Keras implementation of Google BERT(Bidirectional Encoder Representations from Transformers) and OpenAI's Transformer LM capable of loading pretrained models with a finetuning API.
How to use it?
# this is a pseudo code you can read an actual working example in tutorial.ipynb or the colab notebook text_encoder = MyTextEncoder(**my_text_encoder_params) # you create a text encoder (sentence piece and openai's bpe are included) lm_generator = lm_generator(text_encoder, **lm_generator_params) # this is essentially your data reader (single sentence and double sentence reader with masking and is_next label are included) task_meta_datas = [lm_task, classification_task, pos_task] # these are your tasks (the lm_generator must generate the labels for these tasks too) encoder_model = create_transformer(**encoder_params) # or you could simply load_openai() or you could write your own encoder(BiLSTM for example) trained_model = train_model(encoder_model, task_meta_datas, lm_generator, **training_params) # it does both pretraing and finetuning trained_model.save_weights('my_awesome_model') # save it model = load_model('my_awesome_model', encoder_model) # load it later and use it!
- The general idea of this library is to use OpenAI's/Google's pretrained model for transfer learning
- In order to see how the BERT model works, you can check this colab notebook
- In order to be compatible with both BERT and OpenAI I had to assume a standard ordering for the vocabulary, I'm using OpenAI's so in the loading function of BERT there is a part to change the ordering; but this is an implementation detail and you can ignore it!
- Loading OpenAI model is tested with both tensorflow and theano as backend
- Loading a Bert model is not possible on theano backend yet but the tf version is working and it has been tested
- Training and fine-tuning a model is not possible with theano backend but works perfectly fine with tensorflow
- You can use the data generator and task meta data for most of the NLP tasks and you can use them in other frameworks
- There are some unit tests for both dataset and transformer model (read them if you are not sure about something)
- Even tough I don't like my keras code, it's readable :)
- You can use other encoders, like LSTM or BiQRNN for training if you follow the model contract (have the same inputs and outputs as transformer encoder)
- Why should I use this instead of the official release?, first this one is in Keras and second it has a nice abstraction over token-level and sentence-level NLP tasks which is framework independent
- Why keras? pytorch version is already out! (BTW you can use this data generator for training and fine-tuning that model too)
- I strongly advise you to read the tutorial.ipynb (I don't like notebooks so this is a poorly designed notebook, but read it anyway)
Important code concepts
- Task: there are two general tasks, sentence level tasks(like is_next and sentiment analysis), and token level tasks(like PoS and NER)
- Sentence: a sentence represents an example with it's labels and everything, for each task it provides a target(single one for sentence level tasks and per token label for token level tasks) and a mask, for token levels we need to not only ignore paddings but also we might want to predict class on first char of a word (like the BERT paper(first piece of a multi piece word)) and for sentence levels we want a extraction point(like start token in BERT paepr)
- TaskWeightScheduler: for training we might want to start with language modeling and smoothly move to classification, they can be easily implemented with this class
- attention_mask: with this you can 1.make your model causal 2.ignore paddings 3.do your crazy idea :D
- special_tokens: pad, start, end, delimiter, mask