-
Notifications
You must be signed in to change notification settings - Fork 2
/
Copy pathSEG2020.py
444 lines (418 loc) · 25 KB
/
SEG2020.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
423
424
425
426
427
428
429
430
431
432
433
434
435
436
437
438
439
440
441
442
443
444
'''
SEG2020 is a script that runs all the computations for the different results
presented at the SEG2020 conference (Houston, Texas).
***Improving BEL1D accuracy for geophysical imaging of the subsurface***
It runs at first the numerical benchmark for a dataset that is created directly.
- Creating the dataset
- Running BEL1D (initialization + first iteration)
- Presenting graphs of the results
- Applying IPR
- Presenting graphs of the improved result
- Comparing to McMC results from DREAM (results provided in github)
Then, it runs the same for the Mont Rigi dataset and presents simply the obtained profile.
Author:
Hadrien MICHEL
ULiège, UGent, F.R.S.-FNRS
hadrien[dot]michel[at]uliege[dot]be
(c) August 2020
'''
def round_to_5(x,n=1):
import math as mt
# Modified from: https://stackoverflow.com/questions/3410976/how-to-round-a-number-to-significant-figures-in-python
tmp = [(round(a, -int(mt.floor(mt.log10(abs(a)))) + (n-1)) if a != 0.0 else 0.0) for a in x]
return tmp
if __name__=='__main__':
import numpy as np
from pyBEL1D import BEL1D
from pyBEL1D.utilities import Tools
from matplotlib import pyplot
from pathos import multiprocessing as mp
from pathos import pools as pp
import time
################################################################################
### ###
### Numerical Benchmarking ###
### ###
################################################################################
Benchmark = True
if Benchmark:
# 1) Creating the dataset:
from pygimli.physics import sNMR
KernelBench = "Data/sNMR/Tx50Rx50.mrsk" # A kernel file generated by MRSMatlab (Mueller-Petke et al., 2012)
nbLayer = 3 # 2 layers and a half-space
TimingBench = np.arange(0.005,0.5,0.001) # The time vector for the model
# KFile = sNMR.MRS()
# KFile.loadKernel(KernelBench)
# ModellingMethod = sNMR.MRS1dBlockQTModelling(nlay=nbLayer,K=KFile.K,zvec=KFile.z,t=TimingBench)
ModelBench = np.asarray([25, 25, 0.05, 0.25, 0.1, 0.1, 0.2, 0.05])
Noise = 10 # nV
# DatasetBench = ModellingMethod.response(ModelBench)
# DatasetBench += np.random.normal(loc=0, scale=Noise*1e-9,size=DatasetBench.shape) # Adding noise to the dataset
# np.save('sNMR_Bench_Dataset',DatasetBench)# To use in the McMC algorithm
DatasetBench = np.load('sNMR_Bench_Dataset.npy') # Load the dataset that was already created and run in McMC
# 2) Initializing BEL1D:
priorBench = np.array([[0.0, 50.0, 0.0, 0.15, 0.0, 0.5], [0.0, 50.0, 0.15, 0.50, 0.0, 0.5], [0.0, 0.0, 0.0, 0.15, 0.0, 0.5]])
# Initialsizing the parameters:
start = time.time()
ModelParam = BEL1D.MODELSET.SNMR(prior=priorBench, Kernel=KernelBench, Timing=TimingBench)
# 3) Running pre-BEL operations:
nbSampled = 10000
PreBEL_Bench = BEL1D.PREBEL(ModelParam,nbModels=nbSampled)
PreBEL_Bench.run()
# 4) Sampling 10000 models from the posterior:
PostBEL_Bench = BEL1D.POSTBEL(PreBEL_Bench)
PostBEL_Bench.run(Dataset=DatasetBench,nbSamples=nbSampled,NoiseModel=Noise)
end = time.time()
PostBEL_Bench.KDE.ShowKDE(Xvals=PostBEL_Bench.CCA.transform(PostBEL_Bench.PCA['Data'].transform(np.reshape(DatasetBench,(1,-1)))))
PostBEL_Bench.ShowDataset(RMSE=True)
CurrentGraph = pyplot.gcf()
CurrentGraph = CurrentGraph.get_axes()[0]
CurrentGraph.plot(TimingBench,DatasetBench[:len(TimingBench)],'k',linestyle='None',marker='o',markerfacecolor='None')
# Graph for the CCA space parameters loads
_, ax = pyplot.subplots()
B = PostBEL_Bench.CCA.y_loadings_
B = np.divide(np.abs(B).T,np.repeat(np.reshape(np.sum(np.abs(B),axis=0),(1,B.shape[0])),B.shape[0],axis=0).T)
ind = np.asarray(range(B.shape[0]))+1
ax.bar(x=ind,height=B[0],label=r'${}$'.format(PostBEL_Bench.MODPARAM.paramNames["NamesSU"][0]))
for i in range(B.shape[0]+1)[1:-1]:
ax.bar(x=ind,height=B[i],bottom=np.reshape(np.sum(B[0:i],axis=0),(B.shape[0],)),label=r'${}$'.format(PostBEL_Bench.MODPARAM.paramNames["NamesSU"][i]))
box = ax.get_position()
ax.set_position([box.x0, box.y0, box.width, box.height*0.8])
ax.legend(loc='upper center', bbox_to_anchor=(0.5, 1.4), ncol=3)
ax.set_ylabel('Relative contribution')
ax.set_xlabel('CCA dimension')
pyplot.show(block=False)
PostBEL_Bench.ShowPostModels(TrueModel=ModelBench,RMSE=True)
PostBEL_Bench.ShowPostCorr(TrueModel=ModelBench)
pyplot.show(block=False)
MODELS_1stITER = PostBEL_Bench.SAMPLES
MODELS_1stITER_DATA = PostBEL_Bench.SAMPLESDATA
PRE_MODS = PreBEL_Bench.MODELS
PRE_DATA = PreBEL_Bench.FORWARD
Postbel = PostBEL_Bench
Prebel = PreBEL_Bench
# 5) Applying IPR
nbIter = 100 # maximum number of iterations
tolerance = 5e-3 # Tolerance on the normalized difference between the distributions
nbParam = int(priorBench.size/2 - 1)
means = np.zeros((nbIter,nbParam))
stds = np.zeros((nbIter,nbParam))
timings = np.zeros((nbIter,))
MODELS_ITER = np.zeros((nbIter,nbSampled,nbParam)) # Store the models that have been computed
diverge = True
distancePrevious = 1e10
MixingUpper = 0
MixingLower = 1
for idxIter in range(nbIter):
PostbelLast = Postbel
PrebelLast = Prebel
if idxIter == 0: # Initialization: already done (see 2 and 3)
# PostbelTest.KDE.ShowKDE(Xvals=PostbelTest.CCA.transform(PostbelTest.PCA['Data'].transform(np.reshape(Dataset,(1,-1)))))
means[idxIter,:], stds[idxIter,:] = Postbel.GetStats()
timings[idxIter] = end-start
ModLastIter = Prebel.MODELS
else:
ModLastIter = Postbel.SAMPLES
# Here, we will use the POSTBEL2PREBEL function that adds the POSTBEL from previous iteration to the prior (Iterative prior resampling)
# However, the computations are longer with a lot of models, thus you can opt-in for the "simplified" option which randomely select up to 10 times the numbers of models
MixingUpper += 1
MixingLower += 1
Mixing = MixingUpper/MixingLower
Prebel = BEL1D.PREBEL.POSTBEL2PREBEL(PREBEL=Prebel,POSTBEL=Postbel,Dataset=DatasetBench,NoiseModel=Noise,Simplified=True,nbMax=nbSampled,MixingRatio=Mixing)
# Since when iterating, the dataset is known, we are not computing the full relationship but only the posterior distributions directly to gain computation timing
print(idxIter+1)
Postbel = BEL1D.POSTBEL(Prebel)
Postbel.run(DatasetBench,nbSamples=nbSampled,NoiseModel=None)
means[idxIter,:], stds[idxIter,:] = Postbel.GetStats()
end = time.time()
timings[idxIter] = end-start
# The distance is computed on the normalized distributions. Therefore, the tolerance is relative.
diverge, distance = Tools.ConvergeTest(SamplesA=ModLastIter,SamplesB=Postbel.SAMPLES, tol=tolerance)
print('Wasserstein distance: {}'.format(distance))
if not(diverge) or (abs((distancePrevious-distance)/distancePrevious)*100<1):# If the distance between the distributions is not changing, we converged as well
# Convergence acheived if:
# 1) Distance below threshold
# 2) Distance does not vary significantly (less than 2.5%)
print('Model has converged at iter {}!'.format(idxIter+1))
MODELS_ITER[idxIter,:,:] = Postbel.SAMPLES
break
distancePrevious = distance
MODELS_ITER[idxIter,:,:] = Postbel.SAMPLES
start = time.time()
timings = timings[:idxIter+1]
means = means[:idxIter+1,:]
stds = stds[:idxIter+1,:]
MODELS_ITER = MODELS_ITER[:idxIter+1,:,:]
np.save('ModelsIteration',MODELS_ITER)
# 6) Graphs for the results:
Postbel.ShowDataset(RMSE=True)
CurrentGraph = pyplot.gcf()
CurrentGraph = CurrentGraph.get_axes()[0]
CurrentGraph.plot(TimingBench,DatasetBench[:len(TimingBench)],'k',linestyle='None',marker='o',markerfacecolor='None')
#Postbel.ShowPostCorr(TrueModel=ModelBench,OtherMethod=PRE_MODS)
Postbel.ShowPostModels(TrueModel=ModelBench,RMSE=True)
# Graph for the CCA space parameters loads
_, ax = pyplot.subplots()
B = Postbel.CCA.y_loadings_
B = np.divide(np.abs(B).T,np.repeat(np.reshape(np.sum(np.abs(B),axis=0),(1,B.shape[0])),B.shape[0],axis=0).T)
ind = np.asarray(range(B.shape[0]))+1
ax.bar(x=ind,height=B[0],label=r'${}$'.format(Postbel.MODPARAM.paramNames["NamesSU"][0]))
for i in range(B.shape[0]+1)[1:-1]:
ax.bar(x=ind,height=B[i],bottom=np.reshape(np.sum(B[0:i],axis=0),(B.shape[0],)),label=r'${}$'.format(Postbel.MODPARAM.paramNames["NamesSU"][i]))
box = ax.get_position()
ax.set_position([box.x0, box.y0, box.width, box.height*0.8])
ax.legend(loc='upper center', bbox_to_anchor=(0.5, 1.4), ncol=3)
ax.set_ylabel('Relative contribution')
ax.set_xlabel('CCA dimension')
pyplot.show(block=False)
# Add KDE graph at last iteration:
PrebelGraph = BEL1D.PREBEL.POSTBEL2PREBEL(PREBEL=PrebelLast,POSTBEL=PostbelLast, NoiseModel=Noise,RemoveOutlier=True,Simplified=True,nbMax=nbSampled,MixingRatio=Mixing)
PostbelGraph = BEL1D.POSTBEL(PREBEL=PrebelGraph)
print('Postbel for graphs initialized')
PostbelGraph.run(Dataset=DatasetBench, nbSamples=10000)
print('Printing KDE Graphs')
PostbelGraph.KDE.ShowKDE(Xvals=PostbelGraph.CCA.transform(PostbelGraph.PCA['Data'].transform(np.reshape(DatasetBench,(1,-1)))))
print('Total computation time: {} seconds'.format(np.sum(timings)))
# Comparison with McMC:
OtherMethod = np.load('Data/sNMR/SEG2020_Bench.npy')
fig = pyplot.figure(figsize=[10,10])# Creates the figure space
axs = fig.subplots(nbParam, nbParam)
for i in range(nbParam):
for j in range(nbParam):
if i == j: # Diagonal
if i != nbParam-1:
axs[i,j].get_shared_x_axes().join(axs[i,j],axs[-1,j])# Set the xaxis limit
axs[i,j].hist(np.squeeze(MODELS_ITER[0,:,j]),color='gray',density=True,alpha=0.5)
axs[i,j].hist(np.squeeze(MODELS_ITER[3,:,j]),color='b',density=True,alpha=0.5)
axs[i,j].hist(np.squeeze(MODELS_ITER[6,:,j]),color='m',density=True,alpha=0.5)
axs[i,j].hist(OtherMethod[:,j],color='y',density=True,alpha=0.5)
axs[i,j].hist(np.squeeze(MODELS_ITER[-1,:,j]),color='g',density=True,alpha=0.5)
axs[i,j].plot([ModelBench[i],ModelBench[i]],np.asarray(axs[i,j].get_ylim()),'r')
if nbParam > 8:
axs[i,j].set_xticks([])
axs[i,j].set_yticks([])
elif i > j: # Below the diagonal -> Scatter plot
if i != nbParam-1:
axs[i,j].get_shared_x_axes().join(axs[i,j],axs[-1,j])# Set the xaxis limit
if j != nbParam-1:
if i != nbParam-1:
axs[i,j].get_shared_y_axes().join(axs[i,j],axs[i,-1])# Set the yaxis limit
else:
axs[i,j].get_shared_y_axes().join(axs[i,j],axs[i,-2])# Set the yaxis limit
axs[i,j].plot(np.squeeze(MODELS_ITER[0,:,j]),np.squeeze(MODELS_ITER[0,:,i]),color='gray',linestyle='None',marker='.')
axs[i,j].plot(np.squeeze(MODELS_ITER[3,:,j]),np.squeeze(MODELS_ITER[3,:,i]),'.b')
axs[i,j].plot(np.squeeze(MODELS_ITER[6,:,j]),np.squeeze(MODELS_ITER[6,:,i]),'.m')
axs[i,j].plot(np.squeeze(MODELS_ITER[-1,:,j]),np.squeeze(MODELS_ITER[-1,:,i]),'.g')
axs[i,j].plot(ModelBench[j],ModelBench[i],'.r')
if nbParam > 8:
axs[i,j].set_xticks([])
axs[i,j].set_yticks([])
elif OtherMethod is not None:
if i != nbParam-1:
axs[i,j].get_shared_x_axes().join(axs[i,j],axs[-1,j])# Set the xaxis limit
if j != nbParam-1:
if i != 0:
axs[i,j].get_shared_y_axes().join(axs[i,j],axs[i,-1])# Set the yaxis limit
else:
axs[i,j].get_shared_y_axes().join(axs[i,j],axs[i,-2])# Set the yaxis limit
axs[i,j].plot(np.squeeze(MODELS_ITER[0,:,j]),np.squeeze(MODELS_ITER[0,:,i]),color='gray',linestyle='None',marker='.')
