Ecosyste.ms: Awesome
An open API service indexing awesome lists of open source software.
https://github.com/rish-16/gpt2client
✍🏻 gpt2-client: Easy-to-use TensorFlow Wrapper for GPT-2 117M, 345M, 774M, and 1.5B Transformer Models 🤖 📝
https://github.com/rish-16/gpt2client
gpt-2 machine-learning openai tensorflow text-generation transformer
Last synced: 18 days ago
JSON representation
✍🏻 gpt2-client: Easy-to-use TensorFlow Wrapper for GPT-2 117M, 345M, 774M, and 1.5B Transformer Models 🤖 📝
- Host: GitHub
- URL: https://github.com/rish-16/gpt2client
- Owner: rish-16
- License: mit
- Archived: true
- Created: 2019-08-04T06:35:45.000Z (over 5 years ago)
- Default Branch: master
- Last Pushed: 2021-06-24T04:59:19.000Z (over 3 years ago)
- Last Synced: 2024-08-11T09:54:47.221Z (3 months ago)
- Topics: gpt-2, machine-learning, openai, tensorflow, text-generation, transformer
- Language: Python
- Homepage:
- Size: 3.2 MB
- Stars: 372
- Watchers: 16
- Forks: 74
- Open Issues: 7
-
Metadata Files:
- Readme: README.md
- License: LICENSE.txt
Awesome Lists containing this project
README
gpt2-client (Archived)
Easy-to-use Wrapper for GPT-2 124M, 345M, 774M, and 1.5B Transformer Models
What is it •
Installation •
Getting Started
Made by Rishabh Anand • https://rish-16.github.io
What is it
GPT-2 is a Natural Language Processing model [developed by OpenAI](https://openai.com/blog/better-language-models/) 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](https://arxiv.org/abs/1706.03762) paper in 2017. The model has 4 versions - `124M`, `345M`, `774M`, and `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.
Finally, `gpt2-client` is a wrapper around the original [`gpt-2` repository](https://github.com/openai/gpt-2) 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.\*
Installation
Install client via `pip`. Ideally, `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.
```bash
pip install gpt2-client
```
> **_Note:_** `gpt2-client` is **not** compatible with TensorFlow 2.0 , try TensorFlow 1.14.0
Getting started
**1. Download the model weights and checkpoints**
```python
from gpt2_client import GPT2Client
gpt2 = GPT2Client('124M') # This could also be `355M`, `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 `models` directory.
> **_Note:_** Set `force_download=True` to overwrite the existing cached model weights and checkpoints
**2. Start generating text!**
```python
from gpt2_client import GPT2Client
gpt2 = GPT2Client('124M') # This could also be `355M`, `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**
```python
from gpt2_client import GPT2Client
gpt2 = GPT2Client('124M') # This could also be `355M`, `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**
```python
from gpt2_client import GPT2Client
gpt2 = GPT2Client('124M') # This could also be `355M`, `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](http://colab.research.google.com) is one such tool you can make use of to re-train/fine-tune your custom model.
**5. Encoding and decoding text sequences**
```python
from gpt2_client import GPT2Client
gpt2 = GPT2Client('124M') # This could also be `355M`, `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
```
Contributing
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!
Donations
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!
Licence
[MIT](https://github.com/rish-16/gpt2client/blob/master/LICENSE.txt)
