Neural network for solving regression problem with deployment

This repo contents two following sections:

1. Visualization of training process and dataset analysis

Dataset analysis, preprocessing, metrics and other analyses are shown in Jupyter notebook regression.ipynb.

2. Source files for neural network deployment

Three Python files can be found in src/ folder:

1. server.py - contains logic running local server, loading model and predicting target variables,
2. client.py - contains logic for sending GET and POST requests to local server,
3. neural_network - this folder contains train.py script for training neural network and inference.py implementing NeuralNetwork class for loading and predicting target variables,
4. neural_network/helpers - this folder contains some useful functions and variables for generating training data, preprocessing functions and so on,
5. data - this folder contains trained neural network configs.

NOTE: data/ folder contains three files:

1. neural_network.keras (TRAINED_MODEL variable in helpers.py) - contains trained neural network,
2. pipeline.json (PIPELINE variable in config.py) - contains configs for MinMaxScaler and PolynomialFeatures for features,
3. scaler.json (SCALER variable in config.py) - contains MinMaxScaler configs for target variables.

Usage

All requirements need for running scripts can be found in requirements.txt.

Train the neural network

To train the neural network run:

cd src/
python neural_network/train.py

Use trained neural network configs for predicting targets

Running local server:

Firs run local server:

python server.py

this command start Uvicorn local server at port 5000. Also, server.py contains business logic for handling:

1. GET '/' request - this endpoint returns HTML with this README.md,
2. GET '/func' - this endpoint returns the neural network prediction based on the passed URL parameters,
3. POST '/func' - this endpoint returns the neural network prediction based on the data from JSON request body.

Sending requests

Using python script client.py:

python client.py

next you can choose from three options:

1. home - for receiving home page '/' with GET request,
2. vector - for sending one feature vector with GET request on endpoint 'func/' and receiving prediction as text/plain response, feature vector can be specified as follows:

x1=0.1, x3=-1.1, x2=0.02, x4=0

after what the GET request http://127.0.0.1:5000/func?x1=0.1&x3=-1.1&x2=0.02&x4=0 will be sent, 3. matrix - for sending multiple feature vectors with POST request on endpoint 'func/' and receiving predictions as application/json response, for more details see realization in the client.py, feature matrix can be specified as follows:

[[2.647, 2.144, 0.112, 6.978], [0.4, -2.1, 0.07, -0.9]]

after what the POST request http://127.0.0.1:5000/func will be sent.