A scalable Keras + deep learning REST API

In today’s blog post we are going to create a deep learning REST API that wraps a Keras model in an efficient, scalable manner.

Our Keras + deep learning REST API will be capable of batch processing images, scaling to multiple machines (including multiple web servers and Redis instances), and round-robin scheduling when placed behind a load balancer.

To accomplish this we will be using:

• Keras
• Redis (an in-memory data structure store)
• Flask (a micro web framework for Python)
• Message queuing and message broker programming paradigms

This blog post is a bit more advanced than other tutorials on PyImageSearch and is intended for readers:

• Who are familiar with the Keras deep learning library
• Who have an understanding of web frameworks and web services (and ideally coded a simple website/web service before)
• Who understand basic data structures, such as hash tables/dictionaries, lists, along with their associated asymptotic complexities

For a more simple Keras + deep learning REST API, please refer to this guest post I did on the official Keras.io blog.

To learn how to create your own scalable Keras + deep learning REST API, just keep reading!

A scalable Keras + deep learning REST API

2020-06-16 Update: This blog post is now TensorFlow 2+ compatible!

Today’s tutorial is broken into multiple parts.

We’ll start with a brief discussion of the Redis data store and how it can be used to facilitate message queuing and message brokering.

From there, we’ll configure our Python development environment by installing the required Python packages to build our Keras deep learning REST API.

Once we have our development environment configured we can implement our actual Keras deep learning REST API using the Flask web framework. After implementing, we’ll start the Redis and Flask servers, follow by submitting inference requests to our deep learning API endpoint using both cURL and Python.

Finally, we’ll end with a short discussion on the considerations you should keep in mind when building your own deep learning REST API.

A short introduction to Redis as a REST API message broker/message queue

Redis is an in-memory data store. It is different than a simple key/value store (such as memcached) as it can can store actual data structures.

To read more about Redis, I encourage you to review this short introduction.

Configuring and installing Redis for our Keras REST API

Redis is very easy to install. Below you’ll find the commands to download, extract, and install Redis on your system:

$ wget http://download.redis.io/redis-stable.tar.gz
$ tar xvzf redis-stable.tar.gz
$ cd redis-stable
$ make
$ sudo make install

To start the Redis server, use the following command:

$ redis-server

Leave this terminal open to keep the Redis data store running.

In another terminal, you can validate Redis is up and running:

$ redis-cli ping
PONG

Provided that you get a PONG back from Redis, you’re ready to go.

Configuring your Python development environment to build a Keras REST API

I recommend that you work on this project inside of a Python virtual environment so that it does not impact system level Python and projects.

To do this, you’ll need to install pip, virtualenv, and virtualenvwrapper (provided you haven’t already). Instructions on configuring these tools in your environment are described in:

• How to install TensorFlow 2.0 on Ubuntu
• How to install TensorFlow 2.0 on macOS

Please note that PyImageSearch does not recommend or support Windows for CV/DL projects.

You’ll also need the following packages installed into your virtual environment:

$ workon dl4cv
$ pip install flask
$ pip install gevent
$ pip install requests
$ pip install redis

That’s it!

Implementing a scalable Keras REST API

Let’s get started building our server script. For convenience I’ve implemented the server in a single file, however it can be modularized as you see fit.

For best results and to avoid copy/paste errors, I encourage you to use the “Downloads” section of this blog post to grab the associated scripts and images.

Let’s open up run_keras_server.py and walk through it together:

# import the necessary packages
from tensorflow.keras.applications import ResNet50
from tensorflow.keras.preprocessing.image import img_to_array
from tensorflow.keras.applications.resnet50 import preprocess_input
from tensorflow.keras.applications.resnet50 import decode_predictions
from threading import Thread
from PIL import Image
import numpy as np
import base64
import flask
import redis
import uuid
import time
import json
import sys
import io

There are quite a few imports listed above, notably ResNet50 , flask , and redis .

For the sake of simplicity, we’ll be using ResNet pre-trained on the ImageNet dataset. I’ll point out where you can swap out ResNet for your own models.

The flask module contains the Flask library (used to build our web API). The redis module will enable us to interface with the Redis data store.

From there, let’s initialize constants which will be used throughout run_keras_server.py :

# initialize constants used to control image spatial dimensions and
# data type
IMAGE_WIDTH = 224
IMAGE_HEIGHT = 224
IMAGE_CHANS = 3
IMAGE_DTYPE = "float32"

# initialize constants used for server queuing
IMAGE_QUEUE = "image_queue"
BATCH_SIZE = 32
SERVER_SLEEP = 0.25
CLIENT_SLEEP = 0.25

We’ll be passing float32 images to the server with dimensions of 224 x 224 and containing 3 channels.

Our server can handle a BATCH_SIZE = 32 . If you have GPU(s) on your production system, you’ll want to tune your BATCH_SIZE for optimal performance.

I’ve found that setting both SERVER_SLEEP and CLIENT_SLEEP to 0.25 seconds (the amount of time the server and client will pause before polling Redis again, respectively) will work well on most systems. Definitely adjust these constants if you’re building a production system.

Let’s kick off our Flask app and Redis server:

# initialize our Flask application, Redis server, and Keras model
app = flask.Flask(__name__)
db = redis.StrictRedis(host="localhost", port=6379, db=0)
model = None

Here you can see how easy it is to start Flask.

I’ll assume that before you run this server script that your Redis server is running. Our Python script connect to the Redis store on our localhost on port 6379 (the default host and port values for Redis).

Don’t forget to initialize a global Keras model to None here as well.

From there let’s handle serialization of images:

def base64_encode_image(a):
	# base64 encode the input NumPy array
	return base64.b64encode(a).decode("utf-8")

def base64_decode_image(a, dtype, shape):
	# if this is Python 3, we need the extra step of encoding the
	# serialized NumPy string as a byte object
	if sys.version_info.major == 3:
		a = bytes(a, encoding="utf-8")

	# convert the string to a NumPy array using the supplied data
	# type and target shape
	a = np.frombuffer(base64.decodestring(a), dtype=dtype)
	a = a.reshape(shape)

	# return the decoded image
	return a

Redis will act as our temporary data store on the server. Images will come in to the server via a variety of methods such as cURL, a Python script, or even a mobile app.

Furthermore, images could come in only every once in awhile (a few every hours or days) or at a very high rate (multiple per second). We need to put the images somewhere as they queue up prior to being processed. Our Redis store will act as the temporary storage.

In order to store our images in Redis, they need to be serialized. Since images are just NumPy arrays, we can utilize base64 encoding to serialize the images. Using base64 encoding also has the added benefit of allowing us to use JSON to store additional attributes with the image.

Our base64_encode_image function handles the serialization and is defined on Lines 36-38.

Similarly, we need to deserialize our image prior to passing them through our model. This is handled by the base64_decode_image function on Lines 40-52.

Let’s pre-process our image:

def prepare_image(image, target):
	# if the image mode is not RGB, convert it
	if image.mode != "RGB":
		image = image.convert("RGB")

	# resize the input image and preprocess it
	image = image.resize(target)
	image = img_to_array(image)
	image = np.expand_dims(image, axis=0)
	image = preprocess_input(image)

	# return the processed image
	return image

On Line 54, I’ve defined a prepare_image function which pre-processes our input image for classification using the ResNet50 implementation in Keras.. When utilizing your own models I would suggest modifying this function to perform any required pre-processing, scaling, or normalization.

From there we’ll define our classification method:

def classify_process():
	# load the pre-trained Keras model (here we are using a model
	# pre-trained on ImageNet and provided by Keras, but you can
	# substitute in your own networks just as easily)
	print("* Loading model...")
	model = ResNet50(weights="imagenet")
	print("* Model loaded")

The classify_process function will be kicked off in its own thread as we’ll see in __main__ below. This function will poll for image batches from the Redis server, classify the images, and return the results to the client.

Line 73 loads the model . I’ve sandwiched this action with terminal print messages — depending on the size of your Keras model, loading be instantaneous or it could take a few seconds.

Loading the model happens only once when this thread is launched — it would be terribly slow if we had to load the model each time we wanted to process an image and furthermore it could lead to a server crash due to memory exhaustion.

After loading the model, this thread will continually poll for new images and then classify them:

	# continually pool for new images to classify
	while True:
		# attempt to grab a batch of images from the database, then
		# initialize the image IDs and batch of images themselves
		queue = db.lrange(IMAGE_QUEUE, 0, BATCH_SIZE - 1)
		imageIDs = []
		batch = None

		# loop over the queue
		for q in queue:
			# deserialize the object and obtain the input image
			q = json.loads(q.decode("utf-8"))
			image = base64_decode_image(q["image"], IMAGE_DTYPE,
				(1, IMAGE_HEIGHT, IMAGE_WIDTH, IMAGE_CHANS))

			# check to see if the batch list is None
			if batch is None:
				batch = image

			# otherwise, stack the data
			else:
				batch = np.vstack([batch, image])

			# update the list of image IDs
			imageIDs.append(q["id"])

Here we’re first using the Redis database’s lrange function to get, at most, BATCH_SIZE images from our queue (Line 80).

From there we initialize our imageIDs and batch (Lines 81 and 82) and begin looping over the queue beginning on Line 85.

In the loop, we first decode the object and deserialize it into a NumPy array, image (Lines 87-89).

Next, on Lines 91-97, we’ll add the image to the batch (or if the batch is currently None we just set the batch to the current image ).

We also append the id of the image to imageIDs (Line 100).

Let’s finish out the loop and function:

		# check to see if we need to process the batch
		if len(imageIDs) > 0:
			# classify the batch
			print("* Batch size: {}".format(batch.shape))
			preds = model.predict(batch)
			results = decode_predictions(preds)

			# loop over the image IDs and their corresponding set of
			# results from our model
			for (imageID, resultSet) in zip(imageIDs, results):
				# initialize the list of output predictions
				output = []

				# loop over the results and add them to the list of
				# output predictions
				for (imagenetID, label, prob) in resultSet:
					r = {"label": label, "probability": float(prob)}
					output.append(r)

				# store the output predictions in the database, using
				# the image ID as the key so we can fetch the results
				db.set(imageID, json.dumps(output))

			# remove the set of images from our queue
			db.ltrim(IMAGE_QUEUE, len(imageIDs), -1)

		# sleep for a small amount
		time.sleep(SERVER_SLEEP)

In this code block, we check if there are any images in our batch (Line 103).

If we have a batch of images, we make predictions on the entire batch by passing it through the model (Line 106).

From there, we loop over a the imageIDs and corresponding prediction results (Lines 111-123). These lines append labels and probabilities to an output list and then store the output in the Redis database using the imageID as the key (Lines 117-123).

We remove the set of images that we just classified from our queue using ltrim on Line 126.

And finally, we sleep for the set SERVER_SLEEP time and await the next batch of images to classify.

Let’s handle the /predict endpoint of our REST API next:

@app.route("/predict", methods=["POST"])
def predict():
	# initialize the data dictionary that will be returned from the
	# view
	data = {"success": False}

	# ensure an image was properly uploaded to our endpoint
	if flask.request.method == "POST":
		if flask.request.files.get("image"):
			# read the image in PIL format and prepare it for
			# classification
			image = flask.request.files["image"].read()
			image = Image.open(io.BytesIO(image))
			image = prepare_image(image, (IMAGE_WIDTH, IMAGE_HEIGHT))

			# ensure our NumPy array is C-contiguous as well,
			# otherwise we won't be able to serialize it
			image = image.copy(order="C")

			# generate an ID for the classification then add the
			# classification ID + image to the queue
			k = str(uuid.uuid4())
			d = {"id": k, "image": base64_encode_image(image)}
			db.rpush(IMAGE_QUEUE, json.dumps(d))

As you’ll see later, when we POST to the REST API, we’ll be using the /predict endpoint. Our server could, of course, have multiple endpoints.

We use the @app.route decorator above our function in the format shown on Line 131 to define our endpoint so that Flask knows what function to call. We could easily have another endpoint which uses AlexNet instead of ResNet and we’d define the endpoint with associated function in a similar way. You get the idea, but for our purposes today, we just have one endpoint called /predict .

Our predict method defined on Line 132 will handle the POST requests to the server. The goal of this function is to build the JSON data that we’ll send back to the client.

If the POST data contains an image (Lines 138 and 139) we convert the image to PIL/Pillow format and preprocess it (Lines 142-144).

While developing this script, I spent considerable time debugging my serialization and deserialization functions, only to figure out that I needed Line 148 to convert the array to C-contiguous ordering (which is something you can read more about here). Honestly, it was a pretty big pain in the ass to figure out, but I hope it helps you get up and running quickly.

If you were wondering about the id mentioned back on Line 100, it is actually generated here using uuid , a universally unique identifier, on Line 152. We use a UUID to prevent hash/key conflicts.

Next, we append the id as well as the base64 encoding of the image to the d dictionary. It’s very simple to push this JSON data to the Redis db using rpush (Line 154).

Let’s poll the server to return the predictions:

			# keep looping until our model server returns the output
			# predictions
			while True:
				# attempt to grab the output predictions
				output = db.get(k)

				# check to see if our model has classified the input
				# image
				if output is not None:
 					# add the output predictions to our data
 					# dictionary so we can return it to the client
					output = output.decode("utf-8")
					data["predictions"] = json.loads(output)

					# delete the result from the database and break
					# from the polling loop
					db.delete(k)
					break

				# sleep for a small amount to give the model a chance
				# to classify the input image
				time.sleep(CLIENT_SLEEP)

			# indicate that the request was a success
			data["success"] = True

	# return the data dictionary as a JSON response
	return flask.jsonify(data)

We’ll loop continuously until the model server returns the output predictions. We start an infinite loop and attempt to get the predictions Lines 158-160.

From there, if the output contains predictions, we deserialize the results and add them to data which will be returned to the client.

We also delete the result from the db (since we have pulled the results form the database and no longer need to store them in the database) and break out of the loop (Lines 164-173).

Otherwise, we don’t have any predictions and we need to sleep and continue to poll (Line 177).

If we reach Line 180, we’ve successfully acquired our predictions. In this case we add a success value of True to the client data.

Note: For this example script, I didn’t bother adding timeout logic in the above loop which would ideally add a success value of False to the data. I’ll leave that up to you to handle and implement.

Lastly we call flask.jsonify on data and return it to the client (Line 183). This completes our predict function.

To demo our Keras REST API, we need a __main__ function to actually start the server:

# if this is the main thread of execution first load the model and
# then start the server
if __name__ == "__main__":
	# load the function used to classify input images in a *separate*
	# thread than the one used for main classification
	print("* Starting model service...")
	t = Thread(target=classify_process, args=())
	t.daemon = True
	t.start()

	# start the web server
	print("* Starting web service...")
	app.run()

Lines 187-197 define the __main__ function which will kick off our classify_process thread (Lines 191-193) and run the Flask app (Line 197).

To test our Keras deep learning REST API, be sure to download the source code + example images using the “Downloads” section of this blog post.

From there, let’s start the Redis server if it isn’t already running:

$ redis-server

Then, in a separate terminal, let’s start our REST API Flask server:

$ python run_keras_server.py 
Using TensorFlow backend.
 * Loading Keras model and Flask starting server...please wait until server has fully started
...
 * Running on http://127.0.0.1:5000

Additionally, I would suggest waiting until your model is loaded completely into memory before submitting requests to the server.

Now we can move on to testing the server with both cURL and Python.

Using cURL to access our Keras REST API

The cURL tool is available pre-installed on most (Unix-based) operating systems. We can POST an image file to our deep learning REST API at the /predict endpoint by using the following command:

$ curl -X POST -F image=@jemma.png 'http://localhost:5000/predict'

You’ll receive the predictions back in JSON format right in your terminal:

{
  "predictions": [
    {
      "label": "beagle", 
      "probability": 0.9461546540260315
    }, 
    {
      "label": "bluetick", 
      "probability": 0.031958919018507004
    }, 
    {
      "label": "redbone", 
      "probability": 0.006617196369916201
    }, 
    {
      "label": "Walker_hound", 
      "probability": 0.0033879687543958426
    }, 
    {
      "label": "Greater_Swiss_Mountain_dog", 
      "probability": 0.0025766862090677023
    }
  ], 
  "success": true
}

Let’s try passing another image, this time a space shuttle:

$ curl -X POST -F image=@space_shuttle.png 'http://localhost:5000/predict'
{
  "predictions": [
    {
      "label": "space_shuttle", 
      "probability": 0.9918227791786194
    }, 
    {
      "label": "missile", 
      "probability": 0.006030891090631485
    }, 
    {
      "label": "projectile", 
      "probability": 0.0021368064917623997
    }, 
    {
      "label": "warplane", 
      "probability": 1.980597062356537e-06
    }, 
    {
      "label": "submarine", 
      "probability": 1.8291866581421345e-06
    }
  ], 
  "success": true
}

The results of which can be seen below:

Once again our Keras REST API has correctly classified the input image.

