def figures_of_merit(X, maxPIndex, C, St, j):
# return %explained variance and stdev of residuals when the jth compound is added
C[:, j] = X[:, maxPIndex[j]]
St[0:j + 1, :] = np.linalg.lstsq(C.data[:, 0:j + 1],
X.data,
rcond=None)[0]
Xhat = dot(C[:, 0:j + 1], St[0:j + 1, :])
res = Xhat - X
stdev_res = np.std(res)
rsquare = 1 - np.linalg.norm(res)**2 / np.linalg.norm(X)**2
return rsquare, stdev_res
def test_npy(ds1):
# functions that keep units
# DIAG
with pytest.raises(ValueError):
df = diag(ds1) # work only for 1d or 2D dataset
ds = ds1[0].squeeze()
assert ds.ndim == 2
df = diag(ds)
assert df.units == ds1.units
assert df.ndim == 1
assert df.size == ds.x.size
d = ds[0].squeeze()
assert d.ndim == 1
df = diag(d)
assert df.units == ds1.units
assert df.ndim == 2
assert df.size == d.x.size**2
df = diag(ds.data)
assert df.implements("NDDataset")
# DOT
a = ds # 2D dataset
b = ds1[3].squeeze() # second 2D dataset
b.ito("km", force=True) # put some units to b
x = dot(a.T, b)
assert x.units == a.units * b.units
assert x.shape == (a.x.size, b.x.size)
# allow mixing numpy object with dataset
x = dot(a.T, b.data)
assert x.units == a.units
# if no dataset then is it equivalent to np.dot
x = dot(a.data.T, b.data)
assert isinstance(x, np.ndarray)
def _generate_2D_spectra(concentrations, spectra):
"""
Generate a fake 2D experimental spectra
Parameters
----------
concentrations : |NDDataset|
spectra : |NDDataset|
Returns
-------
|NDDataset|
"""
from spectrochempy.core.dataset.npy import dot
return dot(concentrations.T, spectra)
def generate_fake():
"""
Generate a fake 2D experimental spectra
returns
-------
datasets:
2D spectra, individual spectra and concentrations
"""
# define properties of the spectra and concentration profiles
# ----------------------------------------------------------------------------------------------------------------------
from spectrochempy.core.dataset.npy import dot
# data for four peaks (one very broad)
POS = (6000.0, 4000.0, 2000.0, 2500.0)
WIDTH = (6000.0, 1000.0, 250.0, 800.0)
AMPL = (100.0, 70.0, 10.0, 50.0)
RATIO = (0.1, 0.5, 0.2, 1.0)
ASYM = (0.0, 0.0, 0, 4)
MODEL = ("gaussian", "voigt", "voigt", "asymmetricvoigt")
def C1(t):
return t * 0.05 + 0.01 # linear evolution of the baseline
def C2(t):
return scp.sigmoidmodel().f(t, 1.0, max(t) / 2.0, 1, 2)
def C3(t):
return scp.sigmoidmodel().f(t, 1.0, max(t) / 5.0, 1, -2)
def C4(t):
return 1.0 - C2(t) - C3(t)
specs = _make_spectra_matrix(MODEL, AMPL, POS, WIDTH, RATIO, ASYM)
concs = _make_concentrations_matrix(C1, C2, C3, C4)
# make 2D
d = dot(concs.T, specs)
# add some noise
d.data = np.random.normal(d.data, 0.005 * d.data.max())
# d.plot()
return d, specs, concs
def reconstruct(self):
"""
Transform data back to the original space.
The following matrix operation is performed: :math:`X'_{hat} = C'.S'^t`
Returns
-------
X_hat
The reconstructed dataset based on the SIMPLISMA Analysis.
"""
# reconstruct from concentration and spectra profiles
X_hat = dot(self.C, self.St)
X_hat.description = "Dataset reconstructed by SIMPLISMA\n" + self.logs
X_hat.title = "X_hat: " + self.X.title
return X_hat
def reconstruct(self):
"""
Transform data back to the original space.
The following matrice operation is performed : :math:`X'_{hat} = C'.S'^t`.
