Welcome to Espresso!#
You’ve just come up with a new optimisation algorithm, inversion strategy, or machine learning-based inference framework. Now you want to see how it performs on a real-world problem…
Espresso (Earth science problems for the evaluation of strategies, solvers and optimizers) aims to make this as easy as possible. It provides access to a range of exemplars via a standardized Python interface, including domain-expert–curated datasets and corresponding simulation codes. Swapping from one test problem to the next is just one line of code.
Here’s a simple illustration:
1import numpy as np 2 3# Select the test problem to be used -- change this line to any 4# other test problem and everything else should still work! 5from espresso import SimpleRegression as test_problem 6 7# Create an instance of the test problem class 8tp = test_problem() 9 10# The test problem provides... 11# ... an initial (null) model vector 12model = tp.starting_model 13# ... the ability to compute simulated data for an arbitrary 14# model vector, and the corresponding jacobian (i.e. derivatives 15# of data wrt model parameters) 16predictions, G = tp.forward(model, return_jacobian = True) 17# ... a data vector, which matches the output from `tp.forward()` 18residuals = tp.data - predictions 19# ... and much more! 20 21# Compute a Gauss-Newton model update 22model += np.linalg.solve(G.T.dot(G), G.T.dot(residuals)) 23 24# Compare our result to the answer suggested by Espresso 25print("Our result:", model) 26print("Espresso:", tp.good_model) 27 28# And let's visualise both to see where the differences are: 29my_fig = tp.plot_model(model) 30espresso_fig = tp.plot_model(tp.good_model) 31data_fig = tp.plot_model(tp.data, tp.forward(model))
If this looks interesting, you can:
Espresso is an open-source community effort, currently supported and coordinated by InLab.