tf_classification 基于字符级中文文本常见BaseLine汇集

针对相同数据集分别采用CNN, RNN, RNN Attention机制进行分类

环境　推荐Anaconda mini管理

python3.6.5

tensorflow==1.8.0

项目文件.

1. data_utils.py 为数据预处理文件. 可以有Document*Sentence*Word 模式 和Document*Word模式. 预处理时采用收尾截取一定数量文本作为训练样本(总分总).
2. cnn_model.py 为cnn模型文件, 取固定句子长度. 采用Document*Word模式
3. rnn_model.py 为rnnbilstm模型文件, 每批次句子长度不固定. 采用Document*Word模式
4. rnn_attention.py 为rnngru attention模型文件, 句子长度不固定. 采用Document*Sentence*Word模式
5. train_model.py 为模型加载文件, 分为公共超参数和自定义参数.
6. predict.py 为预测文件.

效果比对(分为训练集和测试集, 超参数一样的情况下, cnn和rnn_attention只需要3轮就可以达到很好的效果, rnn 需要10轮甚至更多)

python train_model.py --model=cnn

2338 <Train>  Epoch: [3] Iter: [2334] Step: [149279] Loss: [0.044]    Acc: [1.000]
2339 <Train>  Epoch: [3] Iter: [2335] Step: [149343] Loss: [0.163]    Acc: [0.953]
2340 <Train>  Epoch: [3] Iter: [2336] Step: [149407] Loss: [0.029]    Acc: [1.000]
2341 <Train>  Epoch: [3] Iter: [2337] Step: [149471] Loss: [0.032]    Acc: [1.000]
2342 <Train>  Epoch: [3] Iter: [2338] Step: [149535] Loss: [0.140]    Acc: [0.938]
2343 <Train>  Epoch: [3] Iter: [2339] Step: [149599] Loss: [0.134]    Acc: [0.953]
2344 <Train>  Epoch: [3] Iter: [2340] Step: [149663] Loss: [0.120]    Acc: [0.953]
2345 <Train>  Epoch: [3] Iter: [2341] Step: [149727] Loss: [0.112]    Acc: [0.953]
2346 <Train>  Epoch: [3] Iter: [2342] Step: [149791] Loss: [0.048]    Acc: [0.969]
2347 <Train>  Epoch: [3] Iter: [2343] Step: [149855] Loss: [0.045]    Acc: [0.984]
2348 <Train>  Epoch: [3] Iter: [2344] Step: [149919] Loss: [0.350]    Acc: [0.969]
2349 <Train>  Epoch: [3] Iter: [2345] Step: [149983] Loss: [0.103]    Acc: [0.984]
2350 <Train>  Epoch: [3] Iter: [2346] Step: [150000] Loss: [0.004]    Acc: [1.000]

2354 <Test>   Iter: [1] Loss: [0.030]     Acc: [1.000]
2355 <Test>   Iter: [2] Loss: [0.152]     Acc: [0.938]
2356 <Test>   Iter: [3] Loss: [0.060]     Acc: [0.984]
2357 <Test>   Iter: [4] Loss: [0.101]     Acc: [0.953]
2358 <Test>   Iter: [5] Loss: [0.017]     Acc: [1.000]
2359 <Test>   Iter: [6] Loss: [0.088]     Acc: [0.969]
2360 <Test>   Iter: [7] Loss: [0.018]     Acc: [1.000]
2361 <Test>   Iter: [8] Loss: [0.071]     Acc: [0.969]
2362 <Test>   Iter: [9] Loss: [0.190]     Acc: [0.953]
2363 <Test>   Iter: [10] Loss: [0.062]    Acc: [0.969]

python train_model.py --model=rnn_attention

<Train>  Epoch: [3] Iter: [2334] Step: [149279] Loss: [0.079]    Acc: [0.953]
2339 <Train>  Epoch: [3] Iter: [2335] Step: [149343] Loss: [0.203]    Acc: [0.938]
2340 <Train>  Epoch: [3] Iter: [2336] Step: [149407] Loss: [0.214]    Acc: [0.922]
2341 <Train>  Epoch: [3] Iter: [2337] Step: [149471] Loss: [0.100]    Acc: [0.984]
2342 <Train>  Epoch: [3] Iter: [2338] Step: [149535] Loss: [0.068]    Acc: [0.984]
2343 <Train>  Epoch: [3] Iter: [2339] Step: [149599] Loss: [0.054]    Acc: [0.984]
2344 <Train>  Epoch: [3] Iter: [2340] Step: [149663] Loss: [0.059]    Acc: [0.969]
2345 <Train>  Epoch: [3] Iter: [2341] Step: [149727] Loss: [0.065]    Acc: [0.984]
2346 <Train>  Epoch: [3] Iter: [2342] Step: [149791] Loss: [0.209]    Acc: [0.969]
2347 <Train>  Epoch: [3] Iter: [2343] Step: [149855] Loss: [0.092]    Acc: [0.953]
2348 <Train>  Epoch: [3] Iter: [2344] Step: [149919] Loss: [0.113]    Acc: [0.953]
2349 <Train>  Epoch: [3] Iter: [2345] Step: [149983] Loss: [0.059]    Acc: [0.984]
2350 <Train>  Epoch: [3] Iter: [2346] Step: [150000] Loss: [0.062]    Acc: [1.000]

