import numpy as np
from KS import KS
import matplotlib.pyplot as plt
import matplotlib.animation as animation
L = 16 # domain is 0 to 2.*np.pi*L
N = 128 # number of collocation points
dt = 0.5 # time step
diffusion = 1.0
ks = KS(L=L,diffusion=diffusion,N=N,dt=dt) # instantiate model
# define initial condition
#u = np.cos(x/L)*(1.0+np.sin(x/L)) # smooth IC
u = 0.01*np.random.normal(size=N) # noisy IC
# remove zonal mean
u = u - u.mean()
# spectral space variable.
ks.xspec[0] = np.fft.rfft(u)
# time stepping loop.
nmin = 1000; nmax = 5000
uu = []; tt = []
vspec = np.zeros(ks.xspec.shape[1], np.float)
x = np.arange(N)
fig, ax = plt.subplots()
line, = ax.plot(x, ks.x.squeeze())
ax.set_xlim(0,N-1)
ax.set_ylim(-3,3)
#Init only required for blitting to give a clean slate.
def init():
global line
# for gaussian, smooth_len is standard deviation.
thresh = 0.99 # threshold for modulated ensemble eigenvalue truncation.
# model parameters...
# for truth run
dt = 0.5; npts = 128
diffusion_truth = 1.0
# for forecast model (same as above for perfect model expt)
# for simplicity, assume dt and npts stay the same.
#diffusion = 0.9
diffusion = diffusion_truth
rstruth = np.random.RandomState(42) # fixed seed for truth run
rsens = np.random.RandomState() # varying seed for ob noise and ensemble initial conditions
# model instance for truth (nature) run
model = KS(N=npts,dt=dt,diffusion=diffusion_truth,rs=rstruth)
# mode instance for forecast ensemble
ensemble = KS(N=npts,members=nens,dt=dt,diffusion=diffusion,rs=rsens)
for nt in range(ntstart): # spinup truth run
model.advance()
# sample obs from truth, compute climo stats for model.
xx = []; tt = []
for nt in range(ntimes):
model.advance()
xx.append(model.x[0]) # single member
tt.append(float(nt)*model.dt)
xtruth = np.array(xx,np.float)
timetruth = np.array(tt,np.float)
xtruth_mean = xtruth.mean()
xprime = xtruth - xtruth_mean
def XGBoost_part(dtrain=None,
test=None,
dtest_X=None,
test_y=None,
k=0,
gamma=0.02,
min_child_weight=1.1,
max_depth=5,
lamda=100,
subsamp=0.7,
col_bytree=0.7,
col_bylevel=0.7,
eta=0.01):
param = {
'booster': 'gbtree',
'objective': 'binary:logistic',
'eval_metric': 'auc',
'gamma': gamma,
'min_child_weight': min_child_weight,
'max_depth': max_depth,
'lambda': lamda,
'subsample': subsamp,
'colsample_bytree': col_bytree,
'colsample_bylevel': col_bylevel,
'eta': eta,
'tree_method': 'exact',
'seed': 0,
'nthread': 12
}
cv_log = xgb.cv(param,
dtrain,
num_boost_round=3500,
nfold=5,
early_stopping_rounds=50,
seed=0)
num_round = cv_log.shape[0]
cf = './featurescore/cvg{}.csv'.format(str(num_round))
cv_log.to_csv(cf)
watchlist = [(dtrain, 'train')]
#auc = cv_log['test-auc-mean'].max()
bst = xgb.train(param,
dtrain,
num_round,
evals=watchlist,
early_stopping_rounds=50)
# make prediction
dtest = xgb.DMatrix(test, missing=-9999)
preds = bst.predict(dtest)
scores = bst.predict(dtrain, ntree_limit=bst.best_ntree_limit)
fp, tp, thresholds = metrics.roc_curve(test_y, scores, pos_label=1)
ks = KS(y=test_y, score=scores)
kk = int(ks * 10000000000) % 1000
print "K-S:{}".format(ks)
