Fine tuning on custom datasets

Reference: "Beginner’s Guide to Retrain GPT-2 (117M) to Generate Custom Text Content"

To retrain GPT-2 117M model on a custom text dataset:

PYTHONPATH=src ./train.py --dataset <file|directory|glob>

If you want to precompute the dataset's encoding for multiple runs, you can instead use:

PYTHONPATH=src ./encode.py <file|directory|glob> /path/to/encoded.npz
PYTHONPATH=src ./train.py --dataset /path/to/encoded.npz

Make sure cudnn is installed. Some have reported that train.py runs without it but has worse memory usage and might OOM.

Tensor Rematerialization

Experimental: a rematerialization rewriter based on Efficient Rematerialization for Deep Networks https://papers.nips.cc/paper/9653-efficient-rematerialization-for-deep-networks.pdf, which unlike gradient checkpointing works in tensorflow 2.0 and is able to automatically select checkpoints in arbitrary graphs. Using this I was able to finetune GPT-2 1.5B on a single graphics card using slightly less than 12G of video ram with very little slowdown.

To use this is a little involved, because the graph optimization algorithm is offloaded to an optimized Haskell program. First, go into subdirectory twremat, and build it by invoking:

cabal v2-install --installdir=../bin

(You'll need to install cabal if you haven't already -- but setting up ghc and haskell compilation is beyond the scope of this README.)

Then run train.py as normal, enabling --twremat and setting --twremat_memlimit to an appropriate value -- this sets the amount of memory assumed to be available for computation of gradients, so it should be roughly the memory size of your graphics card minus whatever is taken up by the gpt-2 weights, and any other bookkeeping variables. You may need to experiment with the memlimit until you find the largest value that doesn't OOM.

(You probably also want to use SGD as optimizer instead of Adam to minimize those bookkeeping variables, of which Adam uses a lot).

Gradient Checkpointing

https://github.com/openai/gradient-checkpointing is included to reduce the memory requirements of the model, and can be enabled by --memory_saving_gradients. The checkpoints are currently chosen manually (poorly) by just adding layer 10 to the 'checkpoints' collection in model.py.

Gradient checkpointing doesn't work in tensorflow v2.0 and later due to the removal of tf.contrib. You should use tensor rematerialization instead if possible.

Validation loss

Set --val_every to a number of steps N > 0, and "validation" loss against a fixed sample of the dataset will be calculated every N steps to get a better sense of training progress. N around 200 suggested. You can set --val_dataset to choose a separate validation dataset, otherwise it defaults to a sample from the train dataset (so not a real cross-validation loss!).

Optimizer

You can use SGD instead of Adam with --optimizer sgd. This also helps conserve memory when training larger models. Note: the learning rate needs to be adjusted for SGD, due to not having Adam's gradient normalization (0.0006 seems to be a good number from some experiments).

Original README

Status: Archive (code is provided as-is, no updates expected)

gpt-2

Code and models from the paper "Language Models are Unsupervised Multitask Learners".

You can read about GPT-2 and its staged release in our original blog post, 6 month follow-up post, and final post.

We have also released a dataset for researchers to study their behaviors.

^* Note that our original parameter counts were wrong due to an error (in our previous blog posts and paper). Thus you may have seen small referred to as 117M and medium referred to as 345M.

Usage

This repository is meant to be a starting point for researchers and engineers to experiment with GPT-2.

For basic information, see our model card.

Some caveats

• GPT-2 models' robustness and worst case behaviors are not well-understood. As with any machine-learned model, carefully evaluate GPT-2 for your use case, especially if used without fine-tuning or in safety-critical applications where reliability is important.
• The dataset our GPT-2 models were trained on contains many texts with biases and factual inaccuracies, and thus GPT-2 models are likely to be biased and inaccurate as well.
• To avoid having samples mistaken as human-written, we recommend clearly labeling samples as synthetic before wide dissemination. Our models are often incoherent or inaccurate in subtle ways, which takes more than a quick read for a human to notice.

Work with us

Please let us know if you’re doing interesting research with or working on applications of GPT-2! We’re especially interested in hearing from and potentially working with those who are studying

• Potential malicious use cases and defenses against them (e.g. the detectability of synthetic text)
• The extent of problematic content (e.g. bias) being baked into the models and effective mitigations

Development

See DEVELOPERS.md

Contributors

See CONTRIBUTORS.md

Citation

Please use the following bibtex entry:

@article{radford2019language,
  title={Language Models are Unsupervised Multitask Learners},
  author={Radford, Alec and Wu, Jeff and Child, Rewon and Luan, David and Amodei, Dario and Sutskever, Ilya},
  year={2019}
}

Future work

We may release code for evaluating the models on various benchmarks.

We are still considering release of the larger models.

License

Modified MIT