Implemenation of Latent Factor Analysis via Dynamical Systems(Pandarinath ... and Sussillo et al., Nature Methods 2018) with Chainer. Official implementation by tensorflow is here. LFADS is a sequential adaptation of a VAE (Variational Auto-Encoder). It can de-noise and find a low-dimensional representations for time series data. Although it is especially designed for investigating neuroscience data, it can be applied to any time series data.
MIT license. Contributions welcome.
This is the blog post (only in Japanese) about this repository.
python 2.x, chainer 3.3.0, numpy, h5py
Before training, you have to prepare the dataset. This is an example dataset.
python synth_data.py
It generates mixed signals from low-dimensional factors.
An example time series of a latent factor.
Corresponding mixed time series (i.e., observation).
Train LFADS without controller.
python run_lfads.py /path/to/dataset --gpu 0 --con-h-dims 0 --epoch 1000 --batch-size 64
Train LFADS with controller (Full model).
python run_lfads.py /path/to/dataset --gpu 0 --epoch 1000 --batch-size 64
python run_posterior_sampling.py /path/to/trained/model --gpu 0 --nsample 5
An example reconstructed time series. The dimension of latent factor is 3, which is much lower than the dimension of observed time course (20).
An example estimated factor.
python run_prior_sampling.py /path/to/trained/model --gpu 0 --nsample 300
- Sampling from poisson distribution
- External inputs
- Dynamic stitchng