fix assess in DynamicTimeSeriesBucket when max_total_size is used

[JanekEbb]

No description provided.

[codecov-commenter]

Codecov Report

All modified and coverable lines are covered by tests ✅

Comparison is base (545ae34) 76.49% compared to head (84b1b8b) 76.49%.

❗ Your organization needs to install the Codecov GitHub app to enable full functionality.

Additional details and impacted files

@@           Coverage Diff           @@
##           master      #64   +/-   ##
=======================================
  Coverage   76.49%   76.49%           
=======================================
  Files           5        5           
  Lines        1736     1736           
=======================================
  Hits         1328     1328           
  Misses        408      408           

☔ View full report in Codecov by Sentry.
📢 Have feedback on the report? Share it here.

[boeddeker]

Can you explain, when and why the code failed? The maybe_append has an assert, that is_completed is False, before calling assess.

Could you add a small test, that ensures, that the code now works as intended?

[JanekEbb]

so far a padded batch could exceed max_total_size when a newly accepted example has the maximum length in the batch. I now added a test case for this. Without the update in assess, the output would be [[6, 7, 9], [5, 6, 3], [7, 4]], where the first batch would exceed max_total_size=21 after padding as 3*9=27.

[boeddeker]

I overlooked that assess has max(self.max_len, seq_len), while is_completed has only seq_len.