Easy-to-use Wrapper for GPT-2 117M, 345M, 774M, and 1.5B Transformer Models
Made by Rishabh Anand • https://rish-16.github.io
What is it
GPT-2 is a Natural Language Processing model developed by OpenAI for text generation. It is the successor to the GPT (Generative Pre-trained Transformer) model trained on 40GB of text from the internet. It features a Transformer model that was brought to light by the Attention Is All You Need paper in 2017. The model has 4 versions -
1558M - that differ in terms of the amount of training data fed to it and the number of parameters they contain.
The 1.5B model is currently the largest available model released by OpenAI.
gpt2-client is a wrapper around the original
gpt-2 repository that features the same functionality but with more accessiblity, comprehensibility, and utilty. You can play around with all four GPT-2 models in less than five lines of code.
Note: This client wrapper is in no way liable to any damage caused directly or indirectly. Any names, places, and objects referenced by the model are fictional and seek no resemblance to real life entities or organisations. Samples are unfiltered and may contain offensive content. User discretion advised.
Install client via
gpt2-client is well supported for Python >= 3.5 and TensorFlow >= 1.X. Some libraries may need to be reinstalled or upgraded using the
--upgrade flag via
pip if Python 2.X is used.
pip install gpt2-client
gpt2-clientis not compatible with TensorFlow 2.0
1. Download the model weights and checkpoints
from gpt2_client import GPT2Client gpt2 = GPT2Client('117M') # This could also be `345M`, `774M`, or `1558M`. Rename `save_dir` to anything. gpt2.load_model(force_download=False) # Use cached versions if available.
This creates a directory called
models in the current working directory and downloads the weights, checkpoints, model JSON, and hyper-parameters required by the model. Once you have called the
load_model() function, you need not call it again assuming that the files have finished downloading in the
force_download=Trueto overwrite the existing cached model weights and checkpoints
2. Start generating text!
from gpt2_client import GPT2Client gpt2 = GPT2Client('117M') # This could also be `345M`, `774M`, or `1558M` gpt2.load_model() gpt2.generate(interactive=True) # Asks user for prompt gpt2.generate(n_samples=4) # Generates 4 pieces of text text = gpt2.generate(return_text=True) # Generates text and returns it in an array gpt2.generate(interactive=True, n_samples=3) # A different prompt each time
You can see from the aforementioned sample that the generation options are highly flexible. You can mix and match based on what kind of text you need generated, be it multiple chunks or one at a time with prompts.
3. Generating text from batch of prompts
from gpt2_client import GPT2Client gpt2 = GPT2Client('117M') # This could also be `345M`, `774M`, or `1558M` gpt2.load_model() prompts = [ "This is a prompt 1", "This is a prompt 2", "This is a prompt 3", "This is a prompt 4" ] text = gpt2.generate_batch_from_prompts(prompts) # returns an array of generated text
4. Fine-tuning GPT-2 to custom datasets
from gpt2_client import GPT2Client gpt2 = GPT2Client('117M') # This could also be `345M`, `774M`, or `1558M` gpt2.load_model() my_corpus = './data/shakespeare.txt' # path to corpus custom_text = gpt2.finetune(my_corpus, return_text=True) # Load your custom dataset
In order to fine-tune GPT-2 to your custom corpus or dataset, it's ideal to have a GPU or TPU at hand. Google Colab is one such tool you can make use of to re-train/fine-tune your custom model.
5. Encoding and decoding text sequences
from gpt2_client import GPT2Client gpt2 = GPT2Client('117M') # This could also be `345M`, `774M`, or `1558M` gpt2.load_model() # encoding a sentence encs = gpt2.encode_seq("Hello world, this is a sentence") # [15496, 995, 11, 428, 318, 257, 6827] # decoding an encoded sequence decs = gpt2.decode_seq(encs) # Hello world, this is a sentence
Suggestions, improvements, and enhancements are always welcome! If you have any issues, please do raise one in the Issues section. If you have an improvement, do file an issue to discuss the suggestion before creating a PR.
All ideas – no matter how outrageous – welcome!
Open-source is really fun. Your donations motivate me to bring fresh ideas to life. If interested in supporting my open-source endeavours, please do donate – it means a lot to me!