fraserwg / elepyant Goto Github PK

A command line interface allowing operations on NetCDF files (but not xarray objects) could be useful. I'm thinking something that would work along the lines of

ncround -i /path/to/in/file.nc -o /path/to/out/file.nc --UVEL 2 --VVEL 2 --WVEL 6

AFAIK netCDFs default compression algorithm is zlib. Zlib is not bad, but I highly recommend to use Zstandard as I achieve much better compression factors with it. In general benchmarks, Zstd is better than Zlib in every aspect. Could you check whether there is a way that the netCDF python implementation supports it?

Just FYI, the unidata netCDF4 documentation contains the following section: https://unidata.github.io/netcdf4-python/netCDF4/index.html#section9 which has a similar approach to the one of elePyant.

Be aware that rounding in decimal representation does is general not round in binary representation too, as the former is not an integer power of the latter. Currently, you seem to use np.around(x,decimals=d), but this does not yield zeroed trailing bits.

import struct 
def binary(num): 
    return ''.join('{:0>8b}'.format(c) for c in struct.pack('!f', num))

In [12]: binary(np.around(np.pi,decimals=2))                                    
Out[12]: '01000000010010001111010111000011'

In [13]: binary(np.around(np.pi,decimals=3))                                    
Out[13]: '01000000010010010001011010000111'

In [14]: binary(np.around(np.pi,decimals=4))                                    
Out[14]: '01000000010010010000111111111001'

Note that you can convert significant digits (i.e. decimal) to significant bits via

"""Number of significant bits `nsb` given the number of significant digits `nsd`."""
nsb(nsd::Integer) = Integer(ceil(log(10)/log(2)*nsd))

which is 7 significant bits to guarantee 2 significant (decimal) digits, 10 for 3, 14 for 3, etc. So what you want is a rounding function that does this:

julia> bitstring(round(Float32(π),sigdigits=7,base=2))
"01000000010010100000000000000000"

julia> bitstring(round(Float32(π),sigdigits=10,base=2))
"01000000010010010000000000000000"

julia> bitstring(round(Float32(π),sigdigits=14,base=2))
"01000000010010010001000000000000"

Question is obviously why the compression still works well? I assume this is because if you round two numbers x,y that are fairly close to each other, to the same decimal places, you get the same bitpatterns in the significant:

In [15]: binary(np.around(np.pi,decimals=2))                                    
Out[15]: '01000000010010001111010111000011'

In [16]: binary(np.around(np.pi*1.001,decimals=2))                              
Out[16]: '01000000010010001111010111000011'

Which is presumably also somewhat compressable. But I guess 0's everywhere are more compressible than these seemingly random, but repeating bitpatterns.