so-vits-svc-fork
so-vits-svc-fork copied to clipboard
voicepaw
Fail to run on colab
Describe the bug
~~EDIT: Issue seems solved. Waitingfor confirmation. DO NOT USE MY WORKAROUND! Well, they updated the repo. There is no need for my workaround anymore. Do NOT run my two lines anymore! END EDIT~~
There are alot of tickets saying that so-vits-svc-fork is bugged or broken and cannot be used especially with the example Colab Notebook. What happened was, that a package from the requirements for this repo was changed and you manually should install
!pip install pyworld==0.3.2 !pip install numpy==1.23.5
After that continue with the first step "Check GPU" and you should be able to clone the repo and use it for training. For the last step "Inference" you might needto run this command:
!pip install numpy==1.24
Check your logs for errors and remember this is only a workaround. There might be more elegant solutions and I hope @voicepaw will update the repo soon.
https://github.com/voicepaw/so-vits-svc-fork/issues/843 https://github.com/voicepaw/so-vits-svc-fork/issues/840 https://github.com/voicepaw/so-vits-svc-fork/issues/838 https://github.com/voicepaw/so-vits-svc-fork/issues/837
To Reproduce
!pip install pyworld==0.3.2 !pip install numpy==1.23.5
Additional context
No response
Version
4.1.1
Platform
Colab
Code of Conduct
- [X] I agree to follow this project's Code of Conduct.
No Duplicate
- [X] I have checked existing issues to avoid duplicates.
pusemuckel888, thank you for trying to fix things in an elegant matter. and thanks again for clearing things up. Now I don't know if this issue is the same, but when I am trying to train a model that I already trained to 900 epochs. Even when I still have the G_0 files for both. it finish the process but do not start training again.
pusemuckel888, thank you for trying to fix things in an elegant matter. and thanks again for clearing things up. Now I don't know if this issue is the same, but when I am trying to train a model that I already trained to 900 epochs. Even when I still have the G_0 files for both. it finish the process but do not start training again.
Might need to restart the runtime inbetween. OR Try without Tensorboard: add a "#" before the lines to comment them out and exclude from execution, like so:
#@title Train #%load_ext tensorboard #%tensorboard --logdir drive/MyDrive/so-vits-svc-fork/logs/44k !svc train --model-path drive/MyDrive/so-vits-svc-fork/logs/44k
Maybe tensorboard causes issues for you. Check if Training works by looking at the log folder in drive: New G_ and D_ fIles should be created. OR If this does not help, in the top menu, click on "View" and then "Executed code history" to show the full log of what is happening. OR Delete your latest D_ and G_ files. Maybe those are corrupted, when Colab happened to disconnect just intime when creating these checkpoints.
Thank you a million for updating. I went through the steps you mentioned and kept doing again and again, but it seems not work. I might be doing. I want to make it clear. I left training then it reached 900 epochs. Today I watn to continue training again. So I restarted the runtime and went again step by step as to train it from the start, the same steps I mean. Should I cancel a step or so? maybe I am doing something wrong. I attached a video ..
https://github.com/voicepaw/so-vits-svc-fork/assets/12218136/051584d0-34a3-4aad-b21e-827aa4cfb2d5
This is the log file
Looks ok so far to me. Try this: Copy your files from the log/44k directory, to keep them safe. Try deleting both your G_929.pth and D_929.pth. Then system will continue with 922 files. If that does not help, delete all the files in log/44k and try from 0, just to make sure the script works correctly.
I believe the colab is broken or something. I went through all the steps you mentioned, I even decided to start a fresh new colab, choose different dataset, When I reach train, it's not working. don't know what happened to it.
I believe the colab is broken or something. I went through all the steps you mentioned, I even decided to start a fresh new colab, choose different dataset, When I reach train, it's not working. don't know what happened to it.
I wanted to ask you, if you even had your graphics card activated, but I could tell that you have from your video :) Well, they updated the repo. There is no need for my workaround anymore. Do NOT run my two lines anymore!
I have been trying for like 3 hours now with a fresh new model and ran through all the steps multiple times and just now it works again. As my workaround is redundant, there must have been changes made by some devs I guess. Maybe they fixed it just now.
Pls try running without my workaround and tell us if works for you as well now. UPDATE: I tested a new colab notebook with a fresh model and on kaggle i could successfully continue train an existing model.
I believe the colab is broken or something. I went through all the steps you mentioned, I even decided to start a fresh new colab, choose different dataset, When I reach train, it's not working. don't know what happened to it.
I wanted to ask you, if you even had your graphics card activated, but I could tell that you have from your video :) Well, they updated the repo. There is no need for my workaround anymore. Do NOT run my two lines anymore!
I have been trying for like 3 hours now with a fresh new model and ran through all the steps multiple times and just now it works again. As my workaround is redundant, there must have been changes made by some devs I guess. Maybe they fixed it just now.
Pls try running without my workaround and tell us if works for you as well now. UPDATE: I tested a new colab notebook with a fresh model and on kaggle i could successfully continue train an existing model.
Hi [pusemuckel888],
Sorry for the late response, had to work on whatever I reached at that point and I would like to share with you what I created in here. By the way, I just tried it and it's still not working. The same thing happening, I don't know why when I reach Train. that's it done.
Please see the link below for the video, Cillian Murphy voice cloned and changed the language to arabic, while deepfaking him on.
https://www.instagram.com/p/CvIi4mjNs0o/
The issue is still there, again. Step "Train" stops executing without error. I will try to debug "svc train" later.
Well, training works for me again. One little change I made was from this:
#@title Copy configs file
!cp configs/44k/config.json drive/MyDrive/so-vits-svc-fork
to this:
#@title Copy configs file
!cp /content/configs/44k/config.json /content/drive/MyDrive/so-vits-svc-fork
Can the colab itself be repaired and updated or forked rather than requiring the end user to modify scripts up front? This colab worked perfectly a few months ago despite the initial learning curve.
There are now numerous roadblocks; the loading of dependencies stage is riddled with errors, the configs folder and .json file creation step doesnt create either, and the Training/Tensorflow step fails.
Thanks to anyone who is able to get it working again.