Returns
-------
X_hat : |NDDataset|
The reconstructed dataset based on the MCS-ALS optimization.
"""
# reconstruct from concentration and spectra profiles
C = self.C
St = self.St
X_hat = dot(C, St)
X_hat.history = "Dataset reconstructed by MCS ALS optimization"
X_hat.title = "X_hat: " + self.X.title
return X_hat
def reconstruct(self, n_pc=None):
"""
Transform data back to the original space using the given number of
PC's.
The following matrice operation is performed : :math:`X' = S'.L'^T`
where S'=S[:, n_pc] and L=L[:, n_pc].
Parameters
----------
n_pc : int, optional
The number of PC to use for the reconstruction.
Returns
-------
X_reconstructed : |NDDataset|
The reconstructed dataset based on n_pc principal components.
"""
# get n_pc (automatic or determined by the n_pc arguments)
n_pc = self._get_n_pc(n_pc)
# reconstruct from scores and loadings using n_pc components
S = self._S[:, :n_pc]
LT = self._LT[:n_pc]
X = dot(S, LT)
# try to reconstruct something close to the original scaled, standardized or centered data
if self._scaled:
X *= self._ampl
X += self._min
if self._standardized:
X *= self._std
if self._centered:
X += self._center
X.history = f'PCA reconstructed Dataset with {n_pc} principal components'
X.title = self._X.title
return X
def __init__(self, dataset, centered=True, standardized=False, scaled=False):
"""
Parameters
----------
dataset : |NDDataset| object
The input dataset has shape (M, N). M is the number of
observations (for examples a series of IR spectra) while N
is the number of features (for example the wavenumbers measured
in each IR spectrum).
centered : bool, optional, default:True
If True the data are centered around the mean values: :math:`X' = X - mean(X)`.
standardized : bool, optional, default:False
If True the data are scaled to unit standard deviation: :math:`X' = X / \\sigma`.
scaled : bool, optional, default:False
If True the data are scaled in the interval [0-1]: :math:`X' = (X - min(X)) / (max(X)-min(X))`
"""
self.prefs = dataset.preferences
self._X = X = dataset
Xsc = X.copy()
# mean center the dataset
# -----------------------
self._centered = centered
if centered:
self._center = center = np.mean(X, axis=0)
Xsc = X - center
Xsc.title = "centered %s" % X.title
# Standardization
# ---------------
self._standardized = standardized
if standardized:
self._std = np.std(Xsc, axis=0)
Xsc /= self._std
Xsc.title = "standardized %s" % Xsc.title
# Scaling
# -------
self._scaled = scaled
if scaled:
self._min = np.min(Xsc, axis=0)
self._ampl = np.ptp(Xsc, axis=0)
Xsc -= self._min
Xsc /= self._ampl
Xsc.title = "scaled %s" % Xsc.title
self._Xscaled = Xsc
# perform SVD
# -----------
svd = SVD(Xsc)
sigma = svd.s.diag()
U = svd.U
VT = svd.VT
# select n_pc loadings & compute scores
# --------------------------------------------------------------------
# loadings
LT = VT
LT.title = 'loadings (L^T) of ' + X.name
LT.history = 'Created by PCA'
# scores
S = dot(U, sigma)
S.title = 'scores (S) of ' + X.name
S.set_coordset(y=X.y,
x=Coord(None, labels=['#%d' % (i + 1) for i in range(svd.s.size)], title='principal component'))
S.description = 'scores (S) of ' + X.name
S.history = 'Created by PCA'
self._LT = LT
self._S = S
# other attributes
# ----------------
self._sv = svd.sv
self._sv.x.title = 'PC #'
self._ev = svd.ev
self._ev.x.title = 'PC #'
self._ev_ratio = svd.ev_ratio
self._ev_ratio.x.title = 'PC #'
self._ev_cum = svd.ev_cum
self._ev_cum.x.title = 'PC #'
return
def __init__(self, dataset, guess, **kwargs):
# list all default arguments:
tol = kwargs.get("tol", 0.1)
maxit = kwargs.get("maxit", 50)
maxdiv = kwargs.get("maxdiv", 5)
nonnegConc = kwargs.get("nonnegConc", "all")
