def _train(X_train, model, y_train, iteration):

# TODO: Train based on model from previous iteration instead of from scratch (although it's not a real bottleneck)

model.get_model().set_params(n_estimators=(iteration * val_at_num_epoch))

model.fit(X_train, y_train)

extract_test_func = lambda df: df[df['split'] == 'valid']

X_train = train[feats]

y_train = train[target]

epochs_without_improvement = 0

best_score = np.inf

best_iter = 0

iter = 1

while epochs_without_improvement < early_stopping:

_train(X_train, model, y_train, iter)

new_valid = predict_one_by_one(train=train, test=valid, feats=feats, model=model, agg_function=agg_function,

extract_test_func=extract_test_func)

score = rmse(valid[target].values, new_valid[target].values)

print(f'RMSE on valid: {score}')

if score < best_score:

best_score = score

best_iter = iter

epochs_without_improvement = 0

else:

epochs_without_improvement += 1

iter += 1

model.get_model().set_params(n_estimators=(best_iter * val_at_num_epoch))

model.fit(X_train, y_train)

print(f'score didn\'t improve for {epochs_without_improvement} epochs - finished training with best score of {best_score}')