In this article, we will learn how to expose ML/DL model as flask APIs.
Frameworks we’ll be using:
Keras is a Deep Learning library, built on top of backends such as Tensorflow, Theano or CNTK. It provides abstraction and allows rapid development of ML/DL models.
Flask is a micro web framework in Python, which is used to quickly spin up servers to serve pages. Refer Introduction to Flask.
Since the focus of the article is to serve a ML/DL model using an API, we wouldn’t get deeper into the making of the Convolutional model. We will be using the ResNet50 Convolutional Neural Network.
Install Tensorflow and Keras
| Step | Linux/Mac | Windows |
|---|---|---|
| Update pip | pip install -U pip | python -m pip install -U pip |
| Install virtualenv | sudo pip install virtualenv | pip install virtualenv |
| Create a new folder ml-api | mkdir ml-api && cd ml-api | mkdir ml-api && cd ml-api | Create a virtual environment(Good practice) and activate it | virtualenv –system-site-packages -p python3 ./venv | virtualenv –system-site-packages -p python3 ./venv | Activate the virtual environment | source ./venv/bin/activate | .\venv\Scripts\activate | Install TensorFlow | pip install –upgrade tensorflow | pip install –upgrade tensorflow | Install Keras | pip install keras | pip install keras | Install other dependencies | pip install flask gevent requests pillow | pip install flask gevent requests pillow |
Examples:
Creating the REST API :
# keras_server.py # Python program to expose a ML model as flask REST API # import the necessary modules from keras.applications import ResNet50 # pre-built CNN Model from keras.preprocessing.image import img_to_array from keras.applications import imagenet_utils import tensorflow as tf from PIL import Image import numpy as np import flask import io # Create Flask application and initialize Keras model app = flask.Flask(__name__) model = None # Function to Load the model def load_model(): # global variables, to be used in another function global model model = ResNet50(weights ="imagenet") global graph graph = tf.get_default_graph() # Every ML/DL model has a specific format # of taking input. Before we can predict on # the input image, we first need to preprocess it. def prepare_image(image, target): if image.mode != "RGB": image = image.convert("RGB") # Resize the image to the target dimensions image = image.resize(target) # PIL Image to Numpy array image = img_to_array(image) # Expand the shape of an array, # as required by the Model image = np.expand_dims(image, axis = 0) # preprocess_input function is meant to # adequate your image to the format the model requires image = imagenet_utils.preprocess_input(image) # return the processed image return image # Now, we can predict the results. @app.route("/predict", methods =["POST"]) def predict(): data = {} # dictionary to store result data["success"] = False # Check if image was properly sent to our endpoint if flask.request.method == "POST": if flask.request.files.get("image"): image = flask.request.files["image"].read() image = Image.open(io.BytesIO(image)) # Resize it to 224x224 pixels # (required input dimensions for ResNet) image = prepare_image(image, target =(224, 224)) # Predict ! global preds, results with graph.as_default(): preds = model.predict(image) results = imagenet_utils.decode_predictions(preds) data["predictions"] = [] for (ID, label, probability) in results[0]: r = {"label": label, "probability": float(probability)} data["predictions"].append(r) data["success"] = True # return JSON response return flask.jsonify(data) if __name__ == "__main__": print(("* Loading Keras model and Flask starting server..." "please wait until server has fully started")) load_model() app.run() |
Run the flask server
python keras_server.py
Note #1: For the first time you run it, it takes time to download the weights of the model. In subsequent runs, this won’t be the case.
Note #2: Save images in the same directory (ml-api) – dog.jpg, cat.jpg, lion.jpg .
Make requests to the server:
Method #1 : use cURL (Download cURL for Windows from here.)
$ curl -X POST -F image=@dog.jpg "http://localhost:5000/predict"
Method #2 : (Suitable if making a Full-stack web app with proper UI)
Create a simple HTML form
<!-- index.html --> enctype = "multipart/form-data"> <input type = "file" name = "image" /> <input type = "submit"/> </form> |
Method #3 : Create a simple python script to make HTTP requests to flask server.
# simple_request.py import requests import sys # provide image name as command line argument IMAGE_PATH = sys.argv[1] image = open(IMAGE_PATH, "rb").read() payload = {"image": image} # make request to the API request = requests.post(URL, files = payload).json() if request["success"]: # Print formatted Result print("% s % 15s % s"%("Rank", "Label", "Probability")) for (i, result) in enumerate(request["predictions"]): print("% d. % 17s %.4f"%(i + 1, result["label"], result["probability"])) else: print("Request failed") |
Run it as :
python simple_request.py dog.jpg
Output:
To deactivate after you are done(both Windows and Linux) :
> deactivate
So the workflow, in general for any model is
- Build the model and save it.
- Create a Flask app.
- Load the Model
- Define endpoints of API for each type of requests.
- Preprocess input according to the Model’s architecture and feed it to the model
- Process the input and send back the output to client.
