def main():
configs = ["IS09", "IS10", "IS11", "IS12", "ComParE", "GeMAPS", "eGeMAPS"]
database = "IEMOCAP"
emotionsTest = ["Neutral", "Happiness", "Sadness", "Anger"]
for config in configs:
print("config: " + config)
features = pd.read_csv("features/{}_{}.csv".format(database, config))
features.set_index(["speaker", "emotion", "actType", "name"],
inplace=True)
cm = np.zeros((len(emotionsTest), len(emotionsTest)), dtype=int)
nSpeaker = len(features.index.unique('speaker'))
sps_val = [0, 0, 2, 2, 4, 4, 6, 6, 8, 8]
for sp in range(nSpeaker):
print("Testing speaker {}".format(sp))
(x_train_scaled, y_train,
x_val_scaled, y_val,
x_test_scaled, y_test) = preprocess(sp, sps_val[sp],
features, emotionsTest)
modelfile = "models/SVM/{}_{}.sav".format(config, sp)
svm = pickle.load(open(modelfile, "rb"))
y_p = svm.predict(x_test_scaled)
cm += confusion_matrix(y_test, y_p)
wa, ua = waua(cm)
cmp = cm / np.reshape(np.sum(cm, 1), (4, 1))
imageName = "results/SVM/Test_{}.png".format(config)
wa, ua = np.around(waua(cm), decimals=4)
title = "wa={}, ua={}".format(wa, ua)
plot_wauacm(title, cmp, emotionsTest, imageName)
print("wa: " + str(wa) + ", ua: " + str(ua))
print(cmp)
def main():
if not os.path.exists("results"):
os.mkdir("results")
if not os.path.exists("models"):
os.mkdir("models")
specPath = "spectrograms/"
if socket.getfqdn(socket.gethostname()) == "d8":
os.environ["CUDA_VISIBLE_DEVICES"] = "2,3"
batch_size = 64 * 2
else:
batch_size = 64
# parameters
t_uplim = 200 # 2 s
f_uplim = 400 # 4 kHz
t_shift_test = 40 # 0.4 s
# load the DataFrame of IEMOCAP
dataDf = pd.read_csv(specPath + "IEMOCAP.csv", index_col=0)
speakers = list(dataDf.speaker.unique())
emotionsTest = list(dataDf.emotion.unique())
num_classes = len(emotionsTest)
# cross-validation
cmTest = np.zeros((num_classes, num_classes))
# %%
for sp_test in range(len(speakers)):
speakers_train = speakers.copy()
speaker_test = speakers_train.pop(sp_test)
print("Test speaker: {}".format(speaker_test))
# preprocessing
# test set
(x_test, y_test, y_test_raw) = utils.prepro(specPath,
dataDf,
speaker_test,
emotionsTest,
"test",
t_shift_test,
f_uplim,
t_uplim)
# class weight
K.clear_session()
# %% test
model = load_model("models/Inception_{}.h5".format(sp_test))
y_test_p = model.predict(x_test,
batch_size=batch_size,
verbose=1)
cmTest += utils.y_raw_evaluate(y_test_p, y_test_raw)
del model
# %% print results
waTest, uaTest = utils.waua(cmTest)
# test set
print("Results on test set: wa={}, ua={}".format(waTest, uaTest))
print(cmTest)
title = "wa={}, ua={}".format(waTest, uaTest)
imageName = "results/Inception_Test_T.png"
cmTestP = cmTest / np.reshape(np.sum(cmTest, 1), (4, 1))
utils.plot_wauacm(title, cmTestP, emotionsTest, imageName, ".4f")
def main():
if socket.getfqdn(socket.gethostname()) == "d8":
os.environ["CUDA_VISIBLE_DEVICES"] = "3"
batch_size = 128
configs = ["IS09", "IS10", "IS11", "IS12", "ComParE", "GeMAPS", "eGeMAPS"]
# IEMOCAP imp NHSA
emotionsTest = ["Neutral", "Happiness", "Sadness", "Anger"]
database = "IEMOCAP"
for config in configs:
# Load features
features = pd.read_csv("features/{}_{}.csv".format(database, config))
features.set_index(["speaker", "emotion", "actType", "name"],
inplace=True)
nSpeaker = len(features.index.unique('speaker'))
cmV = np.zeros((len(emotionsTest), len(emotionsTest)), dtype=int)
cmT = np.zeros((len(emotionsTest), len(emotionsTest)), dtype=int)
for sp in range(nSpeaker):
(x_train_scaled, y_train_C,
x_val_scaled, y_val_C,
x_test_scaled, y_test_C) = preprocess(sp, sp-1,
features,
emotionsTest)
num_classes = len(emotionsTest)
model = load_model("models/DNN/{}_{}.h5".format(config, sp))
y_val_p = model.predict(x_val_scaled,
batch_size=batch_size,
verbose=1)
y_val_p_C = [emoC.argmax() for emoC in y_val_p]
cmV += confusion_matrix(y_val_C,
y_val_p_C,
labels=list(range(num_classes)))
y_test_p = model.predict(x_test_scaled,
batch_size=batch_size,
verbose=1)
cmT += confusion_matrix(y_test_C,
[emoC.argmax() for emoC in y_test_p],
labels=list(range(num_classes)))
del model
K.clear_session()
print("speaker: " + str(sp))
print(cmV)
print(cmT)
waV, uaV = np.around(waua(cmV), decimals=4)
