我正在尝试使用Tensorflow学习用于情感分析的LSTM模型,我已经浏览了LSTM model。
以下代码(create_sentiment_featuresets.py)会从 5000个正面句子和5000个否定句子生成词典。
import nltk
from nltk.tokenize import word_tokenize
import numpy as np
import random
from collections import Counter
from nltk.stem import WordNetLemmatizer
lemmatizer = WordNetLemmatizer()
def create_lexicon(pos, neg):
lexicon = []
with open(pos, 'r') as f:
contents = f.readlines()
for l in contents[:len(contents)]:
l= l.decode('utf-8')
all_words = word_tokenize(l)
lexicon += list(all_words)
f.close()
with open(neg, 'r') as f:
contents = f.readlines()
for l in contents[:len(contents)]:
l= l.decode('utf-8')
all_words = word_tokenize(l)
lexicon += list(all_words)
f.close()
lexicon = [lemmatizer.lemmatize(i) for i in lexicon]
w_counts = Counter(lexicon)
l2 = []
for w in w_counts:
if 1000 > w_counts[w] > 50:
l2.append(w)
print("Lexicon length create_lexicon: ",len(lexicon))
return l2
def sample_handling(sample, lexicon, classification):
featureset = []
print("Lexicon length Sample handling: ",len(lexicon))
with open(sample, 'r') as f:
contents = f.readlines()
for l in contents[:len(contents)]:
l= l.decode('utf-8')
current_words = word_tokenize(l.lower())
current_words= [lemmatizer.lemmatize(i) for i in current_words]
features = np.zeros(len(lexicon))
for word in current_words:
if word.lower() in lexicon:
index_value = lexicon.index(word.lower())
features[index_value] +=1
features = list(features)
featureset.append([features, classification])
f.close()
print("Feature SET------")
print(len(featureset))
return featureset
def create_feature_sets_and_labels(pos, neg, test_size = 0.1):
global m_lexicon
m_lexicon = create_lexicon(pos, neg)
features = []
features += sample_handling(pos, m_lexicon, [1,0])
features += sample_handling(neg, m_lexicon, [0,1])
random.shuffle(features)
features = np.array(features)
testing_size = int(test_size * len(features))
train_x = list(features[:,0][:-testing_size])
train_y = list(features[:,1][:-testing_size])
test_x = list(features[:,0][-testing_size:])
test_y = list(features[:,1][-testing_size:])
return train_x, train_y, test_x, test_y
def get_lexicon():
global m_lexicon
return m_lexicon
以下代码(sentiment_analysis.py)是使用简单神经网络模型进行情感分析并且工作正常
from create_sentiment_featuresets import create_feature_sets_and_labels
from create_sentiment_featuresets import get_lexicon
import tensorflow as tf
import numpy as np
# extras for testing
from nltk.tokenize import word_tokenize
from nltk.stem import WordNetLemmatizer
lemmatizer = WordNetLemmatizer()
#- end extras
train_x, train_y, test_x, test_y = create_feature_sets_and_labels('pos.txt', 'neg.txt')
# pt A-------------
n_nodes_hl1 = 1500
n_nodes_hl2 = 1500
n_nodes_hl3 = 1500
n_classes = 2
batch_size = 100
hm_epochs = 10
x = tf.placeholder(tf.float32)
y = tf.placeholder(tf.float32)
hidden_1_layer = {'f_fum': n_nodes_hl1,
'weight': tf.Variable(tf.random_normal([len(train_x[0]), n_nodes_hl1])),
'bias': tf.Variable(tf.random_normal([n_nodes_hl1]))}
hidden_2_layer = {'f_fum': n_nodes_hl2,
'weight': tf.Variable(tf.random_normal([n_nodes_hl1, n_nodes_hl2])),
'bias': tf.Variable(tf.random_normal([n_nodes_hl2]))}
hidden_3_layer = {'f_fum': n_nodes_hl3,
'weight': tf.Variable(tf.random_normal([n_nodes_hl2, n_nodes_hl3])),
'bias': tf.Variable(tf.random_normal([n_nodes_hl3]))}
