Skip to content

Latest commit

 

History

History
102 lines (69 loc) · 2.5 KB

README.md

File metadata and controls

102 lines (69 loc) · 2.5 KB
Raw

TRANSFORMABLE

simple command line utility for text with HuggingFace transformer pipelines (supports CPU inference)

FEATURES

  • works with or without a GPU for almost any transformer model
  • by default runs completely offline, on your local machine
  • automate an unlimited sequence of generation tasks for set-and-forget use
  • simple TOML configuration file for storing parameters and prompts
  • can be configured to auto-download models from HuggingFace
  • simple and fun command line interface

USAGE

clone this repository.

install python requirements:

pip install -r requirements.txt

install git-lfs for large file storage support.

clone huggingface repositories into ~/src/huggingface.co/:

git clone --recurse-submodules \
    https://huggingface.co/distilbert-base-uncased-distilled-squad/ \
    ~/src/huggingface.co/distilbert-base-uncased-distilled-squad/

generate a few responses using the default model (distilbert-base-uncased-distilled-squad):

python ./transformable.py -c ''

enable automatic fetching of a model from huggingface and use a manual seed:

python ./transformable.py \
    --download_models \
    --model=xlnet-large-cased \
    --seed=31911 \
    'The meaning of life is'

automatic model downloads will be stored in ~/.cache/huggingface/

to dump the configuration for inspection, append the --dump flag to any command.

to repeat tasks (useful with random seed), use the --repeat flag.

for more detailed usage, see help:

python ./transformable.py --help

CONFIGURATION

transformable tasks can be configured in TOML format.

by default, tasks will be read from ./transformable.toml if it exists.

the config file uses the same keys as the flag names above.

see transformable.example.toml for an example.

to execute a task from the config:

$ python ./transformable.py -c ./transformable.example.toml -t introductions

by default, the TOML section is used as the output name.

you can override config options with flags:

python ./transformable.py \
    -c ./transformable.example.toml \
    -t home-planet \
    -x 'Most humans are currently living on Mars.'

it is also possible to run multiple tasks in sequence:

python ./transformable.py \
    -c ./transformable.example.toml \
    -t introductions \
    -t meaning-of-life

to run all tasks from the config in sequence:

python ./transformable.py -c ./transformable.example.toml -a