Tensorflow Object Detection API - Inference Web Service

It is great to create REST API to access from anywhere to use your machine learning model. Flask framework makes it easy to create a web service with python. All the code is include in object_detection_app_test1.py.


Request Example
Type: POST
Address: http://xxxxxx:5000/inference_app
KEY: file
VALUE: IMG_0000.jpg

Set up configuration
In this example, you open 5000 port so you can access the web service, and THRESHOLD allows only inference results higher than 0.6 to be responded. An image sent through POST request is going to be saved at UPLOAD_DIR. Define flask, and write UPLOAD_DIR to app.config so flask app can obtain later when it receives http request.

``` THRESHOLD = 0.6 PORT = 5000 UPLOAD_DIR = 'object_detection/test1/flask_pics/' LABEL_PATH = 'object_detection/test1/pascal_label_map.pbtxt' MODEL_PATH = 'object_detection/test1/output/frozen_inference_graph.pb' app = Flask(__name__) app.config['UPLOAD_DIR'] = UPLOAD_DIR ```

Write model in app.config and run flask app
You load model from frozen_inference_graph.pb and obtain category index from pascal_label_map.pbtxt in the main function. It also writes the model of 'session', 'image_tensor', 'tensor_dict', and 'category_index' in app.config['MODEL'] so the flask app can obtain the model when it receives http requests.

``` if __name__ == '__main__': app.config['MODEL'] = load_model(MODEL_PATH, LABEL_PATH) app.run(host='0.0.0.0', port=PORT, debug=False) ```

Upload an image file and load model
When the flask app receives a http request, it loads the model from app.config['MODEL'] and upload the image file to a directory in the web server.

``` model = app.config['MODEL'] file = request.files['file'] filename = secure_filename(file.filename) filepath = os.path.join(app.config['UPLOAD_DIR'], filename) file.save(filepath) ```

Run the inference

``` response = Response(response=inference(model, filepath), status=200, mimetype='application/json') ```

Delete the uploaded image file

``` os.remove(filepath) ```