table_builder_io defines a minimal API for reading CSVs downloaded from ABS TableBuilder without manual editing of the raw data.
It serves to avoid/ replace bespoke ways of preparing table builder data e.g.
- Cleaning the header and footer data manually
- Trying to be clever with magic arguments to pandas read_csv skipheader and skipfooter that may or may not need to be adjusted every time
- Realising your magic arguments to skipheader and skipfooter are only part of the problem when you have defined wafers and resort to manually cleaning CSVs
- Hacky flattening of row level index labels and column labels into a single set of column headers that definitely works every time
The recommendation is to install table_builder_io with pip,
python -m pip install table_builder_io
Besides python itself, the only dependency for table_builder_io is pandas. It has been tested on pandas 1.1.x but
does not use any special functionality, so may work on older versions as well. The light requirements mean that pip
installing into a conda environment after pandas has already been installed should be relatively safe.
To install for local development in your virtual environemnt tool of choice, active the environment then,
git clone [email protected]:vlc/table_builder_io.git
cd table_builder_io
python -m pip install -e .table_builder_io requires python >=3.6 as it uses f-strings and standard library type hints. It has been
explicitly tested on Python 3.6, 3.8 and 3.10.
Lets say you have a table builder file that looks something like this
Australian Bureau of Statistics
"2016 Census - Counting Persons, Place of Enumeration (MB)"
"SEXP Sex and FMGF - 1 Digit Level by STATE"
"Counting: Persons Location on Census Night"
Filters:
"Default Summation","Persons Location on Census Night"
,"STATE","New South Wales","Victoria","Queensland","South Australia","Western Australia","Tasmania","Northern Territory","Australian Capital Territory","Other Territories","Total",
"SEXP Sex","FMGF - 1 Digit Level",
"Male","Couple family with grandchildren",20710,12307,14166,4066,7151,1435,1926,702,10,62463,
,"Lone grandparent",10617,6127,6351,2085,3369,671,1486,302,13,31019,
,"Not applicable",3692904,2892405,2362975,817562,1250871,244515,132578,196526,2853,11593188,
"Female","Couple family with grandchildren",19712,11688,13790,3723,7000,1364,1820,723,10,59827,
,"Lone grandparent",15730,9441,9534,3135,5042,961,1827,462,13,46152,
,"Not applicable",3805273,3014087,2437722,844224,1244420,255233,119476,201935,2410,11924766,
"Total","Couple family with grandchildren",40422,23999,27950,7780,14154,2793,3742,1423,21,122290,
,"Lone grandparent",26351,15572,15892,5219,8409,1629,3317,761,27,77165,
,"Not applicable",7498170,5906487,4800703,1661786,2495294,499744,252053,398458,5265,23517955,
"Data Source: Census of Population and Housing, 2016, TableBuilder"
"INFO","Cells in this table have been randomly adjusted to avoid the release of confidential data. No reliance should be placed on small cells."
"Copyright Commonwealth of Australia, 2018, see abs.gov.au/copyright"
"ABS data licensed under Creative Commons, see abs.gov.au/ccby"
table_builder_io (for now) defines a single public class TableBuilderReader which is used like so
In[1]: from table_builder_io import TableBuilderReader
In[2]: reader = TableBuilderReader.from_file("test/mini_testfile.csv")
In[3]: df = reader.read_table(as_index=True)
In[4]: df.iloc[:, :4].head()
Out[4]:
STATE New South Wales Victoria Queensland South Australia
SEXP Sex FMGF - 1 Digit Level
Male Couple family with grandchildren 20710 12307 14166 4066
Lone grandparent 10617 6127 6351 2085
Not applicable 3692904 2892405 2362975 817562
Female Couple family with grandchildren 19712 11688 13790 3723
Lone grandparent 15730 9441 9534 3135
# Or alternatively as a flat dataframe
In[5]: df2 = reader.read(as_index=False)
In[6]: df2.iloc[:, :6].head()
Out[6]:
SEXP Sex FMGF - 1 Digit Level New South Wales Victoria Queensland South Australia
0 Male Couple family with grandchildren 20710 12307 14166 4066
1 Male Lone grandparent 10617 6127 6351 2085
2 Male Not applicable 3692904 2892405 2362975 817562
3 Female Couple family with grandchildren 19712 11688 13790 3723
4 Female Lone grandparent 15730 9441 9534 3135
Int[7]: reader.read_header_metadata()
Out[7]:
HeaderInfo(authority='Australian Bureau of Statistics',
dataset='2016 Census - Counting Persons, Place of Enumeration (MB)',
variables='SEXP Sex and FMGF - 1 Digit Level by STATE',
counting='Persons Location on Census Night',
filters='',
summation='Persons Location on Census Night')For more examples, see Examples on Github
Currently should support
- CSVs with multilevel / hierarchical row headers (as in the example above)
- CSVs with multilevel / hierarchical row headers (e.g. the transpose of the above data)
- Wafers: TableBuilderReader.read returns a
Dict[str, pd.DataFrame]where the keys are the wafer names if wafers are found - Currently only intending to support CSV format from Table Builder
- Support for datasets with filters
- support for NVS TableBuilder headers/ footers
- extraction of header/ footer metadata in a retrievable way
- Standard utils after loading the table into memory
-
Not a super optimised implementation, need to be able to read everything into memory
-
File is scanned twice - once to look for header/ footer/ wafers and then to read the csvs
-
First scan is python, second scan is pandas csv reader (c engine)
-
So maybe not the best if you have data sizes near the cell limit
-
Internals are still messy because I haven't cleaned them up yet, waiting since I expect stuff to break
- My colleague @edavisau for articulating that the skipfooter/ skipheader way of reading Table Builder data is error prone, less than optimal and for pointing out that in R, people already have nice things for reading Table Builder data.
- Prior Art (have not investigated the implementation of either):