unimodConc = kwargs.get("unimodConc", "all")
unimodConcTol = kwargs.get("unimodConcTol", 1.1)
unimodConcMod = kwargs.get("unimodMod", "strict")
if "unimodTol" in kwargs.keys():
warnings.warn("unimodTol deprecated, use unimodConcTol instead",
DeprecationWarning)
unimodConcTol = kwargs.get("unimodTol", 1.1)
if "unimodMod" in kwargs.keys():
warnings.warn("unimodMod deprecated, use unimodConcMod instead",
DeprecationWarning)
unimodConcMod = kwargs.get("unimodConcMod", "strict")
monoDecConc = kwargs.get("monoDecConc", None)
monoIncTol = kwargs.get("monoIncTol", 1.1)
monoIncConc = kwargs.get("monoIncConc", None)
monoDecTol = kwargs.get("monoDecTol", 1.1)
closureConc = kwargs.get("closureConc", None)
closureTarget = kwargs.get("closureTarget", "default")
closureMethod = kwargs.get("closureMethod", "scaling")
hardConc = kwargs.get("hardConc", None)
getConc = kwargs.get("getConc", None)
argsGetConc = kwargs.get("argsGetConc", None)
hardC_to_C_idx = kwargs.get("hardC_to_C_idx", "default")
unimodSpec = kwargs.get("unimodSpec", None)
unimodSpecTol = kwargs.get("unimodSpecTol", 1.1)
unimodSpecMod = kwargs.get("unimodSpecMod", "strict")
nonnegSpec = kwargs.get("nonnegSpec", "all")
normSpec = kwargs.get("normSpec", None)
if "verbose" in kwargs.keys():
warnings.warn(
"verbose deprecated. Instead, use set_loglevel(INFO) before launching MCRALS",
DeprecationWarning,
)
set_loglevel(INFO)
# Check initial data
# ------------------------------------------------------------------------
initConc, initSpec = False, False
if type(guess) is np.ndarray:
guess = NDDataset(guess)
X = dataset
if X.shape[0] == guess.shape[0]:
initConc = True
C = guess.copy()
C.name = "Pure conc. profile, mcs-als of " + X.name
nspecies = C.shape[1]
elif X.shape[1] == guess.shape[1]:
initSpec = True
St = guess.copy()
St.name = "Pure spectra profile, mcs-als of " + X.name
nspecies = St.shape[0]
else:
raise ValueError("the dimensions of guess do not match the data")
ny, _ = X.shape
# makes a PCA with same number of species for further comparison
Xpca = PCA(X).reconstruct(n_pc=nspecies)
# reset default text to indexes
# ------------------------------
if nonnegConc == "all":
nonnegConc = np.arange(nspecies)
elif nonnegConc is None:
nonnegConc = []
elif nonnegConc != [] and (len(nonnegConc) > nspecies
or max(nonnegConc) + 1 > nspecies):
raise ValueError(
f"The guess has only {nspecies} species, please check nonnegConc"
)
if unimodConc == "all":
unimodConc = np.arange(nspecies)
elif unimodConc is None:
unimodConc = []
elif unimodConc != [] and (len(unimodConc) > nspecies
or max(unimodConc) + 1 > nspecies):
raise ValueError(
f"The guess has only {nspecies} species, please check unimodConc"
)
if closureTarget == "default":
closureTarget = np.ones(ny)
elif len(closureTarget) != ny:
raise ValueError(
f"The data contain only {ny} observations, please check closureTarget"
)
if hardC_to_C_idx == "default":
hardC_to_C_idx = np.arange(nspecies)
elif len(hardC_to_C_idx
) > nspecies or max(hardC_to_C_idx) + 1 > nspecies:
raise ValueError(
f"The guess has only {nspecies} species, please check hardC_to_C_idx"
)
# constraints on spectra
if unimodSpec == "all":
unimodSpec = np.arange(nspecies)
elif unimodSpec is None:
unimodSpec = []
elif unimodSpec != [] and (len(unimodSpec) > nspecies
or max(unimodSpec) + 1 > nspecies):
raise ValueError(
f"The guess has only {nspecies} species, please check unimodSpec"
)
if nonnegSpec == "all":
nonnegSpec = np.arange(nspecies)
elif nonnegSpec is None:
nonnegSpec = []
elif nonnegSpec != [] and (len(nonnegSpec) > nspecies
or max(nonnegSpec) + 1 > nspecies):
raise ValueError(
f"The guess has only {nspecies} species, please check nonnegSpec"
)
