You can deploy your own model by
- Copying the contents of this repo to a new directory
- Replace
pipeline.pickle,dtypes.pickle, andcolumns.jsonwith your own - Deploy to railway
You'll probably run into a few issues along the way which is why you'll at least want to skim the contents of the notebooks and this README, in order to have an idea of where to look when you hit a bump in the road.
This is a very simplistic yet effective way to deploy a scikit binary classifier behind a HTTP server on railway.
There are 4 main topics to cover here:
- Serialization
- This is covered in notebooks
- Flask
- Covered here in the README
- Database connection
- Covered here in the README
- Deployment to railway
- Also covered here in the README
Clone this repo and copy the contents to a new directory - if you don't remember how, head to the initial setup readme.
Topic #1 is the only one that is not covered here in this README. It is covered in two notebooks that you must read before moving on with the rest of this README.
Notebook #1 has to do with training and serializing a scikit model as well as how to prepare a new observation that arrives for prediction.
Notebook #2 has to do with deserialization so that you can re-use a model on new observations without having to re-train it.
You've probably noticed that we have two requirement files in this repo: requirements_dev.txt and requirements_prod.txt.
The requirements_dev.txt file has the packages that are needed while working on the predictive model, which include jupyter and matplotlib.
The requirements_prod.txt file has the packages that are needed when we deploy our model. At that time, we won't need jupyter or matplotlib, so we can save some resources by not installing them.
Now go ahead and create a python virtual env using the requirements in requirements_dev.txt, in order to follow this tutorial.
Have you already read and understood the notebooks on serialization? Have you already tested your understanding by pickling and un-pickling your scikit model? Yes yes? Alrighty then, you may continue.
Flask is an HTTP micro-framework. It is a very minimal code library that allows for quick and simple HTTP server development and is a great alternative to bigger frameworks like Django. However, be wary before moving forward with a big project using flask - it can get out of hand very quickly without the enforced structure that other heavier frameworks like Django provide.
For us, since we only need a total of two endpoints (an endpoint is the URL that is used to request an action from the server, in our case it will be two types of actions - requesting a prediction for an observation and updating an observation's true class) and it doesn't even need to be RESTful, we can stick with flask and be reasonably justified in it.
In order to use flask, you will need to be writing some code in a regular python file - no more notebooks here. So open a file in your favourite text editor and start writing.
The first step (assuming you have already created an virtual environment and installed the requirements in requirements_dev.txt), is to import flask at the top of the file. Let's pretend that we are working in a file called app.py in our newly created
virtual environment.
# The Flask object is for creating an HTTP server - you'll
# see this a few lines down.
# The request object does exactly what the name suggests: holds
# all of the contents of an HTTP request that someone is making.
# The jsonify function is useful for when we want to return
# json from the function we are using.
from flask import Flask, request, jsonify
# Here we use the Flask constructor to create a new
# application that we can add routes to.
app = Flask(__name__)This server doesn't do anything yet. In order to make it do stuff we will need to add HTTP endpoints to it.
With flask, creating an HTTP endpoint is incredibly simple, assuming that we already
have the app object created from the Flask constructor. Let's make a single
endpoint that will serve the predictions:
@app.route('/predict', methods=['POST'])
def predict():
prediction = 0.5
return jsonify({
'prediction': prediction
})The above route that we have isn't very smart in that it returns the same prediction every time (0.5) and it doesn't actually care about the input that you sent it. But hey, with just a few lines of code we've almost created an entire server that serves a prediction!
Putting it all together with a few lines of code at the end (in order to start the server in development mode), we've created an entire server that can be run by executing python app.py:
# these contents can be put into a file called app.py
# and run by executing:
# python app.py
from flask import Flask, request, jsonify
app = Flask(__name__)
@app.route('/predict', methods=['POST'])
def predict():
prediction = 0.5
return jsonify({
'prediction': prediction
})
if __name__ == "__main__":
app.run(debug=True)If you want to experiment with sending requests to your server, open a new terminal window and execute the following command in order to get a prediction. It's not necessary to activate the virtual environment for this step.
curl -X POST http://localhost:5000/predict
This should be the output you get:
{
"prediction": 0.5
}We are using the curl library here to transfer data to our local server.
