radis.io.cdsd module

Parser for CDSD-HITEMP, CDSD-4000 format.

Routine Listing


CDSD-4000 manual

cdsd2df(fname, version='hitemp', cache=True, verbose=True, drop_non_numeric=True, load_wavenum_min=None, load_wavenum_max=None)[source]

Convert a CDSD-HITEMP 1 or CDSD-4000 2 file to a Pandas dataframe.

fname: str

CDSD file name

version: str (‘4000’, ‘hitemp’)

CDSD version

cache: boolean, or ‘regen’

if True, a pandas-readable HDF5 file is generated on first access, and later used. This saves on the datatype cast and conversion and improves performances a lot (but changes in the database are not taken into account). If False, no database is used. If ‘regen’, temp file are reconstructed. Default True.

Other Parameters
  • drop_non_numeric (boolean) – if True, non numeric columns are dropped. This improves performances, but make sure all the columns you need are converted to numeric formats before hand. Default True. Note that if a cache file is loaded it will be left untouched.

  • load_wavenum_min, load_wavenum_max (float) – if not 'None', only load the cached file if it contains data for wavenumbers above/below the specified value. See :py:func`~radis.io.cache_files.load_h5_cache_file`. Default 'None'.


df – dataframe containing all lines and parameters

Return type

pandas Dataframe


CDSD-4000 Database can be downloaded from 3

Performances: I had huge performance trouble with this function, because the files are huge (500k lines) and the format is to special (no space between numbers…) to apply optimized methods such as pandas’s. A line by line reading isn’t so bad, using struct to parse each line. However, we waste typing determining what every line is. I ended up using the fromfiles functions from numpy, not considering n (line return) as a special character anymore, and a second call to numpy to cast the correct format. That ended up being twice as fast.

  • initial: 20s / loop

  • with mmap: worse

  • w/o readline().rstrip(’n’): still 20s

  • numpy fromfiles: 17s

  • no more readline, 2x fromfile 9s

Think about using cache mode too:

  • no cache mode 9s

  • cache mode, first time 22s

  • cache mode, then 2s

Moving to HDF5:

On cdsd_02069_02070 (56 Mb)


cdsd2df(): 9.29 s
cdsd2df(cache=True [old .txt version]): 2.3s
cdsd2df(cache=True [new h5 version, table]): 910ms
cdsd2df(cache=True [new h5 version, fixed]): 125ms


%timeit df.to_hdf("cdsd_02069_02070.h5", "df", format="fixed")  337ms
%timeit df.to_hdf("cdsd_02069_02070.h5", "df", format="table")  1.03s


Note that CDSD-HITEMP is used as the line database for CO2 in HITEMP 2010


HITEMP 2010, Rothman et al., 2010


CDSD-4000 article, Tashkun et al., 2011


CDSD-4000 database

See also