def on_epoch_end(self, epoch, logs={}):
y_pred_tr = self.model.predict_proba(self.X_tr)
logs['gini_tr'] = gini.gini_sklearn(self.y_tr, y_pred_tr)
y_pred_val = self.model.predict_proba(self.X_val)
logs['gini_val'] = gini.gini_sklearn(self.y_val, y_pred_val)
# if logs['gini_val'] > self.best_lap:
# self.best_lap = logs['gini_val']
# global pred_val, pred_test
# pred_val = y_pred_val
# pred_test = self.model.predict_proba(X_test)
print('Gini Score in training set: {}, test set: {}'.format(
logs['gini_tr'], logs['gini_val']))
return
def on_epoch_end(self, epoch, logs={}):
y_pred_tr = self.model.predict_proba(self.X_tr)
# roc = roc_auc_score(self.y_tr, y_pred_tr)
# logs['roc_auc'] = roc
# logs['gini_tr'] = (roc * 2 ) - 1
logs['gini_tr'] = gini.gini_sklearn(self.y_tr, y_pred_tr)
y_pred_val = self.model.predict_proba(self.X_val)
# roc = roc_auc_score(self.y_val, y_pred_val)
# logs['roc_auc_val'] = roc
# logs['gini_val'] = (roc * 2 ) - 1
logs['gini_val'] = gini.gini_sklearn(self.y_val, y_pred_val)
print('Gini Score in training set: {}, test set: {}'.format(
logs['gini_tr'], logs['gini_val']))
return
epochs=epochs,
verbose=99,
callbacks=callbacks)
# delete current model
del model
# load best model of each run
model = create_model()
model.load_weights(checkpoint_path, by_name=False)
# For train and valid only
pred_val = model.predict_proba(X_val).reshape(-1, )
# Store average score for evaluate model
score_fold.append(gini.gini_sklearn(y_val, pred_val))
oof_train[val_index] += pred_val / runs_per_fold
print('Run {}: {} \n'.format(j + 1, score_fold[j]))
pred_test_lap = model.predict_proba(X_test).reshape(-1, )
pred_test += pred_test_lap / runs_per_fold
# Store test predictions for submissions
oof_test += pred_test / KFOLDS
scores.append(gini.gini_sklearn(y_val, oof_train[val_index]))
print('[Fold {}/{} Gini score: {} \n]'.format(i + 1, KFOLDS,
scores[i]))
verbose=99,
callbacks=callbacks)
# delete current model
del model
# load best model of each run
model = create_model()
model.load_weights(checkpoint_path, by_name=False)
# For train and valid only
pred_val = model.predict_proba(X_val)
oof_train[val_index] += pred_val / runs_per_fold
# Store average score for evaluate model
score_fold.append(gini.gini_sklearn(y_val, pred_val))
print('Run {}: {}'.format(j + 1, score_fold[j]))
pred_test_lap = model.predict_proba(X_test)
pred_test += pred_test_lap / runs_per_fold
# Store test predictions for submissions
oof_test += pred_test / KFOLDS
scores.append(np.mean(score_fold))
print('[Fold {}/{} Gini score: {}]'.format(i + 1, KFOLDS, scores[i]))
gc.collect()
print('[Fold {}/{} END]'.format(i + 1, KFOLDS))
epochs=epochs,
verbose=99,
callbacks=callbacks)
model.fit(X_tr, y_tr, shuffle=False)
# We want the best model from checkpoint
del model
model = load_model(checkpoint_path)
# For train and valid only
pred_val = model.predict_proba(X_val)
oof_train[val_index] = pred_val
# Store average score for evaluate model
scores.append(gini.gini_sklearn(y_val, pred_val))
# Store test predictions for submissions
pred_test = model.predict_proba(X_test) / KFOLDS
oof_test += pred_test
print('[Fold {}/{} Gini score: {}]'.format(i + 1, KFOLDS, scores[i]))
gc.collect()
print('[Fold {}/{} END]'.format(i + 1, KFOLDS))
print('Average score: {}'.format(np.mean(scores)))
print('Total run time: {} seconds'.format(time.time() - tmp))
# Export oof_train
file_path = os.path.join(OOF_PATH, MODEL_NAME + '_train.csv')