Skip to content

Implementation of Differentiable Digital Signal Processing (DDSP) in Pytorch

License

Notifications You must be signed in to change notification settings

acids-ircam/ddsp_pytorch

Repository files navigation

Differentiable Digital Signal Processing

ddsp_tilde_screenshot

Implementation of the DDSP model using PyTorch. This implementation can be exported to a torchscript model, ready to be used inside a realtime environment (see this video).

Pretrained models

instrument realtime preprocessing sampling rate link
saxophone true sigmund~ 48kHz download
violin true sigmund~ 48kHz download

Usage

Edit the config.yaml file to fit your needs (audio location, preprocess folder, sampling rate, model parameters...), then preprocess your data using

python preprocess.py

You can then train your model using

python train.py --name mytraining --steps 10000000 --batch 16 --lr .001

Once trained, export it using

python export.py --run runs/mytraining/

It will produce a file named ddsp_pretrained_mytraining.ts, that you can use inside a python environment like that

import torch

model = torch.jit.load("ddsp_pretrained_mytraining.ts")

pitch = torch.randn(1, 200, 1)
loudness = torch.randn(1, 200, 1)

audio = model(pitch, loudness)

Realtime usage

Be sure that the block_size defined in config.yaml is a power of 2 if you want to use the model in realtime!

If you want to use DDSP in realtime (yeah), we provide a pure data external wrapping everything. Export your trained model using

python export.py --run runs/mytraining/ --realtime true

This will disable the reverb and enable the use of the model in realtime. For now the external works on CPU, but you can enable GPU accelerated inference by changing realtime/ddsp_tilde/ddsp_model.h DEVICE to torch::kCUDA. Inside Pd, simply send load your_model.ts to the ddsp~ object. The first inlet must be a pitch signal, the second a loudness signal. It can be directly plugged to the sigmund~ object for real-time timbre transfer.

You can then apply the exported impulse response using a convolution reverb (such as partconv~ from the bsaylor library).

Compilation

You will need cmake, a C++ compiler, and libtorch somewhere on your computer. Then, run

cd realtime
mkdir build
cd build
cmake ../ -DCMAKE_PREFIX_PATH=/path/to/libtorch -DCMAKE_BUILD_TYPE=Release
make install

If you already have pytorch installed via pip inside a virtual environment, you can use the following PREFIX_PATH:

cmake ../ -DCMAKE_PREFIX_PATH=~/miniconda3/lib/python3.X/site-packages/torch -DCMAKE_BUILD_TYPE=Release
make install

By default, it will install the external in ~/Documents/Pd/externals.