Alright, now that we can run a full flask server, let's try to make something a bit more useful by receiving new data.
So now that we've got a way to build an entire server, let's try to actually use the server to receive new information. There's a pretty nice way to do this via the get_json flask function.
For this server, let's say that the model only takes a single field called unemployed
and returns true if unemployed is true and false otherwise. The server would now
look like this:
from flask import Flask, request, jsonify
app = Flask(__name__)
@app.route('/predict', methods=['POST'])
def predict():
payload = request.get_json()
at_risk = payload['unemployed']
return jsonify({
'prediction': at_risk
})
if __name__ == "__main__":
app.run(debug=True)You can notice that the server has automatically restarted when you made changes to the app.py file (unless you made a syntax error in which case it shut down).
Now call the following two commands and see the output:
curl -X POST http://localhost:5000/predict -d '{"unemployed": true}' -H "Content-Type:application/json"The output will be:
{
"prediction": true
}For this:
curl -X POST http://localhost:5000/predict -d '{"unemployed": false}' -H "Content-Type:application/json"the output will be:
{
"prediction": false
}Take a quick note that we had to supply a header of Content-Type:application/json
and json data of {"unemployed": false}.
Now that we know how to get a python dictionary via the flask get_json
function, we're at a point in which we can pick up where the last tutorial
notebook left off! Let's tie it all together by:
- Deserializing the model, columns, and dtypes
- Turn the new observation into a pandas dataframe
- Call
predict_probato get the likelihood of survival of the new observation
import joblib
import json
import pickle
import pandas as pd
from flask import Flask, request, jsonify
app = Flask(__name__)
with open('columns.json') as fh:
columns = json.load(fh)
with open('dtypes.pickle', 'rb') as fh:
dtypes = pickle.load(fh)
pipeline = joblib.load('pipeline.pickle')
@app.route('/predict', methods=['POST'])
def predict():
payload = request.get_json()
obs = pd.DataFrame([payload], columns=columns).astype(dtypes)
proba = pipeline.predict_proba(obs)[0, 1]
return jsonify({
'prediction': proba
})
if __name__ == "__main__":
app.run(debug=True)Check out how we have now taken the payload and turned it into a new observation that is a single entry in a dataframe, and can be consumed by the pipeline to be turned into a prediction of survival. You can see the output with the following:
curl -X POST http://localhost:5000/predict -d '{"Age": 22.0, "Cabin": NaN, "Embarked": "S", "Fare": 7.25, "Parch": 0, "Pclass": 3, "Sex": "male", "SibSp": 1}' -H "Content-Type:application/json"The expected output is
{
"prediction": 0.09194812456330213
}
Okay, now that you can get data, produce predictions, and return those predictions, you will need to keep track of what you've been saying about who. Said another way: you can't just provide predictions and then forget about it all. You need to take record of what you have predicted about whom, so that later on you can do some additional analysis on your "through the door" population.
In order to do this, we will need to start working with a database. The database will keep track of the observations, the predictions we have provided for them, and the true outcomes (should we be lucky enough to find out about them).
When working with databases in code, you generally want to be using a layer of abstraction called an ORM. For this exercise we will use a very simplistic ORM called peewee. This will allow us to use a local database called sqlite (which is basically a file) when we are developing on our laptops, and use a more production-ready database called postgresql when deploying to railway, with very little change to our code.
One cool thing that ORMs allow us to do is define the data model that we want to use in code. So let's use peewee to create a data model to keep track of predictions and the probabilities we have assigned to them. Once again, we can take care of this in a few lines of code:
from peewee import (
SqliteDatabase, Model, IntegerField,
FloatField, TextField,
)
DB = SqliteDatabase('predictions.db')
class Prediction(Model):
observation_id = IntegerField(unique=True)
observation = TextField()
proba = FloatField()
true_class = IntegerField(null=True)
class Meta:
database = DB
DB.create_tables([Prediction], safe=True)Now we need to take a moment to understand exactly how much these few lines of code have done for us because it is A LOT.
