Add new environment for model backwards compatibility tests

which uses latest dependences except for sklearn (which breaks compatibility otherwise)

tox.ini

@@ -4,7 +4,7 @@
 # and then run "tox" from this directory.to
 
 [tox]
-envlist = py, py_latest_dependencies, docs, report
+envlist = py, py_latest_dependencies, py_backwardscompat, docs, report
 isolated_build = True
 
 [testenv:py]
@@ -33,17 +33,36 @@ deps =
     -rrequirements_geoanalytics.txt
     -rrequirements_tensorflow.txt
 
-# NOTE: notebooks are not executed here again (the build would become very slow) so they are not tested with
-#   latest dependencies. Hopefully, this is not a problem
+# This environment tests the latest dependencies
 [testenv:py_latest_dependencies]
 commands =
     pytest
+; NOTE: notebooks are not executed here again (the build would become very slow) so they are not tested with
+;   latest dependencies. Hopefully, this is not a problem
 deps =
     pytest
 ;   NOTE: For some reason including this into setup.py makes things slow locally, fail silently on gitlab
 ;     and lead to non-ending builds (2.5h+) on github. This dependency should be installable with pip install sensai[torch] though...
     pytorch-lightning>=1.1
-; no pinned requirements included, so relaxed versions created in by setup.py apply
+; no pinned requirements included, so relaxed versions created by setup.py apply
+extras =
+    full
+
+# Special environment to test backward compatibility of persisted models, where we test with the latest
+# dependencies except sklearn, which we must keep at an older version to ensure compatibility
+# (because sklearn does break compatibility!)
+[testenv:py_backwardscompat]
+commands =
+; We test only backward compatibility in this environment, so we do not need to run the full test suite
+    pytest tests/backwardscompat
+deps =
+    pytest
+;   NOTE: For some reason including this into setup.py makes things slow locally, fail silently on gitlab
+;     and lead to non-ending builds (2.5h+) on github. This dependency should be installable with pip install sensai[torch] though...
+    pytorch-lightning>=1.1
+; No pinned requirements included except sklearn, so relaxed versions created by setup.py apply.
+; We use sklearn version 1.0.2, because it is the latest version that is still compatible with the lowest version we support.
+    scikit-learn==1.0.2
 extras =
     full