def BoxPlot(x, path='none', title='none', label='none'):
"""
x: numpy [perticles,id]
"""
sns.set_style("dark")
# numpy -> list
x_tuple = tuple(x[:, i].tolist() for i in np.arange(x.shape[1]))
idx = []
for i in np.arange(x.shape[-1]): # data [peticles]
idx = np.append(idx, str(i + 1))
fig, ax = plt.subplots()
ax.boxplot(x_tuple,
sym='d',
patch_artist=True,
boxprops=dict(facecolor='lightblue', color='gray'),
medianprops=dict(color='gray'))
ax.set_xticklabels(idx)
myData.isDirectory(path)
plt.savefig(os.path.join(path, f"{label}.png"))
plt.close()
def scatter3D_heatmap(x,
y,
z,
var,
rangeP,
path="none",
title="none",
label="none"):
"""
Args
val: variable for heatmap
"""
#pdb.set_trace()
sns.set_style("dark")
#var = np.hstack([var[0],var[1]])
# normalize variable(var) into 0 to 1
nlVar = list((var - min(var)) / (max(var) - min(var)))
# red(row,<<0) -> yellow(high)
colors = plt.cm.autumn(nlVar)
fig = plt.figure()
ax = Axes3D(fig)
#pdb.set_trace()
ax.scatter(x, y, z, c=colors, marker="o", alpha=0.5)
#ax.scatter(x[0],y[0],z[0],c=colors[:x[0].shape[0]],marker=".",alpha=0.5)
#ax.scatter(x[1],y[1],z[1],s=10,c=colors[x[0].shape[0]:],marker="x",alpha=0.5)
"""
xmin = np.min([np.min(x[0]),np.min(x[1])])
xmax = np.max([np.max(x[0]),np.max(x[1])])
ymin = np.min([np.min(y[0]),np.min(y[1])])
ymax = np.max([np.max(y[0]),np.max(y[1])])
zmin = np.min([np.min(z[0]),np.min(z[1])])
zmax = np.max([np.max(z[0]),np.max(z[1])])
ax.set_xlim(xmin,xmax)
ax.set_ylim(ymin,ymax)
ax.set_zlim(zmin,zmax)
"""
ax.set_xlim(rangeP[0][ntI], rangeP[1][ntI])
ax.set_ylim(rangeP[0][tntI], rangeP[1][tntI])
ax.set_zlim(rangeP[0][ttI], rangeP[1][ttI])
ax.set_title(f"{title}")
myData.isDirectory(path)
plt.savefig(os.path.join(path, f"{label}.png"))
plt.close()
def HistLikelihood(weights, path, label="auto", color="black"):
#pdb.set_trace()
# mean & var for label
lhMean = np.mean(weights, 0)
lhVar = np.var(weights, 0)
sns.set_style("dark")
sns.distplot(weights, kde=False, rug=False, color=color)
#plt.xlim([0,0.12])
#plt.ylim([0,175])
plt.suptitle(f"mean:{lhMean}\n var:{lhVar}")
myData.isDirectory(path)
plt.savefig(os.path.join(path, f"{label}.png"))
plt.close()
def scatter3D(x, y, z, rangeP, path="none", title="none", label="none"):
sns.set_style("dark")
fig = plt.figure()
ax = Axes3D(fig)
ax.scatter(x, y, z, c="black", marker="o", alpha=0.5, linewidths=0.5)
ax.set_xlabel("nk")
ax.set_ylabel("tnk")
ax.set_zlabel("tk")
ax.set_xlim(rangeP[0][ntI], rangeP[1][ntI])
ax.set_ylim(rangeP[0][tntI], rangeP[1][tntI])
ax.set_zlim(rangeP[0][ttI], rangeP[1][ttI])
ax.set_title(f"{title}")
myData.isDirectory(path)
plt.savefig(os.path.join(path, f"{label}.png"))
plt.close()
def NumberLine(gt, pred, path, label="auto"):
"""
発生年数がどうなってるかを確認したくって
Args
gt: 真値t list[nk,tnk,tk]
pred: 予測値t [perticles,cells]
"""
for cell in np.arange(3):
#pdb.set_trace()
# predict year [perticles,]
x = gt[cell]
xhat = pred[:, cell]
y = [0] * 1 # y = 0
yhat = [0] * xhat.shape[0]
# 数直線 ---------------------------------------------------------------
fig, ax = plt2.subplots(figsize=(10, 10)) #画像サイズ
fig.set_figheight(1) #高さ調整
ax.tick_params(labelbottom=True, bottom=False) #x軸設定
ax.tick_params(labelleft=False, left=False) #y軸設定
# ---------------------------------------------------------------------
# グラフの体裁 -----------------------------------------------------------
#xMin, xMax = np.min(np.append(x,xhat)), np.max(np.append(x,xhat))
xMin, xMax = 0, 1400
plt2.tight_layout() #グラフの自動調整
plt2.hlines(y=0, xmin=xMin, xmax=xMax, color="silver") #横軸
pylab.box(False) #枠を消す
# -----------------------------------------------------------------
# 散布図 -----------------------------------------------------------
plt2.scatter(xhat, yhat, c='skyblue') # 予測値
plt2.scatter(x, y[0], c='coral') # 真値
plt2.title(f"min:{int(np.min(xhat))} max:{int(np.max(xhat))}")
# -----------------------------------------------------------------
myData.isDirectory(path)
plt2.savefig(os.path.join(path, f"{label}_{cellname[cell]}.png"),
bbox_inches="tight")
plt2.close()
#------------------------------------------------------------------------------
if __name__ == "__main__":
# 1 確実に起きた地震
# 190 全て起きた地震
for tfID in [190]:
print("-----------------------------------")
print("------ {} historical eq data ------".format(tfID))
# dirpath for each logs
dirPath = f"{tfID}"
# full path for each logs
filePath = os.path.join(logsPath, dirPath, fileName)
# path exist or not exist
myData.isDirectory(os.path.join(logsPath, dirPath))
myData.isDirectory(os.path.join('parFile', dirPath))
# ----------------- 真の南海トラフ巨大地震履歴 V------------------------- #
with open(os.path.join(featuresPath, "nankairireki.pkl"), "rb") as fp:
nkfiles = pickle.load(fp)
# 発生年数取得 & slip velocity (all 30)
gtV = nkfiles[tfID, :, :]
# 重複なしの真値地震
#gtJ = np.unique(np.where(gtV>0)[0])
gtJ_nk = np.where(
gtV[:, ntI] > 0)[0] # [84,287,499,761,898,1005,1107,1254,1346]
gtJ_tnk = np.where(
gtV[:, tntI] > 0)[0] # [84,287,496,761,898,1005,1107,1254,1344]
# update paramb
updateBs = np.array([sb for sb in standB])
# plot
meanB, medianB = np.mean(updateBs, 0), np.median(updateBs, 0)
# Num.of perticle for label
numBs = updateBs.shape[0]
# index max perticle
maxBsind = [i for i, x in enumerate(ratebs) if x == max(ratebs)]
# max perticle
maxBs = np.array([np.array(updateBs[ind]) for ind in maxBsind])
# rate of max perticle
maxrate = int(np.max(ratebs[maxBsind]))
# save paramters (high rate of perticle)
maxbpath = os.path.join(savetxtPath, f'maxB_{mode}')
myData.isDirectory(maxbpath)
np.savetxt(os.path.join(maxbpath, f'maxB_{iS+1}_{maxrate}.txt'),
maxBs * 1000000,
fmt='%d',
delimiter=',')
# plot 3D heatmap scatter
s3hpath = os.path.join(imgPath, f'PF_{mode}')
myPlot.scatter3D_heatmap(updateBs[:, ntI],
updateBs[:, tntI],
updateBs[:, ttI],
ratebs,
rangeP=[minB, maxB],
path=s3hpath,
title=f'mean:{meanB}\n median:{medianB}',
label=f'Bheatmap_{iS+1}_{numBs}')
# -------------------------------------------------------------------------
def simulate(features,y,x,mode=0,t=0,pTime=0,nP=0,nCell=3,isSavetxt=False,isPlot=False):
"""
Args
features: システムモデル値xt. th[1400,perticles,3], V[1400,perticles,3], b[perticles,3]
y: 観測モデル値yt. [eq.years,]
