# parameters for the simulation
m   = 1.
sig = 0.5

# initialization of the grid
grd = gl.DbGrid.create(x0=(0.0,0.0), dx=(0.01,0.01), nx=(100,100))

# simulation from a model
model = gl.Model.createFromParam(gl.ECov.EXPONENTIAL, range=0.2)
gl.simtub(dbin = None, dbout = grd, model = model, nbsimu = 1)
grd.setName("Simu", "Y")
grd["Z"] = m * np.exp(sig * grd["Y"].squeeze() - sig**2 / 2)

# Data set (10% of the grid)
data = gl.Db.createSamplingDb(grd, 0.1, flagAddSampleRank=0)
data.useSel = True

# plots
fig, (ax1,ax2) = gp.init(1,2, figsize=(15,6))
ax1.raster(grd,"Z")
ax1.decoration(title="Initial variable")
ax2.histogram(grd, name='Z',  bins = 25, color="orange")
ax2.decoration(xlabel = "Raw variable", title="Histogram of the initial variable")

fig, ax3 = gp.init(figsize=(6,6))
ax3.raster(grd,"Z")
ax3.symbol(data,nameSize="Z", c='yellow')
ax3.decoration(title="Random data subsample")

zcut = np.quantile(data['Z'], q = [0.3, 0.5, 0.7, 0.9])

print("\n{:^40}".format("Coupures sur la variable Z"))
print("{:^10}{:^10}{:^10}{:^10}".format(0.3, 0.5, 0.7, 0.9))
print("{:^10.3f}{:^10.3f}{:^10.3f}{:^10.3f}".format(*zcut),'\n')

mylimits = gl.Limits(zcut, True) #defines limits based on the cutoff values 
# True for addFromZero, so that we add the interval [0, z_1[
mylimits.display()

# Compute the indicators (4 new variables)
iud_Indic = mylimits.toIndicator(data, name='Z', OptionIndicator=1) 
# Compute the discretized version of the variable
iud_Mean  = mylimits.toIndicator(data, name='Z', OptionIndicator=0) 

# statistics on the indicators
w = gl.dbStatisticsMono(data, ["Indicator.Z.Class*"], [gl.EStatOption.MEAN]).getValues()
w = list(w) + [1 - np.sum(w)]
print("Proportions = ", np.round(w,3))
Nclass = len(w) # total number of indicators, including the fifth one which is not computed

print(data)
gp.symbol(data, nameColor="*Mean")
gp.decoration(title="Discretized variable")

# Locate Z
data.clearLocators(gl.ELoc.Z)
data.setLocator("Z", gl.ELoc.Z)

# Variogram parameters (omnidirectional)
varioParam = gl.VarioParam.createOmniDirection(nlag=10, dlag=0.05)
# Compute variogram
var_Z = gl.Vario.computeFromDb(varioParam, data)

# fit model
mod_Z = gl.Model()
opt=gl.Option_AutoFit()
opt.setWmode(2) # weighted proportional to the number of pairs and inverse proportional to the distance
mod_Z.fit(var_Z, [gl.ECov.NUGGET, gl.ECov.EXPONENTIAL, gl.ECov.EXPONENTIAL], mauto = opt)

# plot
gp.varmod(var_Z, mod_Z, flagLegend=True)
gp.decoration(title = "Variogram of the data sample")

mod_Z.display()

data.clearLocators(gl.ELoc.Z)
data.setLocator("Indicator.Z.Mean", gl.ELoc.Z)

var_Z = gl.Vario.computeFromDb(varioParam, data)

mod_Z = gl.Model()
mod_Z.fit(var_Z, [gl.ECov.NUGGET, gl.ECov.EXPONENTIAL, gl.ECov.EXPONENTIAL], mauto = opt)

gp.varmod(var_Z, mod_Z, flagLegend=True)
gp.decoration(title = "Variogram of the discretized variable")

mod_Z.display()

gp.varmod(var_Z, mod_Z)
gp.decoration(title = "Simple and cross variograms of the indicators of Z", fontsize=20)
gp.geometry(dims=(15,15))

data.clearLocators(gl.ELoc.Z)
data.setLocator("Indicator.Z.Class*", gl.ELoc.Z)

maf_I = gl.PCA()
maf_I.maf_compute_interval(data, hmin=0.045, hmax=0.055)
maf_I.display()

#extract matrices of conversion between factors and indicators
#Mz2f = np.reshape(maf_I.getZ2Fs(),(Nclass-1,Nclass-1)).T
#Mf2z = np.reshape(maf_I.getF2Zs(),(Nclass-1,Nclass-1)).T

maf_I.mafOfIndex()

# Calcul des maf comme fonction de l'index
maf = np.reshape(maf_I.mafOfIndex(), (Nclass, Nclass)).T

## Check that the factors are orthogonal
tab = np.dot(np.dot(maf.T, np.diag(w)),maf).round(10)
print("Correlation between factors MAF")
print(tab)

## plot MAF as a function of the index
fig, ax = gp.init(figsize = (10,6))
for k in range(Nclass-1):
    ax.plot(range(Nclass), np.sign(maf[0, k+1]) * maf[:,k+1], marker='.', label="MAF-"+str(k+1))
ax.set_xticks(range(Nclass), range(1, Nclass+1))
ax.hlines(0, 0, Nclass-1, color='black', linestyle='dashed', linewidth=0.7)
ax.set_xlabel("Indice des classes")
ax.set_ylabel("MAF")
l = ax.legend()

gp.varmod(vario_maf, model_maf)
gp.decoration(ax, title = "Variograms of the MAFs", fontsize=20)
gp.geometry(dims=(15,15))