axs[i,j].plot(OtherMethod[:,j],OtherMethod[:,i],'.y')
axs[i,j].plot(ModelBench[j],ModelBench[i],'.r')
if nbParam > 8:
axs[i,j].set_xticks([])
axs[i,j].set_yticks([])
else:
axs[i,j].set_visible(False)
if j == 0: # First column of the graph
if ((i==0)and(j==0)) or not(i==j):
axs[i,j].set_ylabel(r'${}$'.format(Postbel.MODPARAM.paramNames["NamesSU"][i]))
if i == nbParam-1: # Last line of the graph
axs[i,j].set_xlabel(r'${}$'.format(Postbel.MODPARAM.paramNames["NamesSU"][j]))
if j == nbParam-1:
if not(i==j):
axs[i,j].yaxis.set_label_position("right")
axs[i,j].yaxis.tick_right()
axs[i,j].set_ylabel(r'${}$'.format(Postbel.MODPARAM.paramNames["NamesSU"][i]))
if i == 0:
axs[i,j].xaxis.set_label_position("top")
axs[i,j].xaxis.tick_top()
axs[i,j].set_xlabel(r'${}$'.format(Postbel.MODPARAM.paramNames["NamesSU"][j]))
fig.suptitle("Posterior model space visualtization")
from matplotlib.lines import Line2D
custom_lines = [Line2D([0], [0], color='gray', lw=4),
Line2D([0], [0], color='y', lw=4),
Line2D([0], [0], color='b', lw=4),
Line2D([0], [0], color='m', lw=4),
Line2D([0], [0], color='g', lw=4),
Line2D([0], [0], color='r', lw=4)]
fig.legend(custom_lines, ['Prior', 'DREAM', '3rd Iteration', '6th Iteration', 'Last Iteration', 'Benchmark'],loc='lower center',ncol=6)
for ax in axs.flat:
ax.label_outer()
pyplot.show(block=False)
# Graph with the models:
DREAM = OtherMethod # Already loaded
Prior = PRE_MODS
Iter5 = np.squeeze(MODELS_ITER[3,:,:])
Iter10 = np.squeeze(MODELS_ITER[6,:,:])
IterLast = np.squeeze(MODELS_ITER[-1,:,:])
pltidx = [1,4,2,5,8,3,6,9]
fig = pyplot.figure(figsize=[10,10])
for idx in range(8):
ax = pyplot.subplot(3,3,pltidx[idx])
pyplot.hist(Prior[:,idx].ravel(),bins=50,density=True,alpha=0.5,label='Prior')
pyplot.hist(Iter5[:,idx].ravel(),bins=50,density=True,alpha=0.5,label='3rd iteration')
pyplot.hist(Iter10[:,idx].ravel(),bins=50,density=True,alpha=0.5,label='6th iteration')
pyplot.hist(IterLast[:,idx].ravel(),bins=50,density=True,alpha=0.5,label='Last iteration')
pyplot.hist(DREAM[:,idx].ravel(),bins=50,density=True,alpha=0.5,label='DREAM')
ax.plot([ModelBench[idx],ModelBench[idx]],np.asarray(ax.get_ylim()),label='Benchmark')
pyplot.plot()
if pltidx[idx]==1:
ax.set_ylabel('Layer 1')
elif pltidx[idx]==4:
ax.set_ylabel('Layer 2')
elif pltidx[idx]==8:
ax.set_xlabel('Water content [/]')
elif pltidx[idx]==9:
ax.set_xlabel('Relaxation time [sec]')
handles, labels = ax.get_legend_handles_labels()
ax = pyplot.subplot(3,3,7)# Not used but labels needed
ax.set_xlabel('Thickness [m]')
ax.set_ylabel('Half-space')
ax.spines['bottom'].set_color('None')
ax.spines['top'].set_color('None')
ax.spines['right'].set_color('None')
ax.spines['left'].set_color('None')
ax.xaxis.set_ticks([])
ax.yaxis.set_ticks([])
#pyplot.axis('off')
CurrentGraph = pyplot.gcf()
CurrentGraph.legend(handles, labels,loc='lower center', ncol=6)
# Graph for the i-th pusle moment:
import matplotlib
from matplotlib import colors
from scipy import stats
RMS = np.sqrt(np.square(np.subtract(DatasetBench,Postbel.SAMPLESDATA)).mean(axis=-1))
quantiles = np.divide([stats.percentileofscore(RMS,a,'strict') for a in RMS],100)
