Question

我将当前数据放入了post.embedding文件中，并使用Bidirectional_LSTM进行了测试。但是准确性并不令人满意。如果我想提高准确性，再训练post.embedding更好吗？还是训练更多Bidirectional_LSTM更好？下面是我当前正在使用的代码。

createEmbedding.py

from gensim.models import Word2Vec
import csv

token = loadCsv("test_data")
embeddingmodel = []
for i in range(len(token)):
temp_embeddingmodel = []
for k in range(len(token[i][0])):
    temp_embeddingmodel.append(token[i][0][k])
embeddingmodel.append(temp_embeddingmodel)
embedding = Word2Vec(embeddingmodel, size=300, window=5, min_count=3, iter=100, sg=1,workers=4, max_vocab_size = 360000000)
embedding.save('post.embedding')

Bi_LSTM_train_csv.py

# -*- coding: utf-8 -*-

import time
import tensorflow as tf
import numpy as np
import Bi_LSTM as Bi_LSTM
import Word2Vec as Word2Vec
import csv
import os


tokens = token = loadCsv("test_data")
train_X = tokens[:, 0]
train_Y = tokens[:, 1]

train_Y_ = W2V.One_hot(train_Y)  ## Convert to One-hot
train_X_ = W2V.Convert2Vec("Data/post.embedding",train_X)  ## import word2vec model where you have trained before

Batch_size = 64
Total_size = len(train_X)
Vector_size = 300
seq_length = [len(x) for x in train_X]
Maxseq_length = max(seq_length)
learning_rate = 0.001
lstm_units = 128
num_class = 2
training_epochs = 5
keep_prob = 0.75

X = tf.placeholder(tf.float32, shape = [None, Maxseq_length, Vector_size], name = 'X')
Y = tf.placeholder(tf.float32, shape = [None, num_class], name = 'Y')
seq_len = tf.placeholder(tf.int32, shape = [None])

BiLSTM = Bi_LSTM.Bi_LSTM(lstm_units, num_class, keep_prob)

with tf.variable_scope("loss", reuse = tf.AUTO_REUSE):
logits = BiLSTM.logits(X, BiLSTM.W, BiLSTM.b, seq_len)
loss, optimizer = BiLSTM.model_build(logits, Y, learning_rate)

prediction = tf.nn.softmax(logits)
correct_pred = tf.equal(tf.argmax(prediction, 1), tf.argmax(Y, 1))
accuracy = tf.reduce_mean(tf.cast(correct_pred, tf.float32))

init = tf.global_variables_initializer()

total_batch = int(Total_size / Batch_size)

print("Start training!")


saver = tf.train.Saver()
ckpt = tf.train.get_checkpoint_state("")
print(ckpt.model_checkpoint_path)


with tf.Session() as sess:

start_time = time.time()
saver.restore(sess, ckpt.model_checkpoint_path) # search for checkpoint file
graph = tf.get_default_graph()
train_writer = tf.summary.FileWriter('Bidirectional_LSTM', sess.graph)
i = 0
for epoch in range(training_epochs):

    avg_acc, avg_loss = 0. , 0.
    for step in range(total_batch):
        train_batch_X = train_X_[step*Batch_size : step*Batch_size+Batch_size]
        train_batch_Y = train_Y_[step*Batch_size : step*Batch_size+Batch_size]
        batch_seq_length = seq_length[step*Batch_size : step*Batch_size+Batch_size]

        train_batch_X = W2V.Zero_padding(train_batch_X, Batch_size, Maxseq_length, Vector_size)

        sess.run(optimizer, feed_dict={X: train_batch_X, Y: train_batch_Y, seq_len: batch_seq_length})
        # Compute average loss
        loss_ = sess.run(loss, feed_dict={X: train_batch_X, Y: train_batch_Y, seq_len: batch_seq_length})
        avg_loss += loss_ / total_batch

        acc = sess.run(accuracy , feed_dict={X: train_batch_X, Y: train_batch_Y, seq_len: batch_seq_length})
        avg_acc += acc / total_batch
        print("epoch : {:02d} step : {:04d} loss = {:.6f} accuracy= {:.6f}".format(epoch+1, step+1, loss_, acc))

    summary = sess.run(BiLSTM.graph_build(avg_loss, avg_acc))
    train_writer.add_summary(summary, i)
    i += 1

duration = time.time() - start_time
minute = int(duration / 60)
second = int(duration) % 60
print("%dminutes %dseconds" % (minute,second))
save_path = saver.save(sess, "/Users/aa/Downloads/Bi_LSTM/Data/Bi_LSTM")

train_writer.close()
print('save_path',save_path)

我仍然不知道post.embedding和Bidirectional_LSTM在做什么。如何提高此过程的准确性？

我应该再训练Bidirectional_LSTM吗？还是我应该训练更多post.embedding？我应该做这件事还是两者都做？

将分类器与Word2Vec和Bidirectional_LSTM一起使用时，如何提高准确性？

0 个答案: