tests: Simplify huggingface tests.

[daavoo]

Use simple dataset and model from transformers/tests/trainer.

[codecov-commenter]

Codecov Report

Base: 94.98% // Head: 94.82% // Decreases project coverage by -0.15% ⚠️

Coverage data is based on head (9611927) compared to base (b8b4ca8).
Patch coverage: 92.45% of modified lines in pull request are covered.

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #321      +/-   ##
==========================================
- Coverage   94.98%   94.82%   -0.16%     
==========================================
  Files          36       36              
  Lines        1734     1758      +24     
  Branches      161      164       +3     
==========================================
+ Hits         1647     1667      +20     
- Misses         64       66       +2     
- Partials       23       25       +2     

Impacted Files	Coverage Δ
src/dvclive/huggingface.py	90.47% <33.33%> (-9.53%)	⬇️
tests/test_huggingface.py	97.72% <96.00%> (-2.28%)	⬇️

Help us with your feedback. Take ten seconds to tell us how you rate us. Have a feature suggestion? Share it here.

☔ View full report at Codecov.
📢 Do you have feedback about the report comment? Let us know in this issue.

[pared]

Seems like windows has some problems with new model.

[daavoo]

Seems like windows has some problems with new model.

I have no idea whats going on 😅

[skshetry]

Trainer is using time.time() which has a very low resolution in Windows. So I think time.time() - start_time becomes 0. Try adding a time.sleep() to the minimum resolution in the training loop.

time.sleep(time.get_clock_info('time').resolution)

[skshetry]

Or, propose a fix in transformers to move away from time.time() to higher precision clocks perf_counter/monotonic etc.. 😅