[MNT] full test run with minimal dependencies #4016
Conversation
fkiraly
added
the
maintenance
fkiraly
requested review from
lmmentel,
freddyaboulton,
aiwalter and
GuzalBulatova
as code owners
This PR contains downwards compatibility fixes for the minimal version dependency set currently supported by `sktime`. Because CI was not testing for minimal version dependencies, there have been punctual divergences in compatibility with the minimal dependency set. This PR aims to fix this. PR #4016 introduces a CI element to test on the minimal dependency set, against which the fixes in this PR are developed. Fixes: * old `sklearn` versions require an experimental enablement flag for use of `HistGradientBoostingRegressor`. The fix is to include that flag conditional on `sklearn` version. * the `DateTimeFeatures` transformer depends on `pandas` `DatetimeProperties` which is only available in its current form since `pandas 1.2.0`. The fix is to include this as a dependency requirement in the estimator and isolate the tests conditionally. Depends on: * #4044 as PEP 440 specifiers are introduced in estimator dependencies
…e#4041) This PR contains downwards compatibility fixes for the minimal version dependency set currently supported by `sktime`. Because CI was not testing for minimal version dependencies, there have been punctual divergences in compatibility with the minimal dependency set. This PR aims to fix this. PR sktime#4016 introduces a CI element to test on the minimal dependency set, against which the fixes in this PR are developed. Fixes: * old `sklearn` versions require an experimental enablement flag for use of `HistGradientBoostingRegressor`. The fix is to include that flag conditional on `sklearn` version. * the `DateTimeFeatures` transformer depends on `pandas` `DatetimeProperties` which is only available in its current form since `pandas 1.2.0`. The fix is to include this as a dependency requirement in the estimator and isolate the tests conditionally. Depends on: * sktime#4044 as PEP 440 specifiers are introduced in estimator dependencies
fkiraly
requested review from
achieveordie,
benHeid and
yarnabrina
as code owners
|
any ideas about the strange error in the test-minimaldeps? |
|
This is just a guess: If this is really the problem, then we might need to separate the installation into two commands. First, we need to install numpy. Afterwards, we can install the rest. |
|
yes, I also got that - somehow it seems two This wasn't the case ca 6 months ago when this PR first got opened though, so I also am not sure on what would have changed here. |
|
I guess that the runner has a numpy version installed via apt-get. During the last six months this numpy version was updated. If the updated version now has a new feature which the minimal dependency does not have. Then this would mean that scipy is built with a new feature that is not available in the older numpy version which is installed after building the wheels. |
This adds a full test run with minimal versions and minimal dependencies, under python 3.7 and ubuntu, to test at the lower end of
sktime's support range.Depends on:
statsmodelsandpmdarima#4443