X, Y = generate_dataset(output_dim=nclass, num_examples=time_step)
all_X.append(X.T)
all_Y.append(Y)
all_X, all_Y = np.array(all_X), np.array(all_Y)
hist = model.fit([all_Y, all_X], all_Y, nb_epoch=200, batch_size=5)
# calc EER: we cut MFoM head, up to sigmoid output
input = model.get_layer(name='main_input').output
out = model.get_layer(name='output').output
cut_model = Model(input=input, output=out)
y_pred_sig = cut_model.predict(all_X)
eer_val = MT.eer(all_Y.flatten(), y_pred_sig.flatten())
print('sigma_EER: %.4f' % eer_val)
y_pred = model.predict([all_Y, all_X])
eer_val = MT.eer(all_Y.flatten(), 1. - y_pred.flatten())
print('l_EER: %.4f' % eer_val)
print(model.evaluate([all_Y, all_X], all_Y))
# TODO notice from the experiments:
# when we minimize obj.mfom_microf1 with psi = y_pred or psi = -y_pred + 0.5 in
# UvZMisclassification() layer, the smoothF1 is minimized !!! but EER is not at all.
# When we minimize obj.mfom_microf1 with psi = -y_pred + y_neg * unit_avg + y_true * zeros_avg,
# then both smoothF1 and EER are minimized :)
# history plot, alpha and beta params
m = model.get_layer('smooth_error_counter')
print('alpha: ', K.get_value(m.alpha))
print('beta: ', K.get_value(m.beta))
# print stats of psi misclassification measure
# misclassification layer, feed Y
y_true = Input(shape=(nclass, ), name='y_true')
psi = mfom.UvZMisclassification(name='uvz_misclass')([y_true, y_pred])
# class Loss function layer
out = mfom.SmoothErrorCounter(name='smooth_error_counter')(psi)
# compile model
model = Model(input=[y_true, feat_input], output=out)
model.compile(loss=obj.mfom_eer_normalized, optimizer='Adam')
model.summary()
# train
X, Y = generate_dataset(output_dim=nclass)
hist = model.fit([Y, X], Y, nb_epoch=100, batch_size=16)
# calc accuracy: we cut MFoM head, up to sigmoid output
input = model.get_layer(name='main_input').output
out = model.get_layer(name='output').output
cut_model = Model(input=input, output=out)
y_pred = cut_model.predict(X)
eer_val = MT.eer(y_true=Y.flatten(), y_pred=y_pred.flatten())
print('EER: %.4f' % eer_val)
# history plot, alpha and beta params
m = model.get_layer('smooth_error_counter')
print('alpha: ', K.get_value(m.alpha))
print('beta: ', K.get_value(m.beta))
plt.plot(hist.history['loss'])
plt.show()