# somata
Github: https://github.com/mh105/somata
**State-space Oscillator Modeling And Time-series Analysis (SOMATA)** is a Python library for state-space neural signal
processing algorithms developed in the [Purdon Lab](https://purdonlab.stanford.edu).
Basic state-space models are introduced as class objects for flexible manipulations.
Classical exact and approximate inference algorithms are implemented and interfaced as class methods.
Advanced neural oscillator modeling techniques are brought together to work synergistically.
[](https://github.com/mh105/pot/commits/master)
[](https://www.python.org/)
[](https://spdx.org/licenses/BSD-3-Clause-Clear.html)
[](https://zenodo.org/badge/latestdoi/556083594)
---
## Table of Contents
* [Requirements](#requirements)
* [Install](#install)
* [Basic state-space models](#basic-state-space-models)
* [StateSpaceModel](#class-statespacemodel)
* [OscillatorModel](#class-oscillatormodelstatespacemodel)
* [AutoRegModel](#class-autoregmodelstatespacemodel)
* [GeneralSSModel](#class-generalssmodelstatespacemodel)
* [Advanced neural oscillator methods](#advanced-neural-oscillator-methods)
* [Authors](#authors)
* [Citation](#citation)
* [License](#license)
---
## Requirements
[`somata`](https://pypi.org/project/somata/) is built on [`numpy`](https://numpy.org) arrays for computations. [`joblib`](https://joblib.readthedocs.io/en/stable/) is used for multithreading.
Additional dependencies include [`scipy`](https://scipy.org), [`matplotlib`](https://matplotlib.org), and [`spectrum`](https://pyspectrum.readthedocs.io/en/latest/index.html).
The source localization module also requires [`pytorch`](https://pytorch.org) and [`MNE-python`](https://mne.tools/stable/index.html).
Full requirements for each release version will be updated under
[`install_requires`](https://setuptools.pypa.io/en/latest/userguide/dependency_management.html#platform-specific-dependencies) in the [`setup.cfg`](setup.cfg) file.
If the [`environment.yml`](environment.yml) file is used to [create a new conda environment](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file),
all and only the required packages will be installed.
## Install
```
$ pip install somata
```
### conda-forge channel
While [`pip install`](https://pip.pypa.io/en/stable/cli/pip_install/) usually works, [an alternative way](https://pythonspeed.com/articles/conda-vs-pip/) to install `somata` is through the [conda-forge](https://conda-forge.org/docs/index.html) [channel](https://docs.conda.io/projects/conda/en/latest/user-guide/concepts/channels.html#what-is-a-conda-channel), which utilizes [continuous integration (CI)](https://conda-forge.org/docs/user/ci-skeleton.html) [across OS platforms](https://conda-forge.org/docs/user/introduction.html#why-conda-forge).
This means [conda-forge packages](https://conda-forge.org/feedstock-outputs/index.html) are more [compatible with each other](https://conda-forge.org/docs/maintainer/adding_pkgs.html#avoid-external-dependencies) compared to [Python Package Index (PyPI) packages](https://pypi.org) installed via [`pip` by default](https://packaging.python.org/en/latest/tutorials/installing-packages/#installing-from-pypi).
When `somata` is installed into an [existing conda environment](https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#viewing-a-list-of-your-environments), unmet dependencies are automatically searched, downloaded, and installed from the same repository of packages containing `somata`.
If `pip install somata` fails to resolve some dependencies, the [conda-forge somata](https://github.com/conda-forge/somata-feedstock) [feedstock](https://github.com/conda-forge/conda-feedstock#terminology) can be used to install:
```
$ conda install somata -c conda-forge
```
### torch requirement
If the [`pytorch`](https://pytorch.org) dependency is not resolved correctly for your [OS](https://whatsmyos.com),
first [install `pytorch` manually](https://pytorch.org/get-started/locally/) in a [conda environment](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html) that you want to install `somata` in, and then rerun either of the above two lines to install `somata`.
### (For development only)
- ### Fork this repo to personal git
[How to: GitHub fork](https://docs.github.com/en/get-started/quickstart/fork-a-repo)
- ### Clone forked copy to local computer
[How to: GitHub clone](https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository)
- ### Install conda
[Recommended conda distribution: Miniforge3](https://github.com/conda-forge/miniforge#miniforge3)
_[Apple silicon Mac](https://support.apple.com/en-us/HT211814): choose Miniforge3 native to the [ARM64 architecture](https://www.anaconda.com/blog/new-release-anaconda-distribution-now-supporting-m1) instead of [Intel x86](https://en.wikipedia.org/wiki/X86)._
- ### Create a new conda environment
``` $ cd <repo root directory with environment.yml> ```\
``` $ mamba env create -f environment.yml ```\
``` $ conda activate somata ```
_You may also [install `somata` in an existing environment](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#using-pip-in-an-environment) by skipping this step._
- ### Install somata as a package in development mode
``` $ cd <repo root directory with setup.py> ```\
``` $ pip install -e . --config-settings editable_mode=compat ```
_[What is: Editable Installs](https://setuptools.pypa.io/en/latest/userguide/development_mode.html)_
- ### Configure IDEs to use the conda environment
[How to: Configure an existing conda environment](https://www.jetbrains.com/help/pycharm/conda-support-creating-conda-virtual-environment.html#existing-conda-environment)
---
## Basic state-space models
`somata`, much like a neuron body supported by dendrites, is built on a set of basic state-space models introduced as class objects.
The motivations are to:
- develop a standardized format to store model parameters of state-space equations
- override Python dunder methods so `__repr__` and `__str__` return something useful
- define arithmetic-like operations such as `A + B` and `A * B`
- emulate `numpy.array()` operations including `.append()`
- implement inference algorithms like Kalman filtering and parameter update (m-step) equations as callable class methods
At present, and in the near future, `somata` will be focused on **time-invariant Gaussian linear dynamical systems**.
This limit on models we consider simplifies basic models to avoid nested classes such as `transition_model` and
`observation_model`, at the cost of restricting `somata` to classical algorithms with only some extensions to
Bayesian inference and learning. This is a deliberate choice to allow easier, faster, and cleaner applications of
`somata` in neural data analysis, instead of to provide a full-fledged statistical inference package.
---
### _class_ StateSpaceModel
```python
somata.StateSpaceModel(components=None, F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)
```
`StateSpaceModel` is the parent class of basic state-space models. The corresponding Gaussian linear dynamical system is:
$$
\mathbf{x}_ {t} = \mathbf{F}\mathbf{x}_{t-1} + \boldsymbol{\eta}_t, \boldsymbol{\eta}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q})
$$
$$
\mathbf{y}_ {t} = \mathbf{G}\mathbf{x}_{t} + \boldsymbol{\epsilon}_t, \boldsymbol{\epsilon}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R})
$$
$$
\mathbf{x}_0 \sim \mathcal{N}(\mathbf{\mu}_0, \mathbf{Q}_0)
$$
Most of the constructor input arguments correspond to these model parameters, which are stored as instance attributes.
There are two additional arguments: `Fs` and `components`.
