solanovisitor/ForeML

A Python package for time series forecasting using Tensorflow and Keras

ForeML

A package for training and evaluating time series forecasting using Tensorflow.

🧠 Set up the environment

1. Install Poetry
2. Set up the environment:

make activate
make setup

🦾 Install Tensorflow using pip

pip install tensorflow

To install new PyPI packages, run:

poetry add <package-name>

🧙 Basic usage

📁 You will find the configurations for your runs in the config folder.

Change the paths in the main.yaml file to your personal directories that contain your data. For example:

raw:
  path: /home/user/ForeML/data/raw/test_data.csv

📋 The main.yaml file points to the parameters regarding the data processing and model training:

defaults:
  - process: lstm
  - model: lstm
  - forecast: lstm
  - _self_

hypertune: False

If you change the hypertune parameter to True, it will run a tunable model with predefined parameters.

✂️ The process YAML file listed in the main.yaml should look like this:

delimiter: ','
target_index: 5
date_index: 3
n_steps_in: 12
n_steps_out: 8
n_features: 1

• First, select the delimiter for your .csv file. After, you specify your target and datetime column indexes.
• Finally, you can input the number of timesteps your model will be trained on (n_steps_in) and the timesteps it will forecast (n_steps_out).
• If your model have more than one feature, you can specify it in the last parameter.

📌 The model YAML file listed in the main.yaml should look like this:

name: lstm
type: lstm
n_units: 100
activation: relu
dropout: 0.7
epochs: 10
batch_size: 500
validation_split: 0.2
learning_rate: 0.001
lossfunction: msle

You can change the parameters in this file as you desire. Please note some of those won't cause any effect if you choose to hypertune.

🏃‍♂️ When you have all ready and set, you can just run the files you desire:

python3 /src/process.py
python3 /src/train_model.py
python3 /src/forecaster.py

• process.py will return the .csv with processed data in the right format and the X and y ready to serve as input to the model.
• train_model.py will train the model with your input data and return a trained model as selected (specified architecture and hypertuning option in the config files).
• forecaster.py will take your test data, generate predictions on the dataset and compare with actual values, returning a plot for each timestep.