x: 地震年数(1400年). [(eq.years zero padding),perticles]
"""
#pdb.set_trace()
# 1. 初期化 ----------------------------------------------------------------
# 状態ベクトル theta,v,year ※1セルの時おかしいかも
# リサンプリング後の特徴量ベクトル
xResampled = np.zeros((nP,nCell))
# all norm-likelihood
maxgW = np.zeros((nP))
maxpW = np.zeros((nP))
# weight for eq. year in each cell + penalty
gw = np.zeros((nP,nCell+1))
# weight for eq. times
pw = np.zeros((nP,nCell+1))
# -------------------------------------------------------------------------
#pdb.set_trace()
# -------------------------------------------------------------------------
flag = False
for i in np.arange(nP): # アンサンブル分まわす
# =====================================================================
# 尤度計算
# =====================================================================
# zero-paddingしたpredを予測した年数だけにする [地震発生年数(可変),]
# predicted eq.year
yhat_nk = (x[x[:,i,ntI]>0,i,ntI]).astype(int)
yhat_tnk = (x[x[:,i,tntI]>0,i,tntI]).astype(int)
yhat_tk = (x[x[:,i,ttI]>0,i,ttI]).astype(int)
yhat = [yhat_nk,yhat_tnk,yhat_tk]
#pdb.set_trace()
# 尤度は地震発生年数、重みとかけるのは状態ベクトル
# 2.c & 2.d 各粒子の尤度と重み -------------------------------------------
# ground truth eq.year (time=t)
standYs = [y[ntI][t],y[tntI][t],y[ttI][t]]
# nearist -----
if mode == 100:
weight, maxweight, years = norm_likelihood.norm_likelihood_nearest(y,yhat,standYs=standYs,time=t)
gw[i] = weight
maxgW[i] = maxweight
if mode == 13:
weight, maxweight, years = norm_likelihood.norm_likelihood_nearest_penalty(y,yhat,standYs=standYs,time=t)
gw[i] = weight
maxgW[i] = maxweight
if mode == 101:
weight, maxweight, years = norm_likelihood.norm_likelihood_nearest_safetypenalty(y,yhat,standYs=standYs,time=t)
gw[i] = weight
maxgW[i] = maxweight
if mode == 102:
gweight, gmaxweight, years = norm_likelihood.norm_likelihood_nearest_safetypenalty(y,yhat,standYs=standYs,time=t)
pweight = norm_likelihood.norm_likelihood_alltimes(y,yhat)
gw[i] = gweight
pw[i] = pweight
maxgW[i] = gmaxweight
maxpW[i] = pweight
# ---------------------------------------------------------------------
# for plot ------------------------------------------------------------
if not flag:
yearInds = years
flag = True
else:
# [perticle,3]
yearInds = np.vstack([yearInds,years])
# ---------------------------------------------------------------------
# 規格化 -------------------------------------------------------------------
# only eq.years
if mode == 100 or mode == 101:
# 全セルがぴったりの時
if any(maxgW==0):
zeroind = np.where(maxgW==0)[0].tolist()
maxgW[zeroind] = -1
tmpgW = 1/-maxgW
wNorm = tmpgW/np.sum(tmpgW)
# eq.years & eq.times
elif mode == 102:
if any(maxgW==0):
zeroind = np.where(maxgW==0)[0].tolist()
maxgW[zeroind] = -1
tmpgW = 1/-maxgW
maxW = tmpgW + maxpW
wNorm = maxW/np.sum(maxW)
print(wNorm)
# -------------------------------------------------------------------------
#pdb.set_trace()
# =========================================================================
# リサンプリング
# =========================================================================
initU = np.random.uniform(0,1/nP)
# ※3セル同じ組み合わせのbが選ばれる
# index for resampling
k = resampling(initU,wNorm,nP=nP)
