Flask api
Flask API to serve predictions on the Iris dataset. The dataset consists 4 measurements for 150 samples of 3 different varieties of Irises.
Takes a Post request containing one or more of the feature measurements and returns the predicted class as a string in JSON format
Example
# Input
{
"sepal_width":4,
"sepal_length":3.5,
"petal_width":2.5,
"petal_length":4
}# Output
{
"Prediction": "versicolor"
}The application.py file can be run to initialize the server and requests submitted via postman.
If one or more of the predictors is missing it is replaced with the sample mean. For example if no value is specified for sepal_width, the model will use the average value of 3.06 to return a prediction. If all values are missing then an error is raised. The model used is an SVM but many models can be used with great sucess on the iris dataset.
The Iris dataset is very small and thus fitting models does not lead to any memory or computing issues. Handling datasets which do not fit into system memory would require some alternative aproaches. For development of a model one can use a smaller random sample of the data, allowing for rapid prototyping. The final model should then be fit on the full dataset using one of the following approaches. One is to use a cloud computing platform to provide the necessary memory. One could also progressively load the data or use a relational database. The streaming approach might require a choice of model that can handle iterative learning, for example one trained using stochastic gradient descent.
A versioning strategy for machine learning APIs shoudl make retraining and or updating the model as seamless as possible. The versioning should distinguish between model versions used in production and iterations of models used during development or retraining. Training models should maintain a version history of any data transformations or hyperparameter changes. After training it should be easy to do AB testing of the new and old model before full deployment of the new model. The process should be able to automatically revert to the origninal model if something goes wrong. Therefore the versioning should store a history of each deployment model as well as a history of the steps used to train each model.
I choose to use an SVM model on the data because it was a model I hadn't used before and because it had a number of advantages for this dataset. I expected the SVM to be effective for a small dataset without a large degree of noise and with a small number of classes to predict. In addition this data was clearly linearly seperable. Indeed, the SVM was able to distinguish between the 3 varieties of iris with a high degree of accuracy. An SVM can also make predictions extremely quickly. If the dataset were to scale vastly in size it may become expensive to fit the SVM and require using faster algorithms such as sub-gradient descent.