Create a python 3.8 virtual environment into which to install the library
conda create -n ts_automl_test python=3.8clone the github branch to download the installation files
git clone https://github.com/MMurilloGlez/ts-automl.git ts-automl
activate your environment and install wheel
conda activate ts_automl_test
pip install wheelinstall the library using pip install
pip install './ts-automl/dist/ts_automl-0.2.0-py3-none-any.whl'installation will take a while depending on the number of dependencies already on your system
import the prediction models and execute the scripts (changing anything needed for the particular series)
from ts_automl.pipelines import Pipeline- for slow prediction (LSTM + LightGBM + KNN):
model=Pipeline(filename='./example.csv',
type='slow'
freq='15T',
targetcol='TARGET',
datecol='DATES',
sep=',',
decimal='.',
date_format="%d/%m/%Y %H:%M")- for balanced prediction (LightGBM + KNN):
model=Pipeline(filename='./example.csv',
type='balanced'
freq='15T',
targetcol='TARGET',
datecol='DATES',
sep=',',
decimal='.',
date_format="%d/%m/%Y %H:%M")- for fast prediction (KNN Model):
model=Pipeline(filename='./example.csv',
type='slow'
freq='15T',
targetcol='TARGET',
datecol='DATES',
sep=',',
decimal='.',
date_format="%d/%m/%Y %H:%M")- for naïve prediction (mean of last 50 values) (not recommended except for comparison with other models):
from ts_automl.pipelines import naive_prediction
naive_prediction(filename='./example.csv',
freq='15T',
targetcol='TARGET',
datecol='DATES',
sep=',',
decimal='.',
date_format="%d/%m/%Y %H:%M")Once a Pipeline has been initialized with a given model and a given dataset, the next step is to fit the model to the dataset
model.fit()Optionally optimization on the fitted models can be applied using Hyperopt:
model.fit_opt()After fitting the models a prediction can be made using the predict method of the class
model.predict()Details of the fit validation error, the training data or the error relative to the naïve prediction can be found inside the Pipeline class object, by accessing the following parameters:
filename: str
file path of a csv or xls relative to python working path
type: str ['fast','balanced','slow']:
type of model to be fit.
freq: str
Frequency to interpolate the time series data
targetcol: str
Name of target column of data i.e. the values column
datecol: str
Name of the timestamp column of the data
sep: str
Separator for csv data
decimal: str
Decimal point or comma for data
date_format: str
Date format for the timestamps values
features: list of str
Features to be computed for the creation of te ddataset
selected_feat: int
Number of features to preserve from the ones created
plot: bool
Whether or not to plot the training data
window_length: int
Number of lags to include in dataset
horizon: int
Step to calculate in the future for prediction
rolling_window: list of int
Rolling window for which to calculate features
points: int
Number of points in the future to calculate prediction / test size
error: list of str
Metric(s) to calculate for characterizing predictions
rel_metrics: bool
Indicates if error relative to naive is to be calculated.
num_datapoints: int
Number of datapoints from dataset for which to generate predictions
plot_train: bool
Whether to plot the training time series
df: pd.DataFrame
Full generated dtaframe of the input time series with lags and features
y_train, y_test: pd.Series
Time series containing the training and test data for the series respectively
X_Train_selec: pd.DataFrame
DataFrame containing the best features according to the selected_feat argument
best_features: List of str
List containing names of the best features in the dataframe.
naive: List of float
Prediction for test uding the Naive model (if rel_metrics=True)
model: object
Model fitted to training data
opt_runs: int
Number of optimization passes to run on the hyperopt version of a given regressor- for fast, balanced and slow prediction:
points: int
Size of the horizon to predict in the future
window_length: int
Number of points in the past to obtain lags for
rolling_window: list of int
rolling window size to calculate the features
selected_feat: int
Number of features to retain in feature selection.
num_datapoints: int
Size of the training set for the models
plot: bool
Whether to plot or not the results
error: list of str ['mape','mse','r2', 'exp_var']
Error metrics to calculate with respect to the test data
rel_error: bool
Whether or not to calculate the error relative to naive forecast- for naïve prediction:
points: int
Size of the horizon to predict in the future
num_datapoints: int
Size of the training set for the models
plot: bool
Whether to plot or not the results
error: list of str ['mape','mse','r2', 'exp_var']
Error metrics to calculate with respect to the test data to launch the API the module launch the following line of code once the library is installed:
from ts_automl import api
api.run_api()This will automatically launch the api on import, which can be reached at:
localhost:2021
The interactive SwaggerUI version can be reached at:
localhost:2021/docs