`Fs` is the sampling frequency of observed data `y`.
`components` is a list of independent components underlying the hidden states $\mathbf{x}$. The independent components are
assumed to appear in block-diagonal form in the state equation. For example, $\mathbf{x}_t$ might have two independent autoregressive
models (AR) of order 1, and the observation matrix is simply $[1, 1]$ that sums these two components. In this case, `components`
would be a list of two AR1 models. Note that each element of the `components` list should be an instance of one of basic model
class objects. To break the recursion, often the `components` attribute of a component is set to `None`, i.e.,
`components[0].components = None`.
1. `StateSpaceModel.__repr__()`
The double-under method `__repr__()` is overwritten to provide some unique identification info:
```python
>>> s1 = StateSpaceModel()
>>> s1
Ssm(0)<f4c0>
```
where the number inside parenthesis indicates **the number of components** (the `ncomp` attribute) in the model, and the four-digits in `<>` are the last four digits of the memory address of the object instance.
2. `StateSpaceModel.__str__()`
The double-under method `__str__()` is overwritten so `print()` returns useful info:
```python
>>> print(s1)
<Ssm object at 0x102a8f4c0>
nstate = 0 ncomp = 0
nchannel = 0 ntime = 0
nmodel = 1
components = None
F .shape = None Q .shape = None
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = None
y .shape = None Fs = None
```
3. Model _stacking_ in `StateSpaceModel`
In many applications, there are several possible parameter values for a given state-space model structure. Instead of duplicating
the same values in multiple instances, somata uses _stacking_ to store multiple model values in the same object instance. Stackable
model parameters are `F, Q, mu0, Q0, G, R`. For example:
```python
>>> s1 = StateSpaceModel(F=1, Q=2)
>>> s2 = StateSpaceModel(F=2, Q=2)
>>> print(s1)
<Ssm object at 0x11fd7bfa0>
nstate = 1 ncomp = 0
nchannel = 0 ntime = 0
nmodel = 1
components = None
F .shape = (1, 1) Q .shape = (1, 1)
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = None
y .shape = None Fs = None
>>> print(s2)
<Ssm object at 0x102acc130>
nstate = 1 ncomp = 0
nchannel = 0 ntime = 0
nmodel = 1
components = None
F .shape = (1, 1) Q .shape = (1, 1)
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = None
y .shape = None Fs = None
>>> s3 = s1+s2
>>> print(s3)
<Ssm object at 0x102acc280>
nstate = 1 ncomp = 0
nchannel = 0 ntime = 0
nmodel = 2
components = None
F .shape = (1, 1, 2) Q .shape = (1, 1)
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = None
y .shape = None Fs = None
```
Invoking the arithmetic operator `+` stacks the two instances `s1` and `s2` into a new instance, where the third dimension of the
`F` attribute is now `2`, with the two values from `s1` and `s2`. The `nmodel` attribute is also updated to `2`.
```python
>>> s3.F
array([[[1., 2.]]])
```
Notice how the third dimension of the `Q` attribute is still `None`. This is because the `+` operator has a built-in duplication check
such that the identical model parameters will not be stacked. This behavior of `__add__` and `__radd__` generalizes to all model parameters, and it is convenient when bootstrapping or testing different parameter values during neural data analysis. Manual stacking of a particular
model parameter is also possible with `.stack_attr()`.
4. Model _expanding_ in `StateSpaceModel`
Similar to _stacking_, there is a related concept called _expanding_. Expanding a model is useful when we want to permutate multiple model
parameters each with several possible values. For example:
```python
>>> s1 = StateSpaceModel(F=1, Q=3, R=5)
>>> s2 = StateSpaceModel(F=2, Q=4, R=5)
>>> print(s1+s2)
<Ssm object at 0x1059626b0>
nstate = 1 ncomp = 0
nchannel = 1 ntime = 0
nmodel = 2
components = None
F .shape = (1, 1, 2) Q .shape = (1, 1, 2)
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = (1, 1)
y .shape = None Fs = None
>>> s3 = s1*s2
>>> print(s3)
<Ssm object at 0x1059626b0>
nstate = 1 ncomp = 0
nchannel = 1 ntime = 0
nmodel = 4
components = None
F .shape = (1, 1, 4) Q .shape = (1, 1, 4)
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = (1, 1)
y .shape = None Fs = None
>>> s3.F
array([[[1., 1., 2., 2.]]])
>>> s3.Q
array([[[3., 4., 3., 4.]]])
```
Multiplying two `StateSpaceModel` instances with more than one differing model parameters results in expanding these parameters into all possible combinations while keeping other identical attributes intact.
5. Arrays of `StateSpaceModel`
Building on _stacking_ and _expanding_, we can easily form an array of `StateSpaceModel` instances using `.stack_to_array()`:
```python
>>> s_array = s3.stack_to_array()
>>> s_array
array([Ssm(0)<4460>, Ssm(0)<4430>, Ssm(0)<4520>, Ssm(0)<4580>],
dtype=object)
```
Note that a `StateSpaceModel` array is duck-typing with a Python `list`, which means one can also form a valid array with `[s1, s2]`.
6. `StateSpaceModel.__len__()`
Invoking `len()` returns the number of stacked models:
```python
>>> len(s2)
1
>>> len(s3)
4
```
7. `StateSpaceModel.append()`
Another useful class method on `StateSpaceModel` is `.append()`. As one would expect, appending a model to another results in
combining them in block-diagonal form in the state equation. Compatibility checks happen in the background to make sure no conflict
exists on the respective observation equations and observed data, if any.
```python
>>> s1 = StateSpaceModel(F=1, Q=3, R=5)
>>> s2 = StateSpaceModel(F=2, Q=4, R=5)
>>> s1.append(s2)
>>> print(s1)
<Ssm object at 0x1057cb4c0>
nstate = 2 ncomp = 0
nchannel = 1 ntime = 0
nmodel = 1
components = None
F .shape = (2, 2) Q .shape = (2, 2)
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = (1, 1)
y .shape = None Fs = None
>>> s1.F
array([[1., 0.],
[0., 2.]])
>>> s1.Q
array([[3., 0.],
[0., 4.]])
```
Notice that the `nstate` attribute is now updated to `2`, which is different from the `+` operator that updates the `nmodel` attribute to `2`.
8. Inference and learning with `StateSpaceModel`
Two different implementations of Kalman filtering and fixed-interval smoothing are callable class methods:
```python
.kalman_filt_smooth(y=None, R_weights=None, return_dict=False, EM=False, skip_interp=True, seterr=None)
.dejong_filt_smooth(y=None, R_weights=None, return_dict=False, EM=False, skip_interp=True, seterr=None)
```
With an array of `StateSpaceModel`, one can easily run Kalman filtering and smoothing on all array elements with multithreading using the **static** method `.par_kalman()`:
```python
.par_kalman(ssm_array, y=None, method='kalman', R_weights=None, skip_interp=True, return_dict=False)
```
M-step updates are organized using `m_estimate()` that will recurse into each element of the `components` list and use
the appropriate m-step update methods associated with different types of state-space models.
**Below we explain three kinds of basic state-space models currently supported in somata.**
---
### _class_ OscillatorModel(StateSpaceModel)
```python
somata.OscillatorModel(a=None, freq=None, w=None, sigma2=None, add_dc=False,
components='Osc', F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)
```
`OscillatorModel` is a child class of `StateSpaceModel`, which means it inherits all the class methods explained above. It has a particular form of the state-space model:
$$
\begin{bmatrix}x_{t, 1}\newline x_{t, 2}\end{bmatrix} = \mathbf{F}\begin{bmatrix}x_{t-1, 1}\newline x_{t-1, 2}\end{bmatrix} + \mathbf{\eta}_t, \mathbf{\eta}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q})
$$
$$
\mathbf{y}_ {t} = \mathbf{G}\begin{bmatrix}x_{t, 1}\newline x_{t, 2}\end{bmatrix} + \mathbf{\epsilon}_t, \mathbf{\epsilon}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R})
$$
$$
\begin{bmatrix}x_{0, 1}\newline x_{0, 2}\end{bmatrix} \sim \mathcal{N}(\mathbf{\mu}_0, \mathbf{Q}_0)
$$
$$
\mathbf{F} = a\begin{bmatrix}\cos\omega & -\sin\omega\newline \sin\omega & \cos\omega\end{bmatrix}, \mathbf{Q} = \begin{bmatrix}\sigma^2 & 0\newline 0 & \sigma^2\end{bmatrix}, \mathbf{G} = \begin{bmatrix}1 & 0 \end{bmatrix}
$$
To create a simple oscillator model with rotation frequency $15$ Hz (under $100$ Hz sampling frequency) and damping factor $0.9$:
```python
>>> o1 = OscillatorModel(a=0.9, freq=15, Fs=100)
>>> o1
Osc(1)<81f0>
>>> print(o1)
<Osc object at 0x1058081f0>
nstate = 2 ncomp = 1
nchannel = 0 ntime = 0
nmodel = 1
components = [Osc(0)<4b50>]
F .shape = (2, 2) Q .shape = (2, 2)
mu0.shape = (2, 1) Q0 .shape = (2, 2)
G .shape = (1, 2) R .shape = None
y .shape = None Fs = 100.0 Hz
damping a = [0.9]
freq Hz = [15.]
sigma2 = [3.]
obs noise R = None
dc index = None
```
Notice the `components` attribute auto-populates with a spaceholder `OscillatorModel` instance, which is different from the `o1` instance
as can be recognized by different memory addresses. State noise variance $\sigma^2$ defaults to $3$ when not specified and can be changed
with the `sigma2` argument to the constructor method.
### _class_ AutoRegModel(StateSpaceModel)
```python
somata.AutoRegModel(coeff=None, sigma2=None,
components='Arn', F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)
```
`AutoRegModel` is a child class of `StateSpaceModel`, which means it inherits all the class methods explained above. It has a particular form of the state-space model. For example, an auto-regressive model of order 3 can be expressed as:
$$
\begin{bmatrix}x_{t}\newline x_{t-1}\newline x_{t-2}\end{bmatrix} = \mathbf{F}\begin{bmatrix}x_{t-1}\newline x_{t-2}\newline x_{t-3}\end{bmatrix} + \mathbf{\eta}_t, \mathbf{\eta}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q})
$$
$$
\mathbf{y}_ {t} = \mathbf{G}\begin{bmatrix}x_{t}\newline x_{t-1}\newline x_{t-2}\end{bmatrix} + \mathbf{\epsilon}_t, \mathbf{\epsilon}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R})
$$
$$
\begin{bmatrix}x_{0}\newline x_{-1}\newline x_{-2}\end{bmatrix} \sim \mathcal{N}(\mathbf{\mu}_0, \mathbf{Q}_0)
$$
$$
\mathbf{F} = \begin{bmatrix} a_1 & a_2 & a_3 \newline 1 & 0 & 0 \newline 0 & 1 & 0 \end{bmatrix}, \mathbf{Q} = \begin{bmatrix}\sigma^2 & 0 & 0\newline 0 & 0 & 0\newline 0 & 0 & 0\end{bmatrix}, \mathbf{G} = \begin{bmatrix}1 & 0 & 0\end{bmatrix}
$$
To create an AR3 model with parameters $a_1=0.5, a_2=0.3, a_3=0.1$ and $\sigma^2=1$:
```python
>>> a1 = AutoRegModel(coeff=[0.5,0.3,0.1], sigma2=1)
>>> a1
Arn=3<24d0>
>>> print(a1)
<Arn object at 0x1035524d0>
nstate = 3 ncomp = 1
nchannel = 0 ntime = 0
nmodel = 1
components = [Arn=3<2680>]
F .shape = (3, 3) Q .shape = (3, 3)
mu0.shape = (3, 1) Q0 .shape = (3, 3)
G .shape = (1, 3) R .shape = None
y .shape = None Fs = None
AR order = [3]
AR coeff = ([0.5 0.3 0.1])
sigma2 = [1.]
```
Note that `__repr__()` is slightly different for `AutoRegModel`, since the key information is not how many components but rather the AR order. We display the order of the auto-regressive model with an `=` sign as shown above instead of showing the number of components in
`()` as for `OscillatorModel` and `StateSpaceModel`.
### _class_ GeneralSSModel(StateSpaceModel)
```python
somata.GeneralSSModel(components='Gen', F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)
```
`GeneralSSModel` is a child class of `StateSpaceModel`, which means it inherits all the class methods explained above. The same general Gaussian linear dynamic system as before is followed:
$$
\mathbf{x}_ t = \mathbf{F}\mathbf{x}_{t-1} + \boldsymbol{\eta}_t, \boldsymbol{\eta}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{Q})
$$
$$
\mathbf{y}_ {t} = \mathbf{G}\mathbf{x}_{t} + \boldsymbol{\epsilon}_t, \boldsymbol{\epsilon}_t \sim \mathcal{N}(\mathbf{0}, \mathbf{R})
$$
$$
\mathbf{x}_0 \sim \mathcal{N}(\mathbf{\mu}_0, \mathbf{Q}_0)
$$
`GeneralSSModel` is added to somata so that one can perform the most general Gaussian updates for a state-space model without special structures as specified in `OscillatorModel` and `AutoRegModel`. In other words, with non-sparse structures in the model parameters
`F, Q, Q0, G, R`. To create a simple general state-space model:
```python
>>> g1 = GeneralSSModel(F=[[1,2],[3,4]])
>>> g1
Gen(1)<2440>
>>> print(g1)
<Gen object at 0x103552440>
nstate = 2 ncomp = 1
nchannel = 0 ntime = 0
nmodel = 1
components = [Gen(0)<2710>]
F .shape = (2, 2) Q .shape = None
mu0.shape = None Q0 .shape = None
G .shape = None R .shape = None
y .shape = None Fs = None
```
### For more in-depth working examples with the basic models in somata
Look at the demo script [basic_models_demo_08302023.py](examples/basic_models_demo_08302023.py) and execute the code line by line to get familiar with class objects and methods of somata basic models.
