mickwar/ml-deploy

Model deployment

This repository presents a simple example for deploying a machine learning model to a (local) server. A server will load the model and wait for user input. Data can be sent to the server with an HTTP POST request and model predictions will be returned.

Code for the server is written using flask, and the server is hosted using gunicorn.

Docker can be used as a sort of portable, lightweight hosting environment. The Docker image used in this example uses Ubuntu installed with Python3 and contains the trained machine learning model.

Props to this tutorial.

Install the required packages

It is recommended to use a virtual environment.

# bash
cd program                          # note: the virtual environment doesn't have to be in this directory
pip install virtualenv              # install virtualenv module
virtualenv env                      # initialize a virtual environment named env
source env/bin/activate             # activate the environment
pip install -r requirements.txt     # install the required packages of this example to the environment
pip list                            # view list of installed packages

(Note that the virtual environment can be named whatever, but the local .gitignore will exclude the env directory.)

When within the virtual environment, packages were installed via pip install. But when the environment is set up as desired, future intializations are completed using the requirements.txt file as above.

The requirements.txt file was generated with

pip freeze > requirements.txt

while within the virtual environment. The command pip freeze takes a snapshot of the currently installed modules and their version numbers.

The virtual environment is activated with

source env/bin/activate

and will prepend the shell prompt to include (env).

The virtual environment can be deactivated simply with

deactivate

The remainder of this example expects the user to have the virtual environment activated with the packages listed under requirements.txt installed. The Docker image does not use the virtual environment, instead the packages are installed globally (for the image).

Get the server ready

Generating fake data, training a linear model, and spinning up the server can be done using the bash scripts inside program.

cd program      # This is necessary to execute setup.sh and run.sh
./setup.sh      # generate data, fit the model
./run.sh bg     # start the server in the background

These two scripts are simply wrapper functions for the process described in the next three subsections.

Generate data

Execute the gen_data.py script

python data/gen_data.py

This will create four files:

data/train_x.txt
data/train_y.txt
data/test_x.txt
data/test_y.txt

Fit a model

Execute the lm_train.py script to fit a linear model to the training data. Note that this will be a bad model since the response is fairly non-linear.

python model/lm_train.py

The model will be saved to model/lm_mod.pickle using pickle.

Run the server

Start hosting the server with

gunicorn --bind 0.0.0.0:8000 --daemon server:app

This will run the run the flask app (called app) found in server.py and can be accessed at http://0.0.0.0:8000. The app itself only has one endpoint /predict and expects a POST method. So, if you go to http://0.0.0.0:8000/predict in your browser, you'll receive an error saying the method is not allowed (the method in this case would be the GET method).

When running gunicorn in the background (i.e. as a daemon), the process can be killed with

pkill gunicorn

Make predictions

With the server up and running---listening on 0.0.0.0:8000---we are ready to send data to the server.

Technically, anything that lets you make a POST request---accompanied with some properly formatted JSON data---would be an acceptable way of communicating with the server. For this example, we can use the pred.py script.

The pred.py script sends the test data (in JSON format) to the server via the /predict endpoint. The data is processed within server.apicall() and the predictions are sent back as JSON to the user. Notice that some conversion between JSON and numpy array is done since the model expects a numpy array, but the HTTP request expects JSON.

It may be most helpful to play with pred.py interactively, but running

python pred.py

is also acceptable. You should see a table print out for the predicted values, followed by the MSE:

           0
0  -0.155496
1   0.098487
2   2.130050
3  -2.077463
4   2.562407
5  -0.085750
6  -0.352975
7  -1.987332
8  -0.351220
9   0.951248
10 -1.636224
11 -3.547744
12 -1.239803
13 -0.666220
14 -0.059957
15 -1.367055
16  0.370481
17 -1.266943
18 -0.540817
19 -2.843020

[2.09111105]

Docker

Preliminaries

Install Docker

Use Docker without sudo

Set up

Build the Docker image with

./build_docker.sh

This may take a minute or two since the base Ubuntu image will be pulled from Docker Hub and Python3 will have to be installed on the image.

Once the image is built, you are ready to run a container (something like a temporary copy of Docker image). Run the container with

docker run -p 8181:8000 ml-deploy

The 8181 is the port that the host will listen on, meaning we will send requests to 8181 (note the commented out section in pred.py). The 8000 is the port used internally by the Docker container for the gunicorn server. And ml-deploy is the name of the image created from the build_docker.sh script.

At this point, you can use the pred.py script again to send requests to the container. As noted earlier, the container listens on port 8181, so that change would need to be accounted for when running pred.py.

To run the container interactively, comment out the ENTRYPOINT line in the dockerfile when building the image. This shouldn't really be the case, but I couldn't figure out how to be able to run the container normally and also have access to it interactively. Running the container interactively puts you in a shell for the container. You can do whatever you want, such as installing packages or running some Python code. But after the container is stopped, no changes are saved (recall the temporary nature of Docker containers).

Useful Docker commands

command	description
`docker build -t <tag> .`	build an image from a dockerfile (found in current directory)
`docker run -it <image>`	run interactively (opens bash, at least on my system when running on a base Ubuntu image)
`docker images`	list images
`docker image rm <image>`	remove
`docker image prune`	remove unused (including intermediary) images
`docker ps`	list containers (`-a` to show all, including stopped)
`docker container prune`	remove all stopped containers

mickwar / ml-deploy