sortIndex = np.argsort(RMS)
sortIndex = np.flip(sortIndex)
fig = pyplot.figure()
ax = fig.add_subplot(1, 1, 1)
colormap = matplotlib.cm.get_cmap('jet')
for j in sortIndex:
ax.plot(Postbel.MODPARAM.forwardFun["Axis"],np.squeeze(Postbel.SAMPLESDATA[j,-5*len(Postbel.MODPARAM.forwardFun["Axis"]):-4*len(Postbel.MODPARAM.forwardFun["Axis"])]),color=colormap(quantiles[j]))
ax.plot(TimingBench,DatasetBench[-5*len(Postbel.MODPARAM.forwardFun["Axis"]):-4*len(Postbel.MODPARAM.forwardFun["Axis"])],'k',linestyle='None',marker='o',markerfacecolor='None')
ax.set_xlabel(r'${}$'.format(Postbel.MODPARAM.paramNames["DataAxis"]),fontsize=14)
ax.set_ylabel(r'${}$'.format(Postbel.MODPARAM.paramNames["DataName"]),fontsize=14)
fig.subplots_adjust(bottom=0.30)
ax_colorbar = fig.add_axes([0.10, 0.15, 0.80, 0.05])
nb_inter = 1000
color_for_scale = colormap(np.linspace(0,1,nb_inter,endpoint=True))
cmap_scale = colors.ListedColormap(color_for_scale)
scale = [stats.scoreatpercentile(RMS,a,limit=(np.min(RMS),np.max(RMS)),interpolation_method='lower') for a in np.linspace(0,100,nb_inter,endpoint=True)]
norm = colors.BoundaryNorm(scale,len(color_for_scale))
data = np.atleast_2d(np.linspace(np.min(RMS),np.max(RMS),nb_inter,endpoint=True))
ax_colorbar.imshow(data, aspect='auto',cmap=cmap_scale,norm=norm)
ax_colorbar.set_xlabel('Root Mean Square Error {}'.format(Postbel.MODPARAM.paramNames["DataUnits"]),fontsize=12)
ax_colorbar.yaxis.set_visible(False)
nbTicks = 5
ax_colorbar.set_xticks(ticks=np.linspace(0,nb_inter,nbTicks,endpoint=True))
ax_colorbar.set_xticklabels(labels=round_to_5([stats.scoreatpercentile(RMS,a,limit=(np.min(RMS),np.max(RMS)),interpolation_method='lower') for a in np.linspace(0,100,nbTicks,endpoint=True)],n=5),rotation=15,ha='center')
pyplot.show()
################################################################################
### ###
### Case study: Mont Rigi ###
### ###
################################################################################
MtRigi = False
if MtRigi:
# Load the field data
from pygimli.physics import sNMR
Dataset = "Data/sNMR/SEG2020_MtRigi.mrsd"
Kernel = "Data/sNMR/SEG2020_MtRigi.mrsk"
ModelParam = sNMR.MRS()
sNMR.MRS.loadKernel(ModelParam,Kernel)
sNMR.MRS.loadMRSI(ModelParam,Dataset)
FieldData = np.ravel(ModelParam.dcube)
TimingField = ModelParam.t
Noise = 18 #nV
# Initialize BEL1D:
nbSampled = 5000
priorMtRigi = np.asarray([[0.0, 7.5, 0.30, 0.80, 0.0, 0.200], [0, 0, 0.0, 0.15, 0.100, 0.400]])
start = time.time()
MODEL_MtRigi = BEL1D.MODELSET().SNMR(prior=priorMtRigi,Kernel=Kernel, Timing=TimingField)
PREBEL_MtRigi = BEL1D.PREBEL(MODEL_MtRigi,nbModels=nbSampled)
PREBEL_MtRigi.run()
POSTBEL_MtRigi = BEL1D.POSTBEL(PREBEL_MtRigi)
POSTBEL_MtRigi.run(FieldData,nbSamples=nbSampled,NoiseModel=Noise)
end = time.time()
POSTBEL_MtRigi.ShowDataset(RMSE=True,Prior=True)
CurrentGraph = pyplot.gcf()
CurrentGraph = CurrentGraph.get_axes()[0]
CurrentGraph.plot(TimingField,FieldData[:len(TimingField)],'k',linestyle='None',marker='o',markerfacecolor='None')
pyplot.show(block=False)
POSTBEL_MtRigi.ShowPostModels(RMSE=True)
# Iterations:
Postbel = POSTBEL_MtRigi