---
## Advanced neural oscillator methods
1. [Oscillator Model Learning](#1-oscillator-model-learning)
2. [Phase Amplitude Coupling Estimation](#2-phase-amplitude-coupling-estimation)
3. [Iterative Oscillator Algorithm](#3-iterative-oscillator-algorithm)
4. [Switching State-Space Inference](#4-switching-state-space-inference)
5. [Multi-channel Oscillator Component Analysis](#5-multi-channel-oscillator-component-analysis)
6. [State-Space Event Related Potential](#6-state-space-event-related-potential)
7. [Dynamic Source Localization](#7-dynamic-source-localization)
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Functional-success.svg?logo=Python">
</picture>
### 1. Oscillator Model Learning
For fitting data with oscillator models, it boils down to three steps:
- Initialize an oscillator model object
- Perform state estimation, i.e., E-step
- Update model parameters, i.e., M-step
Given some time series `data`, we can fit an oscillator to the data using the expectation-maximization (EM) algorithm.
```python
from somata.basic_models import OscillatorModel as Osc
o1 = Osc(freq=1, Fs=100, y=data) # create an oscillator object instance
_ = [o1.m_estimate(**o1.kalman_filt_smooth(EM=True))for x in range(50)] # run 50 steps of EM
```
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Missing-critical.svg?logo=Python">
</picture>
### 2. Phase Amplitude Coupling Estimation
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Functional-success.svg?logo=Python">
</picture>
### 3. Iterative Oscillator Algorithm
For a well-commented example script, see [IterOsc_example.py](examples/IterOsc_example.py).
_**N.B.:** We recommend downsampling to 120 Hz or less, depending on the oscillations present in your data. Highly oversampled data will make it more difficult to identify oscillatory components, increase the computational time, and could also introduce high frequency noise._
One major goal of this method was to produce an algorithm that requires minimal user intervention, if any. This algorithm is designed to fit well automatically in most situations, but there will still be some data sets where it does not fit well without intervention. We recommend starting with the algorithm as is, but in the case of poor fitting, we suggest the following modifications:
1. If the model does not choose the correct number of oscillations, we recommend looking at all fitted models and selecting the best fitting model based on other selection criteria or using your best judgement. You can also choose a subset of well-fitted oscillations and run `kalman_filt_smooth()` to estimate oscillations using those fitted parameters.
2. This algorithm assumes stationary parameters, and therefore a stationary signal. Although the Kalman smoothing allows the model to work with some time-varying signal, the success of the method depends on the strength and duration of the signal components. The weaker and more brief the time-varying component is, the more poorly the model will capture it, if at all. We recommend decreasing the length of your window until you have a more stationary signal.
When using this module, please cite the following [paper](https://www.biorxiv.org/content/10.1101/2022.10.30.514422):
Beck, A. M., He, M., Gutierrez, R. G., & Purdon, P. L. (2022). An iterative search algorithm to identify oscillatory dynamics in neurophysiological time series. bioRxiv, 2022-10.
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Functional-success.svg?logo=Python">
</picture>
### 4. Switching State-Space Inference
When using this module, please cite the following [paper](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011395):
He, M., Das, P., Hotan, G., & Purdon, P. L. (2023). Switching state-space modeling of neural signal dynamics. PLOS Computational Biology, 19(8), e1011395.
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Missing-critical.svg?logo=Python">
</picture>
### 5. Multi-channel Oscillator Component Analysis
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Missing-critical.svg?logo=Python">
</picture>
### 6. State-Space Event Related Potential
---
<picture>
<img align="right" src="https://img.shields.io/badge/Status-Functional-success.svg?logo=Python">
</picture>
### 7. Dynamic Source Localization
---
## Authors
Mingjian He, Proloy Das, Amanda Beck, Patrick Purdon
## Citation
Use different citation styles at: https://doi.org/10.5281/zenodo.7242130
## License
SOMATA is licensed under the [BSD 3-Clause Clear license](https://spdx.org/licenses/BSD-3-Clause-Clear.html).\
Copyright © 2023. All rights reserved.
Raw data
{
"_id": null,
"home_page": "https://github.com/mh105/somata",
"name": "somata",
"maintainer": null,
"docs_url": null,
"requires_python": ">=3.8",
"maintainer_email": null,
"keywords": "state-space oscillator time-series",
"author": "Mingjian He",
"author_email": "mh1@stanford.edu",
"download_url": "https://files.pythonhosted.org/packages/84/c3/c0643dcf10ada26e0044d802052c7210eba19c812140d1883503fc2f2c3b/somata-0.4.1.tar.gz",
"platform": null,
"description": "# somata\n\nGithub: https://github.com/mh105/somata\n\n**State-space Oscillator Modeling And Time-series Analysis (SOMATA)** is a Python library for state-space neural signal\nprocessing algorithms developed in the [Purdon Lab](https://purdonlab.stanford.edu).\nBasic state-space models are introduced as class objects for flexible manipulations.\nClassical exact and approximate inference algorithms are implemented and interfaced as class methods.\nAdvanced neural oscillator modeling techniques are brought together to work synergistically.\n\n[](https://github.com/mh105/pot/commits/master)\n[](https://www.python.org/)\n[](https://spdx.org/licenses/BSD-3-Clause-Clear.html)\n[](https://zenodo.org/badge/latestdoi/556083594)\n\n---\n\n## Table of Contents\n* [Requirements](#requirements)\n* [Install](#install)\n* [Basic state-space models](#basic-state-space-models)\n * [StateSpaceModel](#class-statespacemodel)\n * [OscillatorModel](#class-oscillatormodelstatespacemodel)\n * [AutoRegModel](#class-autoregmodelstatespacemodel)\n * [GeneralSSModel](#class-generalssmodelstatespacemodel)\n* [Advanced neural oscillator methods](#advanced-neural-oscillator-methods)\n* [Authors](#authors)\n* [Citation](#citation)\n* [License](#license)\n\n---\n\n## Requirements\n[`somata`](https://pypi.org/project/somata/) is built on [`numpy`](https://numpy.org) arrays for computations. [`joblib`](https://joblib.readthedocs.io/en/stable/) is used for multithreading. \nAdditional dependencies include [`scipy`](https://scipy.org), [`matplotlib`](https://matplotlib.org), and [`spectrum`](https://pyspectrum.readthedocs.io/en/latest/index.html).\nThe source localization module also requires [`pytorch`](https://pytorch.org) and [`MNE-python`](https://mne.tools/stable/index.html).\nFull requirements for each release version will be updated under \n[`install_requires`](https://setuptools.pypa.io/en/latest/userguide/dependency_management.html#platform-specific-dependencies) in the [`setup.cfg`](setup.cfg) file. \nIf the [`environment.yml`](environment.yml) file is used to [create a new conda environment](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#creating-an-environment-from-an-environment-yml-file), \nall and only the required packages will be installed.\n\n## Install\n```\n$ pip install somata\n```\n\n### conda-forge channel\nWhile [`pip install`](https://pip.pypa.io/en/stable/cli/pip_install/) usually works, [an alternative way](https://pythonspeed.com/articles/conda-vs-pip/) to install `somata` is through the [conda-forge](https://conda-forge.org/docs/index.html) [channel](https://docs.conda.io/projects/conda/en/latest/user-guide/concepts/channels.html#what-is-a-conda-channel), which utilizes [continuous integration (CI)](https://conda-forge.org/docs/user/ci-skeleton.html) [across OS platforms](https://conda-forge.org/docs/user/introduction.html#why-conda-forge).\nThis means [conda-forge packages](https://conda-forge.org/feedstock-outputs/index.html) are more [compatible with each other](https://conda-forge.org/docs/maintainer/adding_pkgs.html#avoid-external-dependencies) compared to [Python Package Index (PyPI) packages](https://pypi.org) installed via [`pip` by default](https://packaging.python.org/en/latest/tutorials/installing-packages/#installing-from-pypi).\nWhen `somata` is installed into an [existing conda environment](https://docs.conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#viewing-a-list-of-your-environments), unmet dependencies are automatically searched, downloaded, and installed from the same repository of packages containing `somata`. \nIf `pip install somata` fails to resolve some dependencies, the [conda-forge somata](https://github.com/conda-forge/somata-feedstock) [feedstock](https://github.com/conda-forge/conda-feedstock#terminology) can be used to install:\n```\n$ conda install somata -c conda-forge\n```\n\n### torch requirement\nIf the [`pytorch`](https://pytorch.org) dependency is not resolved correctly for your [OS](https://whatsmyos.com), \nfirst [install `pytorch` manually](https://pytorch.org/get-started/locally/) in a [conda environment](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html) that you want to install `somata` in, and then rerun either of the above two lines to install `somata`.\n\n### (For development only)\n\n- ### Fork this repo to personal git\n [How to: GitHub fork](https://docs.github.com/en/get-started/quickstart/fork-a-repo) \n\n- ### Clone forked copy to local computer\n [How to: GitHub clone](https://docs.github.com/en/repositories/creating-and-managing-repositories/cloning-a-repository)\n\n- ### Install conda\n [Recommended conda distribution: Miniforge3](https://github.com/conda-forge/miniforge#miniforge3)\n\n _[Apple silicon Mac](https://support.apple.com/en-us/HT211814): choose Miniforge3 native to the [ARM64 architecture](https://www.anaconda.com/blog/new-release-anaconda-distribution-now-supporting-m1) instead of [Intel x86](https://en.wikipedia.org/wiki/X86)._\n\n- ### Create a new conda environment\n ``` $ cd <repo root directory with environment.yml> ```\\\n ``` $ mamba env create -f environment.yml ```\\\n ``` $ conda activate somata ```\n\n _You may also [install `somata` in an existing environment](https://conda.io/projects/conda/en/latest/user-guide/tasks/manage-environments.html#using-pip-in-an-environment) by skipping this step._\n\n- ### Install somata as a package in development mode\n ``` $ cd <repo root directory with setup.py> ```\\\n ``` $ pip install -e . --config-settings editable_mode=compat ```\n\n _[What is: Editable Installs](https://setuptools.pypa.io/en/latest/userguide/development_mode.html)_\n\n- ### Configure IDEs to use the conda environment\n [How to: Configure an existing conda environment](https://www.jetbrains.com/help/pycharm/conda-support-creating-conda-virtual-environment.html#existing-conda-environment)\n\n---\n\n## Basic state-space models\n`somata`, much like a neuron body supported by dendrites, is built on a set of basic state-space models introduced as class objects.\n\nThe motivations are to:\n- develop a standardized format to store model parameters of state-space equations\n- override Python dunder methods so `__repr__` and `__str__` return something useful\n- define arithmetic-like operations such as `A + B` and `A * B`\n- emulate `numpy.array()` operations including `.append()`\n- implement inference algorithms like Kalman filtering and parameter update (m-step) equations as callable class methods\n\nAt present, and in the near future, `somata` will be focused on **time-invariant Gaussian linear dynamical systems**.\nThis limit on models we consider simplifies basic models to avoid nested classes such as `transition_model` and\n`observation_model`, at the cost of restricting `somata` to classical algorithms with only some extensions to\nBayesian inference and learning. This is a deliberate choice to allow easier, faster, and cleaner applications of\n`somata` in neural data analysis, instead of to provide a full-fledged statistical inference package.\n\n---\n\n### _class_ StateSpaceModel\n```python\nsomata.StateSpaceModel(components=None, F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)\n```\n`StateSpaceModel` is the parent class of basic state-space models. The corresponding Gaussian linear dynamical system is:\n\n$$\n\\mathbf{x}_ {t} = \\mathbf{F}\\mathbf{x}_{t-1} + \\boldsymbol{\\eta}_t, \\boldsymbol{\\eta}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{Q})\n$$\n\n$$\n\\mathbf{y}_ {t} = \\mathbf{G}\\mathbf{x}_{t} + \\boldsymbol{\\epsilon}_t, \\boldsymbol{\\epsilon}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{R})\n$$\n\n$$\n\\mathbf{x}_0 \\sim \\mathcal{N}(\\mathbf{\\mu}_0, \\mathbf{Q}_0)\n$$\n\nMost of the constructor input arguments correspond to these model parameters, which are stored as instance attributes.\nThere are two additional arguments: `Fs` and `components`.\n\n`Fs` is the sampling frequency of observed data `y`.\n\n`components` is a list of independent components underlying the hidden states $\\mathbf{x}$. The independent components are\nassumed to appear in block-diagonal form in the state equation. For example, $\\mathbf{x}_t$ might have two independent autoregressive\nmodels (AR) of order 1, and the observation matrix is simply $[1, 1]$ that sums these two components. In this case, `components`\nwould be a list of two AR1 models. Note that each element of the `components` list should be an instance of one of basic model\nclass objects. To break the recursion, often the `components` attribute of a component is set to `None`, i.e.,\n`components[0].components = None`.\n\n1. `StateSpaceModel.__repr__()`\n\nThe double-under method `__repr__()` is overwritten to provide some unique identification info:\n\n```python\n>>> s1 = StateSpaceModel()\n>>> s1\nSsm(0)<f4c0>\n```\nwhere the number inside parenthesis indicates **the number of components** (the `ncomp` attribute) in the model, and the four-digits in `<>` are the last four digits of the memory address of the object instance.\n\n2. `StateSpaceModel.__str__()`\n\nThe double-under method `__str__()` is overwritten so `print()` returns useful info:\n```python\n>>> print(s1)\n<Ssm object at 0x102a8f4c0>\n nstate = 0 ncomp = 0\n nchannel = 0 ntime = 0\n nmodel = 1\n components = None\n F .shape = None Q .shape = None\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = None\n y .shape = None Fs = None\n```\n\n3. Model _stacking_ in `StateSpaceModel`\n\nIn many applications, there are several possible parameter values for a given state-space model structure. Instead of duplicating\nthe same values in multiple instances, somata uses _stacking_ to store multiple model values in the same object instance. Stackable\nmodel parameters are `F, Q, mu0, Q0, G, R`. For example:\n\n```python\n>>> s1 = StateSpaceModel(F=1, Q=2)\n>>> s2 = StateSpaceModel(F=2, Q=2)\n>>> print(s1)\n<Ssm object at 0x11fd7bfa0>\n nstate = 1 ncomp = 0\n nchannel = 0 ntime = 0\n nmodel = 1\n components = None\n F .shape = (1, 1) Q .shape = (1, 1)\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = None\n y .shape = None Fs = None\n\n>>> print(s2)\n<Ssm object at 0x102acc130>\n nstate = 1 ncomp = 0\n nchannel = 0 ntime = 0\n nmodel = 1\n components = None\n F .shape = (1, 1) Q .shape = (1, 1)\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = None\n y .shape = None Fs = None\n\n>>> s3 = s1+s2\n>>> print(s3)\n<Ssm object at 0x102acc280>\n nstate = 1 ncomp = 0\n nchannel = 0 ntime = 0\n nmodel = 2\n components = None\n F .shape = (1, 1, 2) Q .shape = (1, 1)\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = None\n y .shape = None Fs = None\n```\nInvoking the arithmetic operator `+` stacks the two instances `s1` and `s2` into a new instance, where the third dimension of the\n`F` attribute is now `2`, with the two values from `s1` and `s2`. The `nmodel` attribute is also updated to `2`.\n```python\n>>> s3.F\narray([[[1., 2.]]])\n```\nNotice how the third dimension of the `Q` attribute is still `None`. This is because the `+` operator has a built-in duplication check\nsuch that the identical model parameters will not be stacked. This behavior of `__add__` and `__radd__` generalizes to all model parameters, and it is convenient when bootstrapping or testing different parameter values during neural data analysis. Manual stacking of a particular\nmodel parameter is also possible with `.stack_attr()`.\n\n4. Model _expanding_ in `StateSpaceModel`\n\nSimilar to _stacking_, there is a related concept called _expanding_. Expanding a model is useful when we want to permutate multiple model\nparameters each with several possible values. For example:\n\n```python\n>>> s1 = StateSpaceModel(F=1, Q=3, R=5)\n>>> s2 = StateSpaceModel(F=2, Q=4, R=5)\n>>> print(s1+s2)\n<Ssm object at 0x1059626b0>\n nstate = 1 ncomp = 0\n nchannel = 1 ntime = 0\n nmodel = 2\n components = None\n F .shape = (1, 1, 2) Q .shape = (1, 1, 2)\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = (1, 1)\n y .shape = None Fs = None\n\n>>> s3 = s1*s2\n>>> print(s3)\n<Ssm object at 0x1059626b0>\n nstate = 1 ncomp = 0\n nchannel = 1 ntime = 0\n nmodel = 4\n components = None\n F .shape = (1, 1, 4) Q .shape = (1, 1, 4)\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = (1, 1)\n y .shape = None Fs = None\n\n>>> s3.F\narray([[[1., 1., 2., 2.]]])\n>>> s3.Q\narray([[[3., 4., 3., 4.]]])\n```\nMultiplying two `StateSpaceModel` instances with more than one differing model parameters results in expanding these parameters into all possible combinations while keeping other identical attributes intact.\n\n5. Arrays of `StateSpaceModel`\n\nBuilding on _stacking_ and _expanding_, we can easily form an array of `StateSpaceModel` instances using `.stack_to_array()`:\n\n```python\n>>> s_array = s3.stack_to_array()\n>>> s_array\narray([Ssm(0)<4460>, Ssm(0)<4430>, Ssm(0)<4520>, Ssm(0)<4580>],\n dtype=object)\n```\n\nNote that a `StateSpaceModel` array is duck-typing with a Python `list`, which means one can also form a valid array with `[s1, s2]`.\n\n6. `StateSpaceModel.__len__()`\n\nInvoking `len()` returns the number of stacked models:\n\n```python\n>>> len(s2)\n1\n>>> len(s3)\n4\n```\n\n7. `StateSpaceModel.append()`\n\nAnother useful class method on `StateSpaceModel` is `.append()`. As one would expect, appending a model to another results in\ncombining them in block-diagonal form in the state equation. Compatibility checks happen in the background to make sure no conflict\nexists on the respective observation equations and observed data, if any.\n\n```python\n>>> s1 = StateSpaceModel(F=1, Q=3, R=5)\n>>> s2 = StateSpaceModel(F=2, Q=4, R=5)\n>>> s1.append(s2)\n>>> print(s1)\n<Ssm object at 0x1057cb4c0>\n nstate = 2 ncomp = 0\n nchannel = 1 ntime = 0\n nmodel = 1\n components = None\n F .shape = (2, 2) Q .shape = (2, 2)\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = (1, 1)\n y .shape = None Fs = None\n\n>>> s1.F\narray([[1., 0.],\n [0., 2.]])\n>>> s1.Q\narray([[3., 0.],\n [0., 4.]])\n```\n\nNotice that the `nstate` attribute is now updated to `2`, which is different from the `+` operator that updates the `nmodel` attribute to `2`.\n\n8. Inference and learning with `StateSpaceModel`\n\nTwo different implementations of Kalman filtering and fixed-interval smoothing are callable class methods:\n\n```python\n.kalman_filt_smooth(y=None, R_weights=None, return_dict=False, EM=False, skip_interp=True, seterr=None)\n\n.dejong_filt_smooth(y=None, R_weights=None, return_dict=False, EM=False, skip_interp=True, seterr=None)\n```\n\nWith an array of `StateSpaceModel`, one can easily run Kalman filtering and smoothing on all array elements with multithreading using the **static** method `.par_kalman()`:\n\n```python\n.par_kalman(ssm_array, y=None, method='kalman', R_weights=None, skip_interp=True, return_dict=False)\n```\n\nM-step updates are organized using `m_estimate()` that will recurse into each element of the `components` list and use\nthe appropriate m-step update methods associated with different types of state-space models.