- tensorflow实现基于LSTM的文本分类方法
- LSTM Networks for Sentiment Analysis
This code belongs to the "implement a RNN for text classification in Tensorflow" POST
Contrast with IMPLEMENTING A CNN FOR TEXT CLASSIFICATION IN TENSORFLOW
#How to run
bob@bob-OptiPlex-3020:~/workspace/github/RNN_Text_Classify$ python3 train_rnn_classify.py
loading the dataset...
load data from %s data/subj0.pkl
begin training
WARNING:tensorflow:From train_rnn_classify.py:145 in train_step.: all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2016-03-02.
Instructions for updating:
Please use tf.global_variables instead.
WARNING:tensorflow:From train_rnn_classify.py:148 in train_step.: initialize_all_variables (from tensorflow.python.ops.variables) is deprecated and will be removed after 2017-03-02.
Instructions for updating:
Use tf.global_variables_initializer instead.
the 1 epoch training...
the 100 step, train cost is: 0.684807 and the train accuracy is 0.593750 and the valid accuracy is 0.495556
Saved model chechpoint to/home/bob/workspace/github/RNN_Text_Classify/runs/checkpoints/model-127
tensorboard --logdir ./runs/summaries/</code?
#Requirements
Python 2.7/3.5
Tensorflow > 1.1/0.12.0rc0
Numpy
#Training
train_rnn_classify.py
#note
this code save the model checkpoints and save the summary at given epoch, these actions will slow down the training speed,so if you want to do a quick job, just remove these codes if you need
dataset_path='data/subj0.pkl' 请问文件格式和内容
是文本词表的单词索引，比如“I love the movie”对应的就是[1,2,3]等（举个例子），索引是按照词频的排序从高往低排列的
train_set_y里面是类别的索引，比如4分类索引就是0，1，2，3这四个数，因为我损失函数里面用到的是sparse_softmax_loss，当然你这种one-hot的写法也是可以的，对应的损失函数就是softmax_loss，对结果没有影响（当然代码调整下维度就行）
代码中data的target，也就是y是怎么设的呢，是最终的分类目标吗？ 我能拿到mean_pooling的输出h吗
target就是分类的目标，也就是类别标签，可以拿到mean-pooling的h，在代码中加上fetches加上h就行了
mask的作用是什么
tensorflow输入矩阵是指定的，但是一个batch的句子长度可能是不定的，所以需要进行padding，但是padding的数据是最后不能参与分类的，所以需要mask标记一下