`models`¶

`counter`()
`atleast_2d`(arr)
`check_response_dimensionality`(train, test[, …])	Make sure matrices are 2D arrays before running models
`clean_results_dict`(results)	Make sure we return arrays, and ndim is at least 2D
`crossval_stem_wmvnp`((features_train, …[, …])	Cross-validation procedure for spatio-temporal encoding models with MVN priors.
`cvridge`(Xtrain, Ytrain[, Xtest, Ytest, …])	Cross-validation procedure for tikhonov regularized regression.
`dual2primal_weights`(kernel_weights, …[, …])	Recover the feature weights from the kernel weights for any one or all feature spaces.
`dual2primal_weights_banded`(kernel_weights, …)	WIP.
`estimate_simple_stem_wmvnp`(features_train, …)	Estimate model with given hyper-parameters
`estimate_stem_wmvnp`((features_train, …[, …])
`featurespace_dual2primal`(kernel_weights, …)
`find_optimum_mvn`(response_cvmean, …)
`generalized_tikhonov`(X, Y, Li[, ridge])	Implementation fo tikhonov regression using the standard transform (cf.
`hyperopt_crossval_stem_wmvnp`(features_train, …)	Use `hyperopt` to cross-validate all hyper-parameters parameters.
`hyperopt_estimate_stem_wmvnp`(features_train, …)
`hyperopt_trials2cvperf`(Trials)
`kernel_banded_temporal_prior`(kernel, …)
`kernel_cvridge`(Ktrain, Ytrain[, Ktest, …])
`kernel_spatiotemporal_prior`(Xtrain, …[, …])	Compute the kernel matrix of a model with a spatio-temporal prior
`loo_ols`(xtrain_samples, ytrain_samples[, rcond])	Leave-one out OLS Return the mean weight across head-out folds
`nan_to_num`(\args, \\*kwargs)
`ols`(X, Y[, rcond])	Perform OLS fit, return weight estimates
`olspred`(X, Y[, Xtest])	Fit OLS, return predictions `Yhat`
`should_solve_dual`(X, kernel)	Answer whether we should solve the regression problem in the dual (kernel) space.
`simple_generalized_tikhonov`(X, Y, L[, ridge])	Direct implementation of generalized tikhonov regression
`simple_ridge_dual`(X, Y[, ridge])	Return weights for linear kernel ridge regression
`simple_ridge_primal`(X, Y[, ridge])	Return weights for ridge regression
`solve_l2`(Xtrain, Ytrain[, Xtest, Ytest, …])	Solve L2 regularized regression problem
`solve_l2_dual`(Ktrain, Ytrain[, Ktest, …])	Solve the dual (kernel) L2 regression problem for each L2 parameter.
`solve_l2_primal`(Xtrain, Ytrain[, Xtest, …])	Solve the (primal) L2 regression problem for each L2 parameter.
`voxelwise_weights2preds`(kernel_weights, …)	feature_prior : one per voxel! temporal_prior : one for all voxels
`zscore`(\args, \\*kwargs)

`counter`¶

class tikreg.models.counter¶

Bases: object

__init__(self)¶: x.__init__(…) initializes x; see help(type(x)) for signature

update(self)¶

atleast_2d¶

tikreg.models.atleast_2d(arr)¶

check_response_dimensionality¶

tikreg.models.check_response_dimensionality(train, test, allow_test_none=True)¶: Make sure matrices are 2D arrays before running models

clean_results_dict¶

tikreg.models.clean_results_dict(results)¶: Make sure we return arrays, and ndim is at least 2D

crossval_stem_wmvnp¶

crossval_stem_wmvnp(features_train, responses_train, ridges=array([ 1. , 2.15443469, 4.64158883, 10. ,
21.5443469 , 46.41588834, 100. , 215.443469 ,
464.15888336, 1000. ]), temporal_prior=None, feature_priors=None, population_mean=False, folds=(1, 5), method='SVD', verbosity=1, chunklen=True, kernel_features=False, normalize_kernel=False, normalize_hyparams=False, metric='correlation', zscore_ytrain=False, zscore_yval=False, weights=False, predictions=False): Cross-validation procedure for spatio-temporal encoding models with MVN priors.

