aitextgen is a Python package that leverages PyTorch, Huggingface Transformers and pytorch-lightning with specific optimizations for text generation using GPT-2, plus many added features. It is the successor to textgenrnn and gpt-2-simple, taking the best of both packages:
- Finetunes on a pretrained 124M GPT-2 model from OpenAI...or create your own GPT-2 model + tokenizer and train from scratch!
- Generates text faster than gpt-2-simple and with better memory efficiency! (even from the 1.5B GPT-2 model!)
- With Transformers, aitextgen preserves compatibility with the base package, allowing you to use the model for other NLP tasks, download custom GPT-2 models from the Huggingface model repository, and upload your own models! Also, it uses the included
generate()function to allow a massive amount of control over the generated text.
- With pytorch-lightning, aitextgen trains models not just on CPUs and GPUs, but also multiple GPUs and (eventually) TPUs! It also includes a pretty training progress bar, with the ability to add optional loggers.
- The input dataset is its own object, allowing you to not only easily encode megabytes of data in seconds, cache, and compress it on a local computer before transporting to a remote server, but you are able to merge datasets without biasing the resulting dataset, or cross-train on multiple datasets to create blended output.
You can read more about aitextgen in the documentation!
You can play with aitextgen for free with powerful GPUs using these Colaboratory Notebooks!
- Finetune OpenAI's 124M GPT-2 model on your own dataset (GPU)
- Train a GPT-2 model + tokenizer from scratch (GPU)
aitextgen can be installed from PyPI:
pip3 install aitextgen
Here's how you can quickly test out aitextgen on your own computer, even if you don't have a GPU!
For generating text from a pretrained GPT-2 model:
from aitextgen import aitextgen # Without any parameters, aitextgen() will download, cache, and load the 124M GPT-2 "small" model ai = aitextgen() ai.generate() ai.generate(n=3, max_length=100) ai.generate(n=3, prompt="I believe in unicorns because", max_length=100) ai.generate_to_file(n=10, prompt="I believe in unicorns because", max_length=100, temperature=1.2)
You can also generate from the command line:
aitextgen generate aitextgen generate --prompt "I believe in unicorns because" --to_file False
Want to train your own mini GPT-2 model on your own computer? Download this text file of Shakespeare plays, cd to that directory in a Teriminal, open up a
python3 console and go:
from aitextgen.TokenDataset import TokenDataset from aitextgen.tokenizers import train_tokenizer from aitextgen.utils import GPT2ConfigCPU from aitextgen import aitextgen # The name of the downloaded Shakespeare text for training file_name = "input.txt" # Train a custom BPE Tokenizer on the downloaded text # This will save two files: aitextgen-vocab.json and aitextgen-merges.txt, # which are needed to rebuild the tokenizer. train_tokenizer(file_name) vocab_file = "aitextgen-vocab.json" merges_file = "aitextgen-merges.txt" # GPT2ConfigCPU is a mini variant of GPT-2 optimized for CPU-training # e.g. the # of input tokens here is 64 vs. 1024 for base GPT-2. config = GPT2ConfigCPU() # Instantiate aitextgen using the created tokenizer and config ai = aitextgen(vocab_file=vocab_file, merges_file=merges_file, config=config) # You can build datasets for training by creating TokenDatasets, # which automatically processes the dataset with the appropriate size. data = TokenDataset(file_name, vocab_file=vocab_file, merges_file=merges_file, block_size=64) # Train the model! It will save pytorch_model.bin periodically and after completion. # On a 2016 MacBook Pro, this took ~25 minutes to run. ai.train(data, batch_size=16, num_steps=5000) # Generate text from it! ai.generate(10, prompt="ROMEO:")
Want to run aitextgen and finetune GPT-2? Use the Colab notebooks in the Demos section, or follow the documentation to get more information and learn some helpful tips!
- TPUs cannot be used to train a model: although you can train an aitextgen model on TPUs by setting
n_tpu_cores=8in an appropriate runtime, and the training loss indeed does decrease, there are a number of miscellaneous blocking problems. [Tracking GitHub Issue].
- TorchScript exporting, although it works with
ai.export(), behaves oddly when reloaded back into Python, and is therefore not supported (yet). [Tracking GitHub Issue]
- Finetuning the 355M GPT-2 model or larger on a GPU will cause the GPU to go OOM, even 16 GB VRAM GPUs (355M does work with FP16 + 16 GB VRAM however). This is a known issue with the Transformers GPT-2 implementation, unfortunately. Gradient checkpointing may need to be implemented within the training loop of aitextgen. [Tracking GitHub Issue]
The current release (v0.1.X) of aitextgen is considered to be a beta, targeting the most common use cases. The Notebooks and examples written so far are tested to work, but more fleshing out of the docs/use cases will be done over the next few months in addition to fixing the known issues noted above.
The next versions of aitextgen (and one of the reasons I made this package in the first place) will have native support for schema-based generation. (see this repo for a rough proof-of-concept)
Additionally, I plan to develop an aitextgen SaaS to allow anyone to run aitextgen in the cloud and build APIs/Twitter+Slack+Discord bots with just a few clicks. (the primary constraint is compute cost; if any venture capitalists are interested in funding the development of such a service, let me know)
I've listed more tenative features in the UPCOMING document.
aitextgen is a tool primarily intended to help facilitate creative content. It is not a tool intended to deceive. Although parody accounts are an obvious use case for this package, make sure you are as upfront as possible with the methodology of the text you create. This includes:
- State that the text was generated using aitextgen and/or a GPT-2 model architecture. (a link to this repo would be a bonus!)
- If parodying a person, explicitly state that it is a parody, and reference who it is parodying.
- If the generated text is human-curated, or if it's unsupervised random output
- Indicating who is maintaining/curating the AI-generated text.
- Make a good-faith effort to remove overfit output from the generated text that matches the input text verbatim.
It's fun to anthropomorphise the nameless "AI" as an abstract genius, but part of the reason I made aitextgen (and all my previous text-generation projects) is to make the technology more accessible and accurately demonstrate both its promise, and its limitations. Any AI text generation projects that are deliberately deceptive may be disavowed.
Max Woolf (@minimaxir)
Max's open-source projects are supported by his Patreon and GitHub Sponsors. If you found this project helpful, any monetary contributions to the Patreon are appreciated and will be put to good creative use.