print "AUC:{}".format(metrics.auc(fp, tp))
with open('./featurescore/a.txt', 'a') as f:
S = "gamma= "+str(gamma)+\
" min_child_weight= "+str(min_child_weight)+\
" max_depth= "+str(max_depth)+\
" lamda= "+str(lamda)+\
"\n" + \
"subsamp= "+str(subsamp)+\
" col_bytree= "+str(col_bytree)+\
" col_bylevel= "+str(col_bylevel)+\
" eta= "+str(eta) + \
" ntree= "+str(bst.best_ntree_limit)+ \
"\nfeatures scores: " + str(kk)
f.writelines("{}\n".format(S))
f.writelines("K-S:{}\n".format(ks))
f.writelines("AUC:{}\n\n".format(metrics.auc(fp, tp)))
#f.writelines("AUC:{}\n\n".format(metrics.auc(fp, tp)))
# 写入文件
writeDatas(preds, test, "xgk{}".format(str(ks)))
# get feature score
feature_score = bst.get_fscore()
feature_score = sorted(feature_score.items(),
key=lambda x: x[1],
reverse=True)
fs = []
for (key, value) in feature_score:
fs.append("{0},{1}\n".format(key, value))
print "features scores:", kk
ff = './featurescore/feature_score_{0}.csv'.format(kk)
with open(ff, 'w') as f:
f.writelines("feature,score\n")
f.writelines(fs)
return kk
import numpy as np
from KS import KS
import matplotlib.pyplot as plt
import matplotlib.animation as animation
L = 16 # domain is 0 to 2.*np.pi*L
N = 128 # number of collocation points
dt = 0.5 # time step
diffusion = 1.0
ks = KS(L=L,diffusion=diffusion,N=N,dt=dt) # instantiate model
# define initial condition
#u = np.cos(x/L)*(1.0+np.sin(x/L)) # smooth IC
u = 0.01*np.random.normal(size=N) # noisy IC
# remove zonal mean
u = u - u.mean()
# spectral space variable.
ks.xspec[0] = np.fft.rfft(u)
# time stepping loop.
nmin = 1000; nmax = 5000
uu = []; tt = []
vspec = np.zeros(ks.xspec.shape[1], np.float)
x = np.arange(N)
fig, ax = plt.subplots()
line, = ax.plot(x, ks.x.squeeze())
ax.set_xlim(0,N-1)
ax.set_ylim(-3,3)
#Init only required for blitting to give a clean slate.
def init():
global line
from KS import KS, KSAssim
import numpy as np
"""
Author Benjamin Pachev <*****@*****.**> 2020
"""
def fourier_projector(spec, modes=21):
mod_spec = spec.copy()
mod_spec[:, modes:] = 0
return np.fft.irfft(mod_spec, axis=-1)
if __name__ == "__main__":
#See if the data assimilation works
true = KS()
assimilator = KSAssim(fourier_projector,
mu=1,
diffusion=3,
update_params=True)
max_n = 100
for n in range(max_n):
target = fourier_projector(true.xspec)
assimilator.set_target(target)
assimilator.advance()
true.advance()
print(assimilator.error(true))
# model parameters...
# for truth run
dt = 0.5
npts = 128
diffusion_truth = 1.0
# for forecast model (same as above for perfect model expt)
# for simplicity, assume dt and npts stay the same.
#diffusion = 0.9
diffusion = diffusion_truth
rstruth = np.random.RandomState(42) # fixed seed for truth run
rsens = np.random.RandomState(
) # varying seed for ob noise and ensemble initial conditions
# model instance for truth (nature) run
model = KS(N=npts, dt=dt, diffusion=diffusion_truth, rs=rstruth)
# mode instance for forecast ensemble
ensemble = KS(N=npts, members=nens, dt=dt, diffusion=diffusion, rs=rsens)
for nt in range(ntstart): # spinup truth run
model.advance()
