I got access to a model that I want to use as a baseline for my research, but (i) it is in MATLAB and I prefer to work in Python, and (ii) it consists of a single 1000-line function with no modularity or use of classes/objects.
My first attempt at converting the code to Python was to use matlab2python, which did an OK job at translating the script, but had issues with 0-indexing (in Python) vs 1-indexing (in Matlab) and I believe it didn’t quite get some of the matrix operations right as numpy operations. As a result, after a bit of debugging, I got the Python code to run without throwing errors (hoorray!), but numerically something was off, since the simulation reached a stop condition five iterations in, when the Matlab version was able to run indefintely (irrespective of random seeds). Debugging this would have been a nightmare, so I needed an alternative approach.
My supervisor suggested breaking down the code into small functional components that I could analyse and understand, including data type use, array sizes, etc., and test each component individually.
To do so, I first refactored the single Matlab function into a handful of smaller functions under a main function, and then translated each to Python. To be able to test them, I ran the program saving the state of the model at various points, with
% model code ...
save('test-data/1.mat')
% more model code ...
save('test-data/2.mat')
% ...and then loaded the model state into a Jupyter notebook to test the functions, using
import numpy as np
import scipy.io
mat = scipy.io.loadmat('test-data/1.mat')
t = mat['t'].item() - 1 # int (-1 because zero-index in Python)
stock = mat['stock'].squeeze() # (1, F) -> (F,)
price = mat['price'][t] # (T, F) -> (F,)
a, b = my_python_function(stock, price)
assert np.all(a == mat['a'].squeeze())
assert np.all(b == mat['b'].squeeze())This allows me to work through the model, ensuring consistency at any arbitrary point in the flow of the function.
For larger functions that were harder to debug in one go, I applied the same technique within the function itself, checking the variables for consistency at various stages.