Using Python to submit requests to the Keras REST API

As you can see, verification using cURL was quite easy. Now let’s build a Python script that will POST an image and parse the returning JSON programmatically.

Let’s review simple_request.py :

# import the necessary packages
import requests

# initialize the Keras REST API endpoint URL along with the input
# image path
KERAS_REST_API_URL = "http://localhost:5000/predict"
IMAGE_PATH = "jemma.png"

We use Python requests in this script to handle POSTing data to the server.

Our server is running on the localhost and can be accessed on port 5000 with the endpoint /predict as is specified by the KERAS_REST_API_URL variable (Line 6). If the server is running remotely or on a different machine, be sure to specify the appropriate domain/ip, port, and endpoint.

We also define an IMAGE_PATH (Line 7). In this case, jemma.png is in the same directory as our script. If you want to test with other images, be sure to specify the full path to your input image.

Let’s load the image and send it off to the server:

# load the input image and construct the payload for the request
image = open(IMAGE_PATH, "rb").read()
payload = {"image": image}

# submit the request
r = requests.post(KERAS_REST_API_URL, files=payload).json()

# ensure the request was sucessful
if r["success"]:
	# loop over the predictions and display them
	for (i, result) in enumerate(r["predictions"]):
		print("{}. {}: {:.4f}".format(i + 1, result["label"],
			result["probability"]))

# otherwise, the request failed
else:
	print("Request failed")

We read the image on Line 10 in binary mode and put the it into a payload dictionary.

The payload is POST’ed to the server with requests.post on Line 14.

If we get a success message, we can loop over the predictions and print them to the terminal. I made this script simple, but you could also draw the highest prediction text on the image using OpenCV if you want to get fancy.

Running the simple request script

Putting the script to work is easy. Open up a terminal and execute the following command (provided both our Flask server and Redis server are running, of course).

$ python simple_request.py
1. beagle: 0.9462
2. bluetick: 0.0320
3. redbone: 0.0066
4. Walker_hound: 0.0034
5. Greater_Swiss_Mountain_dog: 0.0026

For the space_shuttle.png , simply modify the IMAGE_PATH variable:

IMAGE_PATH = "space_shuttle.png"

And from there, run the script again:

$ python simple_request.py
1. space_shuttle: 0.9918
2. missile: 0.0060
3. projectile: 0.0021
4. warplane: 0.0000
5. submarine: 0.0000

Considerations when scaling your deep learning REST API

If you anticipate heavy load for extended periods of time on your deep learning REST API you may want to consider a load balancing algorithm such as round-robin scheduling to help evenly distribute requests across multiple GPU machines and Redis servers.

Keep in mind that Redis is an in-memory data store so we can only store as many images in the queue we have available memory.

A single 224 x 224 x 3 image with a float32 data type will consume 60,2112 bytes of memory.

Assuming a server with a modest 16GB of RAM, this implies that we can hold approximately 26,500 images in our queue, but at that point we likely would want to add more GPU servers to burn through the queue faster.

However, there is a subtle problem…

Depending on how you deploy your deep learning REST API, there is a subtle problem with keeping the classify_process function in the same file as the rest of our web API code.

Most web servers, including Apache and nginx, allow for multiple client threads.

If you keep classify_process in the same file as your predict view, then you may load multiple models if your server software deems it necessary to create a new thread to serve the incoming client requests — for every new thread, a new view will be created, and therefore a new model will be loaded.

The solution is to move classify_process to an entirely separate process and then start it along with your Flask web server and Redis server.

In next week’s blog post I’ll build on today’s solution, show how to resolve this problem, and demonstrate:

• How to configure the Apache web server to serve our deep learning REST API
• How to run classify_process as an entirely separate Python script, avoiding “multiple model syndrome”
• Provide stress test results, confirming and verifying that our deep learning REST API can scale under heavy load

Summary

In today’s blog post we learned how to build a scalable Keras + deep learning REST API.

To accomplish this, we:

• Built a simple Flask app to load our Keras model into memory and accept incoming requests.
• Utilized Redis to act as an in-memory message queue/message broker.
• Utilized threading to batch process input images, write them back to the message queue and then return the results to the client.

This method can scale to multiple machines, including multiple web servers and Redis instances.

I hope you enjoyed today’s blog post!

Be sure to enter your email address in the form below to be notified when future tutorials are published here on PyImageSearch!