# Compute initial spectra or concentrations (first iteration...)
# ------------------------------------------------------------------------
if initConc:
if C.coordset is None:
C.set_coordset(y=X.y, x=C.x)
St = NDDataset(np.linalg.lstsq(C.data, X.data, rcond=None)[0])
St.name = "Pure spectra profile, mcs-als of " + X.name
St.title = X.title
cy = C.x.copy() if C.x else None
cx = X.x.copy() if X.x else None
St.set_coordset(y=cy, x=cx)
if initSpec:
if St.coordset is None:
St.set_coordset(y=St.y, x=X.x)
Ct = np.linalg.lstsq(St.data.T, X.data.T, rcond=None)[0]
C = NDDataset(Ct.T)
C.name = "Pure conc. profile, mcs-als of " + X.name
C.title = "concentration"
cx = St.y.copy() if St.y else None
cy = X.y.copy() if X.y else None
C.set_coordset(y=cy, x=cx)
change = tol + 1
stdev = X.std()
niter = 0
ndiv = 0
log = "*** ALS optimisation log***\n"
log += "#iter Error/PCA Error/Exp %change \n"
log += "------------------------------------------------- \n"
info_(log)
while change >= tol and niter < maxit and ndiv < maxdiv:
C.data = np.linalg.lstsq(St.data.T, X.data.T, rcond=None)[0].T
niter += 1
# Force non-negative concentration
# --------------------------------
if nonnegConc is not None:
for s in nonnegConc:
C.data[:, s] = C.data[:, s].clip(min=0)
# Force unimodal concentration
# ----------------------------
if unimodConc != []:
C.data = _unimodal_2D(
C.data,
idxes=unimodConc,
axis=0,
tol=unimodConcTol,
mod=unimodConcMod,
)
# Force monotonic increase
# ------------------------
if monoIncConc is not None:
for s in monoIncConc:
for curid in np.arange(ny - 1):
if C.data[curid + 1,
s] < C.data[curid, s] / monoIncTol:
C.data[curid + 1, s] = C.data[curid, s]
# Force monotonic decrease
# ----------------------------------------------
if monoDecConc is not None:
for s in monoDecConc:
for curid in np.arange(ny - 1):
if C.data[curid + 1,
s] > C.data[curid, s] * monoDecTol:
C.data[curid + 1, s] = C.data[curid, s]
# Closure
# ------------------------------------------
if closureConc is not None:
if closureMethod == "scaling":
Q = np.linalg.lstsq(C.data[:, closureConc],
closureTarget.T,
rcond=None)[0]
C.data[:, closureConc] = np.dot(C.data[:, closureConc],
np.diag(Q))
elif closureMethod == "constantSum":
totalConc = np.sum(C.data[:, closureConc], axis=1)
C.data[:, closureConc] = (C.data[:, closureConc] *
closureTarget[:, None] /
totalConc[:, None])
# external concentration profiles
# ------------------------------------------
if hardConc is not None:
extOutput = getConc(*argsGetConc)
if isinstance(extOutput, dict):
fixedC = extOutput["concentrations"]
argsGetConc = extOutput["new_args"]
else:
fixedC = extOutput
C.data[:, hardConc] = fixedC[:, hardC_to_C_idx]
# stores C in C_hard
Chard = C.copy()
# compute St
St.data = np.linalg.lstsq(C.data, X.data, rcond=None)[0]
# stores St in Stsoft
Stsoft = St.copy()
# Force non-negative spectra
# --------------------------
if nonnegSpec is not None:
St.data[nonnegSpec, :] = St.data[nonnegSpec, :].clip(min=0)
# Force unimodal spectra
# ----------------------------
if unimodSpec != []:
St.data = _unimodal_2D(
St.data,
idxes=unimodSpec,
axis=1,
tol=unimodSpecTol,
mod=unimodSpecMod,
)
# recompute C for consistency(soft modeling)
C.data = np.linalg.lstsq(St.data.T, X.data.T)[0].T
# rescale spectra & concentrations
if normSpec == "max":
alpha = np.max(St.data, axis=1).reshape(nspecies, 1)
St.data = St.data / alpha
C.data = C.data * alpha.T
elif normSpec == "euclid":
alpha = np.linalg.norm(St.data, axis=1).reshape(nspecies, 1)
St.data = St.data / alpha
C.data = C.data * alpha.T
