Introduction

This ModelServer repository contains a simple Flask app for hosting multiple TensorFlow Object Detection models and using them to evaluate incoming images.

Setting Up

Please clone the repository, and install its dependencies:

NVidia Cuda 9.0.x
NVidia CuDNN 7.0.x
Python 3.5.x or later (I use Anaconda)

Once you have these installed, either install the Python dependencies directly or create a virtual environment to install them into. To create a virtual environment with Anaconda and install the dependencies into it:

conda create -n model-server python=3.6
activate model-server # or if on Linux, source activate model-server
pip install -r requirements.txt

Configuring the Server

The server uses environment variables for tuning many of its parameters, as specified below:

UPLOAD_FOLDER: Folder in which to (temporarily) store images to be evaluated. Images are deleted after evaluation. Defaults to ./pics.
DEBUG: True|False value controlling whether to dump all detections found for an image with their confidence scores to stdout. Defaults to False.
PORT: Port on which to start the server. Defaults to 5000.
MODEL_FOLDER: Folder from which to load the models. See Loading Models below. Defaults to ./models.
MIN_CONFIDENCE: Minimum score required for us to include a match in our results. Defaults to 0.8.

Loading Models

On startup, the server looks in MODEL_FOLDER for any files named <prefix>.frozen.pb. It uses those prefixes to load those frozen model files into their own TensorFlow sessions, and looks for category mapping files named <prefix>.label_map.pbtxt and loads those as well, all in a model map indexed by prefix.

When the server is shut down, an atexit hook will close() all open TensorFlow sessions.

The Server API

Start the server using

python app.py

It should start up on port PORT (default: 5000) and begin responding to requests on /detect. POSTing to that address with a multi-part form containing a modelname text field with the model prefix name to use and file data with the image to evaluate. It returns a JSON array containing all matches that meet or exceed the MIN_CONFIDENCE value. For example:

[{
    "class": 2,
    "label": "Seal1",
    "confidence": 0.93756123,
    "bounding_box": [0.2, 0.2, 0.8, 0.8]
}]

The bounding_box is in normalized ([0.0, 1.0]) coordinates in [ymin, xmin, ymax, xmax] format. If DEBUG is true, during evaluation we will dump all detections and their confidences to stdout. We also dump evaluation times to stdout regardless of the debug setting.