Summary of topics:
- How to organise plotting large numbers of plots from heterogenous models in python
- Layout of subplots with matplotlib in python
- Fast percentile calculation in python
Organising workflows in python
cf-xarray is a useful library for writing general code for a heterogenous set of models that might not share the same names for things like dimensions (e.g. lat/latitude/LAT/) and variables. cf-xarray has code for inferring this information and allows you to refer to them in a general way. It is available in the conda environments and the documentation is here:
The python3 pathlib library is an object oriented path manipulation library that makes working with paths a lot simpler and cleaner, e.g. when opening data files, and saving processed output or plots to disk:
As far as how to organise plotting of a large number of different plots from a range of models, there are a range of data structures that might suit this purpose, and it comes down to the specifics of what needs to be done and personal preference, but some options are:
For more information, this is a pretty comprehensive write up of some of the commonly used data structures in python https://realpython.com/python-data-structures/
Subplots in matplotlib
Scott wrote a blog showing a sophisticated use of subplot, but also has some tips for organising the plots by saving references to each in a dictionary named for the plot type: https://climate-cms.org/2018/04/27/subplots.html
Unrelated to the original topics, but some of the attendees didn't know it was possible to connect a jupyter notebook directly to gadi compute nodes, which is useful for anyone who must access data on /scratch, or have workloads that are currently too onerous for VDI or ood.nci.org.au (though ood is designed to cater for larger jobs than VDI). This is all covered on our wiki http://climate-cms.wikis.unsw.edu.au/Running_Jupyter_Notebook#On_Gadi which also covers creating symbolic links to access files in other locations on the file system, e.g. /g/data
Fast percentile calculation in python
The question was on how to calculate a percentile climatology where we calculate the 90th percentile for each day of the year considering the values for the 31 days surrounding each date.
The notebook illustrates:
- the use of rolling() and construct() to build a DataArray with the 31 days windows for each day of the timeseries.
- the use of groupby() to do calculations for each day of the year.
- the use of quantile() to calculate percentiles on DataArrays.
- the use of load() if a function complains about chunking.
- the use of a list and xarray.concat() to create a DataArray of results.
Calculating a percentile climatology where for each day in the year, that day plus 15 days either side are gathered together over all years to make a (dayofyear, lat, lon) dataset of 90th percentiles from each sample Rolling will get a 31 day sample centred around the day of interest (at the start and end of the dataset the values will be NAN) tas.rolling(time=31, center=True) Construct will take the rolling samples and convert them to a new dimension, so the data now has dimensions (time, lat, lon, window) tas.rolling(time=31, center=True).construct(time='window') Groupby will collect all the equivalent days of the year (remember the 'time' axis is the centre of the sample) (tas.rolling(time=31, center=True)
.groupby(time='dayofyear')) Normally you can just add a reduction operation (e.g. mean) to the end here, but that doesn't work with percentiles in this case. Instead do a loop: doy_pct = 
for doy, sample in (tas
doy_pct.append(sample.load().quantile(0.9, dim=['time', 'window'])) xarray.concat(doy_pct, dim='dayofyear') See that we've called `.load()` inside the loop, which avoids a Dask chunking error when doing a percentile. Try it with just a single point to help understand how this is working Note the use of the list to gather the percentiles arrays for each day, then concatenate along a new dimension 'dayofyear'.