# compute residuals
# -----------------
X_hat = dot(C, St)
stdev2 = (X_hat - X.data).std()
change = 100 * (stdev2 - stdev) / stdev
stdev = stdev2
stdev_PCA = (X_hat - Xpca.data).std() #
logentry = "{:3d} {:10f} {:10f} {:10f}".format(
niter, stdev_PCA, stdev2, change)
log += logentry + "\n"
info_(logentry)
if change > 0:
ndiv += 1
else:
ndiv = 0
change = -change
if change < tol:
logentry = "converged !"
log += logentry + "\n"
info_(logentry)
if ndiv == maxdiv:
logline = (
f"Optimization not improved since {maxdiv} iterations... unconverged "
f"or 'tol' set too small ?\n")
logline += "Stop ALS optimization"
log += logline + "\n"
info_(logline)
if niter == maxit:
logline = "Convergence criterion ('tol') not reached after {:d} iterations.".format(
maxit)
logline += "Stop ALS optimization"
log += logline + "\n"
info_(logline)
self._X = X
self._params = {
"tol": tol,
"maxit": maxit,
"maxdiv": maxdiv,
"nonnegConc": nonnegConc,
"unimodConc": unimodConc,
"unimodConcTol": unimodConcTol,
"unimodConcMod": unimodConcMod,
"closureConc": closureConc,
"closureTarget ": closureTarget,
"closureMethod": closureMethod,
"monoDecConc": monoDecConc,
"monoDecTol": monoDecTol,
"monoIncConc": monoIncConc,
"monoIncTol": monoIncTol,
"hardConc": hardConc,
"getConc": getConc,
"argsGetConc": argsGetConc,
"hardC_to_C_idx": hardC_to_C_idx,
"nonnegSpec": nonnegSpec,
"unimodSpec": unimodConc,
"unimodSpecTol": unimodSpecTol,
"unimodSpecMod": unimodSpecMod,
"normSpec": normSpec,
}
self._C = C
if hardConc is not None:
self._fixedC = fixedC
self._extOutput = extOutput
else:
self._fixedC = None
self._extOutput = None
self._St = St
self._log = log
self._Stsoft = Stsoft
self._Chard = Chard
def __init__(self,
dataset,
centered=True,
standardized=False,
scaled=False):
super().__init__()
self.prefs = dataset.preferences
self._X = X = dataset
Xsc = X.copy()
# mean center the dataset
# -----------------------
self._centered = centered
if centered:
self._center = center = X.mean(dim=0)
Xsc = X - center
Xsc.name = f"centered {X.name}"
# Standardization
# ---------------
self._standardized = standardized
if standardized:
self._std = Xsc.std(dim=0)
Xsc /= self._std
Xsc.name = f"standardized {Xsc.name}"
# Scaling
# -------
self._scaled = scaled
if scaled:
self._min = Xsc.min(dim=0)
self._ampl = Xsc.ptp(dim=0)
Xsc -= self._min
Xsc /= self._ampl
Xsc.name = "scaled %s" % Xsc.name
self._Xscaled = Xsc
# perform SVD
# -----------
svd = SVD(Xsc)
sigma = svd.s.diag()
U = svd.U
VT = svd.VT
# select n_pc loadings & compute scores
# --------------------------------------------------------------------
# loadings
LT = VT
LT.title = "loadings (L^T) of " + X.name
LT.history = "Created by PCA"
# scores
S = dot(U, sigma)
S.title = "scores (S) of " + X.name
S.set_coordset(
y=X.y,
x=Coord(
None,
labels=["#%d" % (i + 1) for i in range(svd.s.size)],
title="principal component",
),
)
S.description = "scores (S) of " + X.name
S.history = "Created by PCA"
self._LT = LT
self._S = S
# other attributes
# ----------------
self._sv = svd.sv
self._sv.x.title = "PC #"
self._ev = svd.ev
self._ev.x.title = "PC #"
self._ev_ratio = svd.ev_ratio
self._ev_ratio.x.title = "PC #"
self._ev_cum = svd.ev_cum
self._ev_cum.x.title = "PC #"
return
def __init__(self, dataset, guess, **kwargs): # lgtm [py/missing-call-to-init]
"""
Parameters
----------
dataset : |NDDataset|
The dataset on which to perform the MCR-ALS analysis
guess : |NDDataset|
Initial concentration or spectra
verbose : bool
If set to True, prints a summary of residuals and residuals change at each iteration. default = False.
