rajaiswal / tendies

TensorFlow Distributed Image Serving - A lightweight, RESTful remote inference library for decoupling deep learning development and deployment. Example usage cases with CycleGAN and Faster R-CNN. Includes tutorial ipynbs.

Tendies: TensorFlow Distributed Image Serving

A lightweight, RESTful remote inference library for decoupling deep learning development and deployment. Includes CycleGAN and the TensorFlow Object Detection Faster R-CNN as usage examples. Now supports Keras models.

Motivation

A major challenge in combining deep learning and software engineering as we move towards Software 2.0 is that deep learning development and deployment are tightly coupled. This restricts deployment of models in several ways, for example:

1. A model developed using Google Cloud Platform (GCP) or Amazon Web Services (AWS) cannot be deployed on a classified network, or trained using classified data.
2. A model with a deep, complex architecture cannot be utilized for local inference on low-power Internet of Things (IoT) devices.
3. A model with atypical input/output requirements, such as the TensorFlow Object Detection API, cannot be easily integrated with larger, pre-existing software systems.

Tendies solves this problem by acting as a decoupling between these two domains. In the given scenarios:

1. Tendies provides an encapsulated TensorFlow-Serving environment that does not require the Cloud for remote inference.
2. Tendies decouples the IoT client and the GPU backend, so inference can be run on a powerful machine, then transmitted to the device.
3. Tendies enforces input/output standardization through a RESTful API, encapsulating the model server into an easily-integrable black box architecture.

In a sentence, Tendies allows any image-based deep learning model to be embedded in any production system.

Basic usage steps for a custom localhost server:

1. Clone the repository to your machine.
2. Follow the instructions in full_functionality/tendies-extension-tutorial.ipynb to export your model.
3. Build your server with python ServerBuilder.py --checkpoint_dir $(path)
4. Run the server from bash with tensorflow_model_server --rest_api_port=8501 --model_name=saved_model --model_base_path=$(path)
   • You may have to pip install tensorflow-serving-api-python3
5. Select or write a client, then do inference on a directory of images with python client.py --input_dir $(path)

Please see minimum_working_example/tendies-basic-tutorial.ipynb for a short walkthrough of how this library works.

Please also view full_functionality/tendies-extension-tutorial.ipynb for instructions on configuring Tendies with your specific use case.