search

facebookresearch / audiocraft

Audiocraft is a library for audio processing and generation with deep learning. It features the state-of-the-art EnCodec audio compressor / tokenizer, along with MusicGen, a simple and controllable music generation LM with textual and melodic conditioning.
MIT License
20.5k stars 2.06k forks source link

How to train the audiogen model？ #228

Open chenxinglili opened 1 year ago

chenxinglili commented 1 year ago

I want to train an audiogen with my own data, but I have encountered several confusion in the documentation. I am not familiar with dora, and I also meet several errors when I tried to execute "dora grid audiogen.audiogen_base_16khz". Do we have clearer documentation? Or has anyone successfully run it and can give some advice? Or can I use pytorch ddp to perform training?

tanggang1997 commented 1 year ago

老哥，我目前是训练成功了，但是生成的音频电流声过大，你可以在项目的根目录下，运行 dora run solver=musicgen/musicgen_base_32khz ,其中solver对应的就是config/solver/musicgen文件夹下的musicgen_base_32khz.yaml文件，可以对这个文件进行配置一些参数

chenxinglili commented 1 year ago

Hi, thanks to your reply. I want to re-train the audiogen. if there is any difference between retraining audiogen and musicgen? Besides, I don't have a cluster like slurm. I don’t know if this is the reason why the code reported an error after I executed dora. Is it possible to utilize ddp or accelerate to perform training instead of dora?

tanggang1997 commented 1 year ago

audiogen我没有试过，但是应该和musicgen是差不多的，第二个问题：关于集群，我看文档里面有写会自动去检测集群，所以应该是不需要去设置的

tanggang1997 commented 1 year ago

需要看你报错的问题是什么

chenxinglili commented 1 year ago

May i get one of your contact information？

tanggang1997 commented 1 year ago

vx:15722869549

veryVANYA commented 1 year ago

did anyone find a fix around slurp? biggest blocker!

jmren168 commented 10 months ago

@chenxinglili Hi, Did you seccessfully re-train AudioGen on your own dataset? I would like to fine-tune AudioGen on my own dataset. Any guides are highly appreciated.

chenxinglili commented 10 months ago

yep, I successfully run it, but it is very troublesome.

The training pipeline of the model is based on dora, and I have no knowledge of dora before. I want to rewrite the training code, from dora to ddp, but there is a lot of coupling. And sorry, I give up after trying it.

jmren168 commented 10 months ago

yep, I successfully run it, but it is very troublesome.

The training pipeline of the model is based on dora, and I have no knowledge of dora before. I want to rewrite the training code, from dora to ddp, but there is a lot of coupling. And sorry, I give up after trying it.

Thanks for the quick reply. Since the author doesn't release the training code of AudioGen, I may try other audio generative models, e.g., AutoGPT/AutoGen.