MatchZoo-Lite

基于 MatchZoo 2.0.0 开发，并做了简化

主要修改

增

增加数据加载器 dataloader 可方便进行数据的加载，训练数据和测试数据的文件格式统一为 json 文件，格式为：

{"text_left": "xxx xxx xx", "text_right": "xxx xxx xxx", "label": 1}

其中 text_left 和 text_right 为空格分割的分词文本

删改

• 去除 nltk 相关语料库的调用（如停用词）
• 去除预提供的 datasets
• 更换了测试 tests
• 去除部分模型，只保留以下模型：
  • arci
  • arcii
  • dssm
  • cdssm
  • conv_highway
  • duet
  • match_pyramid
  • mvlstm

Install

MatchZoo is dependent on Keras, please install one of its backend engines: TensorFlow, Theano, or CNTK. We recommend the TensorFlow backend. Two ways to install MatchZoo:

Install matchzoo-lite from the Github source

git clone http://gitlab.alipay-inc.com/niming.lxm/matchzoo-lite.git
cd matchzoo-lite
python setup.py install

Docker

docker pull seanlee97/matchzoo-lite:latest

Train your model

• train_duet.py
• train_arcii.py
• train_dssm.py
• train_mvlstm.py

Get Started in 60 Seconds

To train a Deep Semantic Structured Model, import matchzoo and prepare input data.

import matchzoo as mz

train_pack = mz.datasets.wiki_qa.load_data('train', task='ranking')
valid_pack = mz.datasets.wiki_qa.load_data('dev', task='ranking')
predict_pack = mz.datasets.wiki_qa.load_data('test', task='ranking')

Preprocess your input data in three lines of code, keep track parameters to be passed into the model.

preprocessor = mz.preprocessors.DSSMPreprocessor()
train_pack_processed = preprocessor.fit_transform(train_pack)
valid_pack_processed = preprocessor.transform(valid_pack)
predict_pack_processed = preprocessor.transform(predict_pack)

Make use of MatchZoo customized loss functions and evaluation metrics:

ranking_task = mz.tasks.Ranking(loss=mz.losses.RankCrossEntropyLoss(num_neg=4))
ranking_task.metrics = [
    mz.metrics.NormalizedDiscountedCumulativeGain(k=3),
    mz.metrics.NormalizedDiscountedCumulativeGain(k=5),
    mz.metrics.MeanAveragePrecision()
]

Initialize the model, fine-tune the hyper-parameters.

model = mz.models.DSSM()
model.params['input_shapes'] = preprocessor.context['input_shapes']
model.params['task'] = ranking_task
model.params['mlp_num_layers'] = 3
model.params['mlp_num_units'] = 300
model.params['mlp_num_fan_out'] = 128
model.params['mlp_activation_func'] = 'relu'
model.guess_and_fill_missing_params()
model.build()
model.compile()

Generate pair-wise training data on-the-fly, evaluate model performance using customized callbacks on prediction data.

train_generator = mz.PairDataGenerator(train_pack_processed, num_dup=1, num_neg=4, batch_size=64, shuffle=True)

pred_x, pred_y = predict_pack_processed.unpack()
evaluate = mz.callbacks.EvaluateAllMetrics(model, x=pred_x, y=pred_y, batch_size=len(pred_x))

history = model.fit_generator(train_generator, epochs=20, callbacks=[evaluate], workers=5, use_multiprocessing=False)

References

MatchZoo

Tutorials

English Documentation

中文文档

If you're interested in the cutting-edge research progress, please take a look at awaresome neural models for semantic match.

License

Apache-2.0

MatchZoo License

Apache-2.0 Copyright (c) 2015-present, Yixing Fan (faneshion)