As far as I understand the results should match to machine precision. We are able to match these kinds of models executed on different platforms using different formulation techniques to machine precision (e-14 or more). In fact, we use this to be able to verify that two independent modelers have created the same model.
The fact that my model can't get the same results with two numerical evaluations of sympy code is concerning. My question in this email though, is how to improve the comparison code so that it doesn't crash for large expressions. Jason moorepants.info +01 530-601-9791 On Thu, Oct 22, 2020 at 6:49 PM Oscar Gustafsson <[email protected]> wrote: > How much do they differ? I checked the issue and the latest result wasn't > that bad, although probably too much to blame floating-point errors. > However, if you push the numerical range into e.g. denormalized numbers I > guess it can simply be that. Not very likely though. Is it still that type > of numerical errors? > > BR Oscar > > Den tors 22 okt. 2020 09:37Jason Moore <[email protected]> skrev: > >> Howdy, >> >> I'm trying to track down a bug in this PR: >> >> https://github.com/pydy/pydy/pull/122 >> >> I have quite large SymPy expressions which I evaluate with floating point >> numbers. I also cse those expressions and evaluate those with the same >> floating point numbers. I'm getting discrepancies in the results. I'd like >> to eliminate the possibility that the cse'd expressions are different from >> the original expression. I wrote this naive function to make the comparison: >> >> def compare_cse(expr): >> >> args = list(expr.free_symbols) >> args.remove(TIME) >> args += me.find_dynamicsymbols(expr) >> args = list(sm.ordered(args)) >> vals = list(np.random.random(len(args))) >> >> eval_expr = sm.lambdify(args, expr) >> res = eval_expr(*vals) >> >> replacements, reduced_expr = sm.cse(expr) >> for repl in replacements: >> var = repl[0] >> sub_expr = repl[1] >> eval_sub_expr = sm.lambdify(args, sub_expr) >> vals.append(eval_sub_expr(*vals)) >> args.append(var) >> eval_reduced_expr = sm.lambdify(args, reduced_expr[0]) >> cse_res = eval_reduced_expr(*vals) >> >> np.testing.assert_allclose(res, cse_res) >> >> This seems to work for small expressions, but for the large expressions >> run this actually kills my Python interpreter after some minutes. >> >> Is there a better way to write this? I need some way to verify that >> numerical evaluation of a cse'd expression gives the same result as >> numerical evaluation of the expression that can actually complete in a >> reasonable amount of time. >> >> Jason >> moorepants.info >> +01 530-601-9791 >> >> -- >> You received this message because you are subscribed to the Google Groups >> "sympy" group. >> To unsubscribe from this group and stop receiving emails from it, send an >> email to [email protected]. >> To view this discussion on the web visit >> https://groups.google.com/d/msgid/sympy/CAP7f1Aivs53AcwyaoyO5%2Bzkn%2BPrX3TfoJJ5X85qQNZ6JRbo5JQ%40mail.gmail.com >> <https://groups.google.com/d/msgid/sympy/CAP7f1Aivs53AcwyaoyO5%2Bzkn%2BPrX3TfoJJ5X85qQNZ6JRbo5JQ%40mail.gmail.com?utm_medium=email&utm_source=footer> >> . >> > -- > You received this message because you are subscribed to the Google Groups > "sympy" group. > To unsubscribe from this group and stop receiving emails from it, send an > email to [email protected]. > To view this discussion on the web visit > https://groups.google.com/d/msgid/sympy/CAFjzj-LQNBFoF7Q4Rb%2BZG3J35-Et9b__SED-DMJvwUYqyKYADQ%40mail.gmail.com > <https://groups.google.com/d/msgid/sympy/CAFjzj-LQNBFoF7Q4Rb%2BZG3J35-Et9b__SED-DMJvwUYqyKYADQ%40mail.gmail.com?utm_medium=email&utm_source=footer> > . > -- You received this message because you are subscribed to the Google Groups "sympy" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. To view this discussion on the web visit https://groups.google.com/d/msgid/sympy/CAP7f1AgdkaC0D69s_ohka_AjDFvKCD%2BMoVKyFhs0Vn2b_DtFkw%40mail.gmail.com.
