'colsample_bytree': 0.5,
'target': 'target',
'validation_set': xg_valid,
'num_class' : 9,
'objective': 'multi:softprob',
'eval:metric': 'mlogloss',
'silent': 1,
}
watchlist = [ (xg_train, 'train'), (xg_valid, 'valid') ]
bst = xgb.train(params, xg_train, rounds, watchlist,
early_stopping_rounds=100, evals_result=evals)
return bst, evals
if __name__ == '__main__':
X, y = OttoCompetition.load_data(train=True)
X_test, _ = OttoCompetition.load_data(train=False)
le = LabelEncoder().fit(y)
all_hold = []
all_hold_predict = []
all_test_predict = []
all_weights = []
# 20% holdout
for i, (data_index, hold_index) in enumerate(StratifiedKFold(y, n_folds = 5, random_state=0)):
X_data, X_hold = X[data_index], X[hold_index]
y_data, y_hold = y[data_index], y[hold_index]
y_hold_predict = []
y_test_predict = []
# train with 50%, validation with 5%
for j, (train_index, valid_index) in enumerate(StratifiedShuffleSplit(y_data, 20, test_size = 0.05, train_size = 0.5, random_state=0)):
from lasagne.layers import DenseLayer
from lasagne.layers import InputLayer
from lasagne.layers import DropoutLayer
from lasagne.nonlinearities import softmax
from lasagne.updates import nesterov_momentum
from nolearn.lasagne import NeuralNet
if __name__ == '__main__':
encoder = LabelEncoder()
# Identical to StandardScaler using all train and test data.
scaler = OttoScaler()
# Training data
X, y = OttoCompetition.load_data(train=True)
y = encoder.fit_transform(y).astype('int32')
X = scaler.transform(X).astype('float32')
n_classes = np.unique(y).shape[0]
n_features = X.shape[1]
# Split a holdout set
data_idx, hold_idx = next(iter(StratifiedShuffleSplit(y, 1, test_size = 0.2, random_state=0)))
X_data, X_hold = X[data_idx], X[hold_idx]
y_data, y_hold = y[data_idx], y[hold_idx]
# Test data
X_test, _ = OttoCompetition.load_data(train=False)
X_test = scaler.transform(X_test).astype('float32')
]
Xt, X_test, yt, y_test = train_test_split(train_ffm.values, y.values, test_size = 0.2)
ll = []
for i, (train_index, valid_index) in enumerate(StratifiedKFold(yt, n_folds = 10, random_state=0)):
print('Fold {}'.format(i))
X_train, X_valid = Xt[train_index], Xt[valid_index]
y_train, y_valid = yt[train_index], yt[valid_index]
valid_set = []
lb = LabelBinarizer()
ybin = lb.fit_transform(yt)
for ylabel in lb.classes_:
print(ylabel)
tdf = pd.DataFrame(np.vstack([(y_train == ylabel).T,X_train.T]).T)
vdf = pd.DataFrame(np.vstack([(y_valid == ylabel).T,X_valid.T]).T)
train_file = './ffm/ffm_train_fold_{}_{}.csv'.format(i, ylabel)
valid_file = './ffm/ffm_valid_fold_{}_{}.csv'.format(i, ylabel)
model_file = './ffm/ffm_model_fold_{}_{}.csv'.format(i, ylabel)
predt_file = './ffm/ffm_predt_fold_{}_{}.csv'.format(i, ylabel)
tdf.to_csv(train_file, sep=" ", header=False, index=False,
quote=csv.QUOTE_NONE, quotechar=" ")
vdf.to_csv(valid_file, sep=" ", header=False, index=False,
quote=csv.QUOTE_NONE, quotechar=" ")
check_call(['ffm-train'] + ffm_params + ['-p', valid_file, train_file, model_file])
check_call(['ffm-predict', valid_file, model_file, predt_file])
valid_set.append(np.loadtxt(predt_file))
yp = np.array(valid_set).T
yp = (yp / yp.sum(axis=1)[:, np.newaxis])
ll.append(OttoCompetition.score(y_valid, yp, lb.classes_.tolist()))