Design and Analysis of Computational Experiments as python toolbox.
This project is an adaptation from the work of Hans Bruun Nielsen, SÃ¸ren Nymand and Lophaven Jacob SÃ¸ndergaard.
This is a implementation that relies heavily on linear algebra solvers (least-squares solvers, Cholesky and QR decompositions, etc.). Therefore, it is strongly advised that your numpy library be integrated to a BLAS library (e.g.: Intel-MKL, OpenBLAS, ATLAS, etc.) in order to attain satisfactory performances of calculation.
For the sake of convenience, Anaconda handles the gritty details of how to combine numpy and those libraries natively.
To install through PyPi Repository:
pip install pydace
To install via conda
conda install -c felipes21 pydace
Example with dace model
import numpy as np import scipy.io as sio from pydace import Dace import matplotlib.pyplot as plt # Load the training and validation data. (Here we are using a file from the # github repo located in the folder pydace\tests with the name # 'doe_final_infill_mat' mat_contents = sio.loadmat('doe_final_infill.mat') design_data = mat_contents['MV'] # design sites observed_data = mat_contents['CV'] # experiment results # define the hyperparameters bounds and initial estimate theta0 = 1 * np.ones((design_data.shape,)) lob = 1e-5 * np.ones(theta0.shape) upb = 1e5 * np.ones(theta0.shape) # select the training and validation data design_val = design_data[:99, :] observed_val = observed_data[:99, :] design_train = design_data[100:, :] observed_train = observed_data[100:, :] # build the univariate kriging models with a first order polynomial # regression and a gaussian regression model observed_prediction = np.empty(observed_val.shape) for j in np.arange(design_data.shape): # initialize the dace object dace_obj = Dace('poly1', 'corrgauss', optimizer='boxmin') # fit the training data using the default hyperparameter optimizer dace_obj.fit(design_train, observed_train[:, j], theta0, lob, upb) # predict the validation data observed_prediction[:, [j]], *_ = dace_obj.predict(design_val) # labels for the observed data var_labels = ['L/F', 'V/F', 'xD', 'xB', 'J', 'QR'] # plot the validation data for var in np.arange(design_data.shape): plt.figure(var + 1) plt.plot(observed_val[:, var], observed_prediction[:,var], 'b+') plt.xlabel(var_labels[var] + ' - Observed') plt.ylabel(var_labels[var] + ' - Kriging Prediction') plt.show()
Example of design of experiment data generation
It is also possible to generate design of experiment data with a variation reduction technique called Latin Hypercube Sampling (LHS) that is already implemented in this toolbox.
Lets say we have a 4-th dimensional problem (i.e. 4 design/input variables). They are defined by the following bounds.
If we want to build a latin hypercube within these bounds we would do the following:
import numpy as np from pydace.aux_functions import lhsdesign lb = np.array([8.5, 0., 102., 0.]) ub = np.array([20., 100., 400., 400.]) lhs = lhsdesign(53, lb, ub, include_vertices=False)
Contact/Talk to me
My e-email is email@example.com. Feel free to contact me anytime, or just nag me if I'm being lazy.
Download the file for your platform. If you're not sure which to choose, learn more about installing packages.
|Filename, size||File type||Python version||Upload date||Hashes|
|Filename, size pydace-0.1.3-py3-none-any.whl (18.2 kB)||File type Wheel||Python version py3||Upload date||Hashes View|
|Filename, size pydace-0.1.3.tar.gz (13.9 kB)||File type Source||Python version None||Upload date||Hashes View|