cvridge¶

tikreg.models.cvridge(Xtrain, Ytrain, Xtest=None, Ytest=None, ridges=[0.0], Li=None, kernel_name='linear', kernel_params=None, folds='cv', nfolds=5, blocklen=5, trainpct=0.8, verbose=True, EPS=1e-10, withinset_test=False, performance=False, predictions=False, weights=False, kernel_weights=False, metric='correlation')¶

Cross-validation procedure for tikhonov regularized regression.

Parameters:

Parameters:	Xtrain (n, p): Design matrix Ytrain (n, v): Training set responses Xtest (None, (m, p)): Design matrix for held-out set Ytest (None, (m, v)): Held-out set responses ridges (r,): Ridge parameters to evaluate Li (q,p): Generalized tikhonov regression. This solves the problem with a prior on $eta$ determined by $L^ op L$. The problem is solved in the standard form and in kernel space if necessary. kernel_name (str): Kernel to use kernel_params (None, (k,)): Kernel parameters to cross-validate folds (str, list): (str) Type of cross-validation ‘cv’ - cross-validation with chunks of size `blocklen` ‘nbb’ - block boostrap with chunks of size `blocklen` ‘mbb’ - moving/overlapping block bootstrap chunks of size `blocklen` (list) Can also be a list of (train, test) pairs: [(trn1, test1),…] nfolds (int): Number of learning folds blocklen (int): Chunk data into blocks of this size, and sample these blocks trainpct (float 0-1): Percentage of data to use in training if using a bootstrap sampler. withinset_test (bool): If no `Xtest` or `Ytest` is given and `predictions` and/or `performance` are requested, compute these testues based on training set. performance (bool): Held-out prediction performance predictions (bool): Held-out timecourse predictions weights (bool): Weight estimates on training set (does not depend on (`Xtest`,``Ytest``) kernel_weights (bool): Whether to project kernel weights into feature weights. If True, the kernel weights are returned. This is useful when fitting large models and storing the feature weights is expensive. verbose (bool): Verbosity EPS (float): Testue used to threshold small eigentestues
Returns:	fit (optional; dict): cross-validation results per response for each fold, kernel and L2 parameters cvresults (`nfolds`, len(`kernel_params`), len(`ridges`), nresponses) If a held-out set (`Xtest` and `Ytest`) is specified, performs the fit on the full training set with the optimal L2 and kernel parameters. It returns, as requested, any of: predictions (m, v) `Ytest` prediction for each voxel performance (v,) correlation coefficient of predictions and `Ytest` weights: (p, v) for linear models, (n by v) for non-linear models

Xtrain (n, p):

Design matrix

Ytrain (n, v):

Training set responses

Xtest (None, (m, p)):

Design matrix for held-out set

Ytest (None, (m, v)):

Held-out set responses

ridges (r,):

Ridge parameters to evaluate

Li (q,p):

Generalized tikhonov regression. This solves the problem with a prior on $eta$ determined by $L^ op L$. The problem is solved in the standard form and in kernel space if necessary.

kernel_name (str):

Kernel to use

kernel_params (None, (k,)):

Kernel parameters to cross-validate

folds (str, list):

(str) Type of cross-validation
- ‘cv’ - cross-validation with chunks of size blocklen
- ‘nbb’ - block boostrap with chunks of size blocklen
- ‘mbb’ - moving/overlapping block bootstrap chunks of size blocklen
(list) Can also be a list of (train, test) pairs: [(trn1, test1),…]

nfolds (int):

Number of learning folds

blocklen (int):

Chunk data into blocks of this size, and sample these blocks

trainpct (float 0-1):

Percentage of data to use in training if using a bootstrap sampler.

withinset_test (bool):

If no Xtest or Ytest is given and predictions and/or performance are requested, compute these testues based on training set.

performance (bool):

Held-out prediction performance

predictions (bool):

Held-out timecourse predictions

weights (bool):