2354 <Test>   Iter: [1] Loss: [0.068]     Acc: [0.969]
2355 <Test>   Iter: [2] Loss: [0.173]     Acc: [0.953]
2356 <Test>   Iter: [3] Loss: [0.174]     Acc: [0.969]
2357 <Test>   Iter: [4] Loss: [0.210]     Acc: [0.922]
2358 <Test>   Iter: [5] Loss: [0.214]     Acc: [0.953]
2359 <Test>   Iter: [6] Loss: [0.221]     Acc: [0.922]
2360 <Test>   Iter: [7] Loss: [0.162]     Acc: [0.969]
2361 <Test>   Iter: [8] Loss: [0.013]     Acc: [1.000]
2362 <Test>   Iter: [9] Loss: [0.145]     Acc: [0.953]
2363 <Test>   Iter: [10] Loss: [0.060]    Acc: [0.984]

python train_model.py --model=rnn

 <Train>  Epoch: [3] Iter: [2334] Step: [149279] Loss: [0.567]    Acc: [0.781]
2339 <Train>  Epoch: [3] Iter: [2335] Step: [149343] Loss: [0.869]    Acc: [0.750]
2340 <Train>  Epoch: [3] Iter: [2336] Step: [149407] Loss: [0.661]    Acc: [0.812]
2341 <Train>  Epoch: [3] Iter: [2337] Step: [149471] Loss: [0.689]    Acc: [0.766]
2342 <Train>  Epoch: [3] Iter: [2338] Step: [149535] Loss: [0.556]    Acc: [0.828]
2343 <Train>  Epoch: [3] Iter: [2339] Step: [149599] Loss: [0.600]    Acc: [0.766]
2344 <Train>  Epoch: [3] Iter: [2340] Step: [149663] Loss: [0.932]    Acc: [0.719]
2345 <Train>  Epoch: [3] Iter: [2341] Step: [149727] Loss: [0.604]    Acc: [0.812]
2346 <Train>  Epoch: [3] Iter: [2342] Step: [149791] Loss: [0.535]    Acc: [0.844]
2347 <Train>  Epoch: [3] Iter: [2343] Step: [149855] Loss: [0.680]    Acc: [0.812]
2348 <Train>  Epoch: [3] Iter: [2344] Step: [149919] Loss: [0.506]    Acc: [0.828]
2349 <Train>  Epoch: [3] Iter: [2345] Step: [149983] Loss: [0.635]    Acc: [0.828]
2350 <Train>  Epoch: [3] Iter: [2346] Step: [150000] Loss: [0.332]    Acc: [1.000]

2354 <Test>   Iter: [1] Loss: [0.695]     Acc: [0.797]
2355 <Test>   Iter: [2] Loss: [0.783]     Acc: [0.734]
2356 <Test>   Iter: [3] Loss: [0.768]     Acc: [0.781]
2357 <Test>   Iter: [4] Loss: [0.981]     Acc: [0.703]
2358 <Test>   Iter: [5] Loss: [0.745]     Acc: [0.797]
2359 <Test>   Iter: [6] Loss: [0.606]     Acc: [0.828]
2360 <Test>   Iter: [7] Loss: [0.564]     Acc: [0.828]
2361 <Test>   Iter: [8] Loss: [0.519]     Acc: [0.891]
2362 <Test>   Iter: [9] Loss: [0.687]     Acc: [0.859]
2363 <Test>   Iter: [10] Loss: [0.896]    Acc: [0.781]

参考文献及Demo

主要采用了数据集和CNN相关文档: CNN中文文本分类Demo

分别对字和句子进行Attention处理: Attention机制论文

Seq2Seq图文: Attention机制博客

通俗易懂Attention: Attention机制CSDN

大神博客(神经网络相关): Colah博客