In any case, the same information is returned in self.logs
**kwargs : dict
Optimization parameters : See Other Parameters.
Other Parameters
----------------
tol : float, optional, default=0.1
Convergence criterion on the change of resisuals.
(percent change of standard deviation of residuals).
maxit : int, optional, default=50
Maximum number of ALS minimizations.
maxdiv : int, optional, default=5.
Maximum number of successive non-converging iterations.
nonnegConc : list or tuple, default=Default [0, 1, ...] (only non-negative concentrations)
Index of species having non-negative concentration profiles. For instance [0, 2] indicates that species
#0 and #2 have non-negative conc profiles while species #1 can have negative concentrations.
unimodConc : list or tuple, Default=[0, 1, ...] (only unimodal concentration profiles)
index of species having unimodal concentrationsprofiles.
closureConc : list or tuple, Default=None (no closure)
Index of species subjected to a closure constraint.
externalConc: list or tuple, Default None (no external concentration).
Index of species for which a concentration profile is provided by an external function.
getExternalConc : callable
An external function that will provide `n_ext` concentration profiles:
getExternalConc(C, extConc, ext_to_C_idx, *args) -> extC
or
etExternalConc(C, extConc, ext_to_C_idx, *args) -> (extC, out2, out3, ...)
where C is the current concentration matrix, *args are the parameters needed to completely
specify the function, extC is a nadarray or NDDataset of shape (C.y, n_ext), and out1, out2, ... are
supplementary outputs returned by the function (e.g. optimized rate parameters)
args : tuple, optional.
Extra arguments passed to the external function
external_to_C_idx : array or tuple, Default=np.arange(next)
Indicates the correspondence between the indexes of external chemical
profiles and the columns of the C matrix. [1, None, 0] indicates that the first external profile is the
second pure species (index 1).
