batch_sz * 10, 10, rnn_units))x = self.drop2(x, training=training)x = self.rnn2(x)x = tf.reduce_max(x, 1)x = tf.reshape(x, (batch_sz, 10, rnn_units))x = self.drop3(x, training=training)x = self.rnn3(x)x = tf.reduce_max(x, 1)x = self.drop_fc(x, training=training)x = self.fc(x)x = self.out_linear(x)return xparams = {'vocab_path': './vocab/word.txt','train_path': './data/train.txt','test_path': './data/test.txt','num_samples': 25000,'num_labels': 2,'batch_size': 32,'max_len': 1000,'rnn_units': 200,'dropout_rate': 0.2,'clip_norm': 10.,'num_patience': 3,'lr': 3e-4,
}def is_descending(history: list):history = history[-(params['num_patience']+1):]for i in range(1, len(history)):if history[i-1] <= history[i]:return Falsereturn Trueword2idx = {}
with open(params['vocab_path'],encoding='utf-8') as f:for i, line in enumerate(f):line = line.rstrip()word2idx[line] = i
params['word2idx'] = word2idx
params['vocab_size'] = len(word2idx) + 1model = Model(params)
model.build(input_shape=(None, None))
decay_lr = tf.optimizers.schedules.ExponentialDecay(params['lr'], 1000, 0.95)
optim = tf.optimizers.Adam(params['lr'])
global_step = 0history_acc = []
best_acc = .0t0 = time.time()
logger = logging.getLogger('tensorflow')
logger.setLevel(logging.INFO)while True:for texts, labels in dataset(is_training=True, params=params):with tf.GradientTape() as tape: logits = model(texts, training=True)loss = tf.nn.sparse_softmax_cross_entropy_with_logits(labels=labels, logits=logits)loss = tf.reduce_mean(loss)optim.lr.assign(decay_lr(global_step))grads = tape.gradient(loss, model.trainable_variables)grads, _ = tf.clip_by_global_norm(grads, params['clip_norm']) optim.apply_gradients(zip(grads, model.trainable_variables)) if global_step % 50 == 0:logger.info("Step {} | Loss: {:.4f} | Spent: {:.1f} secs | LR: {:.6f}".format(global_step, loss.numpy().item(), time.time() - t0, optim.lr.numpy().item()))t0 = time.time()global_step += 1m = tf.keras.metrics.Accuracy()for texts, labels in dataset(is_training=False, params=params):logits = model(texts, training=False)y_pred = tf.argmax(logits, axis=-1)m.update_state(y_true=labels, y_pred=y_pred)acc = m.result().numpy()logger.info("Evaluation: Testing Accuracy: {:.3f}".format(acc))history_acc.append(acc)if acc > best_acc:best_acc = acclogger.info("Best Accuracy: {:.3f}".format(best_acc))if len(history_acc) > params['num_patience'] and is_descending(history_acc):logger.info("Testing Accuracy not improved over {} epochs, Early Stop".format(params['num_patience']))break
