# load audio models
# fit audio ensemble
trainy_audio = pd.Series(trainy_audio).factorize(sort=True)[0]
audio_logistic = fit_logistic(all_models, trainX_audio, trainy_audio)
#
# evaluate standalone models on test dataset
from tensorflow.keras.utils import to_categorical
from sklearn.metrics import accuracy_score, confusion_matrix
stackedX_train = stacked_dataset(all_models, trainX_audio)
trainy_audio_pred = audio_logistic.predict(stackedX_train)
print('Stacked Model Accuracy (train): ', accuracy_score(trainy_audio, trainy_audio_pred))
#############
testy_audio = pd.Series(testy_audio).factorize(sort=True)[0]
stackedX_test = stacked_dataset(all_models, testX_audio)
testy_audio_pred = audio_logistic.predict(stackedX_test)
print('Stacked Model Accuracy (test): ', accuracy_score(testy_audio, testy_audio_pred))
# stacking: regression + 4 emotions (treat audio models as one)
from sklearn.linear_model import LogisticRegression
from scipy.special import softmax
import numpy as np
index_map = {'neg': 0, 'neu': 1, 'pos': 2}
rev = { v:k for k,v in index_map.items()}
##### audio
# convert audio into feature
audio = audio2wave(file)
audio_X = prepare_data_librosa(audio,
features='logmel',
scaled=True)
stackedX_test = stacked_dataset(all_models, audio_X)
audio_pred = audio_logistic.predict(stackedX_test)
print('Audio prediction:', [rev[item] for item in audio_pred])
##### text
text_pred = predict_text(bert_model,text)
print('Text prediction:', [rev[item] for item in np.array([np.argmax(text_pred)])])
##### ensemble
ensemble_text_test = softmax(text_pred)
ensemble_audio_test = audio_logistic.predict_proba(stackedX_test)
stack_test = np.dstack((ensemble_text_test, ensemble_audio_test))
stack_test = stack_test.reshape((stack_test.shape[0], stack_test.shape[1]*stack_test.shape[2]))
ensemble_pred = final_logistic.predict(stack_test)