gpt-2

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

For now, we have only released a smaller (117M parameter) version of GPT-2.

See more details in our blog post.

Installation

Download the model data

sh download_model.sh 117M

Install python packages:

pip3 install -r requirements.txt

Unconditional sample generation

WARNING: Samples are unfiltered and may contain offensive content.

To generate unconditional samples from the small model:

python3 src/generate_unconditional_samples.py | tee samples

There are various flags for controlling the samples:

python3 src/generate_unconditional_samples.py --top_k 40 --temperature 0.7 | tee samples

While we have not yet released GPT-2 itself, you can see some unconditional samples from it:

• gpt2-samples.txt (with default settings of temperature 1 and no truncation)
• gpt2-topk40-samples.txt (with temperature 1 and top_k 40 truncation)

Conditional sample generation (Interactive)

To give the model custom prompts, you can use:

python3 src/interactive_conditional_samples.py --top_k 40

Conditional sample generation (Non-interactive)

To give the model custom prompts without confirmation, you can use:

python3 src/generaete_conditional_samples.py "INPUT TEXT" --top_k 40

Docker Usage

To run gpt-2 from a docker container:

# build container and tag as gpt-2
docker build --tag gpt-2 .
# start container in the background and add volume to sync local changes
docker run --runtime=nvidia -d -v $PWD:/gpt-2 --name gpt-2 gpt-2
# open an interactive session
docker exec -it gpt-2 bash

To build gpt-2 on a docker image (default : CPU) :

# build container and tag as xift.ai:gpt-2
./setup_docker.sh

To check all the parameters:

./run_docker.sh 

./run_docker.sh "INPUT TEXT" [PARAMETERS]

Parameters :
--seed (None) : a random value is generated unless specified. give a specific integer value if you want to reproduce same results in the future.
--nsamples (1) : specify the number of samples you want to print
--length (None) : number of tokens (words) to print on each sample.
--batch_size (1) : how many inputs you want to process simultaneously. doesn't seem to affect the results.
--temperature (1) : scales logits before sampling prior to softmax.
--top_k (0) : truncates the set of logits considered to those with the highest values.

Example : ./run_docker.sh "In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains." --nsamples 2 --top_k 40 --temperature .80

To run the container for condicional sample generation on some input text:

# run the container with some text
./run_docker.sh "In a shocking finding, scientist discovered a herd of unicorns living in a remote, previously unexplored valley, in the Andes Mountains." --nsamples 2 --top_k 40 --temperature .80

Future work

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

We are still considering release of the larger models.

License

Coming soon!