Statsmodels Holt-Winter's time-series model

The following notebook are steps to deploy your first time-series model on Seldon. The first step is to install statsmodels on our local system, along with s2i. s2i will be used to convert the source code to a docker image and stasmodels is a python library to build statistical models.

Dependencies:

1. Seldon-core (https://docs.seldon.io/projects/seldon-core/en/v1.1.0/workflow/install.html)

2. s2i - Source to Image

3. statsmodels (https://www.statsmodels.org/stable/index.html)

Assuming you have installed statsmodels and s2i, the next step is to create a joblib file of your time-series model. The sample code is given below . Here we have considered a Holt- Winter's seasonal model and the shampoo sales dataset as a basic example.

The univariate dataset : https://raw.githubusercontent.com/jbrownlee/Datasets/master/shampoo.csv

!pip install statsmodels

Code snippet to create a joblib file :

import joblib
import numpy as np
import pandas as pd
from statsmodels.tsa.holtwinters import ExponentialSmoothing

df = pd.read_csv(
    "https://raw.githubusercontent.com/jbrownlee/Datasets/master/shampoo.csv"
)

# Taking a test-train split of 80 %
train = df[0 : int(len(df) * 0.8)]
test = df[int(len(df) * 0.8) :]

# Pre-processing the  Month  field
train.Timestamp = pd.to_datetime(train.Month, format="%m-%d")
train.index = train.Timestamp
test.Timestamp = pd.to_datetime(test.Month, format="%m-%d")
test.index = test.Timestamp

# fitting the model based on  optimal parameters
model = ExponentialSmoothing(
    np.asarray(train["Sales"]),
    seasonal_periods=7,
    trend="add",
    seasonal="add",
).fit()
joblib.dump(model, "model.sav")

The Next step is to write the code in a format defined by s2i as given below :

%%writefile holt_winter.py

import joblib
class holt_winter(object):
    """
    Model template. You can load your model parameters in __init__ from a location accessible at runtime
    """
    
    def __init__(self):
        
        """
        Add any initialization parameters. These will be passed at runtime from the graph definition parameters defined in your seldondeployment kubernetes resource manifest.
        
        loading the joblib file 
        """
        self.model = joblib.load('model.sav')
        print("Initializing ,inside constructor")


    def predict(self,X,feature_names):
        """
        Return a prediction.
        Parameters
        ----------
        X : array-like
        feature_names : array of feature names (optional)
        
        This space can be used for data pre-processing as well
        """
        print(X)
        print("Predict called - will run idenity function")
        return self.model.forecast(X)

After saving the code, we now create an environment_rest file and add the following lines:

MODEL_NAME=holt_winter API_TYPE=REST SERVICE_TYPE=MODEL PERSISTENCE =0

MODEL_NAME: The name of the class containing the model. Also the name of the python file which will be imported.

API_TYPE: API type to create. Can be REST or GRPC

SERVICE_TYPE: The service type being created. Available options are:

1. MODEL

2. ROUTER

3. TRANSFORMER

4. COMBINER

5. OUTLIER_DETECTOR

PERSISTENCE: Set either to 0 or 1. Default is 0. If set to 1 then your model will be saved periodically to redis and loaded from redis (if exists) or created fresh if not.

%%writefile requirements.txt
joblib
statsmodels
pandas
numpy

%%writefile environment_rest

MODEL_NAME=holt_winter
API_TYPE=REST 
SERVICE_TYPE=MODEL
PERSISTENCE =0

Now we build the image using the s2i command, replace "seldonio/statsmodel-holts:0.1" with the image name of your choice :

!s2i build -E environment_rest . seldonio/seldon-core-s2i-python37-ubi8:1.7.0-dev seldonio/statsmodel-holts:0.1

Running the docker image created:

!docker run --name "holt_predictor" -d --rm -p 5000:5000 seldonio/statsmodel-holts:0.1

The code is now running at the local host at port 5000. It can be tested by sending a curl command, here we are sending a request to the model to predict the sales for the next 3 weeks.

!curl  -s http://localhost:5000/predict -H "Content-Type: application/json" -d '{"data":{"ndarray":3}}'

The next step is to push the code into the docker registry, you are free to use the docker hub or the private registry in your cluster.

!docker push seldonio/statsmodel-holts:0.1

The final step is to deploy the configuration file on your cluster as shown below.

%%writefile model.yaml

apiVersion: machinelearning.seldon.io/v1alpha2
kind: SeldonDeployment
metadata:
  name: holt-predictor
spec:
  name: holt-predictor
  predictors:
  - componentSpecs:
    - spec:
        containers:
        - image: seldonio/statsmodel-holts:0.1
          imagePullPolicy: IfNotPresent
          name: holt-predictor
    graph:
      children: []
      endpoint:
        type: REST
      name: holt-predictor
      type: MODEL
    name: holt-predictor
    replicas: 1

!kubectl apply -f model.yaml

Your model will now be deployed as a service, create a route in order for external traffic to access it . A sample curl request (with a dummy I.P, replace it with the route created by you) for the model is :

!curl -s -d '{"data": {"ndarray":2}}'    -X POST http://160.11.22.334:4556/seldon/testseldon/holt-predictor/api/v1.0/predictions    -H "Content-Type: application/json"

In the above command, we send a request to get a prediction of the sales of the shampoo for the next 2 days. testseldon is the namespace, you can replace it with the namespace created by you where the model is deployed .

The response we get is :

{"data":{"names":[],"ndarray":[487.86681173,415.82743026 ]},"meta":{}}

The data returned is an n-dimensional array with 2 values which is the predicted values by the model, in this case the sales of the shampoo.

Note: it is suggested that you try the model on your local system before deploying it on the cluster.

Model Monitoring

Once the model is deployed, you can now monitor various metrics, the 2 main ones being:

1. Requests per second

2. Latency in serving the request

The model deployed on Seldon can be monitored using build in metrics dashboard on Grafana. Here is the link to deploy metrics dashboard: https://docs.seldon.io/projects/seldon-core/en/v1.1.0/analytics/analytics.html. The screenshot of a sample dashboard is given below:

Summary

This documentation covers deploying time series model on Seldon, this model could be inferenced for forecasting values from a given data set. This is very useful for customers who want to deploy time series alogithm for forecasting models.