# not update b var. ----
if mode == 90:
xResampled = features[k]
# simple var. ----
if mode == 100:
# system noise --------------------------------------------------------
# ※元の値と足してもマイナスになるかも
# array[cell,perticles] V & theta parameterがすべて同じ組み合わせになるのを防ぐため
bnoise = np.array([np.random.normal(0,0.01*np.mean(features[:,cell]),nP) for cell in np.arange(nCell)])
# ---------------------------------------------------------------------
# Add noise
xResampled = features[k] + np.abs(bnoise).T
# 尤度工夫 var. ----
if mode == 101 or mode == 102:
#pdb.set_trace()
# index for mean theta,V,b
muind = np.argmax(wNorm)
# highest of norm likelihood (index) for mean & sigma
muB = features[bInd][muind] * 1000000
# variable & co-variable matrix (xy,yz,zx)
Bin = 10
sigmaB = [[0,Bin,Bin],[Bin,0,Bin],[Bin,Bin,0]]
# 尤度の1番高いところを中心に、次のbの分布決定
# system noise --------------------------------------------------------
# [perticle,cell]
bnoise = np.random.multivariate_normal(muB,sigmaB,nP)
# ---------------------------------------------------------------------
# [perticle,cell]
xResampled = np.abs(bnoise)*0.000001
xResampled[0] = features[muind]
#pdb.set_trace()
print(f"---- 【{t}】 times ----\n")
# 発生年数 plot ------------------------------------------------------------
if isPlot:
nlpath = os.path.join(imgPath,f'numlines_{mode}')
myData.isDirectory(nlpath)
myPlot.NumberLine(standYs,yearInds,path=nlpath,label=f"best_years_{t}")
# -------------------------------------------------------------------------
# save year & likelihood txt ----------------------------------------------
if isSavetxt:
# nearist ----
lhpath = os.path.join(savetxtPath,f"lh_{mode}")
myData.isDirectory(lhpath)
if mode == 100 or 101:
np.savetxt(os.path.join(lhpath,f"lh_{t}.txt"),gw)
np.savetxt(os.path.join(lhpath,f"sum_lh_{t}.txt"),maxgW)
else:
np.savetxt(os.path.join(lhpath,f"lh_g_{t}.txt"),gw)
np.savetxt(os.path.join(lhpath,f"lh_p_{t}.txt"),pw)
np.savetxt(os.path.join(lhpath,f"sum_lh_g_{t}.txt"),maxgW)
np.savetxt(os.path.join(lhpath,f"sum_lh_p_{t}.txt"),maxpW)
np.savetxt(os.path.join(lhpath,f"w_{t}.txt"),wNorm)
# Save param b
xt = features[k]
bpath = os.path.join(savetxtPath,f'B_{mode}')
myData.isDirectory(bpath)
np.savetxt(os.path.join(bpath,f'{t}.txt'),xt,fmt='%6f')
# -------------------------------------------------------------------------
return xResampled, k
def Rireki(gt,
pred,
path='none',
title="none",
label="none",
isShare=False,
isSeparate=True,
isResearch=False,
iseach=False):
"""
Args
gt: gt eq. (best year). [1400,3]
"""
if isResearch:
if iseach:
dists = myData.eachMAEyear(gt, pred)
else:
# degree of similatery
dists = myData.MAEyear(gt, pred)
title = dists
sns.set_style("dark")
# share gt & pred
if isShare:
fig, figInds = plt.subplots(nrows=3, sharex=True)
for figInd in np.arange(len(figInds)):
figInds[figInd].plot(np.arange(pred.shape[0]),
pred[:, figInd],
color="skyblue")
figInds[figInd].plot(np.arange(gt.shape[0]),
gt[:, figInd],
color="coral")
#pdb.set_trace()
if isSeparate:
colors = ["coral", "skyblue", "coral", "skyblue", "coral", "skyblue"]
