seq2vec 0.4.0

Turn sequence of words into a fix-length representation vector

This is a version to refactor all the seq2vec structures and use customed layers in yklz.

Install

pip install seq2vec

or clone the repo, then install:

git clone --recursive https://github.com/Yoctol/seq2vec.git
python setup.py install

Usage

Simple hash:

fromseq2vecimportSeq2VecHashtransformer=Seq2VecHash(vector_length=100)
seqs= [
['我', '有', '一個', '蘋果'],
['我', '有', 'pineapple'],
]
result=transformer.transform(seqs)
print(result)
'''array([[ 0., 0., 0., 0., 0., 0., 0., 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., 1., 0., 0., 0., 0., 0., 0., 0., 0., 0., 0., 1., 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., 1., 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., 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.]])'''

Sequence-to-sequence auto-encoder:

• LSTM to LSTM auto-encoder with word embedding (RNN to RNN architecture)

  fromseq2vec.word2vecimportGensimWord2vecfromseq2vecimportSeq2VecR2RWord# load Gensim word2vec from word2vec_model_pathword2vec=GensimWord2vec(word2vec_model_path)
  transformer=Seq2VecR2RWord(
  word2vec_model=word2vec,
  max_length=20,
  latent_size=300,
  encoding_size=300,
  learning_rate=0.05
  )
  train_seq= [
  ['我', '有', '一個', '蘋果'],
  ['我', '有', '筆'],
  ['一個', '鳳梨'],
  ]
  test_seq= [
  ['我', '愛', '吃', '鳳梨'],
  ]
  transformer.fit(train_seq)
  result=transformer.transform(test_seq)

• CNN to LSTM auto-encoder with word embedding (CNN to RNN architecture)

  fromseq2vec.word2vecimportGensimWord2vecfromseq2vecimportSeq2VecC2RWord# load Gensim word2vec from word2vec_model_pathword2vec=GensimWord2vec(word2vec_model_path)
  transformer=Seq2VecC2RWord(
  word2vec_model=word2vec,
  max_length=20,
  latent_size=300,
  conv_size=5,
  channel_size=10,
  learning_rate=0.05,
  )
  train_seq= [
  ['我', '有', '一個', '蘋果'],
  ['我', '有', '筆'],
  ['一個', '鳳梨'],
  ]
  test_seq= [
  ['我', '愛', '吃', '鳳梨'],
  ]
  transformer.fit(train_seq)
  result=transformer.transform(test_seq)

• CNN to LSTM auto-encoder with char embedding (CNN to RNN architecture)

  fromseq2vec.word2vecimportGensimWord2vecfromseq2vecimportSeq2VecC2RChar# load Gensim word2vec from word2vec_model_pathword2vec=GensimWord2vec(word2vec_model_path)
  transformer=Seq2VecC2RChar(
  word2vec_model=word2vec,
  max_index=1000,
  max_length=20,
  embedding_size=200,
  latent_size=200,
  learning_rate=0.05,
  channel_size=10,
  conv_size=5
  )
  train_seq= [
  ['我', '有', '一個', '蘋果'],
  ['我', '有', '筆'],
  ['一個', '鳳梨'],
  ]
  test_seq= [
  ['我', '愛', '吃', '鳳梨'],
  ]
  transformer.fit(train_seq)
  result=transformer.transform(test_seq)

• LSTM to LSTM auto-encoder with hash word embedding (RNN to RNN architecture)

fromseq2vecimportSeq2VecR2RHashtransformer=Seq2VecR2RHash(
max_index=1000,
max_length=10,
latent_size=20,
embedding_size=200,
encoding_size=300,
learning_rate=0.05
)
train_seq= [
['我', '有', '一個', '蘋果'],
['我', '有', '筆'],
['一個', '鳳梨'], ]
test_seq= [
['我', '愛', '吃', '鳳梨'],
]
transformer.fit(train_seq)
result=transformer.transform(test_seq)

Training with generator on file

We provide an example with LSTM to LSTM auto-encoder (word embedding).

Use the following training method while lack of memory is an issue for you.

The file should be a tokenized txt file splitted by whitespace with a sequence per line.

fromseq2vec.word2vecimportGensimWord2vecfromseq2vec.modelimportSeq2VecR2RWordfromseq2vec.transformerimportWordEmbeddingTransformerfromseq2vec.utilimportDataGenteratorword2vec=GensimWord2vec(word2vec_model_path)
max_length=20transformer=Seq2VecR2RWord(
word2vec_model=word2vec,
max_length=max_length,
latent_size=200,
encoding_size=300,
learning_rate=0.05
)
train_data=DataGenterator(
corpus_for_training_path, transformer.input_transformer,
transformer.output_transformer, batch_size=128
)
test_data=DataGenterator(
corpus_for_validation_path, transformer.input_transformer,
transformer.output_transformer, batch_size=128
)
transformer.fit_generator(
train_data,
test_data,
epochs=10,
batch_number=1250# The number of batch per epoch
)
transformer.save_model(model_path) # save your model# You can reload your model and retrain it.transformer.load_model(model_path)
transformer.fit_generator(
train_data,
test_data,
epochs=10,
batch_number=1250# The number of batch per epoch
)

Customized your seq2vec model with our auto-encoder framework

You can customize your seq2vec model easily with our framework.

importkerasfromseq2vec.modelimportTrainableSeq2VecBaseclassYourSeq2Vec(TrainableSeq2VecBase):
def__init__(selfmax_length,
latent_size,
learning_rate
):
# initialize your setting and set input_transformer# and output_transformer# Input and output transformers transform data from # raw sequence into Keras Layer input format# See seq2vec.transformer for more detailself.input_transformer=YourInputTransformer()
self.output_transformer=YourOutputTransformer()
# add your customized layerself.custom_objects= {}
self.custom_objects[customized_class_name] =customized_classsuper(YourSeq2Vec, self).__init__(
max_length,
latent_size,
learning_rate
)
defcreate_model(self):
# create and compile your model in this function# You should return your model and encoder here# encoder is the one encoded input sequencesmodel.compile(loss)
returnmodel, encoderdefload_model(self, file_path):
# load your seq2vec model here and set its attribute valuesself.model=self.load_customed_model(file_path)

Lint

pylint --rcfile=./yoctol-pylintrc/.pylintrc seq2vec

Test

python -m unittest