amica.AmicaConfig#

class amica.AmicaConfig(num_models=1, num_mix_comps=3, pcakeep=None, dtype='float64', max_iter=2000, lrate=0.01, minlrate=1e-08, lratefact=0.5, rholrate=0.05, rholratefact=0.5, minrho=1.0, maxrho=2.0, rho0=1.5, do_newton=True, newt_start=50, newt_ramp=10, newtrate=1.0, do_mean=True, do_sphere=True, sphere_type='zca', do_pca=True, do_approx_sphere=True, mineig=1e-12, do_reject=False, rejsig=3.0, rejstart=2, rejint=3, numrej=5, min_dll=1e-09, use_min_dll=True, max_decs=3, max_incs=10, invsigmax=100.0, invsigmin=1e-08, doscaling=True, writestep=100, outdir=None, fix_init=False, update_alpha=True, update_mu=True, update_beta=True, update_rho=True, chunk_size=None, estep='auto')[source]#

Bases: object

Configuration for AMICA algorithm.

Parameters:

num_modelsint: Number of ICA models to learn simultaneously. Default is 1.
num_mix_compsint: Number of Gaussian mixture components per source. Default is 3.
dtypestr: Data type to use for computation: “float32” or “float64”. Default is “float64”.
pcakeepint, optional: Number of PCA components to keep. If None, uses data rank.
max_iterint: Maximum number of iterations. Default is 2000.
lratefloat: Initial learning rate. Default is 0.1.
minlratefloat: Minimum learning rate before stopping. Default is 1e-8.
lratefactfloat: Factor to decrease learning rate on likelihood decrease. Default is 0.5.
rholratefloat: Learning rate for shape parameter rho. Default is 0.05.
rholratefactfloat: Factor to decrease rho learning rate on likelihood decrease. Default is 0.5.
minrhofloat: Minimum value for shape parameter (1.0 = Laplacian). Default is 1.0.
maxrhofloat: Maximum value for shape parameter (2.0 = Gaussian). Default is 2.0.
do_newtonbool: Whether to use Newton updates for faster convergence. Default is True.
newt_startint: Iteration to start Newton updates. Default is 50.
newt_rampint: Number of iterations to ramp up Newton. Default is 10.
newtratefloat: Newton learning rate multiplier. Default is 1.0.
do_meanbool: Whether to remove data mean. Default is True.
do_spherebool: Whether to whiten/sphere data. Default is True.
sphere_typestr: Type of sphering: “pca” or “zca”. Default is “zca”.
do_pcabool: Whether to apply PCA dimensionality reduction. Default is True.
do_approx_spherebool: Use approximate sphering. Default is True.
mineigfloat: Minimum eigenvalue threshold for PCA. Default is 1e-12.
do_rejectbool: Whether to reject outlier samples. Default is False.
rejsigfloat: Rejection threshold in standard deviations. Default is 3.0.
rejstartint: Iteration to start rejection. Default is 2 (Klug et al. 2024).
rejintint: Interval between rejection passes. Default is 3 (Klug et al. 2024).
numrejint: Number of rejection passes per interval. Default is 5.
min_dllfloat: Minimum log-likelihood change for convergence. Default is 1e-9.
max_decsint: Number of LL decreases before reducing max learning rates. Default is 3.
max_incsint: Number of small LL increases before stopping. Default is 10.
use_min_dllbool: Whether to use min_dll for convergence. Default is True.
invsigmaxfloat: Maximum inverse sigma for numerical stability. Default is 100.0.
invsigminfloat: Minimum inverse sigma for numerical stability. Default is 0.0.
doscalingbool: Whether to rescale A/mu/sbeta each iteration. Default is True.
writestepint: Interval for writing intermediate results. Default is 100.
outdirPath, optional: Output directory for results.
fix_initbool: Use identity matrix initialization instead of random. Default is False.
update_alphabool: Whether to update mixture weights. Default is True.
update_mubool: Whether to update location parameters. Default is True.
update_betabool: Whether to update scale parameters. Default is True.
update_rhobool: Whether to update shape parameters. Default is True.

