Official implementation of the paper Musika! Fast Infinite Waveform Music Generation, accepted to ISMIR 2022.
This work was conducted as part of Marco Pasini's Master thesis at the Institute of Computational Perception at JKU Linz, with Jan Schlüter as supervisor.
Find the demo samples here (old 22.05 kHz implementation)
Find the paper here (old 22.05 kHz implementation)
The current version of musika has been updated to produce 44.1 kHz higher quality samples! You can find the old implementation that is presented in the ISMIR paper in the 22kHz folder.
Listen to some 44.1 kHz samples here
An online demo is available on Huggingface Spaces. Try it out here!
You can use this Colab notebook to train Musika on custom music, no technical skills needed!
Before starting, make sure to have conda and ffmpeg installed.
First, create a new environment for musika:
conda create -n musika python=3.9Then, activate the environment (do this every time you wish to use musika):
conda activate musikaInstall CUDA 11.2 and CuDNN if you have a Nvidia GPU and you do not already have CUDA 11.2 installed on your system (you can check your CUDA version with the command nvcc --version):
conda install -c conda-forge cudatoolkit=11.2 cudnn=8.1.0And configure the system paths with the two following commands (only for Linux, skip on Windows):
mkdir -p $CONDA_PREFIX/etc/conda/activate.d
echo 'export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:$CONDA_PREFIX/lib/' > $CONDA_PREFIX/etc/conda/activate.d/env_vars.shFinally, clone this repository, move to its directory and install the requirements:
git clone https://github.com/marcoppasini/musika
cd musika
pip install -r requirements.txtYou can conveniently generate samples using a Gradio interface by running the command:
python musika_test.pyBy default the system generates techno samples. To generate misc music samples (a diverse music dataset was used for training), specify a different path for the pretrained weights:
python musika_test.py --load_path checkpoints/miscYou can also generate and save an arbitrary number of samples with a specified length with:
python musika_generate.py --load_path checkpoints/misc --num_samples 10 --seconds 120 --save_path generationsYou can train a musika system using your own custom dataset. A pretrained encoder and decoder are provided to produce training data (in the form of compressed latent sequences) of any arbitrary domain.
Note that using the provided universal autoencoder will sometimes produce low quality samples, for example for samples that mainly contain vocals: training a custom encoder and decoder for a specific dataset would produce higher quality samples, especially for narrow music domains. A training script for custom encoders/decoders will be provided!
Before proceeding, make sure to have a Nvidia GPU and CUDA 11.2 installed. Mixed precision is enabled by default, so if your GPU does not support it make sure to disable it using the --mixed_precision False flag.
Also, you may experience errors when training using XLA (--xla True by default for faster training) with CUDA locally installed in the environment. See this link for the solution to this problem. Alternatively, specify --xla False (the speedup provided by XLA is not substantial).
First of all, encode audio files to training samples (in the form of compressed latent vectors) with:
python musika_encode.py --files_path folder_of_audio_files --save_path folder_of_encodingsfolder_of_encodings will be automatically created if it does not exist.
musika encodes audio samples to chunks of sequences of latent vectors of equal length (--max_lat_len) which by default are double the size of the chunks used during training. During training chunks are randomly cropped as a data augmentation technique. Keep in mind that by default audio samples are required to be at least 47 s long (--max_lat_len 512): if you require to encode shorter samples specify a lower value (the minimum is 256, corresponding to about 23 s, which is the length of samples used for training) both during encoding and during training.
Training a model from scratch generally requires large amounts of training data if the audio domain to generate has substantial timbre diversity. However, feel free to experiment with the data you have available! The results of the system should scale quite well with the amount of data used for training.
To train musika from scratch use:
python musika_train.py --train_path folder_of_encodingsPlease be aware that training musika from scratch can take multiple hours on a powerful GPU (at least 2 million iterations are recommended).
A tensorboard and a gradio link will be generated for you to check losses and results during training. Additionally, generated audio samples and the corresponding spectrograms are saved in the checkpoint folder after each epoch.
To train a model with a shorter context window (6 s instead of 12 s) use:
python musika_train.py --train_path folder_of_encodings --small TrueYou can also increase the number of parameters (capacity) of the model with the --base_channels argument (--base_channels 128 is the default). From our experiments, Musika can greatly benefit from an increase in capacity (at the expense of longer training times). We recommend lowering the learning rate to achieve stable training with a substantially larger model than the default one. For example:
python musika_train.py --train_path folder_of_encodings --base_channels 192 --lr 0.00007Finally, if you wish to resume training from a specific checkpoint folder you can specify it in --load_path:
python musika_train.py --train_path folder_of_encodings --load_path checkpoints/MUSIKA_latlen_x_latdepth_x_sr_x_time_x-x/MUSIKA_iterations-xk_losses-x-x-x-xThe fastest way to train a musika system on custom data is to finetune a provided checkpoint that was pretrained on a diverse music dataset. Since training musika from scratch usually requires large amounts of training data, finetuning can represent a good compromise in some cases. And it is incredibly fast! Note that the perfect dataset to finetune musika on has very limited timbre diversity (metal, piano music, ...). Finetuning on a diverse music dataset will not produce good results! Training a musika system from scratch will produce better results in the majority of cases!
You can finetune the misc musika model (trained on a diverse music collection) on your dataset with:
python musika_train.py --train_path folder_of_encodings --load_path checkpoints/misc --lr 0.00004In case you experience nans or training instabilities try lowering the learning rate further using --lr 0.00002.
In case your target music domain is close to the techno genre, you can finetune the provided techno checkpoint instead of the default misc checkpoint:
python musika_train.py --train_path folder_of_encodings --load_path checkpoints/techno --lr 0.00004Also, we provide a misc_small checkpoint you can finetune if you do not need to generate samples with long-range coherence (the model was trained with a shorter context window):
python musika_train.py --train_path folder_of_encodings --load_path checkpoints/misc_small --small True --lr 0.00004Make sure to check out all the other flags in the parse.py file!
You can finally test and generate samples with your trained model with the same commands that are described in the Generate Samples section above, by specifying the desired checkpoint folder in the --load_path argument.
By default musika_encode.py encodes a single audio sample to multiple training encodings of fixed length (--max_lat_len 512 by default). If you wish to encode a single audio file to a single encoding (for use in other applications), you can specify --whole True in the command:
python musika_encode.py --files_path folder_of_audio_files --save_path folder_of_encodings --whole TrueIf you wish to decode back to waveform a folder of encodings created with the musika_encode.py command, you can use:
python musika_decode.py --files_path folder_of_encodings --save_path folder_of_audio_filesfolder_of_audio_files will be automatically created if it does not exist.
By first encoding and then decoding a collection of samples, you can check what is the upper bound on the quality of generated samples by listening to the waveform reconstructions.