intel / openvino-plugins-ai-audacity

A set of AI-enabled effects, generators, and analyzers for Audacity®.
GNU General Public License v3.0
908 stars 57 forks source link

Query on Download Size of Whisper Transcription Models #228

Open hal8kb opened 3 months ago

hal8kb commented 3 months ago

I'm currently compiling Openvino on Ubuntu 22.04

I've just downloaded the whisper transcription model, and its a massive 16GB ! The expanded size of the whisper.cpp=ovenvino-models is an incredible 31G Does this sound correct?

du -sh whisper.cpp-openvino-models/ 31G whisper.cpp-openvino-models/

The extracted models are much smaller and size of openvino-models is only 4G. So far all plugins have used 20G of data, if this is normal, it may be an idea to advise people, especially if they use metered internet connections.

RyanMetcalfeInt8 commented 3 months ago

Hi @hal8kb,

Yes, this is correct. The repo contains all variants (base, small, medium, large) of the whisper models (https://huggingface.co/Intel/whisper.cpp-openvino-models/tree/main), which are very large.

It's a good point -- I'll add some kind of warning to the documentation that these are very large downloads.

The extracted models are much smaller and size of openvino-models is only 4G.

I didn't really understand this part. Which extracted models were smaller?

hal8kb commented 3 months ago

Hi Ryan Sorry, should have been more clearer, the size of the whisper.cpp-ovenvino-models directory was an incredible 31G. However when the extracted files were copied into openvino-models directory the size of that directory is only 4.3G, and it contains the whisper, deepfilter and musicgen files.

du -sh openvino-models/ 4.3G openvino-models/

RyanMetcalfeInt8 commented 3 months ago

Hi Ryan Sorry, should have been more clearer, the size of the whisper.cpp-ovenvino-models directory was an incredible 31G. However when the extracted files were copied into openvino-models directory the size of that directory is only 4.3G, and it contains the whisper, deepfilter and musicgen files.

du -sh openvino-models/ 4.3G openvino-models/

Ah, got it. Yeah, looks like the documentation shows only base & small variants getting unzipped:

unzip whisper.cpp-openvino-models/ggml-base-models.zip -d openvino-models
unzip whisper.cpp-openvino-models/ggml-small-models.zip -d openvino-models
unzip whisper.cpp-openvino-models/ggml-small.en-tdrz-models.zip -d openvino-models

Whereas there are 'medium' and 'large' models available in the git repo that can also get unzipped / copied. Looks like the documentation just needs some updates.

hal8kb commented 3 months ago

Ryan, the install documented is excellent, and works as described. My previous failures with install was down to me not reading through carefully.

If you edit anything, may I suggest a closing paragraph. Something like:

Once you have confirmed that you have a working audacity in audacity-build/release/bin/audacity copy it to /usr/local/bin

sudo cp audacity-build/release/bin/audacity /usr/local/bin/

RyanMetcalfeInt8 commented 3 months ago

Thanks -- hmm, I don't know if sudo cp audacity-build/release/bin/audacity /usr/local/bin/ is the right way to install it. The reason is, there are probably other things (like a bunch of .so's) that also need to be installed to various locations.

The typical Linux operation is sudo make install -- although I don't know if this is the proper way to install Audacity.

In the cases that I've built / used Audacity in Linux, I've simply used it from the build directory and didn't really have any reason to install it. But, I can see the value in someone wanting to do that..

oxidian24 commented 3 months ago

I am pulling the install script (just now), but I definitely recommend to add a warning about using of the git clone (to install.sh). Users may be warned that complete download of all the models including large models is massive. I do not include this into the current version of the install script.