# scalling var.
predV, gtV = np.zeros([1400, 3]), np.zeros([1400, 3])
# del first year
pred_nk = [s for s in pred[ntI].tolist() if s != 0]
pred_tnk = [s for s in pred[tntI].tolist() if s != 0]
pred_tk = [s for s in pred[ttI].tolist() if s != 0]
gt_tk = [s for s in gt[ttI].tolist() if s != 0]
predV[pred_nk, ntI] = 5
predV[pred_tnk, tntI] = 5
predV[pred_tk, ttI] = 5
gtV[gt[ntI].tolist(), ntI] = 5
gtV[gt[tntI].tolist(), tntI] = 5
gtV[gt_tk, ttI] = 5
#pdb.set_trace()
# [1400,3]
plot_data = [
gtV[:, ntI], predV[:, ntI], gtV[:, tntI], predV[:, tntI],
gtV[:, ttI], predV[:, ttI]
]
# not scalling var. [1400,3]
#plot_data = [gt[:,ntI],pred[:,ntI],gt[:,tntI],pred[:,tntI],gt[:,ttI],pred[:,ttI]]
fig = plt.figure()
fig, axes = plt.subplots(nrows=6, sharex="col")
for row, (color, data) in enumerate(zip(colors, plot_data)):
axes[row].plot(np.arange(1400), data, color=color)
plt.suptitle(f"{title}", fontsize=8)
myData.isDirectory(path)
plt.savefig(os.path.join(path, f"{label}.png"))
plt.close()
if isResearch:
return int(np.sum(dists))
# -----------------------------------------------------------------------------
def simulate(features,
y,
x,
ssYears,
mode=0,
t=0,
pTime=0,
sy=0,
nP=0,
nCell=3,
isSavetxt=False,
isPlot=False):
"""
Args
features: システムモデル値xt. th[1400,perticles,3], V[1400,perticles,3], b[perticles,3]
y: 観測モデル値yt. [eq.years,]
x: 地震年数(1400年). [(eq.years zero padding),perticles]
sy: start of assimilation for perticles.
"""
#pdb.set_trace()
# 1. 初期化 ----------------------------------------------------------------
# 状態ベクトル theta,v,year ※1セルの時おかしいかも
ThVec = np.zeros((nP, nCell))
VVec = np.zeros((nP, nCell))
# リサンプリング後の特徴量ベクトル
xResampled = np.zeros((nParam, nP, nCell))
# all norm-likelihood
maxgW = np.zeros((nP))
maxpW = np.zeros((nP))
# weight for eq. year in each cell + penalty
gw = np.zeros((nP, nCell + 1))
# weight for eq. times
pw = np.zeros((nP, nCell + 1))
# weight
wNorm = np.zeros((nP))
# -------------------------------------------------------------------------
#pdb.set_trace()
# -------------------------------------------------------------------------
flag = False
for i in np.arange(nP): # アンサンブル分まわす
# =====================================================================
# 尤度計算
# =====================================================================
# zero-paddingしたpredを予測した年数だけにする [地震発生年数(可変),]
yhat_nk = (x[x[:, i, ntI] > 0, i, ntI]).astype(int)
yhat_tnk = (x[x[:, i, tntI] > 0, i, tntI]).astype(int)
yhat_tk = (x[x[:, i, ttI] > 0, i, ttI]).astype(int)
if t > 0:
# 2000年 + 同化した年数
standInds = ssYears[i] + state_Year
# 1400年のスケールに合わせる
yhat_nk = yhat_nk - standInds
yhat_tnk = yhat_tnk - standInds
yhat_tk = yhat_tk - standInds
yhat = [yhat_nk, yhat_tnk, yhat_tk]
#pdb.set_trace()
# 尤度は地震発生年数、重みとかけるのは状態ベクトル
# 2.c & 2.d 各粒子の尤度と重み -------------------------------------------
standYs = [y[ntI][t], y[tntI][t], y[ttI][t]]
# nearist -----
# if mode == 'near' or mode == 'simple' or mode == 'b_near'
if mode == 0:
weight, maxweight, years = norm_likelihood.norm_likelihood_nearest(
y, yhat, standYs=standYs, time=t)
gw[i] = weight
maxgW[i] = maxweight
# if mode == 'sp_time_near' or mode == 'b_sp_time_near'
elif mode == 3 or mode == 2:
gweight, gmaxweight, years = norm_likelihood.norm_likelihood_nearest_safetypenalty(
y, yhat, standYs=standYs, time=t)
pweight = norm_likelihood.norm_likelihood_alltimes(y, yhat)
gw[i] = gweight
pw[i] = pweight
maxgW[i] = gmaxweight
maxpW[i] = pweight
elif mode == 4:
weight, maxweight, years = norm_likelihood.norm_likelihood_eachnearest(
y, yhat, standYs=standYs, time=t)
gw[i] = weight
maxgW[i] = maxweight
elif mode == 5 or mode == 6:
weight, maxweight, years = norm_likelihood.norm_likelihood_eachnearest_penalty(
y, yhat, standYs=standYs, time=t)
gw[i] = weight
maxgW[i] = maxweight
elif mode == 7:
pweight = norm_likelihood.norm_likelihood_alltimes(y, yhat)
gweight, gmaxweight, years = norm_likelihood.norm_likelihood_eachnearest_penalty(
y, yhat, standYs=standYs, time=t)
gw[i] = gweight
pw[i] = pweight
maxgW[i] = gmaxweight
maxpW[i] = pweight
# ---------------------------------------------------------------------
#pdb.set_trace()
for indY, indC in zip(years, [ntI, tntI, ttI]):
# 各セルで尤度の一番高かった年数に合わせる 1400 -> 1
# ※ 別々の同化タイミングになる
# ※地震が発生していないときは、tonankaiの地震発生年数を採用
# ※違う年数でも同じ値の時あり
if int(indY) == 0: # for tk
ThVec[i, indC] = features[0][int(years[tntI]), i, indC]
VVec[i, indC] = features[1][int(years[tntI]), i, indC]
else:
ThVec[i, indC] = features[0][int(years[indC]), i, indC]
VVec[i, indC] = features[1][int(years[indC]), i, indC]
if not flag:
yearInds = years
flag = True
else:
# [perticle,3]
yearInds = np.vstack([yearInds, years])
#pdb.set_trace()
# 規格化 -------------------------------------------------------------------
# only eq.years
if mode == 0 or mode == 4 or mode == 5 or mode == 6:
# 全セルがぴったりの時
if any(maxgW == 0):
zeroind = np.where(maxgW == 0)[0].tolist()
maxgW[zeroind] = 1
tmpgW = 1 / maxgW
wNorm = tmpgW / np.sum(tmpgW)
# eq.years & eq.times
elif mode == 3 or mode == 2 or mode == 7:
if any(maxgW == 0):
zeroind = np.where(maxgW == 0)[0].tolist()
maxgW[zeroind] = 1
tmpgW = 1 / maxgW
maxW = tmpgW + maxpW
wNorm = maxW / np.sum(maxW)
# -------------------------------------------------------------------------
#pdb.set_trace()
# =========================================================================
# リサンプリング
# =========================================================================
initU = np.random.uniform(0, 1 / nP)
# ※3セル同じ組み合わせのbが選ばれる
# index for resampling
k = resampling(initU, wNorm, nP=nP)
# simple var. ----
if mode == 0:
# system noise --------------------------------------------------------
# ※元の値と足してもマイナスになるかも
# array[cell,perticles] V & theta parameterがすべて同じ組み合わせになるのを防ぐため
Thnoise = np.array([
np.random.normal(0, 0.01 * np.mean(ThVec[:, cell]), nP)
for cell in np.arange(nCell)
])
Vnoise = np.array([
np.random.normal(0, 0.01 * np.mean(VVec[:, cell]), nP)
for cell in np.arange(nCell)
])
bnoise = np.array([
np.random.normal(0, 0.01 * np.mean(features[bInd][:, cell]), nP)
for cell in np.arange(nCell)
])
# ---------------------------------------------------------------------
xResampled[thInd] = ThVec[k] + np.abs(Thnoise).T
xResampled[vInd] = VVec[k] + np.abs(Vnoise).T
# Add noise
xResampled[bInd] = features[bInd][k] + np.abs(bnoise).T