\n\n**Below we explain three kinds of basic state-space models currently supported in somata.**\n\n---\n### _class_ OscillatorModel(StateSpaceModel)\n```python\nsomata.OscillatorModel(a=None, freq=None, w=None, sigma2=None, add_dc=False,\n components='Osc', F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)\n```\n`OscillatorModel` is a child class of `StateSpaceModel`, which means it inherits all the class methods explained above. It has a particular form of the state-space model:\n\n$$\n\\begin{bmatrix}x_{t, 1}\\newline x_{t, 2}\\end{bmatrix} = \\mathbf{F}\\begin{bmatrix}x_{t-1, 1}\\newline x_{t-1, 2}\\end{bmatrix} + \\mathbf{\\eta}_t, \\mathbf{\\eta}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{Q})\n$$\n\n$$\n\\mathbf{y}_ {t} = \\mathbf{G}\\begin{bmatrix}x_{t, 1}\\newline x_{t, 2}\\end{bmatrix} + \\mathbf{\\epsilon}_t, \\mathbf{\\epsilon}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{R})\n$$\n\n$$\n\\begin{bmatrix}x_{0, 1}\\newline x_{0, 2}\\end{bmatrix} \\sim \\mathcal{N}(\\mathbf{\\mu}_0, \\mathbf{Q}_0)\n$$\n\n$$\n\\mathbf{F} = a\\begin{bmatrix}\\cos\\omega & -\\sin\\omega\\newline \\sin\\omega & \\cos\\omega\\end{bmatrix}, \\mathbf{Q} = \\begin{bmatrix}\\sigma^2 & 0\\newline 0 & \\sigma^2\\end{bmatrix}, \\mathbf{G} = \\begin{bmatrix}1 & 0 \\end{bmatrix}\n$$\n\nTo create a simple oscillator model with rotation frequency $15$ Hz (under $100$ Hz sampling frequency) and damping factor $0.9$:\n\n```python\n>>> o1 = OscillatorModel(a=0.9, freq=15, Fs=100)\n>>> o1\nOsc(1)<81f0>\n>>> print(o1)\n<Osc object at 0x1058081f0>\n nstate = 2 ncomp = 1\n nchannel = 0 ntime = 0\n nmodel = 1\n components = [Osc(0)<4b50>]\n F .shape = (2, 2) Q .shape = (2, 2)\n mu0.shape = (2, 1) Q0 .shape = (2, 2)\n G .shape = (1, 2) R .shape = None\n y .shape = None Fs = 100.0 Hz\n damping a = [0.9]\n freq Hz = [15.]\n sigma2 = [3.]\n obs noise R = None\n dc index = None\n```\n\nNotice the `components` attribute auto-populates with a spaceholder `OscillatorModel` instance, which is different from the `o1` instance\nas can be recognized by different memory addresses. State noise variance $\\sigma^2$ defaults to $3$ when not specified and can be changed\nwith the `sigma2` argument to the constructor method.\n\n### _class_ AutoRegModel(StateSpaceModel)\n```python\nsomata.AutoRegModel(coeff=None, sigma2=None,\n components='Arn', F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)\n```\n`AutoRegModel` is a child class of `StateSpaceModel`, which means it inherits all the class methods explained above. It has a particular form of the state-space model. For example, an auto-regressive model of order 3 can be expressed as:\n\n$$\n\\begin{bmatrix}x_{t}\\newline x_{t-1}\\newline x_{t-2}\\end{bmatrix} = \\mathbf{F}\\begin{bmatrix}x_{t-1}\\newline x_{t-2}\\newline x_{t-3}\\end{bmatrix} + \\mathbf{\\eta}_t, \\mathbf{\\eta}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{Q})\n$$\n\n$$\n\\mathbf{y}_ {t} = \\mathbf{G}\\begin{bmatrix}x_{t}\\newline x_{t-1}\\newline x_{t-2}\\end{bmatrix} + \\mathbf{\\epsilon}_t, \\mathbf{\\epsilon}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{R})\n$$\n\n$$\n\\begin{bmatrix}x_{0}\\newline x_{-1}\\newline x_{-2}\\end{bmatrix} \\sim \\mathcal{N}(\\mathbf{\\mu}_0, \\mathbf{Q}_0)\n$$\n\n$$\n\\mathbf{F} = \\begin{bmatrix} a_1 & a_2 & a_3 \\newline 1 & 0 & 0 \\newline 0 & 1 & 0 \\end{bmatrix}, \\mathbf{Q} = \\begin{bmatrix}\\sigma^2 & 0 & 0\\newline 0 & 0 & 0\\newline 0 & 0 & 0\\end{bmatrix}, \\mathbf{G} = \\begin{bmatrix}1 & 0 & 0\\end{bmatrix}\n$$\n\nTo create an AR3 model with parameters $a_1=0.5, a_2=0.3, a_3=0.1$ and $\\sigma^2=1$:\n\n```python\n>>> a1 = AutoRegModel(coeff=[0.5,0.3,0.1], sigma2=1)\n>>> a1\nArn=3<24d0>\n>>> print(a1)\n<Arn object at 0x1035524d0>\n nstate = 3 ncomp = 1\n nchannel = 0 ntime = 0\n nmodel = 1\n components = [Arn=3<2680>]\n F .shape = (3, 3) Q .shape = (3, 3)\n mu0.shape = (3, 1) Q0 .shape = (3, 3)\n G .shape = (1, 3) R .shape = None\n y .shape = None Fs = None\n AR order = [3]\n AR coeff = ([0.5 0.3 0.1])\n sigma2 = [1.]\n```\n\nNote that `__repr__()` is slightly different for `AutoRegModel`, since the key information is not how many components but rather the AR order. We display the order of the auto-regressive model with an `=` sign as shown above instead of showing the number of components in\n`()` as for `OscillatorModel` and `StateSpaceModel`.\n\n### _class_ GeneralSSModel(StateSpaceModel)\n```python\nsomata.GeneralSSModel(components='Gen', F=None, Q=None, mu0=None, Q0=None, G=None, R=None, y=None, Fs=None)\n```\n`GeneralSSModel` is a child class of `StateSpaceModel`, which means it inherits all the class methods explained above. The same general Gaussian linear dynamic system as before is followed:\n\n$$\n\\mathbf{x}_ t = \\mathbf{F}\\mathbf{x}_{t-1} + \\boldsymbol{\\eta}_t, \\boldsymbol{\\eta}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{Q})\n$$\n\n$$\n\\mathbf{y}_ {t} = \\mathbf{G}\\mathbf{x}_{t} + \\boldsymbol{\\epsilon}_t, \\boldsymbol{\\epsilon}_t \\sim \\mathcal{N}(\\mathbf{0}, \\mathbf{R})\n$$\n\n$$\n\\mathbf{x}_0 \\sim \\mathcal{N}(\\mathbf{\\mu}_0, \\mathbf{Q}_0)\n$$\n\n`GeneralSSModel` is added to somata so that one can perform the most general Gaussian updates for a state-space model without special structures as specified in `OscillatorModel` and `AutoRegModel`. In other words, with non-sparse structures in the model parameters\n`F, Q, Q0, G, R`. To create a simple general state-space model:\n\n```python\n>>> g1 = GeneralSSModel(F=[[1,2],[3,4]])\n>>> g1\nGen(1)<2440>\n>>> print(g1)\n<Gen object at 0x103552440>\n nstate = 2 ncomp = 1\n nchannel = 0 ntime = 0\n nmodel = 1\n components = [Gen(0)<2710>]\n F .shape = (2, 2) Q .shape = None\n mu0.shape = None Q0 .shape = None\n G .shape = None R .shape = None\n y .shape = None Fs = None\n```\n\n### For more in-depth working examples with the basic models in somata\nLook at the demo script [basic_models_demo_08302023.py](examples/basic_models_demo_08302023.py) and execute the code line by line to get familiar with class objects and methods of somata basic models.\n\n---\n\n## Advanced neural oscillator methods\n1. [Oscillator Model Learning](#1-oscillator-model-learning)\n2. [Phase Amplitude Coupling Estimation](#2-phase-amplitude-coupling-estimation)\n3. [Iterative Oscillator Algorithm](#3-iterative-oscillator-algorithm)\n4. [Switching State-Space Inference](#4-switching-state-space-inference)\n5. [Multi-channel Oscillator Component Analysis](#5-multi-channel-oscillator-component-analysis)\n6. [State-Space Event Related Potential](#6-state-space-event-related-potential)\n7. [Dynamic Source Localization](#7-dynamic-source-localization)\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Functional-success.svg?logo=Python\">\n</picture>\n\n### 1. Oscillator Model Learning\n\nFor fitting data with oscillator models, it boils down to three steps:\n - Initialize an oscillator model object\n - Perform state estimation, i.e., E-step\n - Update model parameters, i.e., M-step\n\nGiven some time series `data`, we can fit an oscillator to the data using the expectation-maximization (EM) algorithm.\n```python\nfrom somata.basic_models import OscillatorModel as Osc\no1 = Osc(freq=1, Fs=100, y=data) # create an oscillator object instance\n_ = [o1.m_estimate(**o1.kalman_filt_smooth(EM=True))for x in range(50)] # run 50 steps of EM\n```\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Missing-critical.svg?logo=Python\">\n</picture>\n\n### 2. Phase Amplitude Coupling Estimation\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Functional-success.svg?logo=Python\">\n</picture>\n\n### 3. Iterative Oscillator Algorithm\n\nFor a well-commented example script, see [IterOsc_example.py](examples/IterOsc_example.py).\n\n_**N.B.:** We recommend downsampling to 120 Hz or less, depending on the oscillations present in your data. Highly oversampled data will make it more difficult to identify oscillatory components, increase the computational time, and could also introduce high frequency noise._\n\nOne major goal of this method was to produce an algorithm that requires minimal user intervention, if any. This algorithm is designed to fit well automatically in most situations, but there will still be some data sets where it does not fit well without intervention. We recommend starting with the algorithm as is, but in the case of poor fitting, we suggest the following modifications:\n\n1. If the model does not choose the correct number of oscillations, we recommend looking at all fitted models and selecting the best fitting model based on other selection criteria or using your best judgement. You can also choose a subset of well-fitted oscillations and run `kalman_filt_smooth()` to estimate oscillations using those fitted parameters.\n\n2. This algorithm assumes stationary parameters, and therefore a stationary signal. Although the Kalman smoothing allows the model to work with some time-varying signal, the success of the method depends on the strength and duration of the signal components. The weaker and more brief the time-varying component is, the more poorly the model will capture it, if at all. We recommend decreasing the length of your window until you have a more stationary signal.\n\nWhen using this module, please cite the following [paper](https://www.biorxiv.org/content/10.1101/2022.10.30.514422):\n\nBeck, A. M., He, M., Gutierrez, R. G., & Purdon, P. L. (2022). An iterative search algorithm to identify oscillatory dynamics in neurophysiological time series. bioRxiv, 2022-10.\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Functional-success.svg?logo=Python\">\n</picture>\n\n### 4. Switching State-Space Inference\n\nWhen using this module, please cite the following [paper](https://journals.plos.org/ploscompbiol/article?id=10.1371/journal.pcbi.1011395):\n\nHe, M., Das, P., Hotan, G., & Purdon, P. L. (2023). Switching state-space modeling of neural signal dynamics. PLOS Computational Biology, 19(8), e1011395.\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Missing-critical.svg?logo=Python\">\n</picture>\n\n### 5. Multi-channel Oscillator Component Analysis\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Missing-critical.svg?logo=Python\">\n</picture>\n\n### 6. State-Space Event Related Potential\n\n---\n<picture>\n <img align=\"right\" src=\"https://img.shields.io/badge/Status-Functional-success.svg?logo=Python\">\n</picture>\n\n### 7. Dynamic Source Localization\n\n---\n\n## Authors\nMingjian He, Proloy Das, Amanda Beck, Patrick Purdon\n\n## Citation\nUse different citation styles at: https://doi.org/10.5281/zenodo.7242130\n\n## License\nSOMATA is licensed under the [BSD 3-Clause Clear license](https://spdx.org/licenses/BSD-3-Clause-Clear.html).\\\nCopyright \u00a9 2023. All rights reserved.\n",
"bugtrack_url": null,
"license": null,
"summary": "State-space Oscillator Modeling And Time-series Analysis",
"version": "0.4.1",
"project_urls": {
"Homepage": "https://github.com/mh105/somata"
},
"split_keywords": [
"state-space",
"oscillator",
"time-series"
],
"urls": [
{
"comment_text": "",
"digests": {
"blake2b_256": "3d4cf8747ca7faf41982e54dcead71c5ad3c837dca6c1d8d6556f96ad02645c1",
"md5": "8656411fed3176bcc1be3ef9ee655b14",
"sha256": "cf9961bbf624951d208f2a58e9e90725f3cd1ad7274733f02402088a727030cc"
},
"downloads": -1,
"filename": "somata-0.4.1-py3-none-any.whl",
"has_sig": false,
"md5_digest": "8656411fed3176bcc1be3ef9ee655b14",
"packagetype": "bdist_wheel",
"python_version": "py3",
"requires_python": ">=3.8",
"size": 106779,
"upload_time": "2024-03-24T03:21:36",
"upload_time_iso_8601": "2024-03-24T03:21:36.819912Z",
"url": "https://files.pythonhosted.org/packages/3d/4c/f8747ca7faf41982e54dcead71c5ad3c837dca6c1d8d6556f96ad02645c1/somata-0.4.1-py3-none-any.whl",
"yanked": false,
"yanked_reason": null
},
{
"comment_text": "",
"digests": {
"blake2b_256": "84c3c0643dcf10ada26e0044d802052c7210eba19c812140d1883503fc2f2c3b",
"md5": "4ab6759f2fd1541ad3f8d6f965cbddb7",
"sha256": "92450714664286b6e6f22479f78d9d7c58dcd50bcd4382d98e25666b6d59ee36"
},
"downloads": -1,
"filename": "somata-0.4.1.tar.gz",
"has_sig": false,
"md5_digest": "4ab6759f2fd1541ad3f8d6f965cbddb7",
"packagetype": "sdist",
"python_version": "source",
"requires_python": ">=3.8",
"size": 107861,
"upload_time": "2024-03-24T03:21:38",
"upload_time_iso_8601": "2024-03-24T03:21:38.943683Z",
"url": "https://files.pythonhosted.org/packages/84/c3/c0643dcf10ada26e0044d802052c7210eba19c812140d1883503fc2f2c3b/somata-0.4.1.tar.gz",
"yanked": false,
"yanked_reason": null
}
],
"upload_time": "2024-03-24 03:21:38",
"github": true,
"gitlab": false,
"bitbucket": false,
"codeberg": false,
"github_user": "mh105",
"github_project": "somata",
"travis_ci": false,
"coveralls": false,
"github_actions": true,
"lcname": "somata"
}
JSON
