Understanding ai4water

Most of the technical details of ai4water are available in the ai4water paper . Here we provide the most important technical highlights.

Model Definition

In ai4water, the user can define the machine learning model using the model argument in Model class. For scikit-learn, lightgbm, catboost and xgboost based models, the method is illustrated in quick start. For neural network based models, there are two ways of model definition. For most basic usage cases, the user can adopt the functions available in functional interface for neural network architectures. However, the user can also define models using json style dictionaries. This method of definition style for tensorflow and pytorch is given below

• declarative model definition for tensorflow
  • multi-layer perceptron
  • LSTM based model
  • 1d CNN based model
  • LSTM -> CNN based model
  • ConvLSTM based model
  • CNN -> LSTM
  • LSTM based auto-encoder
  • TCN layer
  • Multiple Inputs
  • Multiple Output Layers
  • Additional call args
  • lambda layers
  • Custom Layers
  • Activation layers
• declarative model definition for pytorch
  • multi-layer perceptron
  • LSTM based model
• declarative model definition for classical ML models
  • sklearn based models
  • catboost, lgbm and xgboost based models
  • Specifying hyperparameters of models
  • custom models

Subclassing vs Functional API

The Model subclassing vs functional API article describes the difference between model subclassing API and functional API in ai4water

Module Structure

Module Hierarchy

At the core of ai4water is the Model class. The HyperOpt class is above it because it it calls the Model class during each of hyperparameter iteration. The Experiments class is further up, because it involves hyperparameter optimization for each of the model considered.

Module linkage

The information flow in the ai4water framework is illustrated in following diagram. The sub-modules on the left are about model building, training, hyperaparameter optimization and comparison of models. On the other hand, the sub-modules on the left are related to data-preprocessing, data-preparation, post-processing of results and visualization.