Parameters:

num_models (int)
num_mix_comps (int)
pcakeep (int | None)
dtype (str)
max_iter (int)
lrate (float)
minlrate (float)
lratefact (float)
rholrate (float)
rholratefact (float)
minrho (float)
maxrho (float)
rho0 (float)
do_newton (bool)
newt_start (int)
newt_ramp (int)
newtrate (float)
do_mean (bool)
do_sphere (bool)
sphere_type (str)
do_pca (bool)
do_approx_sphere (bool)
mineig (float)
do_reject (bool)
rejsig (float)
rejstart (int)
rejint (int)
numrej (int)
min_dll (float)
use_min_dll (bool)
max_decs (int)
max_incs (int)
invsigmax (float)
invsigmin (float)
doscaling (bool)
writestep (int)
outdir (Path | None)
fix_init (bool)
update_alpha (bool)
update_mu (bool)
update_beta (bool)
update_rho (bool)
chunk_size (int | Literal['auto'] | None)
estep (Literal['auto', 'fused', 'classic'])

__init__(num_models=1, num_mix_comps=3, pcakeep=None, dtype='float64', max_iter=2000, lrate=0.01, minlrate=1e-08, lratefact=0.5, rholrate=0.05, rholratefact=0.5, minrho=1.0, maxrho=2.0, rho0=1.5, do_newton=True, newt_start=50, newt_ramp=10, newtrate=1.0, do_mean=True, do_sphere=True, sphere_type='zca', do_pca=True, do_approx_sphere=True, mineig=1e-12, do_reject=False, rejsig=3.0, rejstart=2, rejint=3, numrej=5, min_dll=1e-09, use_min_dll=True, max_decs=3, max_incs=10, invsigmax=100.0, invsigmin=1e-08, doscaling=True, writestep=100, outdir=None, fix_init=False, update_alpha=True, update_mu=True, update_beta=True, update_rho=True, chunk_size=None, estep='auto')#

Parameters:

num_models (int)
num_mix_comps (int)
pcakeep (int | None)
dtype (str)
max_iter (int)
lrate (float)
minlrate (float)
lratefact (float)
rholrate (float)
rholratefact (float)
minrho (float)
maxrho (float)
rho0 (float)
do_newton (bool)
newt_start (int)
newt_ramp (int)
newtrate (float)
do_mean (bool)
do_sphere (bool)
sphere_type (str)
do_pca (bool)
do_approx_sphere (bool)
mineig (float)
do_reject (bool)
rejsig (float)
rejstart (int)
rejint (int)
numrej (int)
min_dll (float)
use_min_dll (bool)
max_decs (int)
max_incs (int)
invsigmax (float)
invsigmin (float)
doscaling (bool)
writestep (int)
outdir (Path | None)
fix_init (bool)
update_alpha (bool)
update_mu (bool)
update_beta (bool)
update_rho (bool)
chunk_size (int | Literal['auto'] | None)
estep (Literal['auto', 'fused', 'classic'])

Return type:

None

Methods

__init__([num_models, num_mix_comps, ...])

Attributes

num_models: int = 1#

num_mix_comps: int = 3#

pcakeep: int | None = None#

dtype: str = 'float64'#

max_iter: int = 2000#

lrate: float = 0.01#

minlrate: float = 1e-08#

lratefact: float = 0.5#

rholrate: float = 0.05#

rholratefact: float = 0.5#

minrho: float = 1.0#

maxrho: float = 2.0#

rho0: float = 1.5#

do_newton: bool = True#

newt_start: int = 50#

newt_ramp: int = 10#

newtrate: float = 1.0#

do_mean: bool = True#

do_sphere: bool = True#

sphere_type: str = 'zca'#

do_pca: bool = True#

do_approx_sphere: bool = True#

mineig: float = 1e-12#

do_reject: bool = False#

rejsig: float = 3.0#

rejstart: int = 2#

rejint: int = 3#

numrej: int = 5#

min_dll: float = 1e-09#

use_min_dll: bool = True#

max_decs: int = 3#

max_incs: int = 10#

invsigmax: float = 100.0#

invsigmin: float = 1e-08#

doscaling: bool = True#

writestep: int = 100#

outdir: Path | None = None#

fix_init: bool = False#

update_alpha: bool = True#

update_mu: bool = True#

update_beta: bool = True#

update_rho: bool = True#

chunk_size: int | Literal['auto'] | None = None#

estep: Literal['auto', 'fused', 'classic'] = 'auto'#

`chunk_size`
`do_approx_sphere`
`do_mean`
`do_newton`
`do_pca`
`do_reject`
`do_sphere`
`doscaling`
`dtype`
`estep`
`fix_init`
`invsigmax`
`invsigmin`
`lrate`
`lratefact`
`max_decs`
`max_incs`
`max_iter`
`maxrho`
`min_dll`
`mineig`
`minlrate`
`minrho`
`newt_ramp`
`newt_start`
`newtrate`
`num_mix_comps`
`num_models`
`numrej`
`outdir`
`pcakeep`
`rejint`
`rejsig`
`rejstart`
`rho0`
`rholrate`
`rholratefact`
`sphere_type`
`update_alpha`
`update_beta`
`update_mu`
`update_rho`
`use_min_dll`
`writestep`