# sample obs from truth, compute climo stats for model.
xx = []
tt = []
for nt in range(ntimes):
model.advance()
xx.append(model.x[0]) # single member
tt.append(float(nt) * model.dt)
xtruth = np.array(xx, np.float)
timetruth = np.array(tt, np.float)
xtruth_mean = xtruth.mean()
def XGBoost_(train=None,
y=None,
test=None,
dtest_X=None,
test_y=None,
k=0,
num_round=3500,
gamma=0.02,
min_child_weight=1.1,
max_depth=5,
lamda=10,
scale_pos_weight=3,
subsamp=0.7,
col_bytree=0.7,
col_bylevel=0.7,
eta=0.01,
file="aac"):
param = {
'booster': 'gbtree',
'objective': 'binary:logistic',
#'eval_metric':'auc',
'gamma': gamma,
'min_child_weight': min_child_weight,
'max_depth': max_depth,
'lambda': lamda,
'subsample': subsamp,
'colsample_bytree': col_bytree,
'colsample_bylevel': col_bylevel,
'eta': eta,
'tree_method': 'exact',
'seed': 0,
'nthread': 12
}
with open('./test/a{}.txt'.format(file), 'a') as f:
S = "gamma= " + str(gamma) + \
" scale_pos_weight= " + str(scale_pos_weight) + \
" min_child_weight= " + str(min_child_weight) + \
" max_depth= " + str(max_depth) + \
" lamda= " + str(lamda) + \
"\n" + \
"subsamp= " + str(subsamp) + \
" col_bytree= " + str(col_bytree) + \
" col_bylevel= " + str(col_bylevel) + \
" eta= " + str(eta)
f.writelines("{}\n".format(S))
dtrain = xgb.DMatrix(train, label=y, missing=-9999)
#cv_log = xgb.cv(param, dtrain,show_stdv=True,verbose_eval=1,feval=evalerror,num_boost_round=3500, nfold=5,early_stopping_rounds=10, seed=0)
#num_round = 21#cv_log.shape[0]
#cf = './featurescore/acvg{}.csv'.format(str(num_round))
#cv_log.to_csv(cf)
watchlist = [(dtrain, 'train'), (dtest_X, 'eval')]
#auc = cv_log['test-auc-mean'].max()
bst = xgb.train(param,
dtrain,
num_round,
watchlist,
maximize=True,
feval=evalerror,
early_stopping_rounds=50)
# make prediction
dtest = xgb.DMatrix(test, missing=-9999)
preds = bst.predict(dtest, ntree_limit=bst.best_ntree_limit)
p = bst.predict(dtrain, ntree_limit=bst.best_ntree_limit)
scores = bst.predict(dtest_X, ntree_limit=bst.best_ntree_limit)
fp, tp, thresholds = metrics.roc_curve(test_y, scores, pos_label=1)
auc = metrics.auc(fp, tp)
ks = KS(y=test_y.label, pred=scores)
kk = int(ks * 10000000000) % 10000
print "K-S:{}".format(ks)
print "AUC:{}".format(auc)
with open('./test/a{}.txt'.format(file), 'a') as f:
S = " best_ntree_limit:" + str(bst.best_ntree_limit) + \
" best_iteration= "+str(bst.best_iteration)+ \
"\nfeatures scores: " + str(kk)
f.writelines("{}\n".format(S))
f.writelines("K-S:{}\n".format(ks))
f.writelines("AUC:{}\n\n".format(metrics.auc(fp, tp)))
res = writeDatas(preds, test, "xgk_{}".format(str(kk)))
res.columns = ['label' + str(kk)]
y['label' + str(kk)] = p
y = pd.concat([y, res])
y.drop('label', axis=1, inplace=True)
y = y.reset_index()
try:
ypred = pd.read_csv("./test/y/a{}.csv".format(file))
y = pd.merge(y, ypred, on='userid')
except:
pass
finally:
y.to_csv("./test/y/a{}.csv".format(file), index=None)
# get feature score
feature_score = bst.get_fscore()
feature_score = sorted(feature_score.items(),
key=lambda x: x[1],
reverse=True)
fs = []
for (key, value) in feature_score:
fs.append("{0},{1}\n".format(key, value))
print "features scores:", kk
ff = './test/feature_score_{0}.csv'.format(kk)
with open(ff, 'w') as f:
f.writelines("feature,score\n")
f.writelines(fs)
def evalerror(preds, d):
labels = d.get_label()
return 'KS', KS(pred=preds, y=labels)
def XGBoost_gbdm(train=None,
y=None,
test=None,
dtest_X=None,
test_y=None,
k=0,
num_round=3500,
gamma=0.02,
min_child_weight=1.1,
max_depth=5,
lamda=10,
scale_pos_weight=3,
subsamp=0.7,
col_bytree=0.7,
col_bylevel=0.7,
eta=0.01,
file="aac"):
param = {
'booster': 'gbtree',
'objective': 'binary:logistic',
#'eval_metric':'auc',
'scale_pos_weight': scale_pos_weight,
'gamma': gamma,
'min_child_weight': min_child_weight,
'max_depth': max_depth,
'lambda': lamda,
'subsample': subsamp,
'colsample_bytree': col_bytree,
'colsample_bylevel': col_bylevel,
'eta': eta,
'tree_method': 'exact',
'seed': 0,
'nthread': 12
}
with open('./findx/af{}.txt'.format(file), 'a') as f:
S = "gamma= " + str(gamma) + \
" scale_pos_weight= " + str(scale_pos_weight) + \
" min_child_weight= " + str(min_child_weight) + \
" max_depth= " + str(max_depth) + \
" lamda= " + str(lamda) + \
"\n" + \
"subsamp= " + str(subsamp) + \
" col_bytree= " + str(col_bytree) + \
" col_bylevel= " + str(col_bylevel) + \
" eta= " + str(eta)
f.writelines("{}\n".format(S))
dtrain = xgb.DMatrix(train, label=y.label, missing=-9999)
#cv_log = xgb.cv(param, dtrain,show_stdv=True,verbose_eval=1,feval=evalerror,num_boost_round=3500, nfold=5,early_stopping_rounds=10, seed=0)
#num_round = 21#cv_log.shape[0]
#cf = './featurescore/acvg{}.csv'.format(str(num_round))
#cv_log.to_csv(cf)
watchlist = [(dtrain, 'train'), (dtest_X, 'eval')]
bst = xgb.train(param,
dtrain,
num_round,
watchlist,
maximize=True,
feval=evalerror,
early_stopping_rounds=50)
scores = bst.predict(dtest_X, ntree_limit=bst.best_ntree_limit)
fp, tp, thresholds = metrics.roc_curve(test_y, scores, pos_label=1)
auc = metrics.auc(fp, tp)
ks = KS(y=test_y.label, pred=scores)
kk = int(ks * 10000000000) % 10000
print "K-S:{}".format(ks)
print "AUC:{}".format(auc)
with open('./findx/af{}.txt'.format(file), 'a') as f:
S = " best_ntree_limit:" + str(bst.best_ntree_limit) + \
" best_iteration= "+str(bst.best_iteration)+ \
"\nfeatures scores: " + str(kk)
f.writelines("{}\n".format(S))
f.writelines("K-S:{}\n".format(ks))
f.writelines("AUC:{}\n\n".format(metrics.auc(fp, tp)))
#f.writelines("AUC:{}\n\n".format(metrics.auc(fp, tp)))
return ks, auc, bst.best_ntree_limit
criterion=c,
warm_start=True,
max_depth=md,
max_features=0.6,
min_samples_leaf=5,
n_jobs=12,
random_state=0)
rf.fit(train_X, train_y)
score = rf.predict_proba(test_X)[:, 1]
fp, tp, thresholds = metrics.roc_curve(test_y.values,
score,
pos_label=1)
ks = KS(y=test_y, score=score)
print "K-S:{}".format(ks)
print "AUC:{}".format(metrics.auc(fp, tp))
ans = rf.predict_proba(test)[:, 1]
with open('./featurescore/a.txt', 'a') as f:
S = "criterion= " + str(c) + \
" n_estimators= " + str(n) + \
" max_depth= " + str(md)
f.writelines("{}\n".format(S))
f.writelines("K-S:{}\n".format(ks))
f.writelines("AUC:{}\n\n".format(metrics.auc(fp, tp)))
writeDatas(ans, test, "rf{}".format(str(ks)))
except: