fixes fieldname issues

[benidis]

The field names of a dataset are not checked for consistency across the different modules. If a field changes in the FieldName class, there is no error raised, the old fields are never there in the new datasets and therefore ignored, and the new fields are not processed.

This caused issue #283 in the dataset statistics calculation dataset.calc_stats().

This PR includes the following changes:

• fixes the calc_stats issue (fixes issue The calculation of dataset statistics reads wrong feature fields #283 )
• moves the FieldName class to a separate module to avoid circular dependencies and to increase visibility for potential future changes (it was in the transform package with all the transformations and it did not really belong there)
• adds tests for potential change on the FieldName class so people are warned at least that they should not touch it without a larger refactoring
• adds tests for the calc_stats method

By submitting this pull request, I confirm that you can use, modify, copy, and redistribute this contribution, under the terms of your choice.

[szha]

Job PR-292/3 is complete.
Docs are uploaded to http://gluon-ts-staging.s3-accelerate.dualstack.amazonaws.com/PR-292/3/index.html

[codecov-io]

Codecov Report

Merging #292 into master will increase coverage by 37.12%.
The diff coverage is 97.5%.

@@             Coverage Diff             @@
##           master     #292       +/-   ##
===========================================
+ Coverage   39.43%   76.56%   +37.12%     
===========================================
  Files         141      142        +1     
  Lines        7904     7940       +36     
===========================================
+ Hits         3117     6079     +2962     
+ Misses       4787     1861     -2926

Impacted Files	Coverage Δ
src/gluonts/transform.py	82.89% <ø> (+49.7%)	⬆️
src/gluonts/model/deepstate/_estimator.py	0% <0%> (ø)	⬆️
src/gluonts/model/deep_factor/_estimator.py	0% <0%> (ø)	⬆️
src/gluonts/model/seq2seq/_seq2seq_estimator.py	53.96% <100%> (+0.74%)	⬆️
src/gluonts/model/canonical/_estimator.py	100% <100%> (+44.44%)	⬆️
src/gluonts/dataset/field_names.py	100% <100%> (ø)
src/gluonts/model/gp_forecaster/_estimator.py	100% <100%> (+53.65%)	⬆️
src/gluonts/model/trivial/mean.py	97.29% <100%> (+45.94%)	⬆️
src/gluonts/model/trivial/identity.py	100% <100%> (+35.29%)	⬆️
src/gluonts/dataset/stat.py	96.32% <100%> (+15.24%)	⬆️
... and 97 more

[lostella]

Could you point to the changes in the tests that would catch #283 without this fix? I assume these are in https://github.com/benidis/gluon-ts/blob/af9c7386a4ed1eaf66c44aa04a0b3f7c77f177f9/test/dataset/test_stat.py

[lostella]

Looks good, thanks!

[benidis]

@lostella: The issue with #283 was that the fields in calc_stats were hard coded so when we changed them in the FieldName class the function kept the old ones. It was not possible to use FieldName.{field} in calc_stat due to circular dependency. The fix basically removed the circular dependency and uses ts[FieldName.{field}] to reference a field instead of the previous hardcoded ts["field"] that can be obsolete. This would prevent an issue like #283 .

The tests you referred to were obsolete since they were for the older fields and the fixes there reflect that. I also added a test that hardcodes the current fields. This test aims to ensure that a change in the fields will throw an error and should only be done as a part of a bigger refactoring.