AI4Water

A uniform and simplified framework for rapid experimentation with deep leaning and machine learning based models for time series and tabular data. To put into Andrej Karapathy's words

Because deep learning is so empirical, success in it is to a large extent proportional to raw experimental throughput, the ability to babysit a large number of experiments at once, staring at plots and tweaking/re-launching what works. This is necessary, but not sufficient.

The specific purposes of the repository are

• compliment the functionality of keras/pytorch/sklearn by making pre and post-processing easier for time-series prediction/classification problems (also holds true for any tabular data).

• save, load/reload or build models from readable json file. This repository provides a framework to build layered models using python dictionary and with several helper tools which fasten the process of modeling time-series forecasting.

• provide a uniform interface for optimizing hyper-parameters for skopt; sklearn based grid and random; hyperopt based tpe, atpe or optuna based tpe, cmaes etc. See example
  using its application.

• cut short the time to write boilerplate code in developing machine learning based models.

• It should be possible to overwrite/customize any of the functionality of the AI4Water's Model by subclassing the Model. So at the highest level you just need to initiate the Model, and then need fit, predict and view_model methods of Model class, but you can go as low as you could go with tensorflow/keras.

• All the above functionalities should be available without complicating keras implementation.

Installation

An easy way to install ai4water is using pip

pip install ai4water

You can also use GitHub link

python -m pip install git+https://github.com/AtrCheema/AI4Water.git

or using setup file, go to folder where repo is downloaded

python setup.py install

The latest code however (possibly with fewer bugs and more features) can be installed from dev branch instead

python -m pip install git+https://github.com/AtrCheema/AI4Water.git@dev

To install the latest branch (dev) with all requirements use the following command

python -m pip install "AI4Water[all] @ git+https://github.com/AtrCheema/AI4Water.git@dev"

installation options

all keyword will install all the dependencies. You can choose the dependencies of particular sub-module by using the specific keyword. Following keywords are available

• hpo if you want hyperparameter optimization
• post_process if you want postprocessing
• exp for experiments sub-module

Sub-modules

AI4Water consists of several submodules, each of wich responsible for a specific tasks. The modules are also liked with each other. For understanding sub-module structure of ai4water, see this article

How to use

Build a Model by providing all the arguments to initiate it.

```python from ai4water import Model from ai4water.models import MLP from ai4water.datasets import mgphotodegradation data, * = mg_photodegradation(encoding="le")

model = Model( # define the model/algorithm model=MLP(units=24, activation="relu", dropout=0.2), # columns in data file to be used as input inputfeatures=data.columns.tolist()[0:-1], # columns in csv file to be used as output outputfeatures=data.columns.tolist()[-1:], lr=0.001, # learning rate batch_size=8, # batch size epochs=500, # number of epochs to train the neural network patience=50, # used for early stopping ) ```

Train the model by calling the fit() method python history = model.fit(data=data)

After training, we can make predictions from it on test/training data python prediction = model.predict_on_test_data(data=data)

The model object returned from initiating AI4Water's Model is same as that of Keras' Model We can verify it by checking its type python import tensorflow as tf isinstance(model, tf.keras.Model) # True

Using your own pre-processed data

You can use your own pre-processed data without using any of pre-processing tools of AI4Water. You will need to provide input output paris to data argument to fit and/or predict methods. ```python import numpy as np from ai4water import Model # import any of the above model from ai4water.models import LSTM

batch_size = 16 lookback = 15 inputs = ['dummy1', 'dummy2', 'dummy3', 'dummy4', 'dummy5'] # just dummy names for plotting and saving results. outputs=['DummyTarget']

model = Model( model = LSTM(units=64), batchsize=batchsize, tsargs={'lookback':lookback}, inputfeatures=inputs, outputfeatures=outputs, lr=0.001 ) x = np.random.random((batchsize10, lookback, len(inputs))) y = np.random.random((batch_size10, len(outputs)))

model.fit(x=x,y=y)

```

using for scikit-learn/xgboost/lgbm/catboost based models

The repository can also be used for machine learning based models such as scikit-learn/xgboost based models for both classification and regression problems by making use of model keyword arguments in Model function. However, integration of ML based models is not complete yet. ```python from ai4water import Model from ai4water.datasets import busan_beach

data = busan_beach() # path for data file

model = Model( # columns in data to be used as input inputfeatures=['tidecm', 'wattempc', 'salpsu', 'relhum', 'pcpmm'], outputfeatures = ['tetxcoppml'], # columns in data file to be used as input seed=1872, valfraction=0.0, split_random=True, # any regressor from https://scikit-learn.org/stable/modules/classes.html model={"RandomForestRegressor": {}}, # set any of regressor's parameters. e.g. for RandomForestRegressor above used, # some of the paramters are https://scikit-learn.org/stable/modules/generated/sklearn.ensemble.RandomForestRegressor.html#sklearn.ensemble.RandomForestRegressor )

history = model.fit(data=data)

model.predictontest_data(data=data) ```

Hyperparameter optimization

For hyperparameter optimization, replace the actual values of hyperparameters with the space. ```python

from ai4water.functional import Model from ai4water.datasets import MtropicsLaos from ai4water.hyperopt import Real, Integer

data = MtropicsLaos().makeregression(lookbacksteps=1)

model = Model( model = {"RandomForestRegressor": { "nestimators": Integer(low=5, high=30, name='nestimators', numsamples=10), "maxleafnodes": Integer(low=2, high=30, prior='log', name='maxleafnodes', numsamples=10), "minweightfractionleaf": Real(low=0.0, high=0.5, name='minweightfractionleaf', numsamples=10), "maxdepth": Integer(low=2, high=10, name='maxdepth', numsamples=10), "minsamplessplit": Integer(low=2, high=10, name='minsamplessplit', numsamples=10), "minsamplesleaf": Integer(low=1, high=5, name='minsamplesleaf', numsamples=10), }}, inputfeatures=data.columns.tolist()[0:-1], outputfeatures=data.columns.tolist()[-1:], crossvalidator = {"KFold": {"nsplits": 5}}, xtransformation="zscore", ytransformation="log", )

First check the performance on test data with default parameters

model.fitonalltrainingdata(data=data) print(model.evaluateontestdata(data=data, metrics=["r2score", "r2"]))

optimize the hyperparameters

optimizer = model.optimizehyperparameters( algorithm = "bayes", # you can choose between random, grid or tpe data=data, numiterations=60, )

Now check the performance on test data with default parameters

print(model.evaluateontestdata(data=data, metrics=["r2score", "r2"])) ```

Running the above code will optimize the hyperparameters and generate following figures

Experiments

The experiments module is for comparison of multiple models on a single data or for comparison of one model under different conditions.

```python from ai4water.datasets import busan_beach from ai4water.experiments import MLRegressionExperiments

data = busan_beach()

comparisons = MLRegressionExperiments( inputfeatures=data.columns.tolist()[0:-1], outputfeatures=data.columns.tolist()[-1:], split_random=True )

train all the available machine learning models

comparisons.fit(data=data)

Compare R2 of models

bestmodels = comparisons.compareerrors( 'r2', data=data, cutofftype='greater', cutoffval=0.1, figsize=(8, 9), colors=['salmon', 'cadetblue'] )

Compare model performance using Taylor diagram

_ = comparisons.taylorplot( data=data, figsize=(5, 9), exclude=["DummyRegressor", "XGBRFRegressor", "SGDRegressor", "KernelRidge", "PoissonRegressor"], legkws={'facecolor': 'white', 'edgecolor': 'black','bboxtoanchor':(2.0, 0.9), 'fontsize': 10, 'labelspacing': 1.0, 'ncol': 2 }, ) ```

For more comprehensive and detailed examples see

Disclaimer

The library is still under development. Fundamental changes are expected without prior notice or without regard of backward compatability.

Related

sktime: A Unified Interface for Machine Learning with Time Series

Seglearn: A Python Package for Learning Sequences and Time Series

Pastas: Open Source Software for the Analysis of Groundwater Time Series

Time Series FeatuRe Extraction on basis of Scalable Hypothesis tests (tsfresh -- A Python package)

MLAir

pyts: A Python Package for Time Series Classification

Tslearn, A Machine Learning Toolkit for Time Series Data

TSFEL: Time Series Feature Extraction Library

catch22

vest

pyunicorn (Unified Complex Network and RecurreNce analysis toolbox

TSFuse Python package for automatically constructing features from multi-view time series data

Catalyst

tsai - A state-of-the-art deep learning library for time series and sequential data