def k_fold_train(self):
if self.config.KFOLD:
k_fold = KFold(n_splits=self.config.N_FOLDS,
shuffle=True,
random_state=712)
else:
k_fold = StratifiedKFold(n_splits=self.config.N_FOLDS,
shuffle=True,
random_state=712)
test_predictions = np.zeros(self.test_features.shape[0])
for train_indices, valid_indices in k_fold.split(
self.train_features,
None if self.config.KFOLD else self.train_labels):
train_x, train_y = self.train_features.iloc[
train_indices, :], self.train_labels.iloc[train_indices, :]
valid_x, valid_y = self.train_features.iloc[
valid_indices, :], self.train_labels.iloc[valid_indices, :]
model = ctrNet.build_model(self.param)
model.train(train_data=(train_x, train_y),
dev_data=(valid_x, valid_y))
print("Training Done! Inference...")
test_predictions = np.add(
test_predictions,
rankdata(
model.infer(dev_data=(self.test_features,
test_predictions))) /
test_predictions.shape[0])
del model
gc.collect()
submission(self.config, test_predictions, True, '')
cross_activation='identity',
init_method='uniform',
init_value=0.1,
feature_nums=len(features),
kfold=5)
utils.print_hparams(hparam)
# # Training model
# In[6]:
kfold = KFold(n_splits=hparam.kfold, shuffle=True, random_state=712)
for i, (train_index, dev_index) in enumerate(kfold.split(train)):
print('Fold', i)
model = ctrNet.build_model(hparam)
model.train(train_data=(train.iloc[train_index][features],
train.iloc[train_index]['HasDetections']),
dev_data=(train.iloc[dev_index][features],
train.iloc[dev_index]['HasDetections']))
print("Training Done! Inference...")
if i == 0:
preds = model.infer(dev_data=(test[features],
test['HasDetections'])) / hparam.kfold
else:
preds += model.infer(dev_data=(test[features],
test['HasDetections'])) / hparam.kfold
# # Inference
# In[7]: