tensorio-bundler

tensorio-bundler

Create TensorIO model bundles

Running the bundler from the command line

NOTE: Working on making a PyPI package. Once that is done, these instructions will change to use whatever binary the corresponding pip install produces.

Requirements

• Python 3

Instructions

The tensorio_bundler module comes with a bundler utility that you can use to create TensorIO zipped tiobundle files directly from your command line.

For more information on how to run the bundler, run:

python -m tensorio_bundler.bundler -h

A sample invocation (using test data, assumed to be run from project root -- same directory as this README):

python -m tensorio_bundler.bundler \
    --tflite-model ./tensorio_bundler/fixtures/test.tflite \
    --model-json ./tensorio_bundler/fixtures/test.tiobundle/model.json \
    --assets-dir ./tensorio_bundler/fixtures/test.tiobundle/assets \
    --bundle-name sample.tiobundle \
    --outfile sample.tiobundle.zip

Calling the bundler locally through the REST API

To run the REST API locally from project root (same directory as this README):

gunicorn tensorio_bundler.rest:api

In a separate terminal window, you can invoke the bundler as follows:

TFLITE_PATH="\"$(mktemp -d)/model.tflite\""

read -r -d '' REQUEST_BODY <<-EOF
    {
        "saved_model_dir": "./tensorio_bundler/fixtures/test-model",
        "build": true,
        "tflite_path": $TFLITE_PATH,
        "model_json_path": "./tensorio_bundler/fixtures/test.tiobundle/model.json",
        "assets_path": "./tensorio_bundler/fixtures/test.tiobundle/assets",
        "bundle_name": "curl-test.tiobundle",
        "bundle_output_path": "curl-test.tiobundle.zip"
    }
EOF

curl -v -X POST \
    -H "Content-Type: application/json" \
    -d "$REQUEST_BODY" \
    http://localhost:8000/bundle

Running the bundler via docker

Requirements

• Docker

If you don't have it, get it

Instructions

You can either bind mount the paths to the inputs into your docker container when you run the bundler or you can bind mount in a service account credentials file and set the GOOGLE_APPLICATION_CREDENTIALS environment variable to point at the mount path in the container.

NOTE: These instructions are extremely sparse at the moment. They will not be so forever.

TensorIO Models repositories

The TensorIO bundler is now integrated with tensorio-models via the Repository REST API. Once a bundle has been built, you can use the tensorio_bundler.bundler.register_bundle method to register it against a TensorIO Models repository. The tensorio_bundler.bundler CLI allows you to do this automatically through the --repository-path argument.

This requires two environment variables to be set in your environment:

1. REPOSITORY -- a URL for a TensorIO models repository API URL (e.g. https://tio-models-test.dev.docai.beer/rest/v1/repository)

2. REPISITORY_API_KEY -- a basic auth token used to authenticate requests against the repository REST API.

Running tests if you want to contribute to this project

Requirements

• Docker

If you don't have it, get it

Instructions

Simply run:

./test.sh