nonnegSpec : list or tuple, Default [1, ..., 1] (only non-negative spectra)
Indicates species having non-negative spectra
unimodSpec : list or tuple, Default [0, ..., 0] (no unimodal concentration profiles)
Indicates species having unimodal spectra
"""
verbose = kwargs.pop('verbose', False)
if verbose:
set_loglevel(INFO)
# Check initial data
# ------------------------------------------------------------------------
initConc, initSpec = False, False
if type(guess) is np.ndarray:
guess = NDDataset(guess)
X = dataset
if X.shape[0] == guess.shape[0]:
initConc = True
C = guess.copy()
C.name = 'Pure conc. profile, mcs-als of ' + X.name
nspecies = C.shape[1]
elif X.shape[1] == guess.shape[1]:
initSpec = True
St = guess.copy()
St.name = 'Pure spectra profile, mcs-als of ' + X.name
nspecies = St.shape[0]
else:
raise ValueError('the dimensions of initial concentration '
'or spectra dataset do not match the data')
ny, nx = X.shape
# makes a PCA with same number of species
Xpca = PCA(X).reconstruct(n_pc=nspecies)
# Get optional parameters in kwargs or set them to their default
# ------------------------------------------------------------------------
# TODO: make a preference file to set this kwargs
# optimization
tol = kwargs.get('tol', 0.1)
maxit = kwargs.get('maxit', 50)
maxdiv = kwargs.get('maxdiv', 5)
# constraints on concentrations
nonnegConc = kwargs.get('nonnegConc', np.arange(nspecies))
unimodConc = kwargs.get('unimodConc', np.arange(nspecies))
unimodTol = kwargs.get('unimodTol', 1.1)
unimodMod = kwargs.get('unimodMod', 'strict')
closureConc = kwargs.get('closureConc', None)
if closureConc is not None:
closureTarget = kwargs.get('closureTarget', np.ones(ny))
closureMethod = kwargs.get('closureMethod', 'scaling')
monoDecConc = kwargs.get('monoDecConc', None)
monoDecTol = kwargs.get('monoDecTol', 1.1)
monoIncConc = kwargs.get('monoIncConc', None)
monoIncTol = kwargs.get('monoIncTol', 1.1)
externalConc = kwargs.get('externalConc', None)
if externalConc is not None:
external_to_C_idx = kwargs.get('external_to_C_idx', np.arange(nspecies))
if externalConc is not None:
try:
getExternalConc = kwargs.get('getExternalConc')
except Exception:
raise ValueError('A function must be given to get the external concentration profile(s)')
external_to_C_idx = kwargs.get('external_to_C_idx', externalConc)
args = kwargs.get('args', ())
# constraints on spectra
nonnegSpec = kwargs.get('nonnegSpec', np.arange(nspecies))
normSpec = kwargs.get('normSpec', None)
# TODO: add unimodal constraint on spectra
# Compute initial spectra or concentrations (first iteration...)
# ------------------------------------------------------------------------
if initConc:
if C.coordset is None:
C.set_coordset(y=X.y, x=C.x)
St = NDDataset(np.linalg.lstsq(C.data, X.data, rcond=None)[0])
St.name = 'Pure spectra profile, mcs-als of ' + X.name
St.title = X.title
cy = C.x.copy() if C.x else None
cx = X.x.copy() if X.x else None
St.set_coordset(y=cy, x=cx)
if initSpec:
if St.coordset is None:
St.set_coordset(y=St.y, x=X.x)
Ct = np.linalg.lstsq(St.data.T, X.data.T, rcond=None)[0]
C = NDDataset(Ct.T)
C.name = 'Pure conc. profile, mcs-als of ' + X.name
C.title = 'concentration'
cx = St.y.copy() if St.y else None
cy = X.y.copy() if X.y else None
C.set_coordset(y=cy, x=cx)
change = tol + 1
stdev = X.std() # .data[0]
niter = 0
ndiv = 0
logs = '*** ALS optimisation log***\n'
logs += '#iter Error/PCA Error/Exp %change\n'
logs += '---------------------------------------------------'
info_(logs)
while change >= tol and niter < maxit and ndiv < maxdiv:
C.data = np.linalg.lstsq(St.data.T, X.data.T, rcond=None)[0].T
niter += 1
# Force non-negative concentration
# --------------------------------
if nonnegConc is not None:
for s in nonnegConc:
C.data[:, s] = C.data[:, s].clip(min=0)
# Force unimodal concentration
# ----------------------------
if unimodConc is not None:
for s in unimodConc:
maxid = np.argmax(C.data[:, s])
curmax = C.data[maxid, s]
curid = maxid
while curid > 0:
curid -= 1
if C.data[curid, s] > curmax * unimodTol:
if unimodMod == 'strict':
C.data[curid, s] = C.data[curid + 1, s]
if unimodMod == 'smooth':
C.data[curid, s] = (C.data[curid, s] + C.data[
curid + 1, s]) / 2
C.data[curid + 1, s] = C.data[curid, s]
curid = curid + 2
curmax = C.data[curid, s]
curid = maxid
while curid < ny - 1:
curid += 1
if C.data[curid, s] > curmax * unimodTol:
if unimodMod == 'strict':
C.data[curid, s] = C.data[curid - 1, s]
if unimodMod == 'smooth':
C.data[curid, s] = (C.data[curid, s] + C.data[
curid - 1, s]) / 2
C.data[curid - 1, s] = C.data[curid, s]
curid = curid - 2
curmax = C.data[curid, s]
# Force monotonic increase
# ------------------------
if monoIncConc is not None:
for s in monoIncConc:
for curid in np.arange(ny - 1):
if C.data[curid + 1, s] < C.data[curid, s] / monoIncTol:
C.data[curid + 1, s] = C.data[curid, s]
# Force monotonic decrease
# ----------------------------------------------
if monoDecConc is not None:
for s in monoDecConc:
for curid in np.arange(ny - 1):
if C.data[curid + 1, s] > C.data[curid, s] * monoDecTol:
C.data[curid + 1, s] = C.data[curid, s]
# Closure
# ------------------------------------------
if closureConc is not None:
if closureMethod == 'scaling':
Q = np.linalg.lstsq(C.data[:, closureConc], closureTarget.T, rcond=None)[0]
C.data[:, closureConc] = np.dot(C.data[:, closureConc], np.diag(Q))
elif closureMethod == 'constantSum':
totalConc = np.sum(C.data[:, closureConc], axis=1)
C.data[:, closureConc] = C.data[:, closureConc] * closureTarget[:, None] / totalConc[:, None]
# external concentration profiles
# ------------------------------------------
if externalConc is not None:
extOutput = getExternalConc(*((C, externalConc, external_to_C_idx,) + args))
if isinstance(extOutput, dict):
extC = extOutput['concentrations']
args = extOutput['new_args']
else:
extC = extOutput
if type(extC) is NDDataset:
extC = extC.data
C.data[:, externalConc] = extC[:, external_to_C_idx]
# stores C in C_hard
Chard = C.copy()
# compute St
St.data = np.linalg.lstsq(C.data, X.data, rcond=None)[0]
# stores St in Stsoft
Stsoft = St.copy()
# Force non-negative spectra
# --------------------------
if nonnegSpec is not None:
St.data[nonnegSpec, :] = St.data[nonnegSpec, :].clip(min=0)
# recompute C for consistency(soft modeling)
C.data = np.linalg.lstsq(St.data.T, X.data.T, rcond=None)[0].T
# rescale spectra & concentrations
if normSpec == 'max':
alpha = np.max(St.data, axis=1).reshape(nspecies, 1)
St.data = St.data / alpha
C.data = C.data * alpha.T
elif normSpec == 'euclid':
alpha = np.linalg.norm(St.data, axis=1).reshape(nspecies, 1)
St.data = St.data / alpha
C.data = C.data * alpha.T
# compute residuals
# -----------------
X_hat = dot(C, St)
stdev2 = (X_hat - X.data).std()
change = 100 * (stdev2 - stdev) / stdev
stdev = stdev2
stdev_PCA = (X_hat - Xpca.data).std() # TODO: Check PCA : values are different from the Arnaud version ?
logentry = '{:3d} {:10f} {:10f} {:10f}'.format(niter, stdev_PCA, stdev2, change)
logs += logentry + '\n'
info_(logentry)
if change > 0:
ndiv += 1
else:
ndiv = 0
change = -change
if change < tol:
logentry = 'converged !'
logs += logentry + '\n'
info_(logentry)
if ndiv == maxdiv:
logline = f"Optimization not improved since {maxdiv} iterations... unconverged " \
f"or 'tol' set too small ?\n"
logline += 'Stop ALS optimization'
logs += logline + '\n'
info_(logline)
if niter == maxit:
logline = 'Convergence criterion (\'tol\') not reached after {:d} iterations.'.format(maxit)
logline += 'Stop ALS optimization'
logs += logline + '\n'
info_(logline)
self._X = X
self._params = kwargs
self._C = C
if externalConc is not None:
self._extC = extC
self._extOutput = extOutput
else:
self._extC = None
self._extOutput = None
self._St = St
self._logs = logs
self._Stsoft = Stsoft
self._Chard = Chard