Noticed that the code was making an exception with the line:
`if version19 == "v1" or sr2 == "40k":`
This approach isn't ideal, as it causes v2 - 40k to fallback to v1 - 40k. While this fix re-adds the necessary code, a better long-term solution might be to copy the v1 - 40k config into the v2 config folder to avoid such exceptions in the future.
This commit addresses two issues on the Train tab:
1. **Inconsistent Target Sample Rates:**
- Resolved an issue where switching versions led to inconsistent Target sample rates. Additionally, it fixes a missing v2 - 40k configuration file. Despite v2 having a 40k pretrained model, configurations were mismatched:
- assets/pretrained (32k, 40k, 48k)
- assets/pretrained_v2 (32k, 40k, 48k)
- configs/v1 (32k, 40k, 48k)
- configs/v2 (32k, 48k)
- Tests showed that a 40k - v2 setting generated an incorrect 40k - v1 config file. The default settings have been updated to use 48k - v2 to align the sample rates correctly.
Now:
- v1 will consistently show 32k, 40k, and 48k
- v2 will consistently show 32k and 48k
**Note:** Ensures that selecting the 40k option for v2 no longer results in an incorrect configuration file, aligning the sample rates across both versions.
2. **Repositioned 'Version' Field:**
- Moved the 'Version' field to appear before the 'Target sample rate' field on the Train tab. This improves the logical flow and user experience by ensuring that the 'Version' selection is made before choosing the 'Target sample rate'.
Changes:
- Updated layout on the Train tab to reposition the 'Version' field.
- Ensured that field updates are reflected correctly.
**Note:** Aligns with a more intuitive user workflow by ensuring users select the version first before setting the target sample rate. No functional changes to the data processing were made.
There is a bug in librosa 0.9.1.
https://github.com/librosa/librosa/pull/1594
As a result, an error occurs when executing the "Vocals/Accompaniment Separation & Reverberation Removal" function.
To address this issue, librosa has been upgraded to version 0.10.2.
Additionally, torchcrepe has been upgraded due to its dependency on librosa.
* for GTX1650
* testing torch-profiler
* no more profiler & change default audio
* longer slice
* fluid container
* cache rmvpe and ui tweaks
* get my changes back after merge
* format code
* only load rmvpe when necessary
* fix rmvpe & config bug
* fix is_half again
* manual sync with upstream
* fix bool() for is_half
* revert changes for PR
* revert again
* typo
* match lower case
