bbux-dev / transformers-spec-demo

Dataspec Demo using Huggingface Transformers Library

Data Spec Transformers Demo

1. Overview
2. Build
3. Example
4. Download Model
5. CSV Example

Overview

This is a more detailed demo for the datacraft library. This make use of the huggingface Masked Language Modeling example to generate tokens from surrounding context.

Build

To install the datacraft library and the huggingface dependencies:

pip install datacraft transformers torch

The datacraft executable should now be on your path

Example

This example takes advantage of the datacraft custom code loading capability. This provides an extension point into the datacraft library to build and define custom types and handlers for those types. The custom_code.py module defines a hf-fill-mask type handler. This example code will load a huggingface fill-mask transformer pipeline. This pipeline is used to generate tokens given the surrounding context. The handler uses the __MASK__ token to denote where the token should be placed. For example if we have the sentence:

I seem to have lost my number. Can I have yours?

We can put the __MASK__ marker in various places to see what possible substitutions the huggingfaces model will provide:

I seem to have lost my __MASK__. Can I have yours?

To run the demo:

$ datacraft --spec demo.json -i 20 --code custom_code.py --log-level error
Go ahead, feel my shirt. It's made of recycled material!
If you were a Transformer you'd be thrilled!
Do you believe in love at first? Or should I walk past you again?
I'm learning about important dates in 2017. Wanna be one of them?
I seem to have lost my mind. Can I have yours?
...
Do you have a name? Or can I call you Bob?

Download Model

By default, every time you run the demo it will download the model from https://huggingface.co. To keep from doing this over and over, use the download_model.py script.

python download_model.py /path/to/model/dir 2>&1 | grep INFO
#INFO: Loading fill-mask pipeline...
#INFO: Saving fill-mask to /path/to/model/dir

# now specify the downloaded dir as the datadir
datacraft -s demo.json -i 20 -c custom_code.py -l debug -d /path/to/model/dir

CSV Example

Many times it is easier to externalize larger values lists into a csv file. This can be referenced in a Data Spec using the csv Field Spec type. The demo-csv.json spec does this with the lines.csv file. To use this example:

# copy csv file to same location as downloaded model:
cp lines.csv /path/to/model/dir
datacraft -s demo-csv.json -i 20 -c custom_code.py -l debug --datadir /path/to/model/dir