Weight estimates on training set (does not depend on (Xtest,``Ytest``)

kernel_weights (bool):

Whether to project kernel weights into feature weights. If True, the kernel weights are returned. This is useful when fitting large models and storing the feature weights is expensive.

verbose (bool):

Verbosity

EPS (float):

Testue used to threshold small eigentestues

Returns:

fit (optional; dict):

cross-validation results per response for each fold, kernel and L2 parameters

cvresults (nfolds, len(kernel_params), len(ridges), nresponses)

If a held-out set (Xtest and Ytest) is specified, performs the fit on the full training set with the optimal L2 and kernel parameters. It returns, as requested, any of:

predictions (m, v) Ytest prediction for each voxel

performance (v,) correlation coefficient of predictions and Ytest

weights: (p, v) for linear models, (n by v) for non-linear models

dual2primal_weights¶

tikreg.models.dual2primal_weights(kernel_weights, features_train, feature_priors, feature_hyparams, temporal_prior, temporal_hhparam=1.0)¶

Recover the feature weights from the kernel weights for any one or all feature spaces.

Parameters:

Parameters:	kernel_weights : 2D np.ndarray features_train : list of np.ndarrays Training set feature spaces each of shape (n, p_i). feature_priors : list of `SpatialPrior` objects One feature prior per feature space. feature_hyparams : list of scalars Scalar for each feature space prior temporal_prior : `TemporalPrior` object A temporal prior object to use. The temporal prior may contain a hyper-prior. temporal_hhparam : scalar Hyper-prior hyperparameter if required. Defaults to 1.0, no effect.
Returns:	weights : list Feature space weights [(p_1, v), …, (p_l, v)]

kernel_weights : 2D np.ndarray
features_train : list of np.ndarrays: Training set feature spaces each of shape (n, p_i).
feature_priors : list of SpatialPrior objects: One feature prior per feature space.
feature_hyparams : list of scalars: Scalar for each feature space prior
temporal_prior : TemporalPrior object: A temporal prior object to use. The temporal prior may contain a hyper-prior.
temporal_hhparam : scalar: Hyper-prior hyperparameter if required. Defaults to 1.0, no effect.

Returns:

weights : list: Feature space weights [(p_1, v), …, (p_l, v)]

dual2primal_weights_banded¶

tikreg.models.dual2primal_weights_banded(kernel_weights, feature_space_train, population_feature_prior, temporal_prior, delays_mean=False, verbose=True)¶: WIP. WILL CHANGE. USE AT OWN RISK.

estimate_simple_stem_wmvnp¶

tikreg.models.estimate_simple_stem_wmvnp(features_train, responses_train, features_test=None, responses_test=None, feature_priors=None, feature_hyparams=None, temporal_prior=None, temporal_hhparam=1.0, ridge_scale=1.0, weights=False, performance=False, predictions=False, kernel_features=False, method='SVD', verbosity=2, metric='correlation')¶

Estimate model with given hyper-parameters

Parameters:

Parameters:	features_train : list of np.ndarrays Training set feature spaces each of shape (n, p_i). responses_train : 2D np.ndarray Population responses to the training set (n, v). features_test : list of np.ndarrays Test set feature spaces each of shape (m, p_i) responses_test : 2D np.ndarray Population responses to the test set (m, v) temporal_prior : `TemporalPrior` object A temporal prior object to use. The temporal prior may contain a hyper-prior. temporal_hhparam : scalar Hyper-prior hyperparameter if required. Defaults to 1.0, no effect. feature_priors : list of `SpatialPrior` objects One feature prior per feature space. feature_hyparams : list of scalars Scalar for each feature space prior ridge_scale : scalar Scalar on the feature prior hyper-parameters. Defaults to 1.0, no effect weights : bool Compute weights performance : bool Compute accuracy of predictions against test set responses predictions: boot Test set prediction time courses kernel_features : boot If True, `features_train` and `features_test` is a list of kernels, one per feature space. This is only allowed if `feature_priors` are spherical. method : str {“SVD”, “Chol”} Solver to use
Returns:	fit : dictionary Estimated model dictionary with keys: * weights : [(p_1, v), …, (p_l, v)] * performance : (1, v) * predictions : (m, v)