Prebel = PREBEL_MtRigi
nbIter = 100 # maximum number of iterations
tolerance = 5e-3 # Tolerance on the normalized difference between the distributions
nbParam = int(priorMtRigi.size/2 - 1)
means = np.zeros((nbIter,nbParam))
stds = np.zeros((nbIter,nbParam))
timings = np.zeros((nbIter,))
MODELS_ITER = np.zeros((nbIter,nbSampled,nbParam)) # Store the models that have been computed
diverge = True
distancePrevious = 1e10
MixingUpper = 0
MixingLower = 1
for idxIter in range(nbIter):
PostbelLast = Postbel
PrebelLast = Prebel
if idxIter == 0: # Initialization: already done (see 2 and 3)
# PostbelTest.KDE.ShowKDE(Xvals=PostbelTest.CCA.transform(PostbelTest.PCA['Data'].transform(np.reshape(Dataset,(1,-1)))))
means[idxIter,:], stds[idxIter,:] = Postbel.GetStats()
timings[idxIter] = end-start
ModLastIter = Prebel.MODELS
else:
ModLastIter = Postbel.SAMPLES
# Here, we will use the POSTBEL2PREBEL function that adds the POSTBEL from previous iteration to the prior (Iterative prior resampling)
# However, the computations are longer with a lot of models, thus you can opt-in for the "simplified" option which randomely select up to 10 times the numbers of models
MixingUpper += 1
MixingLower += 1
Mixing = MixingUpper/MixingLower
Prebel = BEL1D.PREBEL.POSTBEL2PREBEL(PREBEL=Prebel,POSTBEL=Postbel,Dataset=FieldData,NoiseModel=Noise,Simplified=True,nbMax=nbSampled,MixingRatio=Mixing)
# Since when iterating, the dataset is known, we are not computing the full relationship but only the posterior distributions directly to gain computation timing
print(idxIter+1)
Postbel = BEL1D.POSTBEL(Prebel)
Postbel.run(FieldData,nbSamples=nbSampled,NoiseModel=Noise)
means[idxIter,:], stds[idxIter,:] = Postbel.GetStats()
end = time.time()
timings[idxIter] = end-start
# The distance is computed on the normalized distributions. Therefore, the tolerance is relative.
diverge, distance = Tools.ConvergeTest(SamplesA=ModLastIter,SamplesB=Postbel.SAMPLES, tol=tolerance)
print('Wasserstein distance: {}'.format(distance))
if not(diverge) or (abs((distancePrevious-distance)/distancePrevious)*100<1):# If the distance between the distributions is not changing, we converged as well
# Convergence acheived if:
# 1) Distance below threshold
# 2) Distance does not vary significantly (less than 2.5%)
print('Model has converged at iter {}!'.format(idxIter+1))
MODELS_ITER[idxIter,:,:] = Postbel.SAMPLES
break
distancePrevious = distance
MODELS_ITER[idxIter,:,:] = Postbel.SAMPLES
start = time.time()
timings = timings[:idxIter+1]
means = means[:idxIter+1,:]
stds = stds[:idxIter+1,:]
MODELS_ITER = MODELS_ITER[:idxIter+1,:,:]
np.save('ModelsIteration',MODELS_ITER)
Postbel.ShowPostModels(RMSE=True)
Postbel.ShowDataset(RMSE=True,Prior=True)
CurrentGraph = pyplot.gcf()
CurrentGraph = CurrentGraph.get_axes()[0]
CurrentGraph.plot(TimingField,FieldData[:len(TimingField)],'k',linestyle='None',marker='o',markerfacecolor='None')
pyplot.show(block=False)
# i = int(np.ceil(idxIter/2))
# Postbel.ShowPostCorr(OtherMethod=np.squeeze(MODELS_ITER[i,:,:]))
print('CPU time: {} seconds'.format(np.sum(timings)))
pyplot.show()