# models of the MAFs

fig, axs = gp.init(2,2, figsize=(11,7))

varios, models = [], []
for f in range(Nclass-1):
    vario_reduced = gl.Vario(vario_maf)
    vario_reduced.resetReduce(varcols = [f], dircols = [])
    model = gl.Model()
    opt=gl.Option_AutoFit()
    opt.setWmode(0)
    model.fit(vario_reduced, [gl.ECov.NUGGET, gl.ECov.EXPONENTIAL, gl.ECov.EXPONENTIAL], 
              mauto=opt) 
    varios.append(vario_reduced)
    models.append(model)
    
    axs.flat[f].decoration(f"Simple variogram of MAF{f+1}")
    axs.flat[f].varmod(vario_reduced, model)

### outils de gstlearn : a ameliorer ? Ci-dessous permet de trouver l'estimation des indicatrices 
# à partir des MAFs:
# gridData.setLocator("Kriging.F*.estim", gl.ELoc.Z)
# maf_I.dbF2Z(gridData) # estimation of the indicators
### mais 2 manques:
 # avoir directement l'estimation d'une autre variable (ici S15), à partir des coefficient 
 # de la combinaison linéaire des indicatrices
 # ne prendre en compte que les premiers MAFs (eg dans le cas où les derniers MAFs ne sont pas structurés, 
 # c'est inutile de les kriger)
 # en attendant j'utilise la fonction définie ci-dessous

def vecF2myvar(maf, VecF, coefs=None, mean=None, SigmaF=None):
    """Compute values of a variable that is a linear combination of Z variables, from vectors of factors F.
    maf: object containing infos on the MAF computation (class PCA)
    VecF: table containing values of the factors, shape (Ndata*NF), where NF is the number of factors used 
          (if NF is lower than the total number of factors, then only the first ones are used, 
           the others are considered not structured spatially)
    coefs : coefficients of the variable along all variables Z
    mean : mean of the variable.
    SigmaF: table of the standard deviation of the estimation error of the factors, same shape as VecF. 
            If given, the standard deviation of the output variable are also computed and returned.
    
    eg: coefs = [1,0,0,0,0] is the first variable (indicator), with mean = 0 if not specified.
    """
    VecF = np.atleast_2d(VecF)
    Ndata, NF = VecF.shape
    Nvar = maf.getNVar()
    if NF < Nvar:
        print("The number of factors is less than the number of variables. The other factors will be neglected.")
    
    Mf2z      = maf.getF2Zs().toTL()
    sigma     = np.array(maf.getSigmas())
    means     = np.array(maf.getMeans())
    mean_N    = 1 - np.sum(means)
    coefs     = np.atleast_2d(coefs).T # one column = coefs for 1 variable
    
    if len(coefs) > Nvar + 1:
        raise ValueError(f"Too many coefficients were given ({len(coefs)}), coefs should be of length {Nvar} or {Nvar+1}.")
        
    if mean is None and len(coefs) == Nvar + 1:
        all_means = np.array(list(means) + [mean_N])
        mean = np.sum(coefs.T*all_means, axis=1)
    elif mean is None:
        raise ValueError("The mean of the variable should be given when only the N-1 coefficients are given.")
    
    if len(coefs) == Nvar:
        c_N = 1/mean_N*(mean - np.sum(coefs.T*means, axis=1))
        coefs = np.append(coefs, np.atleast_2d(c_N).T, axis=1) # add c_N
    if len(coefs) != Nvar + 1:
        print("Wrong len for coefs : ", len(coefs))
    
    Vweights = (sigma*(coefs[:-1] - coefs[-1]).T).T # w_i = (c_i - c_N)*sigma_i
    M = np.dot(Mf2z[:NF,:], Vweights)
    VecMyvar = np.dot(VecF, M) + mean
    
    if SigmaF is not None:
        VecSigmaMyvar = np.sqrt(np.dot(SigmaF**2, M**2))
        return np.squeeze(VecMyvar), VecSigmaMyvar.squeeze()
    else:
        return VecMyvar.squeeze()

# extract values of the first two MAFs (the other ones have only nugget effect)
grd.useSel=True
VecMAF   =  grd["Kriging.F.*.estim"]
SigmaMAF =  grd["Kriging.F.*.stdev"]

# estimation of the indicators
INDestim, INDstdev = vecF2myvar(maf_I, VecMAF, coefs=np.eye(5), SigmaF=SigmaMAF)
for i_indicator in range(5):
    grd[f"Kriging.Indicator{i_indicator+1}.estim"] = INDestim[:,i_indicator]
    
# estimation of S15
mean_per_class = np.unique(data["*Mean"])
Zestim, Zstdev = vecF2myvar(maf_I, VecMAF, coefs=mean_per_class, SigmaF=SigmaMAF, mean=np.mean(data['Z']))

grd["Kriging.Z.estim"] = Zestim # Z*
grd["Kriging.Z.stdev"] = Zstdev # standard deviation of the estimation error Var(Z*-Z) = S

fig, axs = gp.init(1,2, figsize=(10,5))
axs[0].raster(grd,"Kriging.F*1.estim")
axs[0].decoration(title='Estimation of MAF1')
axs[1].raster(grd,"Kriging.F*2.estim")
axs[1].decoration(title='Estimation of MAF2')

fig, axs = gp.init(1,3, figsize=(15,5))
axs[0].raster(grd,"Z")
axs[0].decoration(title='Initial Variable Z')
axs[1].raster(grd,"Kriging.Z.estim")
axs[1].decoration(title='Estimation of Z with DDK')
axs[2].raster(grd,"Kriging.Z.stdev")
axs[2].decoration(title='Standard deviation of kriging')