features_train : list of np.ndarrays: Training set feature spaces each of shape (n, p_i).
responses_train : 2D np.ndarray: Population responses to the training set (n, v).
features_test : list of np.ndarrays: Test set feature spaces each of shape (m, p_i)
responses_test : 2D np.ndarray: Population responses to the test set (m, v)
temporal_prior : TemporalPrior object: A temporal prior object to use. The temporal prior may contain a hyper-prior.
temporal_hhparam : scalar: Hyper-prior hyperparameter if required. Defaults to 1.0, no effect.
feature_priors : list of SpatialPrior objects: One feature prior per feature space.
feature_hyparams : list of scalars: Scalar for each feature space prior
ridge_scale : scalar: Scalar on the feature prior hyper-parameters. Defaults to 1.0, no effect
weights : bool: Compute weights
performance : bool: Compute accuracy of predictions against test set responses
predictions: boot: Test set prediction time courses
kernel_features : boot: If True, features_train and features_test is a list of kernels, one per feature space. This is only allowed if feature_priors are spherical.
method : str {“SVD”, “Chol”}: Solver to use

Returns:

fit : dictionary: Estimated model dictionary with keys: * weights : [(p_1, v), …, (p_l, v)] * performance : (1, v) * predictions : (m, v)

estimate_stem_wmvnp¶

estimate_stem_wmvnp(features_train, responses_train, features_test=None, responses_test=None, ridges=array([ 1. , 2.15443469, 4.64158883, 10. ,
21.5443469 , 46.41588834, 100. , 215.443469 ,
464.15888336, 1000. ]), normalize_hyparams=False, normalize_kernel=False, temporal_prior=None, feature_priors=None, weights=False, predictions=False, performance=False, folds=(1, 5), method='SVD', verbosity=1, cvresults=None, population_optimal=False, keep_cvfolds=True, chunklen=True, metric='correlation')

featurespace_dual2primal¶

tikreg.models.featurespace_dual2primal(kernel_weights, feature_space, feature_prior, feature_hyparam, temporal_prior, temporal_hhparam=1.0)¶

find_optimum_mvn¶

tikreg.models.find_optimum_mvn(response_cvmean, temporal_hhparams, spatial_hyparams, ridge_hyparams)¶

generalized_tikhonov¶

tikreg.models.generalized_tikhonov(X, Y, Li, ridge=10.0)¶: Implementation fo tikhonov regression using the standard transform (cf. Hansen, 1998).

hyperopt_crossval_stem_wmvnp¶

tikreg.models.hyperopt_crossval_stem_wmvnp(features_train, responses_train, features_test=None, responses_test=None, temporal_prior=None, feature_priors=None, spatial_sampler=True, temporal_sampler=False, ridge_sampler=False, population_optimal=False, folds=(1, 5), method='SVD', ntrials=100, verbosity=1, dumpcrossval=False, normalize_hyparams=False, normalize_kernel=False, weights=False, predictions=False, performance=True, metric='correlation', zscore_ytrain=True, zscore_yval=True, search_algorithm='tpe', trials=None, **kwargs)¶

Use hyperopt to cross-validate all hyper-parameters parameters.

Search the hyper-parameter space to find the population optimum using a cross-validation procedure.

Parameters:

Parameters:	features_train : list of np.ndarrays The feature spaces of shape (n, p_i). responses_train : 2D np.ndarray The population responses to fit (n, v). temporal_prior : `TemporalPrior` object A temporal prior object to use. The temporal prior may contain a hyper-prior. feature_priors : list of ``SpatialPrior``bjects One feature prior per feature space. spatial_sampler : `hyperopt.hp`, or bool Specifies how to sample the hyperparameter space. Defaults to hp.loguniform(0,7). temporal_sampler : `hyperopt.hp`, or bool Used iff `temporal_prior` has a hyper-prior set. ridge_sampler : `hyperopt.hp`, or bool Defaults to False. Use this with caution. Specifies how to sample the scaling on the spatial hyperparameters. However, Specifying a `spatial_sampler` for all feature spaces and a `ridge_sampler` is redundant. population_optimal : bool If True, individual response cross-validation values are not kept. Only the mean across responses is stored. folds : tuple (N,K), or list of tuples [(trn1, val1),…, (trnK, valK)] If tuple, the second element corresponds to the number of cross-validation folds. The first element determines how many times to repeat the cross validation. (1,5) is standard 5-folds cross-validation. (10,5) performs 5-fold cross-validation 10 times. method : str (“SVD”, “Chol”) Solver to use ntrials : int Number of `hyperopt` iterations verbosity : int (1, 2) Level of print statements dumpcrossval : function(iteration_num, crossval_dict) Save the cross-validation results for every iteration. The function takes the iteration number (int) and a dictionary containing the cross-validation results. This is useful for finding the optimum hyper-parameters for each response. The stored data contains all the info. Defaults to False. kwargs : dict Additional arguments passed to `crossval_stem_wmvnp`.
Returns:	crossval_results : hyperopt.Trials object Contains the cross-validation results from hyperopt.

features_train : list of np.ndarrays: The feature spaces of shape (n, p_i).
responses_train : 2D np.ndarray: The population responses to fit (n, v).
temporal_prior : TemporalPrior object: A temporal prior object to use. The temporal prior may contain a hyper-prior.
feature_priors : list of ``SpatialPrior``bjects: One feature prior per feature space.
spatial_sampler : hyperopt.hp, or bool: Specifies how to sample the hyperparameter space. Defaults to hp.loguniform(0,7).
temporal_sampler : hyperopt.hp, or bool: Used iff temporal_prior has a hyper-prior set.
ridge_sampler : hyperopt.hp, or bool: Defaults to False. Use this with caution. Specifies how to sample the scaling on the spatial hyperparameters. However, Specifying a spatial_sampler for all feature spaces and a ridge_sampler is redundant.
population_optimal : bool: If True, individual response cross-validation values are not kept. Only the mean across responses is stored.
folds : tuple (N,K), or list of tuples [(trn1, val1),…, (trnK, valK)]: If tuple, the second element corresponds to the number of cross-validation folds. The first element determines how many times to repeat the cross validation. (1,5) is standard 5-folds cross-validation. (10,5) performs 5-fold cross-validation 10 times.
method : str (“SVD”, “Chol”): Solver to use
ntrials : int: Number of hyperopt iterations
verbosity : int (1, 2): Level of print statements
dumpcrossval : function(iteration_num, crossval_dict): Save the cross-validation results for every iteration. The function takes the iteration number (int) and a dictionary containing the cross-validation results. This is useful for finding the optimum hyper-parameters for each response. The stored data contains all the info. Defaults to False.
kwargs : dict: Additional arguments passed to crossval_stem_wmvnp.

Returns:

crossval_results : hyperopt.Trials object: Contains the cross-validation results from hyperopt.

hyperopt_estimate_stem_wmvnp¶

tikreg.models.hyperopt_estimate_stem_wmvnp(features_train, responses_train, cvmean, hyparams, features_test=None, responses_test=None, temporal_prior=None, feature_priors=None, spatial_sampler=True, temporal_sampler=False, ridge_sampler=False, population_optimal=False, method='SVD', verbosity=1, normalize_hyparams=False, normalize_kernel=False, weights=False, predictions=False, performance=True, kernel_features=False, metric='correlation', **kwargs)¶

hyperopt_trials2cvperf¶

tikreg.models.hyperopt_trials2cvperf(Trials)¶

kernel_banded_temporal_prior¶

tikreg.models.kernel_banded_temporal_prior(kernel, temporal_prior, spatial_prior, delays)¶

kernel_cvridge¶

tikreg.models.kernel_cvridge(Ktrain, Ytrain, Ktest=None, Ytest=None, ridges=[0.0], folds='cv', nfolds=5, blocklen=5, trainpct=0.8, performance=False, predictions=False, weights=False, metric='correlation', verbose=True, EPS=1e-10)¶

kernel_spatiotemporal_prior¶

tikreg.models.kernel_spatiotemporal_prior(Xtrain, temporal_prior, spatial_prior, Xtest=None, delays=None)¶

Compute the kernel matrix of a model with a spatio-temporal prior

temporal_prior (d, d): d = len(delays)

loo_ols¶

tikreg.models.loo_ols(xtrain_samples, ytrain_samples, rcond=1e-08)¶

Leave-one out OLS Return the mean weight across head-out folds

Parameters:	ytrain_samples : np.ndarray (nfolds, time, voxels) xtrain_samples : np.ndarray (nfolds, time, features)

nan_to_num¶

tikreg.models.nan_to_num(*args, **kwargs)¶

ols¶

tikreg.models.ols(X, Y, rcond=1e-08)¶: Perform OLS fit, return weight estimates

olspred¶

tikreg.models.olspred(X, Y, Xtest=False)¶: Fit OLS, return predictions Yhat

should_solve_dual¶

tikreg.models.should_solve_dual(X, kernel)¶: Answer whether we should solve the regression problem in the dual (kernel) space.

simple_generalized_tikhonov¶

tikreg.models.simple_generalized_tikhonov(X, Y, L, ridge=10.0)¶: Direct implementation of generalized tikhonov regression

simple_ridge_dual¶

tikreg.models.simple_ridge_dual(X, Y, ridge=10.0)¶: Return weights for linear kernel ridge regression

simple_ridge_primal¶

tikreg.models.simple_ridge_primal(X, Y, ridge=10.0)¶: Return weights for ridge regression

solve_l2¶

tikreg.models.solve_l2(Xtrain, Ytrain, Xtest=None, Ytest=None, ridge=0.0, verbose=False, kernel_name='linear', kernel_param=None, kernel_weights=False, **kwargs)¶: Solve L2 regularized regression problem

solve_l2_dual¶

tikreg.models.solve_l2_dual(Ktrain, Ytrain, Ktest=None, Ytest=None, ridges=[0.0], method='SVD', EPS=1e-10, verbose=False, performance=False, predictions=False, weights=False, metric='correlation')¶: Solve the dual (kernel) L2 regression problem for each L2 parameter.

solve_l2_primal¶

tikreg.models.solve_l2_primal(Xtrain, Ytrain, Xtest=None, Ytest=None, ridges=[0], method='SVD', zscore_ytrain=False, zscore_ytest=False, EPS=1e-10, verbose=False, performance=False, predictions=False, weights=False, metric='correlation')¶: Solve the (primal) L2 regression problem for each L2 parameter.

voxelwise_weights2preds¶

tikreg.models.voxelwise_weights2preds(kernel_weights, kernel_test, responses_test, feature_prior, feature_hyparam, temporal_prior, temporal_hyparam=1.0, verbose=True, metric='correlation')¶: feature_prior : one per voxel! temporal_prior : one for all voxels

zscore¶

tikreg.models.zscore(*args, **kwargs)¶

models¶

counter¶

atleast_2d¶

check_response_dimensionality¶

clean_results_dict¶

crossval_stem_wmvnp¶

cvridge¶

dual2primal_weights¶

dual2primal_weights_banded¶

estimate_simple_stem_wmvnp¶

estimate_stem_wmvnp¶

featurespace_dual2primal¶

find_optimum_mvn¶

generalized_tikhonov¶

hyperopt_crossval_stem_wmvnp¶

hyperopt_estimate_stem_wmvnp¶

hyperopt_trials2cvperf¶

kernel_banded_temporal_prior¶

kernel_cvridge¶

kernel_spatiotemporal_prior¶

loo_ols¶

nan_to_num¶

ols¶

olspred¶

should_solve_dual¶

simple_generalized_tikhonov¶

simple_ridge_dual¶

simple_ridge_primal¶

solve_l2¶

solve_l2_dual¶

solve_l2_primal¶

voxelwise_weights2preds¶

zscore¶

`models`¶

`counter`¶