updatesy = sy[k]
# 尤度工夫 var. ----
# if mode == 'sp_alltime' or mode == 'b_sp_nl' or mode == 'b_sp_time_nl':
elif mode == 3 or mode == 4 or mode == 5 or mode == 7:
# index for mean theta,V,b
muind = np.argmax(wNorm)
# highest of norm likelihood (index) for mean & sigma
muB = features[bInd][muind] * 1000000
# variable & co-variable matrix (xy,yz,zx)
Bin = 10
sigmaB = [[0, Bin, Bin], [Bin, 0, Bin], [Bin, Bin, 0]]
# 尤度の1番高いところを中心に、次のbの分布決定
# system noise --------------------------------------------------------
# [cell,perticle]
Thnoise = np.array([
np.random.normal(0, 0.01 * np.mean(ThVec[:, cell]), nP)
for cell in np.arange(nCell)
])
Vnoise = np.array([
np.random.normal(0, 0.01 * np.mean(VVec[:, cell]), nP)
for cell in np.arange(nCell)
])
# [perticle,cell]
bnoise = np.random.multivariate_normal(muB, sigmaB, nP)
# ---------------------------------------------------------------------
# [perticle,cell]
xResampled[thInd] = ThVec[k] + np.abs(Thnoise).T
xResampled[vInd] = VVec[k] + np.abs(Vnoise).T
xResampled[bInd] = np.abs(bnoise) * 0.000001
xResampled[bInd][0] = features[bInd][muind]
# 尤度が一番高いperticleの年数に合わせる
updatesy = np.array(sy[muind].tolist() * nP)[:, np.newaxis]
# not update b var. ----
elif mode == 2 or mode == 6:
# system noise --------------------------------------------------------
# ※元の値と足してもマイナスになるかも
# array[cell,perticles] V & theta parameterがすべて同じ組み合わせになるのを防ぐため
Thnoise = np.array([
np.random.normal(0, 0.01 * np.mean(ThVec[:, cell]), nP)
for cell in np.arange(nCell)
])
Vnoise = np.array([
np.random.normal(0, 0.01 * np.mean(VVec[:, cell]), nP)
for cell in np.arange(nCell)
])
# ---------------------------------------------------------------------
xResampled[thInd] = ThVec[k] + np.abs(Thnoise).T
xResampled[vInd] = VVec[k] + np.abs(Vnoise).T
xResampled[bInd] = features[bInd][k]
updatesy = sy[k]
#pdb.set_trace()
print(f"---- 【{t}】 times ----\n")
# 発生年数 plot ------------------------------------------------------------
if isPlot:
nlpath = os.path.join(imgPath, f'numlines_{mode}')
myData.isDirectory(nlpath)
myPlot.NumberLine(standYs,
yearInds,
path=nlpath,
label=f"best_years_{t}")
# -------------------------------------------------------------------------
# save year & likelihood txt ----------------------------------------------
if isSavetxt:
# nearist ----
lhpath = os.path.join(savetxtPath, f"lh_{mode}")
myData.isDirectory(lhpath)
if mode == 0 or mode == 4 or mode == 5 or mode == 6:
np.savetxt(os.path.join(lhpath, f"lh_{t}.txt"), gw)
np.savetxt(os.path.join(lhpath, f"sum_lh_{t}.txt"), maxgW)
else:
np.savetxt(os.path.join(lhpath, f"lh_g_{t}.txt"), gw)
np.savetxt(os.path.join(lhpath, f"lh_p_{t}.txt"), pw)
np.savetxt(os.path.join(lhpath, f"sum_lh_g_{t}.txt"), maxgW)
np.savetxt(os.path.join(lhpath, f"sum_lh_p_{t}.txt"), maxpW)
np.savetxt(os.path.join(lhpath, f"w_{t}.txt"), wNorm)
# Save param b
xt = features[bInd][k]
bpath = os.path.join(savetxtPath, f'B_{mode}')
myData.isDirectory(bpath)
np.savetxt(os.path.join(bpath, f'{t}.txt'), xt, fmt='%6f')
# -------------------------------------------------------------------------
return xResampled, yearInds.astype(int), updatesy, k