Data Types#

class whobpyt.datatypes.timeseries.Timeseries(data, step_size, modality='')[source]#

This class is responsible for holding timeseries of empirical and simulated data. It is:

Part of the input and output of Model_fitting and Model_simulation[future].
It is the format expected by the visualization function of whobpyt.

However, the Recording class is not used within and across the NMM and Modals, as a simpler dictionary of tensors is used in those contexts.

Numerical simulation internals don’t necessarily use this data structure, as calculations may be more efficient as the transpose: ts_length x num_regions.

Attributes:

dataNumpy Array or Tensor of dimensions num_regions x ts_length: The time series data, either empirical or simulated
step_sizeFloat: The step size of the time points in the data class
modalityString: The name of the modality of the time series
numNodesInt: The number of nodes it time series.
lengthInt: Returns

__init__(data, step_size, modality='')[source]#

Parameters:

dataNumpy Array or Tensor of dimensions num_regions x ts_length: The time series data, either empirical or simulated
step_sizeFloat: The step size of the time points in the data class
modalityString: The name of the modality of the time series

length()[source]#

Returns:

The time series length

npNodeByTime()[source]#

Returns:

Numpy Array of num_regions x ts_length

npResample()[source]#: This outputs resampled data used for figures (TODO: Not yet implemented)

npTS()[source]#

Returns:

Numpy Array of num_regions x ts_length

npTimeByNodes()[source]#

Returns:

Numpy Array of ts_length x num_regions

pyTS()[source]#

Returns:

Tensor of num_regions x ts_length

windowedTensor(TPperWindow)[source]#

This method is called by the Model_fitting Class during training to reshape the data into windowed segments (adds another dimension).

Parameters:

TPperWindowInt: The number of time points in the window that will be back propagated

Returns:

Tensor: num_windows x num_regions x window_length: The time series data in a windowed format

class whobpyt.datatypes.parameter.Parameter(val, prior_mean=None, prior_std=None, fit_par=False, asLog=False, asRand=True, lb=0, device=device(type='cpu'))[source]#

Features of this class:

This class contains a global parameter value or array of values (one per node)
It can also contain associated priors (mean and variance)
It also has attributes for whether the parameter and/or priors should be fit during training
It has a method to return the parameter as numpy value
It has a method to set a random val using based on priors
It has functionality to represent val as log(val) so that during training val will be constrained to be positive

Attributes:

valTensor: The parameter value (or an array of node specific parameter values)
prior_meanTensor: Prior mean of the data value
prior_precisionTensor: Prior inverse of variance of the value
fit_par: Bool: Whether the parameter value should be set to as a PyTorch Parameter
fit_hyperBool: Whether the parameter prior mean and prior variance should be set as a PyTorch Parameter
asLogBool: Whether the log of the parameter value will be stored instead of the parameter itself (will prevent parameter from being negative).

__init__(val, prior_mean=None, prior_std=None, fit_par=False, asLog=False, asRand=True, lb=0, device=device(type='cpu'))[source]#

Parameters:

valFloat (or Array): The parameter value (or an array of node specific parameter values)
prior_meanFloat: Prior mean of the data value
prior_stdFloat: Prior std of the value
fit_par: Bool: Whether the parameter value should be set to as a PyTorch Parameter
device: torch.device: Whether to run on CPU or GPU

npValue()[source]#

Returns:

NumPy of Value: The parameter value(s) as a NumPy Array

to(device)[source]#

value()[source]#

Returns:

Tensor of Value: The parameter value(s) as a PyTorch Tensor

class whobpyt.datatypes.outputs.TrainingStats(model)[source]#

This class is responsible for recording stats during training (it replaces OutputNM) including:

The training and validation losses over time
The change in model parameters over time
changing hyper parameters over time like learing rates TODO

These are things typically recorded on a per window/epoch basis

Attributes:

model_infoDict: Information about model being tracked during training.
track_paramsList: List of parameter names being tracked during training.
lossList: A list of loss values over training.
connectivityList: A list of connectivity values over training.
leadfieldList: A list of leadfield matrices over training.
fit_paramsDict: A dictionary of lists where the key is the parameter name and the value is the list of parameter values over training.

__init__(model)[source]#

Parameters:

modelAbstractNMM: A model for which stats will be recorded during training.

appendLF(newValue)[source]#

Append Lead Field Loss

Parameters:

newValueArray: Current state of a lead field matrix being tracked.

appendLoss(newValue)[source]#

Append Trainig Loss

Parameters:

newValueFloat: The loss value of objective function being tracked.

appendParam(newValues)[source]#

Append Fit Parameters

Parameters:

newValuesDict: Dictionary with current states of each model parameter being tracked.

appendSC(newValue)[source]#

Append Network Connections

Parameters:

newValueArray: Current state of the structural connectivity being tracked.

appendSC_p2e(newValue)[source]#

Append Network Connections

Parameters:

newValueArray: Current state of the structural connectivity being tracked.

appendSC_p2i(newValue)[source]#

Append Network Connections

Parameters:

newValueArray: Current state of the structural connectivity being tracked.

appendSC_p2p(newValue)[source]#

Append Network Connections

Parameters:

newValueArray: Current state of the structural connectivity being tracked.

reset()[source]#: Resets the attributes of the model to a pre-training state.

save(filename)[source]#

Parameters:

filenameString: The filename to use to save the TrainingStats as a pickle object.