def f(mi, s, md, r, c):
clf = BoostedTreesClassifier(max_iterations = mi,
step_size = s,
max_depth = md,
row_subsample = r,
column_subsample = c,
verbose = 0)
clf.fit(X[valid_idx], y[valid_idx])
#yhat = clf.predict_proba(X[train_idx])
#return -log_loss(y[train_idx], yhat)
return clf.score(X[train_idx], y[train_idx])
def OptBTC(step_size=.5,
max_iterations=100,
row_subsample=.9,
column_subsample=.9,
max_depth=8):
model = BoostedTreesClassifier(step_size=step_size,
max_iterations=max_iterations,
row_subsample=row_subsample,
column_subsample=column_subsample,
max_depth=max_depth)
logger.info("Params: %s", model.get_params())
return ReportPerfCV(model, "original", y)
def f(params):
mi = params['mi']
md = params['md']
s = params['s']
r = params['r']
c = params['c']
mi = int(mi)
md = int(md)
clf = BoostedTreesClassifier(max_iterations = mi,
step_size = s,
max_depth = md,
row_subsample = r,
column_subsample = c,
verbose = 0)
clf.fit(X[valid_idx], y[valid_idx])
yhat = clf.predict_proba(X[train_idx])
return log_loss(y[train_idx], yhat)
'row_subsample': Uniform(.3, 1.),
'column_subsample': Uniform(.3, 1.),
'min_child_weight': LogUniform(.01, 100),
'min_loss_reduction': Uniform(0.0001, 10)
}
from gl import BoostedTreesClassifier
logger2 = logging.getLogger('graphlab')
logger2.setLevel(logging.CRITICAL)
CONFIG['ensemble_list'] = [
'btc', 'btc2', 'btc3', 'btc4', 'svc', 'svc2', 'svc3', 'nn', 'nn2',
'nic', 'mpc', 'knc', 'etc', 'cccv', 'log', 'crfcbag', 'cetcbag',
'keras'
]
X, Xtest = GetDataset('ensemble', ensemble_list=CONFIG['ensemble_list'])
print "lkjr"
clf = GaussianProcessCV(estimator=BoostedTreesClassifier(verbose=False),
param_distributions=param_distributions,
kernel=DoubleExponential,
scoring=LogLoss,
mu_prior=-1.,
sigma_prior=.30,
sig=.01,
cv=5,
max_iter=55,
random_state=1,
time_budget=24 * 3600)
clf.fit(X, y)
clf = RandomSearchCV(estimator=BoostedTreesClassifier(verbose=False),
param_distributions=param_distributions,
cv=5,
max_iter=100,
res_cv = res_cv + yhat_cv
res_test = res_test + yhat_test
print i, log_loss(y, yhat_cv)
res_cv = res_cv/n_bag
res_test = res_test/n_bag
np.savez_compressed('../Submission/yhat_cetcbag_full.npz', yhat = res_cv)
np.savez_compressed('../Submission/yhat_cetcbag_test.npz', yhat = res_test)
if False:
X, y, Xtest = LoadData()
from gl import BoostedTreesClassifier
import json
import time
params = json.load(open('../Params/Best/BTC:original_saved_params.json'))
params = params['BTC:original']
np.random.seed(1)
clf = BoostedTreesClassifier(**params)
res_cv, res_test = 0, 0
n_bag = 10
time_before = time.time()
for i in xrange(n_bag):
yhat_cv, yhat_test = GetYhat(clf, 'original', y)
res_cv = res_cv + yhat_cv
res_test = res_test + yhat_test
print i, log_loss(y, yhat_cv)
if time.time() - time_before > 9*3600: break
res_cv = res_cv/(i+1)
res_test = res_test/(i+1)
np.savez_compressed('../Submission/yhat_btcbag_full.npz', yhat = res_cv)
np.savez_compressed('../Submission/yhat_btcbag_test.npz', yhat = res_test)
if True:
import numpy as np
import itertools
from scipy.stats import uniform
from scipy.stats import randint
import logging
from ml import *
X, _ = GetDataset('original')
_, y, _ = LoadData()
from gl import BoostedTreesClassifier
clf = BoostedTreesClassifier(verbose = 0)
clf.fit(X[:10], y[:10])
np.random.seed(1)
if False:
Cs = [.001, .01, .1, 1., 10.]
gammas = [.001, .01, .1, 1., 10.]
res = []; i = 0
for (C, gamma) in itertools.product(Cs, gammas):
print i, C, gamma; i += 1
clf = SVC(C = C, gamma = gamma)
clf.fit(X[train], y[train])
res.append(clf.score(X[valid], y[valid]))
res2 = []
for i in xrange(len(res)*100):
C = 10**uniform(-3.5,5).rvs()
gamma = 10**uniform(-3.5,5).rvs()
print i, C, gamma