DB = SqliteDatabase('predictions.db')
Create a sqlite database that will be stored in a file called predictions.db. This may seem trivial right now, but soon enough you will see that swapping this line of code for another one will result in a lot of value for the effort.
Class Prediction(Model)...
Define the data model that we will work with. The model has sections for the following:
observation_id- There must be a unique identifier to all observations and it is the responsibility of the person providing the observation to give this id.
observation- We should record the observation itself when it comes, in case we want to retrain our model later on.
proba- The probability of survival that we assigned to the observation.
true_class- This is for later on, in case we find out what actually happened to the observation for which we supplied a prediction.
DB.create_tables([Prediction], safe=True)
The model that we specified must correspond to a database table. Creation of these tables is a separate non-trivial headache, and with this one line of code we don't have to worry about any of it.
Now that we have a webserver and a data model that we are happy with, the next question is how do we put them together? It's actually pretty straightforward!
import joblib
import json
import pickle
import pandas as pd
from flask import Flask, jsonify, request
from peewee import (
SqliteDatabase, PostgresqlDatabase, Model, IntegerField,
FloatField, TextField, IntegrityError
)
from playhouse.shortcuts import model_to_dict
########################################
# Begin database stuff
DB = SqliteDatabase('predictions.db')
class Prediction(Model):
observation_id = IntegerField(unique=True)
observation = TextField()
proba = FloatField()
true_class = IntegerField(null=True)
class Meta:
database = DB
DB.create_tables([Prediction], safe=True)
# End database stuff
########################################
########################################
# Unpickle the previously-trained model
with open('columns.json') as fh:
columns = json.load(fh)
with open('dtypes.pickle', 'rb') as fh:
dtypes = pickle.load(fh)
pipeline = joblib.load('pipeline.pickle')
# End model un-pickling
########################################
########################################
# Begin webserver stuff
app = Flask(__name__)
@app.route('/predict', methods=['POST'])
def predict():
obs_dict = request.get_json()
_id = obs_dict['id']
observation = obs_dict['observation']
obs = pd.DataFrame([observation], columns=columns).astype(dtypes)
proba = pipeline.predict_proba(obs)[0, 1]
response = {'proba': proba}
p = Prediction(
observation_id=_id,
proba=proba,
observation=request.data
)
try:
p.save()
except IntegrityError:
error_msg = "ERROR: Observation ID: '{}' already exists".format(_id)
response["error"] = error_msg
print(error_msg)
DB.rollback()
return jsonify(response)
# End webserver stuff
########################################
if __name__ == "__main__":
app.run(debug=True)One piece of the code above that might not be clear at first is:
p = Prediction(
observation_id=_id,
proba=proba,
observation=request.data,
)
try:
p.save()
except IntegrityError:
error_msg = "ERROR: Observation ID: '{}' already exists".format(_id)
response["error"] = error_msg
print(error_msg)
DB.rollback()What is this code doing? When we receive a new prediction request, we want to store such a request in our database (to keep track of our model performance). With peewee, we save a new Prediction (basically a new row in our table) with the save() method, which is very neat and convenient.
However, because our table has a unique constraint (no two rows can have the same observation_id, which is a unique field),
if we perform the same prediction request twice (with the same id) the system will crash because pewee can't save
again an already saved observation_id; it will throw an IntegrityError (as in, we would be asking pewee to violate
the integrity of the table's unique id requirement if we saved a duplicated id, right?).
To avoid that, we do a simple try/except block: if we try a request with the same observation_id, peewee will raise the integrity error and we'll catch it, print a nice error message, and do a database rollback (to close the current save transaction that has failed).
Once your app is set up like this, you can test this with the following command:
curl -X POST http://localhost:5000/predict -d '{"id": 0, "observation": {"Age": 22.0, "Cabin": NaN, "Embarked": "S", "Fare": 7.25, "Parch": 0, "Pclass": 3, "Sex": "male", "SibSp": 1}}' -H "Content-Type:application/json"You should an output like this:
{
"proba": 0.09194812456330213
}Now let's take note of the few things that have changed:
- The structure of the json input changed. It now includes two top level entries:
id- This is the unique identifier of the observation;observation- This is the actual observation contents that will be sent through the pipeline we have un-pickled.
- We create an instance of
Predictionwith the 3 fields that we care about. - We call
save()on the prediction to save it to the database. - We return
probaso that the caller of the HTTP endpoint knows what you are saying about the observation.
Now that we have a way to provide predictions AND keep track of them, we should take it to the next level and provide ourselves with a way to receive updates on observations that we have judged with our predictive model.
We can do this with one extra endpoint that is very straightforward and only introduces one new concept: database querying through the ORM.
@app.route('/update', methods=['POST'])
def update():
obs = request.get_json()
try:
p = Prediction.get(Prediction.observation_id == obs['id'])
p.true_class = obs['true_class']
p.save()
return jsonify(model_to_dict(p))
except Prediction.DoesNotExist:
error_msg = 'Observation ID: "{}" does not exist'.format(obs['id'])
return jsonify({'error': error_msg})Assuming that we have already processed an observation with id=0, we can now receive and record the true outcome. Imagine that it is discovered later on that the person with id=0 didn't survive the titanic disaster. They would probably enter something into a content management system that would then trigger a call to your server which would end up looking like the following:
curl -X POST http://localhost:5000/update -d '{"id": 0, "true_class": 0}' -H "Content-Type:application/json"The output confirms new data stored in the database:
{
"id": 1,
"observation": "{\"id\": 0, \"observation\": {\"Age\": 22.0, \"Cabin\": NaN, \"Embarked\": \"S\", \"Fare\": 7.25, \"Parch\": 0, \"Pclass\": 3, \"Sex\": \"male\", \"SibSp\": 1}}",
"observation_id": 0,
"proba": 0.09194812456330213
"true_class": 0
}Similarly to when we saved the prediction requests, we validate that the observation_id we want to update actually exists.
Now to wrap it all up, the way that we can interpret this sequence of events is the following:
- We provided a prediction of 0.091 probability of survival;
- We found out later that the person didn't survive.
It's cool that we can run the server on our own machine. However, it doesn't do much good in terms of making the model available to the rest of the world. All this localhost stuff doesn't help anybody that's not typing on your local machine.
So let's take all of the work we've done and put it on railway where it can generate real business value. For this part, you can use any server that has a static IP address. Since we want to avoid the overhead of administering our own server, we will use a service to do this for us called railway. Railway includes a free-tier (if you connect your GitHub account) that should be enough for our needs throughout this specialization and the capstone project that is fast approaching, but only if you remember to turn off your applications once you're done with them. You have $5.00 of credit each month, but it's your responsibility to manage this. Be careful before you move forward with a big project on railway - it can get CRAZY expensive REALLY fast.
Go to the railway main page and start a new project deployed from a github repo. Alternatively you can use the login button.
Sign in using your GitHub credentials and verify your account.
Once this is all done, go to the dashboard and create a new app:
Then on the next screen, we'll reselect the "deploy from GitHub" option. Grant access to your repositories and select the appropriate one. Afterwards click on "Deploy now".
Once this is done, you'll be taken to the main dashboard where you'll see your application building. Wait until it's completed, signaled by a green box saying "Success". In this tab we can see a button that will allow us to check the logs of our app. This will be very important to check whether our app is responding successfully to the requests sent or not.
Go to "Settings" and scroll down to "Generate Domain". Click on the "Generate domain" button. This will generate a url for your app, and we can use it to make requests to your app.
One last thing before we can move on to the database creation, we need to redefine what port our requests will be coming from. As seen above, we're using port 5000. We can define this by going to the "Variables" tab, and add a new PORT variable with this value.
