README.md

Goal of this library

Tensorflow's official client implementations depend on the tensorflow package as of this writing. For instance, if you want to run their example MNIST client, you'd need to have tensorflow installed even just to construct the prediction request predict_pb2.PredictRequest, and to call their tf.contrib.util.make_tensor_proto function. So you can't easily use their work to make a lightweight client API that is totally free of tensorflow. Tensorflow being a huge library, the estimated RAM overhead of having tensorflow installed is around 150MB which makes the client heavy. Another goal of this project is to expose other models then tensorflow with a GRPC API following the one defined for tensorflow serving (https://github.com/tensorflow/serving). For instance, it allows to serve a XGBoost model with the same API as the one used to request a tensorflow serving server

How to update this lib with a new version of tensorflow_serving

We need to generate the GRPC client code ourselves from the .proto definition in https://github.com/tensorflow/serving/tree/master/tensorflow_serving/apis

To do this,

run "sh generate_pbs.sh X.Y.Z" (you should have an already created pew virtualenv using python3.11 called 'grpc-build'), this will generate a working python package in ./tensorflow_proto from the tag version X.Y.Z of tensorflow repo you can then copy/paste this package inside src/pythie-serving/ and import it as you need.

This will also generate the necessary stubs for type-checking

For Mac users: replace sed command in generate_pbs.sh by GNU sed (in order to prevent the creation of back files): brew install gnu-sed and replace sed by gsed in generate_pbs.sh.

Available model architectures

The following models can be served by pythie-serving:

• LightGBM: https://lightgbm.readthedocs.io/en/latest/
• XGBoost: https://xgboost.readthedocs.io/en/stable/
• scikit-learn: https://scikit-learn.org/stable/
• treelite: https://treelite.readthedocs.io/en/latest/
• csv table

How to run

Python

1. Define a models.config file (see: https://www.tensorflow.org/tfx/serving/serving_config):

    model_config_list {
        config {
            name: <model_name>,
            base_path: <my_base_path>,
            model_platform: <model_platform>
        }
    }
   

2. Run python pythie-serving/src/pythie_serving/run.py <model_config_file_path>.
   The following options are available:
   • --worker-count: Number of concurrent threads for the GRPC server.
   • --max-concurrent-rpcs: The maximum number of concurrent RPCs this server.
   • --port: Port number to listen to.

Environment variables

• OPENTELEMETRY_COLLECTOR_HOST: OpenTelemetry Collector receiver endpoint. If not defined OpenTelemetry will not be activated. See https://opentelemetry.io/docs/what-is-opentelemetry for more details. For a treelite served model:
• TREELITE_NTHREAD: Number of threads to use to compute predictions
• TREELINTE_BIND_THREADS: Set to 0 to deactivate thread pinning. See https://treelite.readthedocs.io/en/latest/treelite-runtime-api.html For an LGBM served model:
• LGBM_NTHREAD: Number of threads to use to compute predictions
• GRPC_SERVER_TIMEOUT: Timeout at which server will stop and the pod will shut down in seconds.

Docker

The project is published on GitHub Container Registry: https://github.com/wiremind/pythie-serving/pkgs/container/pythie-serving

Development

Add a new architecture type

To add a new architecture type, you need to implement a prediction_service_pb2_grpc.PredictionServiceServicer. To facilitate this, pythie-serving implements an abstract AbstractPythieServingPredictionServiceServicer that already implements the necessary Predict method, taking a PredictRequest as input and outputing a PredictResponse. This class leaves 2 abstract methods to be implemented:

• _create_model_specs to read a ModelConfig from the list and instantiate the model and necessary variables
• _predict to make a prediction using the model (as python class) on a numpy array