Out of memory errors and using whisperX
#249
Open
opened
Loading…
Reference in New Issue
There is no content yet.
Delete Branch "%!s(<nil>)"
Deleting a branch is permanent. It CANNOT be undone. Continue?
Hello, I hope its okay if I continue to post questions here about my efforts to get this working since there does not seem to be any more appropriate forum for this software.
I am having two major issues so far.
The first issue I am facing is when I attempt to use whisperx to create the dataset. When I select the option from the dropdown menu it gives me the error 'no module named 'whisperx'. I've installed whisperx. I've tried starting the software in different conda environments with whisperx installed. I've tried activating the venv in the aicloning directory and installing whisperx to no avail. It still gives me the same error. I'm not even using it through Colab which is where others seem to have this problem. Below is the error in the console.
I can import whisperx in a regular python session just fine but it just won't work here for some reason.
The second issue is that I am getting out of memory errors. This happens even if I shrink the dataset size down to just 5mb, the batch size to 2 and the gradient accumulation size to 2. I have a 3070 mobile with 8gb vram which is similar to other people who are successfully running training here so I don't know whats up. Below is a reproduction of the traceback.
Can you run
whisperx
from the command line? (Not though importing it in a python session, just from the prompt.)Just tried and while I can run whisperx directly it still doesn't work through the ai cloning ui.
Try activating the venv in the directory you cloned the repo into and then
git submodule update –remote
Still gives me
whether I run aivoicecloning ui in (base) or (venv)(base) via source venv/bin/activate
After activating the venv does
pip list installed
show whisperx?Okay we're getting somewhere. Whisperx wasn't on the list despite being able to be called under the environment. I reinstalled all the software but while it now detects whisperx from the ui it still does not work properly.
Im still searching for how to include this authorization token the error asks for assuming it actually is the problem. There doesn't seem to be an obvious way.
It's in the Wiki:
Thanks for your help so far. I still haven't completely gotten it to work. If you want to continue picking away at it with me I'll try a few more things. Otherwise I'll look into an alternative way to get better timestamping than base whisper.
Anyway. I got further and it appears that the current Whisperx is not completely compatible with Voiceclonings implementation of it.
I inserted a transcribe_with_vad function found here
https://github.com/m-bain/whisperX/issues/68
into transcribe.py but that doesn't seem to solve it.
There's probably some way to tinker with your current install to get it working but I think the most efficient thing to do is wipe it (save your datasets, of course), reclone the repo, and run the install scripts over again.
I did. aivoicecloning calls for 'transcribe_with_vad' from whisperx through utils.py
But I downloaded the current whisperx repo and there is no 'transcribe_with_vad' phrase anywhere in it. According to grep
As noted in issue #68 from the whisperx repo "with VAD" is now the default, so you could try changing utils.py to just call transcribe() and see if that works. I always prepare my datasets externally so I haven't tried it.
Oh right, I forgot. I think it was v3 is when he broke everything with faster-whisper, which I tried for a day and found it to be un-good to warrant working around.
I suppose the issue is that I'm not actually printing the exception when it's trying to load WhisperX, and it'll remain unloaded.
mmmm freeze it with
I believe. I can't quite remember how to specifically have Pip install from Git from a branch.