Closed alistairewj closed 1 month ago
Adding streamlit as a dependency and some experimentation exposed the following conflicts:
The conflict is caused by:
b2aiprep 0+untagged.1.g5ba01ee depends on numpy==1.26.4
streamlit 1.35.0 depends on numpy<2 and >=1.19.3
transformers 4.39.3 depends on numpy>=1.17
b2aiprep[dev,tts] 0+untagged.1.g5ba01ee depends on numpy==1.26.4
tts 0.22.0 depends on numpy==1.22.0; python_version <= "3.10"
https://github.com/sensein/b2aiprep/actions/runs/9666011062/job/26664513604#step:4:368
After debugging, my view of the most reasonable solution was to remove the TTS package dependency. This means the CLI no longer supports converting a voice, which was a niche use anyway. The TTS package owners have recently shut down anyway: https://coqui.ai/
If we want to reimplement the feature, we probably need to use NeMo or something similar.
Would Torchvision
still be required as a dependency? (merge commit)
Currently, when running streamlit
using the README.md example & the dataset, the b2aiprep-cli
program outputs torchvision is not available - cannot save figures
- I am assuming this might be a remnant library initialization from the TTS & voice conversion features, but asking to be sure!
Yes AFAIK we don't use any of the features from torchvision. This is reported by one of the imports, though not sure which one in particular.
Adds streamlit to the pyproject.toml as it is required now
This is redundant to #62 - but I'm testing the CI to see if requiring np <2 is explicitly required (streamlit says it doesn't support np >= 2 and it's unclear if any of our other packages do)