output_layer = {'f_fum': None,
'weight': tf.Variable(tf.random_normal([n_nodes_hl3, n_classes])),
'bias': tf.Variable(tf.random_normal([n_classes]))}
def nueral_network_model(data):
l1 = tf.add(tf.matmul(data, hidden_1_layer['weight']), hidden_1_layer['bias'])
l1 = tf.nn.relu(l1)
l2 = tf.add(tf.matmul(l1, hidden_2_layer['weight']), hidden_2_layer['bias'])
l2 = tf.nn.relu(l2)
l3 = tf.add(tf.matmul(l2, hidden_3_layer['weight']), hidden_3_layer['bias'])
l3 = tf.nn.relu(l3)
output = tf.matmul(l3, output_layer['weight']) + output_layer['bias']
return output
# pt B--------------
def train_neural_network(x):
prediction = nueral_network_model(x)
cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(logits= prediction, labels= y))
optimizer = tf.train.AdamOptimizer(learning_rate= 0.001).minimize(cost)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(hm_epochs):
epoch_loss = 0
i = 0
while i < len(train_x):
start = i
end = i+ batch_size
batch_x = np.array(train_x[start: end])
batch_y = np.array(train_y[start: end])
_, c = sess.run([optimizer, cost], feed_dict= {x: batch_x, y: batch_y})
epoch_loss += c
i+= batch_size
print('Epoch', epoch+ 1, 'completed out of ', hm_epochs, 'loss:', epoch_loss)
correct= tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct, 'float'))
print('Accuracy:', accuracy.eval({x:test_x, y:test_y}))
# testing --------------
m_lexicon= get_lexicon()
print('Lexicon length: ',len(m_lexicon))
input_data= "David likes to go out with Kary"
current_words= word_tokenize(input_data.lower())
current_words = [lemmatizer.lemmatize(i) for i in current_words]
features = np.zeros(len(m_lexicon))
for word in current_words:
if word.lower() in m_lexicon:
index_value = m_lexicon.index(word.lower())
features[index_value] +=1
features = np.array(list(features)).reshape(1,-1)
print('features length: ',len(features))
result = sess.run(tf.argmax(prediction.eval(feed_dict={x:features}), 1))
print(prediction.eval(feed_dict={x:features}))
if result[0] == 0:
print('Positive: ', input_data)
elif result[0] == 1:
print('Negative: ', input_data)
train_neural_network(x)
我正在尝试修改LSTM模型的上述(sentiment_analysis.py) 在 mnist image dataset 上阅读LSTM的RNN w/ LSTM cell example in TensorFlow and Python之后:
通过许多命中和运行路径,我能够获得以下运行代码(sentiment_demo_lstm.py):
import tensorflow as tf
from tensorflow.contrib import rnn
from create_sentiment_featuresets import create_feature_sets_and_labels
from create_sentiment_featuresets import get_lexicon
import numpy as np
# extras for testing
from nltk.tokenize import word_tokenize
from nltk.stem import WordNetLemmatizer
lemmatizer = WordNetLemmatizer()
#- end extras
train_x, train_y, test_x, test_y = create_feature_sets_and_labels('pos.txt', 'neg.txt')
n_steps= 100
input_vec_size= len(train_x[0])
hm_epochs = 8
n_classes = 2
batch_size = 128
n_hidden = 128
x = tf.placeholder('float', [None, input_vec_size, 1])
y = tf.placeholder('float')
def recurrent_neural_network(x):
layer = {'weights': tf.Variable(tf.random_normal([n_hidden, n_classes])), # hidden_layer, n_classes
'biases': tf.Variable(tf.random_normal([n_classes]))}
h_layer = {'weights': tf.Variable(tf.random_normal([1, n_hidden])), # hidden_layer, n_classes
'biases': tf.Variable(tf.random_normal([n_hidden], mean = 1.0))}
x = tf.transpose(x, [1,0,2])
x = tf.reshape(x, [-1, 1])
x= tf.nn.relu(tf.matmul(x, h_layer['weights']) + h_layer['biases'])
x = tf.split(x, input_vec_size, 0)
lstm_cell = rnn.BasicLSTMCell(n_hidden, state_is_tuple=True)
outputs, states = rnn.static_rnn(lstm_cell, x, dtype= tf.float32)
output = tf.matmul(outputs[-1], layer['weights']) + layer['biases']
return output
def train_neural_network(x):
prediction = recurrent_neural_network(x)
cost = tf.reduce_mean( tf.nn.softmax_cross_entropy_with_logits(logits= prediction, labels= y))
optimizer = tf.train.AdamOptimizer(learning_rate= 0.001).minimize(cost)
with tf.Session() as sess:
sess.run(tf.global_variables_initializer())
for epoch in range(hm_epochs):
epoch_loss = 0
i = 0
while (i+ batch_size) < len(train_x):
start = i
end = i+ batch_size
batch_x = np.array(train_x[start: end])
batch_y = np.array(train_y[start: end])
batch_x = batch_x.reshape(batch_size ,input_vec_size, 1)
_, c = sess.run([optimizer, cost], feed_dict= {x: batch_x, y: batch_y})
epoch_loss += c
i+= batch_size
print('--------Epoch', epoch+ 1, 'completed out of ', hm_epochs, 'loss:', epoch_loss)
correct= tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))
accuracy = tf.reduce_mean(tf.cast(correct, 'float'))
print('Accuracy:', accuracy.eval({x:np.array(test_x).reshape(-1, input_vec_size, 1), y:test_y}))
# testing --------------
m_lexicon= get_lexicon()
print('Lexicon length: ',len(m_lexicon))
input_data= "Mary does not like pizza" #"he seems to to be healthy today" #"David likes to go out with Kary"
current_words= word_tokenize(input_data.lower())
current_words = [lemmatizer.lemmatize(i) for i in current_words]
features = np.zeros(len(m_lexicon))
for word in current_words:
if word.lower() in m_lexicon:
index_value = m_lexicon.index(word.lower())
features[index_value] +=1
features = np.array(list(features)).reshape(-1, input_vec_size, 1)
print('features length: ',len(features))
result = sess.run(tf.argmax(prediction.eval(feed_dict={x:features}), 1))
print('RESULT: ', result)
print(prediction.eval(feed_dict={x:features}))
if result[0] == 0:
print('Positive: ', input_data)
elif result[0] == 1:
print('Negative: ', input_data)
train_neural_network(x)
输出
print(train_x[0])
[0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 1.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0, 0.0]
print(train_y[0])
[0, 1]
len(train_x)= 9596
,len(train_x[0]) = 423
表示train_x
是9596x423的列表?
我现在有一个正在运行的代码,我仍然有很多疑问。
在 sentiment_demo_lstm 中,我无法理解以下内容
x = tf.transpose(x, [1,0,2])
x = tf.reshape(x, [-1, 1])
x = tf.split(x, input_vec_size, 0)
我打印了以下形状:
x = tf.placeholder('float', [None, input_vec_size, 1]) ==> TensorShape([Dimension(None), Dimension(423), Dimension(1)]))
x = tf.transpose(x, [1,0,2]) ==> TensorShape([Dimension(423), Dimension(None), Dimension(1)]))
x = tf.reshape(x, [-1, 1]) ==> TensorShape([Dimension(None), Dimension(1)]))
x = tf.split(x, input_vec_size, 0) ==> ?
这里我将隐藏图层的数量设为128,是否需要与输入数量相同,即len(train_x)= 9596
中的值1
x = tf.placeholder('float', [None, input_vec_size, 1])
和
x = tf.reshape(x, [-1, 1])
是因为train_x[0]
是428x 1 ?
以下是为了匹配占位符
batch_x = np.array(train_x[start: end]) ==> (128, 423)
batch_x = batch_x.reshape(batch_size ,input_vec_size, 1) ==> (128, 423, 1)
x = tf.placeholder('float', [None, input_vec_size, 1])
尺寸,对吧?
如果我修改了代码:
while (i+ batch_size) < len(train_x):
作为
while i < len(train_x):
我收到以下错误:
Traceback (most recent call last):
File "sentiment_demo_lstm.py", line 131, in <module>
train_neural_network(x)
File "sentiment_demo_lstm.py", line 86, in train_neural_network
batch_x = batch_x.reshape(batch_size ,input_vec_size, 1)
ValueError: cannot reshape array of size 52452 into shape (128,423,1)
=&GT;我不能在训练时包括最后124个记录/功能集吗?
答案 0 :(得分:10)
这是加载的问题。让我试着用简单的英语隐藏所有复杂的内部细节:
具有3个步骤的简单展开LSTM模型如下所示。每个LSTM单元获取前一个LSTM单元的输入向量和隐藏输出向量,并为下一个LSTM单元产生输出向量和隐藏输出。
下面显示了相同模型的简明表示。
LSTM模型是序列模型的序列,即,当序列必须用另一个序列标记时,它们用于问题,如句子中每个单词的POS标记或NER标记。
您似乎将其用于分类问题。有两种方法可以使用LSTM模型进行分类
1)获取所有状态的输出(在我们的示例中为O1,O2和O3)并应用softmax层,softmax层输出大小等于类的数量(在您的情况下为2)
2)获取最后一个状态(O3)的输出并对其应用softmax图层。 (这就是你在鳕鱼中所做的事情。输出[-1]返回输出中的最后一行)
所以我们在softmax输出的误差上反向传播(Backpropagation Through Time - BTT)。
使用Tensorflow实现实现,让我们看看LSTM模型的输入和输出是什么。
每个LSTM接受输入,但我们有3个这样的LSTM单元,因此输入(X占位符)应该是大小的(输入大小*时间步长)。但我们不计算单输入和BTT的误差,而是我们在一批输入 - 输出组合上进行计算。所以LSTM的输入将是(batchsize * inputsize *时间步长)。
LSTM单元格定义为隐藏状态的大小。输出的大小和LSTM单元的隐藏输出向量将与隐藏状态的大小相同(检查LSTM内部计算的原因!)。然后,我们使用这些LSTM单元的列表定义LSTM模型,其中列表的大小将等于模型的展开次数。因此,我们定义了每次展开时要完成的展开次数和输入的大小。
我已经跳过很多东西,例如如何处理可变长度序列,序列到序列错误计算,LSTM如何计算输出和隐藏输出等。
开始实施,您将在每个LSTM单元格的输入之前应用relu图层。我不明白为什么你这样做,但我猜你是这样做的,你的输入大小映射到LSTM输入大小。
回答你的问题:
让我们说input_vec_size = 3
你正在传递一个大小为[128 * 3 * 1]
的ndarrayx = tf.transpose(x,[1,0,2]) - &gt; [3 * 128 * 1]
x = tf.reshape(x,[ - 1,1]) - &gt; [384 * 1]
h_layer ['weights'] - &gt; [1,128]
x = tf.nn.relu(tf.matmul(x,h_layer ['weights'])+ h_layer ['biases']) - &gt; [384 * 128]
没有输入尺寸隐藏尺寸不同。 LSTM对输入和先前的隐藏输出执行一组操作,并给出输出和下一个隐藏输出,这两个输出都是隐藏大小的。
x = tf.placeholder('float',[None,input_vec_size,1])
它定义了一个张量或ndarray或可变数量的行,每行有input_vec_size列a,每个值都是单个值向量。
x = tf.reshape(x,[ - 1,1]) - &gt;将输入x重新整形为一个大小固定为1列和任意行数的矩阵。
batch_x.reshape将失败。这可能是最后一批的情况,因为len(train_x)可能不是batch_size的倍数,导致未完全填充的最后一批。
您可以使用
来避免此问题batch_x = batch_x.reshape(-1 ,input_vec_size, 1)
但我仍然不确定你为什么在输入层前面使用Relu。
您正在最后一个单元格的输出处应用逻辑回归,这很好。
您可以使用双向LSTM查看我的玩具示例,它是一个分类器,用于对序列是增加还是减少或混合进行分类。