search

facebookresearch / demucs

Code for the paper Hybrid Spectrogram and Waveform Source Separation
MIT License
8.29k stars 1.05k forks source link

Training from scratch #442

Open ElizavetaSedova opened 1 year ago

ElizavetaSedova commented 1 year ago

Hello! I have some questions regarding the training htdemucs (4 version). 1) I follow the rules and run: dora run -d -f 81de367c. I have a graphics card with 12 GB of memory. Often there is a CUDA out of memory error. I decrease the following parameters: dset.segment=7 batch_size=4. As I understood, I must have a minimum batch_size=4 per card. This is the only way to start training. Is it possible that several 12 GB cards are not enough to start training with the standard dset.segment=11 parameter? I saw you mention that a decrease of the segment could negatively affect learning results. I'm interested to know what capacity is needed to conduct training from scratch or for fine-tuning. Any ideas on optimization are welcome! 2) What is the signature for htdemucs? I’m not sure 81de367c is what is needed. 3) I see that 800 songs were used to training htdemucs. Is it your own dataset or is it a dataset from open sources? Do you provide these songs for training?

ElizavetaSedova commented 1 year ago

Can you help me also with advice on using more than 4 sources? For example, I don't want to use vocals, but I want to add some additional sources from my dataset. Do I need to use musbdHQ18? Can I train on musbdHQ18 and my dataset if it has different sources? What should I do for this?