One last bit is missing here: the database. We are going to use a big boy database called postgreSQL. You should try to be conservative with how you connect to the app and don't go crazy with it. If the database gets full your app will stop working!
You can check railway's postgreSQL guide here.
To add a Database to our app, select "New", "Database", and then "Add PostgreSQL".
After the postgreSQL database is created, we'll need to get the DATABASE_URL so we can connect our two services. We'll copy the Postgres DATABASE_URL available in the "Variables" tab of the Postgres service.
Go back to the dashboard and select your app.
Got to "Variables" and add a "New Variable" called DATABASE_URL with the value that you copied earlier. It should look something like this:
Wait for the app to redeploy after adding the new variable and test that the app is connected to the database like this:
There is no data in the table, but it exists because the query ran successfully.
To stop billing and resource usage, you'll need to stop the application from running. To do so, find your app and database in the dashboard and select the 3 dots in the most recent deployment. Select "Remove".
We have now successfully deployed our app IN THE CLOUD, WAAAT?!?! It is important to remember that only the changes that you have committed (with git add & git commit) will be deployed. So if you change your pipeline and retrain a different model you'll need to commit the changes before rebuilding in railway.
And boom! We're done and deployed! You can actually see this working by executing
some of the curl commands that we saw before but using the URL we've obtained before from railway, rather than http://localhost like we saw earlier. For my app it looks like the following:
curl -X POST https://<given_url>/predict -d '{"id": 0, "observation": {"Age": 22.0, "Cabin": NaN, "Embarked": "S", "Fare": 7.25, "Parch": 0, "Pclass": 3, "Sex": "male", "SibSp": 1}}' -H "Content-Type:application/json"Output:
{
"proba": 0.09194812456330213
}And we can receive updates like the following:
curl -X POST https://<given_url>/update -d '{"id": 0, "true_class": 1}' -H "Content-Type:application/json"Output:
{
"id": 1,
"observation": "{\"id\": 0, \"observation\": {\"Age\": 22.0, \"Cabin\": NaN, \"Embarked\": \"S\", \"Fare\": 7.25, \"Parch\": 0, \"Pclass\": 3, \"Sex\": \"male\", \"SibSp\": 1}}",
"observation_id": 0,
"proba": 0.09194812
"true_class": 1
}How does all of this work you ask? Well, it has to do with the fact that there exists a Dockerfile this repo.
Just be sure that this file is in your repo and that all is committed before you build your app in railway, and everything should work. If you do a deploy and it doesn't work, take a look at the files and see if you might have done something such as change a filename or something else that would break the boilerplate assumptions that we've made here.
If you are using something like a custom transformer and get an import error having to do with your custom code when unpickling, you'll need to do the following.
Let's say that you have a custom transformer, called MyCustomTransformer, that is part of your
pickled pipeline. In that case, you'll want to create a python package
from which you import the custom transformer in both your training and deployment code.
In our example, let's create the following package called custom_transformers by just creating a directory
with the same name and putting two files inside of it so that it looks like this:
└── custom_transformers
├── __init__.py
└── transformer.pyAnd inside of transformer.py you can put the code for MyCustomTransformer. Then, in your training code, you can import them with:
from custom_transformers.transformer import MyCustomTransformerWhen you unpickle your model, python should be able to find the custom transformer too.
The dependencies of your custom transformer should be added to the two requirements.txt files.
There were a few additional changes to app.py and the rest of the repo that we haven't covered yet, so
let's get that out of the way. You probably won't need to know much about them but if you are having
troubleshooting issues, knowing the following may come in handy.
When our app is running on railway, we want to connect to a postgres database, rather than to an sqlite one. Thus, we had to change the database-related code to something that takes care of this:
import os
from playhouse.db_url import connect
# The connect function checks if there is a DATABASE_URL env var.
# If it exists, it uses it to connect to a remote postgres db.
# Otherwise, it connects to a local sqlite db stored in the predictions.db file.
DB = connect(os.environ.get('DATABASE_URL') or 'sqlite:///predictions.db')