'eta': 0.1,
'eval_metric': 'auc',
'max_depth': 4,
'objective': 'binary:logistic',
'silent': 1,
'tree_method': 'hist',
'nthread': 64,
'seed': SEED
}
#dtrain = xgb.DMatrix(X, y)
#col = X.columns
#del X, y; gc.collect()
gc.collect()
yhat, imp, ret = ex.stacking(X, y, param, 9999, nfold=5, esr=30)
imp.to_csv('imp_{}.csv'.format(datetime.today().date()), index=False)
# =============================================================================
# cv
# =============================================================================
#model = xgb.train(param, dbuild, NROUND, watchlist, verbose_eval=10,
# early_stopping_rounds=50)
#
#imp = ex.getImp(model)
#imp.to_csv('imp.csv', index=False)
#==============================================================================
utils.end(__file__)
])
return df
X = pd.concat([question_id] + [utils.read_pickles(f)
for f in train_files] + [vec],
axis=1)
#X = pd.concat([question_id]+[subsample(f) for f in train_files], axis=1)
y = utils.read_pickles('../data/label')['answer_score'].map(np.log1p)
# =============================================================================
# xgb
# =============================================================================
params = {
'max_depth': 5,
'eta': 0.1,
'colsample_bytree': 0.7,
'silent': 1,
'eval_metric': 'rmse',
'objective': 'reg:linear',
'tree_method': 'hist'
}
yhat, imp, ret = ex.stacking(X,
y,
params,
9999,
esr=50,
by='question_id',
seed=SEED)
}
y = X.is_duplicate
col = ['id', 'qid1', 'qid2', 'question1', 'question2', 'is_duplicate']
lls = []
iters = []
for i in range(loop):
X_, y_ = utils.down_sampling(X, y, p)
sub = X_[col]
X_.drop(col, axis=1, inplace=True)
yhat, imp, ret = ex.stacking(X_, y_, param, nround, esr=30, nfold=5)
ll = log_loss(y_, yhat)
iter_ = ret['best_iter']
print('998-1_cv LOGLOSS:', ll)
print('998-1_cv best iter:', iter_)
lls.append(ll)
iters.append(iter_)
print(np.mean(lls), np.mean(iters))
sub['yhat'] = yhat
sub['d'] = abs(sub.is_duplicate - sub.yhat)
sub.sort_values('d', ascending=False, inplace=True)
test_sub = test[['test_id']]
for i in range(1):
train = pd.merge(train, utils.load_010([0, 1, 2], 1), on='id', how='left')
test = pd.merge(test,
utils.load_010([0, 1, 2], 0),
on='test_id',
how='left')
col = get_valid_col(train.columns)
train.drop(col, axis=1, inplace=1)
test.drop(col, axis=1, inplace=1)
yhat, imp, ret = ex.stacking(train.drop('id', axis=1),
y_train,
param,
nround,
esr=30,
test=test.drop('test_id', axis=1))
gc.collect()
print('log_loss:', log_loss(y_train, yhat))
train_sub['yhat'] = yhat
test_sub['yhat'] = ret.get('test')
train_sub.drop('is_duplicate',
axis=1).to_csv('../feature/train_f009_xgb.csv.gz',
index=False,
compression='gzip')
train_sub.to_csv('../feature/test_f009_xgb.csv.gz',
index=False,
compression='gzip')
y_train = X.is_duplicate
sub = X[['id', 'is_duplicate']]
X.drop(col, axis=1, inplace=1)
def get_valid_col(col):
return [
c for c in col if c.count(',') > 0 or c.count('[') > 0
or c.count(']') > 0 or c.count('>') > 0
]
col = get_valid_col(X.columns)
X.drop(col, axis=1, inplace=1)
yhat, imp, ret = ex.stacking(X, y_train, param, nround, esr=30)
sub['yhat'] = yhat
sub[['id', 'yhat']].to_csv('../output/cv.csv.gz',
index=False,
compression='gzip')
imp.to_csv('../output/imp.csv.gz', index=False, compression='gzip')
print(log_loss(y_train, yhat))
"""
eta: 0.02
CV: 0.317928303359 ==> LB: 0.28018
eta: 0.2
CV: 0.363839232148