我的代码是预测句子的情绪。我训练了一个CNN模型并保存了它。当我加载我的模型并尝试预测句子的情绪时，我对同一个句子有不同的预测。我的代码如下，当我尝试通过调用底部的函数predict_cnn_word2vec来预测sentene时会出现问题：
import logging;
import numpy as np;
import tensorflow as tf;
import sklearn as sk
import re;
import json
import string;
import math
import os
from sklearn.metrics import recall_score, f1_score, precision_score;
class CNN(object):
def __init__(self,logger):
self.logger = logger;
def _weight_variable(self,shape):
initial = tf.truncated_normal(shape, stddev = 0.1);
return tf.Variable(initial);
def _bias_variable(self,shape):
initial = tf.constant(0.1, shape = shape);
return tf.Variable(initial);
def _conv2d(self,x, W, b, strides=1):
# convolve and relu activation
x = tf.nn.conv2d(x, W, strides=[1, strides, strides, 1], padding='SAME');
x = tf.nn.bias_add(x, b);
return tf.nn.relu(x);
def _maxpool(self,x, k=2):
return tf.nn.max_pool(x, ksize=[1, k, k, 1], strides=[1, k, k, 1], padding='SAME');
def _init_CNN(self, sentence_width, sentence_height, dropout, learning_rate,n_class,is_training):
self.logger.info("----------Initiating CNN---------");
self.X = tf.placeholder(tf.float32, [None, sentence_height * sentence_width]);
self.Y = tf.placeholder(tf.float32, [None, n_class]);
x = tf.reshape(self.X, shape = [-1, sentence_height, sentence_width, 1]);
#1st convolution layer
wc1 = tf.Variable(tf.random_normal([3, 3, 1, 5]));
bc1 = tf.Variable(tf.random_normal([5]))
stride1 = 2;
pool1 = 2
conv1 = self._conv2d(x, wc1, bc1,stride1);
conv1 = self._maxpool(conv1, pool1);
conv2 = conv1;
pools = [2]#,2,2];
strides = [2]#,1,1];
last_channel = 5;
first_size = self._get_first_connected_size(sentence_height,sentence_width, strides,pools,last_channel);
# #1st fully connected layer
wf1 = tf.Variable(tf.random_normal([first_size, 32]));
bf1 = tf.Variable(tf.random_normal([32]));
fc1 = tf.reshape(conv2, [-1, wf1.get_shape().as_list()[0]]);
fc1 = tf.add(tf.matmul(fc1, wf1), bf1);
fc1 = tf.nn.relu(fc1);
fc1 = tf.nn.dropout(fc1, dropout)
#dropout layer
outw = tf.Variable(tf.random_normal([32, n_class]));
outb = tf.Variable(tf.random_normal([n_class]));
self.pred = tf.add(tf.matmul(fc1, outw), outb);
self.y_p = tf.argmax(self.pred,1);
if is_training is False:
return;
# self.pred = self._predict(self.X,sentence_width, sentence_height, settings, dropout);
self.cost = tf.reduce_mean(tf.nn.softmax_cross_entropy_with_logits(self.pred, self.Y));
self.optimizer = tf.train.AdamOptimizer(learning_rate=learning_rate).minimize(self.cost);
#accuracy
self.y_t = tf.argmax(self.Y,1);
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(self.y_p, self.y_t), "float"))
self.init = tf.initialize_all_variables();
def _get_first_connected_size(self,ih,iw, strides,pools,last_channel):
i = 1;
while i <= len(strides):
iw = math.ceil(float(iw) / float(strides[i-1]));
iw = math.ceil(float(iw) /pools[i-1]);
ih = math.ceil(float(ih) / float(strides[i-1]));
ih = math.ceil(float(ih) /pools[i-1]);
i = i + 1;
first_connected_size = int(ih*iw*last_channel);
return first_connected_size;
def train(self,data_provider,config):
self._init_CNN(config.sentence_width, config.num_word, config.dropout, config.learning_rate,config.n_class,True);
sess = tf.Session();
sess.run(self.init);
self.logger.info("Start Training!");
#saver
saver = tf.train.Saver();
cur_max_accuracy = 0;
cur_max_recall = 0;
cur_max_precision = 0;
if config.model_init_from is not None and os.path.exists(config.model_init_from):
#restore model if exist
saver.restore(sess, config.model_init_from);
for epoch in range(config.epochs):
data_provider.reset_batch_pointer();
for i in range(data_provider.num_batches):
batch_x, batch_y = data_provider.next_batch();
accuracy_score,y_p,y_t, _, cost = sess.run([self.accuracy,self.y_p, self.y_t, self.optimizer, self.cost], feed_dict={self.X: batch_x, self.Y: batch_y});
if i %10 == 0:
self.logger.info("(%d/%d,%d epo) cost = %f, accuracy = %f,precision = %f, recall = %f, f_score = %f" % (i+epoch * data_provider.num_batches, data_provider.num_batches*config.epochs, epoch,cost,accuracy_score,precision_score(y_t, y_p),recall_score(y_t,y_p),f1_score(y_t,y_p)));
self.accuracy = tf.reduce_mean(tf.cast(tf.equal(self.y_p, self.y_t), "float"))
accuracy_score,y_p,y_t, _,cost = sess.run([self.accuracy,self.y_p, self.y_t, self.optimizer, self.cost], feed_dict={self.X: data_provider.get_test_X(), self.Y: data_provider.get_test_Y()})
precision_score1 = precision_score(y_t, y_p);
recall_score1 = recall_score(y_t,y_p);
f1_score1 = f1_score(y_t,y_p);
self.logger.info("#####(%d/%d epoch) cost = %f, accuracy = %f(max: %f), precision = %f(max: %f), recall = %f(max:%f), f_score = %f" % (epoch,config.epochs, cost,accuracy_score,cur_max_accuracy, precision_score1,cur_max_precision,recall_score1,cur_max_recall,f1_score1));
save_loc = saver.save(sess, config.model_save_path);
print("Model has been saved to: %s" % save_loc);
cur_max_accuracy = accuracy_score;
cur_max_recall = recall_score1;
cur_max_precision = precision_score1;
def predict_cnn_word2vec(self, data_provider, config):
if not os.path.exists(config.model_init_from):
self.logger.info("model does not exist!");
sys.exit(2);
self._init_CNN(config.sentence_width, config.num_word, config.dropout, config.learning_rate,config.n_class,False)
while True:
sentence = input("Enter a sentence:");
with tf.Session() as sess:
saver = tf.train.Saver();
saver.restore(sess, config.model_init_from);
batch_x = np.empty((1, config.num_word*300));
batch_x[0,:] = data_provider.get_sentence_vec(sentence,config);
y_p = sess.run([self.y_p], feed_dict={self.X: batch_x});
result = "positive" if y_p == [1] else "negative";
self.logger.info("[%s] is %s" %(sentence,result));
当我在main函数中调用predict_cnn_word2vec（）时，我输入了一个while循环。对于每个句子，将重新加载模型以预测句子的情绪。对于前几个循环，它给出了相同的结果。但随着我做出更多预测，结果似乎发生了变化。以下是我的日志示例：
Enter a sentence:i adore him
2016-08-11 05:26:16,690 : INFO : [i adore him] is positive
Enter a sentence:i adore him
2016-08-11 05:26:19,662 : INFO : [i adore him] is positive
Enter a sentence:i adore him
2016-08-11 05:26:21,176 : INFO : [i adore him] is positive
Enter a sentence:i adore him
2016-08-11 05:26:22,568 : INFO : [i adore him] is positive
Enter a sentence:i adore him
2016-08-11 05:26:23,724 : INFO : [i adore him] is negative
Enter a sentence:i adore him
2016-08-11 05:26:25,791 : INFO : [i adore him] is negative
我已经检查了我的data_provider，它将我的句子放入向量中，并且每次返回相同的句子向量时它都能正常工作。所以这个问题不应该归因于输入数据。我想知道每次tensorflow重新加载模型时，模型中的某些内容都会正确重新加载。有人可以帮我解决这个问题吗？
答案 0 :(得分：5)
你在推理期间没有关闭辍学！
Dropout会导致图层中单位的随机部分将其激活次数降至0.这在训练期间是一个有用的正则化程序，但在验证或测试模型或在生产中运行时，您不希望出现此行为。 / p>
您不需要使用浮动退出初始化网络并继续您的快乐方式，而是需要将丢包作为占位符，就像您的输入和目标一样。在训练期间，通过feed dict将此占位符设置为合理的（例如0.5）。在推理期间，将此占位符设置为1.0。