GPT-2 Pre-training and text generation, implemented in Tensorflow 2.0

Originally implemented in tensorflow 1.14 by OapenAi :- “openai/gpt-2”. OpenAi GPT-2 Paper:-“Language Models are Unsupervised Multitask Learners”

**This repository has OpenAi GPT-2 pre-training and sequence generation implementation in tensorflow 2.0, **

Requirements

• python >= 3.6
• setuptools==41.0.1
• ftfy==5.6
• tqdm==4.32.1
• Click==7.0
• sentencepiece==0.1.83
• tensorflow-gpu==2.3.0
• numpy==1.16.4

Setup

$ git clone https://github.com/akanyaani/gpt-2-tensorflow2.0
$ cd gpt-2-tensorflow2.0
$ pip install -r requirements.txt

You can pre-train the model using sample data available in repository or you can download the data using this github repo https://github.com/eukaryote31/openwebtext

Pre-Training model on sample data available in repository

$ python pre_process.py --help

Options:
  --data-dir TEXT        training data path  [default: /data/scraped]
  --vocab-size INTEGER   byte pair vocab size  [default: 24512]
  --min-seq-len INTEGER  minimum sequence length  [default: 15]
  --max-seq-len INTEGER  maximum sequence length  [default: 512]
  --help                 Show this message and exit.
  
  
>> python pre_process.py

Pre-Training model on openwebtext or any other data

>> python pre_process.py --data-dir=data_directory --vocab-size=32000

$ python train_gpt2.py --help

Options:
  --num-layers INTEGER      No. of decoder layers  [default: 8]
  --embedding-size INTEGER  Embedding size  [default: 768]
  --num-heads INTEGER       Number of heads  [default: 8]
  --dff INTEGER             Filter Size  [default: 3072]
  --max-seq-len INTEGER     Seq length  [default: 515]
  --vocab-size INTEGER      Vocab size  [default: 24512]
  --optimizer TEXT          optimizer type  [default: adam]
  --batch-size INTEGER      batch size  [default: 8]
  --learning-rate FLOAT     learning rate  [default: 0.001]
  --graph-mode BOOLEAN      TF run mode  [default: False]
  --distributed BOOLEAN     distributed training  [default: False]
  --help                    Show this message and exit.
  
  
>> python train_gpt2.py \
  --num-layers=8 \
  --num-heads=8 \
  --dff=3072 \
  --embedding-size=768 \
  --batch-size=32 \
  --learning-rate=5e-5
  --graph-mode=True

Distributed training on multiple gpu.

>> python train_gpt2.py \
  --num-layers=8 \
  --num-heads=8 \
  --dff=3072 \
  --embedding-size=768 \
  --batch-size=32 \
  --learning-rate=5e-5 \
  --distributed=True \
  --graph-mode=True

Start TensorBoard through the command line.

$ tensorboard --logdir /log

After pretraining your model, you can generate sequences by giving some context to model. Open this notebook and load the pretrained model and pass context to model it will return the generated sequence.

$ sequence_generator.ipynb

TO DO

1. Parallel Preprocessing.
2. Shared weights across layers.
3. Factorized embedding.
4. Fine-Tuning wrapper.

References:

• “Openai/gpt-2”
• “Huggingface pytorch-transformers”
• “Tensorflow Transformers”
• “The Illustrated GPT-2 “

Contribution

• Your issues and PRs are always welcome.

Author

• Abhay Kumar
• Author Email : akanyaani@gmail.com
• Follow me on Twitter

License

• MIT

Computation Graph of GPT-2 Model.