26 lines
1.1 KiB
Makefile
26 lines
1.1 KiB
Makefile
# # Prune poorly filled or uninteresting columns
|
|
# languages.pruned1.tsv: languages.original.tsv
|
|
# cut --complement -f1,16-18,22,23,39,40,50,52,53,58-61,64-68,70-75,79,82,84-86,100,103-104,107-109,112,114,116-118,121-136,142,149,150,157,162,163,163,164,165,176,177,185,186,195, $< >$@
|
|
|
|
# Allow user to mark in which columns is interested
|
|
columns.user.tsv: columns.pruned.tsv
|
|
awk 'NR==1{ printf("Keep\t"); print } NR>1{printf("n\t"); print}' $< >$@
|
|
|
|
# Prune columns that are not needed to create specification
|
|
columns.pruned.tsv: columns.original.tsv
|
|
cut --complement -f3,4,5,7,8,9,10 $< >$@
|
|
|
|
# Change data to TSV format since it is easier to process using standard UNIX tools
|
|
%.tsv: %.csv
|
|
go run ./csv2tsv.go <$< >$@
|
|
|
|
# Check while downloading that file is as expected.
|
|
# Otherwise automatic filter mechanism wouldn't work.
|
|
# If hashes differ then user of this repo must migrate columns.user.tsv to a new format
|
|
columns.original.csv:
|
|
wget 'https://pldb.com/columns.csv' -O $@
|
|
sha256sum -c checksums.sha256
|
|
|
|
languages.original.csv: columns.original.csv
|
|
wget 'https://pldb.com/languages.csv' -O $@
|