[StyleTTS2] load pretrained checkpoint from HuggingFace

[roedoejet]

PR Goal?

Loading pretrained models from within the codebase is messy since we need to handle the paths, and the model checkpoints bloat the package/our git history.

Fixes?

An element of #686

Feedback sought?

Does this look reasonable? I think the "local override" option is still not going to be handled well since it will put a potentially absolute path in the config.

Priority?

high

Tests added?

None

How to test?

if you train a model with this branch, you should see the HF models downloading automatically

Confidence?

medium

Version change?

n/a

Related PRs?

EveryVoiceTTS/StyleTTS2#6

Once accepted, we'll need to clean the git history of any *.t7 or *.pth files

[semanticdiff-com]

Review changes with  

Changed Files

File	Status
everyvoice/tests/data/relative/config/everyvoice-text-to-wav.yaml	33% smaller
everyvoice/cli.py	15% smaller
everyvoice/.schema/everyvoice-text-to-wav-0.5.json	4% smaller
everyvoice/model/e2e/StyleTTS2_lightning	0% smaller
everyvoice/wizard/basic.py	0% smaller

[joanise]

the new-project wizard still expected the pre-trained files to exist

[joanise]

The concept is good here.

For the local override, we could put a warning in the field's documentation that it gets saved in the models. Also, we could support $HOME/path/to/... fairly easily with os.path.expandvars() and os.path.expanduser() which could get called after pydantic validation, or maybe in post-validation, as long as the input string is what gets saved in the config.

[github-actions]

CLI load time: 0:00.29
Pull Request HEAD: 149600faf4eeae8730bf13a3cdb68732bd3f2199

Imports that take more than 0.1 s:
import time: self [us] | cumulative | imported package
import time:       974 |     108906 |   typer
import time:      6022 |     107542 |               loguru._logger
import time:      3714 |     112498 |             loguru
import time:      2775 |     162048 |           everyvoice.utils
import time:       903 |     162950 |         everyvoice.model.feature_prediction.FastSpeech2_lightning.fs2.cli.benchmark
import time:       525 |     171552 |       everyvoice.model.feature_prediction.FastSpeech2_lightning.fs2.cli.cli
import time:       316 |     171868 |     everyvoice.model.feature_prediction.FastSpeech2_lightning.fs2.cli
import time:      1333 |     173200 |   everyvoice.model.feature_prediction.FastSpeech2_lightning.fs2.cli.check_data
import time:      8242 |     388353 | everyvoice.cli
import time:      2183 |     119259 |       markdown_it.main
import time:       379 |     119637 |     markdown_it
import time:      5246 |     177500 |   rich.markdown
import time:      4248 |     267395 | typer.rich_utils

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 84.99%. Comparing base (c1875dd) to head (8b813f5).
⚠️ Report is 1 commits behind head on dev.ap/styletts2-improvements.

Additional details and impacted files

@@                      Coverage Diff                       @@
##           dev.ap/styletts2-improvements     #807   +/-   ##
==============================================================
  Coverage                          84.98%   84.99%           
==============================================================
  Files                                 46       46           
  Lines                               4217     4219    +2     
  Branches                             632      632           
==============================================================
+ Hits                                3584     3586    +2     
  Misses                               494      494           
  Partials                             139      139           

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

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[roedoejet]

The concept is good here.

For the local override, we could put a warning in the field's documentation that it gets saved in the models. Also, we could support $HOME/path/to/... fairly easily with os.path.expandvars() and os.path.expanduser() which could get called after pydantic validation, or maybe in post-validation, as long as the input string is what gets saved in the config.

good idea. I've added this.