Releases: EducationalTestingService/skll
Releases · EducationalTestingService/skll
Version 0.22.4
Fix missing import sys
in run_experiment.py
Version 0.22.3
Very minor bug fix release. Changes are:
main
functions for all utility scripts now take optional argument lists to make unit testing simpler (and not require subprocesses).- Fix another bug that was causing missing "ablated features" lists in summary files.
Version 0.22.2
Fix crash with filter_megam
and join_megam
due to references to old API.
Version 0.22.1
Minor bug fix release. Changes are:
- Switch to joblib.dump and joblib.load for serialization (should fix #94)
- Switch to using official drmaa-python release now that it's updated on PyPI
- Fix issue where training examples were being loaded for pre-trained models (#95)
- Change to using
entry_points
to generate scripts instead ofscripts
insetup.py
, and utilities are now in a sub-package.
Version 0.22.0
This release features mostly bug fixes, but also includes a few minor features:
- Change license to BSD 3 clause. Now any of our code could be added back into scikit-learn without licensing issues.
- Add gamma to default paramater search grid for SVC (#84).
- Add
--verbose
flag torun_experiment
to simplify debugging. - Add support for wheel packaging.
- Fixed bug in
_write_summary_file
that prevented writing of summary files for--ablation_all
experiments. - Fixed SVR kernel string type issue (#87).
- Fixed
fit_intercept
default value issue (#88). - Fixed incorrect error message (#86)
- Tweaked
.travis.yml
to make builds a little faster.
Version 0.21.0
- Added support for
ElasticNet
,Lasso
, andLinearRegression
learners. - Reorganized examples, and created new example based on the Kaggle
Titanic data set. - Added ability to easily create multiple files at once when using
write_feature_file
. (#80) - Added support for the
.ndj
file extension for new-line delimited JSON
files. It's the same format as.jsonlines
, just with a different name. - Added support for comments and skipping blank lines in
.jsonlines
files. - Made some efficiency tweaks when creating logging messages.
- Made labels in
.results
files a little clearer for objective function
scores. - Fixed some misleading error messages.
- Fixed issue with backward-compatibility unit test in Python 2.7.
- Fixed issue where predict mode required data to already be labelled.
Version 0.20.0
- Refactored
experiments
module to remove unnecessary child processes,
and greatly simplify ablation code. This should fix issues #73 and #49. - Deprecated
run_ablation
function, as its functionality has been folded
intorun_configuration
. - Removed ability to run multiple configuration files in parallel, since
this lead to too many processes being created most of the time. - Added ability to run multiple ablation experiments from the same
configuration file by adding support for multiple featuresets. - Added
min_feature_count
value to results files, which fixes #62. - Added more informative error messages when we run out of memory while
converting things to dense. They now say why something was converted to
dense in the first place. - Added option to
skll_convert
for creating ARFF files that can be used
for regression in Weka. Previously, files would always contain non-numeric
labels, which would not work with Weka. - Added ability to name relation in output ARFF files with
skll_convert
. - Added
class_map
setting for collapsing multiple classes into one
(or just renaming them). See the
run_experiment documentation for details. - Added warning when using
SVC
withprobability
flag set (#2). - Made logging much less verbose by default and switched to using
QueueHandler
andQueueListener
instances when dealing with
multiple processes/threads to prevent deadlocks (#75). - Added simple no-crash unit test for all learners. We check results with
some, but not all. (#63)
Version 0.19.0
- Added support for running ablation experiments with all combinations of
features (instead of just holding out one feature at a time) via
run_experiment --ablation_all
. As a result, we've also changed the
names of theablated_feature
column in result summary files to
ablated_features
. - Added ARFF and CSV file support across the board. As a result, all
instances of the parametertsv_label
have now been replaced with
label_col
. - Fixed issue #71.
- Fixed process leak that was causing sporadic issues.
- Removed
arff_to_megam
,csv_to_megam
,megan_to_arff
, and
megam_to_csv
because they are all superseded by ARFF and CSV support
inskll_convert
. - Switched to using Anaconda for installing Atlas.
- Switched back to http://skll.readthedocs.org
URLs for documentation, now that readthedocs/readthedocs.org#456 has been fixed.
Version 0.18.1
- Updated
generate_predictions
to use latest API. - Switched to using multiprocessing-compatible logging. This should fix some
intermittent deadlocks. - Switched to using miniconda for install Python on Travis-CI.
Version 0.18.0
- Fixed crash when
modelpath
is blank andtask
is not
cross_validate
. - Fixed crash with
convert_examples
when given a generator. - Refactored
skll.data
's private_*_dict_iter
functions to be
classes to reduce code duplication.