In week 3, I began testing my implementation of the chunksize feature. I found that there was an error being raised in the plotting of graphs, due to a library called publib being used. Upon searching for this library, I was amazed to see that it was built by Dr. Erwan! You can have a look at it here

I initially thought nothing of the error, and was convinced that my implementation was correct. Dr Erwan then guided me to an example in the RADIS Example Gallery, in which you can change the library being used. Hence, I dropped publib and just used the default config used in the example, which used Seaborn. I found it worked for me, hence I continued with it.

I ran the chunksize code again, and found out that I was getting some errors regarding array sizes not being equal, during the initialization of a DataFrame in the code. Since Pandas DataFrames require all rows to be equal in size, this was an unavoidable error. I began debugging the error.

It turned out that, when the chunksize loop was running in the code, and optimization was set as a parameter, certain values were being calculated incorrectly. Let me walk you through the broadening calculations of RADIS:

  • When we want to calculate the broadening without chunksize, i.e. over the entire DataFrame, we set parameters like line_profile, wavenumber, abscoeff, and certain other parameters such as line_profile_LDM, wL, wG, wL_dat, wG_dat which are important to the internal computations of RADIS.
  • The wL, wG, wG_dat and wL_dat parameters then set a few more values that are used in the function _apply_lineshape_LDM(), which is the final step in the broadening process.
  • Usually, when we calculate all these constants over the entire DataFrame, they can be arrays with large lengths. Hence while computing them over chunks of the DataFrame, these individually calculated arrays need to be appended to the initialised constants. This was the step I was doing wrong, which I then fixed.

I found that after this correction, the errors I had encountered while plotting the graphs were gone. However, just as soon as one error was resolved, another one occurred :’) The example that I was testing my code in wasn’t even entering the Chunksize loop. I took a long look at the code, and added chunksize as parameters in two very important functions: calc_spectrum() and calc_spectrum_one_molecule(). These are the two main functions invoked whenever we want to calculate a spectrum with RADIS, given the input parameters such as the temperature, pressure, type of molecule, database to refer etc. I then improved the code a bit, fixed the mistakes I had made during my initial implementation, and voila! The example was atleast entering the right part of the code. This didn’t solve all my issues though.

image

This was the output when I ran the code with the exact same conditions, apart from the chunksize parameter, which you can observe was set to 300 for the black coloured lines and None for the red one. The output should’ve been the exact same, but it wasn’t. And so, I got to debugging the error.

It turns out that my implementation logic was flawed. And so, I got to rethinking my approach.