Week 12 - Final implementation into RADIS, along with a plethora of illustrative examples.
1. Implementation of modules into RADIS
Finally, after being approved by Mr. Erwan, I can implement all of my modules, developed separately in my repo RADIS-Spectrum-Fitting-Benchmark, into RADIS codebase. The implementation features new_fitting.py, the new fitting module that stores all the fitting functions and associated models, whose performance confirmed after a bunch of user-testing cases.
2. Accompanied illustrative examples
They are gallery examples that are added into radis/examples, serving as illustrative scripts for my new fitting module:
(i) plot_newfitting_Tgas.py
The most basic example of how to use new fitting module, including the formats and so on.
(ii) plot_newfitting_Trot-Tvib-molfrac.py
The real-life fitting case provided by Mr. Corentin, featuring non-LTE CO spectrum in which we will fit Trot, Tvib and mole_fraction.
(iii) plot_newfitting_Tgas-molfrac.py
Mr. Minou’s user-testing case of CO absorbance spectrum near 2011 cm-1. This case features spectrum extraction from a .mat MATLAB file. Originally, this file was 1.2 MB, quite large to be added to RADIS. Thus, I removed all fields unnecessary for spectrum generation, and now it only has around 400 kB left.
(iv) plot_newfitting_comparison_oldnew.py
Performance comparison example between current 1-temperature fitting and my new fitting module, under exactly the same ground-truths and settings. The benchmark result shows that, under exactly the same conditions, the new best fitted value differ 0.45% from the old one (1464.1 K from the old 1457.5 K). New fitting module requires half as many iterations as the old one and hence faster, with much smaller residual. In detail:
==================== PERFORMANCE COMPARISON BETWEEN 2 FITTING METHODS ====================
1. LAST RESIDUAL
- Old 1T fitting example: 0.002730027027336094
- New fitting module: 0.0005174179496843629
2. NUMBER OF FITTING LOOPS
- Old 1T fitting example: 32 loops
- New fitting module: 16 loops
3. TOTAL TIME TAKEN (s)
- Old 1T fitting example: 4.881942987442017 s
- New fitting module: 2.7344970703125 s
==========================================================================================I’m not sure this superiority will persist in all cases, but even so, I believe the value of my module still lies in its practical and easy to use/apply.
(v) plot_newfitting_comparison_methods.py
A benchmarking example which compares performance between different LMFIT fitting algorithms. It measures their last residual (for accuracy evaluation) and number of iterations (for robustness evaluation). The benchmark result shows that, under exactly the same conditions, leastsq and lbfgsb work best, with leastsq good at accuracy, while lbfgsb good at speed (and theoretically, memory requirement). Thus, I set leastsq as default method for the module, but also encourage users to switch to lbfgsb in case things turn sour.
======================== BENCHMARKING RESULT ========================
|| METHOD || RESIDUAL || LOOPS ||
||---------------------------||---------------------------||-------||
|| leastsq || 1.4739494411950239e-07 || 24 ||
|| least_squares || 1.2170348021620847e-05 || 1 ||
|| differential_evolution || 1.4739855740762716e-07 || 151 ||
|| brute || 1.2287258962300115e-06 || 20 ||
|| basinhopping || 7.930954059631543e-06 || 151 ||
|| ampgo || 4.105104127826488e-07 || 151 ||
|| nelder || 1.4739942144030064e-07 || 30 ||
|| lbfgsb || 1.4739494411955646e-07 || 28 ||
|| powell || 1.473949441200994e-07 || 43 ||
|| cg || 1.4776331905574135e-07 || 30 ||
|| cobyla || 1.1524288718226295e-05 || 21 ||
|| bfgs || 1.4776331905574135e-07 || 30 ||
|| tnc || 1.4740393115424221e-07 || 28 ||
|| trust-constr || 1.4739494411948182e-07 || 26 ||
|| slsqp || 1.2170348021620847e-05 || 2 ||
|| shgo || 1.0507694502308952e-05 || 5 ||
|| dual_annealing || 1.5455930218501237e-05 || 151 ||
||---------------------------||---------------------------||-------||