waT, uaT = np.around(waua(cmT), decimals=4)
cmpV = cmV / np.reshape(np.sum(cmV, 1), (4, 1))
cmpT = cmT / np.reshape(np.sum(cmT, 1), (4, 1))
imageName = "results/DNN/Test_{}_V.png".format(config)
title = "wa={}, ua={}".format(waV, uaV)
plot_wauacm(title, cmpV, emotionsTest, imageName)
imageName = "results/DNN/Test_{}_T.png".format(config)
title = "wa={}, ua={}".format(waT, uaT)
plot_wauacm(title, cmpT, emotionsTest, imageName)
print("waV: " + str(waV) + ", uaV: " + str(uaV))
print(cmpV)
print("waT: " + str(waT) + ", uaT: " + str(uaT))
print(cmpT)
def main():
dataPath = "../../../Database/IEMOCAP_full_release/"
# hyperparams
tMax = 300
fMax = 400
batch_size = 64 * 2
# IEMOCAP
database = "IEMOCAP"
emotionsTest = ["Neutral", "Happiness", "Sadness", "Anger"]
num_classes = len(emotionsTest)
actTypeToUse = ["impro"]
emoTest = "".join([emo[0] for emo in actTypeToUse+emotionsTest])
dataname = "{}_{}".format(database, emoTest)
dataDf = utils.load_IEMOCAP(dataPath, actTypeToUse, emotionsTest)
# confusion matrix
cmV = np.zeros((num_classes, num_classes), dtype=int)
cmT = np.zeros((num_classes, num_classes), dtype=int)
for sp_test in range(10):
print("test speaker:", sp_test)
(x_train, y_train, y_train_raw,
x_val, y_val, y_val_raw,
x_test, y_test, y_test_raw) = utils.Preprocess(sp_test, emotionsTest,
dataDf, tMax, fMax)
model = load_model("models/CNN/{}.h5".format(sp_test))
y_val_p = model.predict(x_val,
batch_size=batch_size,
verbose=1)
y_val_p_C = utils.y_raw_transform(y_val_p, y_val_raw)
cmV += confusion_matrix(y_val_raw[:, 0], y_val_p_C)
y_test_p = model.predict(x_test,
batch_size=batch_size,
verbose=1)
y_test_p_C = utils.y_raw_transform(y_test_p, y_test_raw)
cmT += confusion_matrix(y_test_raw[:, 0], y_test_p_C)
print("speaker: " + str(sp_test))
print(cmV)
print(cmT)
del model
keras.backend.clear_session()
# Evaluation
waV, uaV = np.around(utils.waua(cmV), decimals=4)
waT, uaT = np.around(utils.waua(cmT), decimals=4)
cmpV = cmV / np.reshape(np.sum(cmV, 1), (4, 1))
cmpT = cmT / np.reshape(np.sum(cmT, 1), (4, 1))
imageName = "results/CNN/Test_{}_CNN_V.png".format(dataname)
title = "wa={}, ua={}".format(waV, uaV)
utils.plot_wauacm(title, cmpV, emotionsTest, imageName)
imageName = "results/CNN/Test_{}_CNN_T.png".format(dataname)
title = "wa={}, ua={}".format(waT, uaT)
utils.plot_wauacm(title, cmpT, emotionsTest, imageName)
print("waV: " + str(waV) + ", uaV: " + str(uaV))
print(cmpV)
print("waT: " + str(waT) + ", uaT: " + str(uaT))
print(cmpT)
def SVM_Train(config, emotionsTest, features, params):
C, gamma = params
cm = np.zeros((len(emotionsTest), len(emotionsTest)), dtype=int)
nSpeaker = len(features.index.unique('speaker'))
sps_val = [0, 0, 2, 2, 4, 4, 6, 6, 8, 8]
for sp in range(nSpeaker):
print("Training with test speaker {}".format(sp))
(x_train_scaled, y_train, x_val_scaled, y_val, x_test_scaled,
y_test) = preprocess(sp, sps_val[sp], features, emotionsTest)
svm = SVC(class_weight="balanced", C=C, gamma=gamma)
svm.fit(x_train_scaled, y_train)
y_p = svm.predict(x_test_scaled)
cm += confusion_matrix(y_test, y_p)
modelfile = "models/SVM/{}_{}.sav".format(config, sp)
pickle.dump(svm, open(modelfile, "wb"))
wa, ua = waua(cm)
cmp = cm / np.reshape(np.sum(cm, 1), (4, 1))
imageName = "results/SVM/Train_{}.png".format(config)
wa, ua = np.around(waua(cm), decimals=4)
title = "C={}, gamma={}, wa={}, ua={}".format(C, gamma, wa, ua)
plot_wauacm(title, cmp, emotionsTest, imageName)
print("wa: " + str(wa) + ", ua: " + str(ua))
print(cmp)
def main():
dataPath = "../../../Database/IEMOCAP_full_release/"
if not os.path.exists("results"):
os.mkdir("results")
if not os.path.exists("results/CNN"):
os.mkdir("results/CNN")
if not os.path.exists("models"):
os.mkdir("models")
if not os.path.exists("models/CNN"):
os.mkdir("models/CNN")
if socket.getfqdn(socket.gethostname()) == "d8":
os.environ["CUDA_VISIBLE_DEVICES"] = "2,3"
# hyperparams
dr = 0.3
lr = 0.006
tMax = 300
fMax = 400
input_shape = (tMax, fMax, 1)
batch_size = 64 * 2
# IEMOCAP
database = "IEMOCAP"
emotionsTest = ["Neutral", "Happiness", "Sadness", "Anger"]
num_classes = len(emotionsTest)
actTypeToUse = ["impro"]
emoTest = "".join([emo[0] for emo in actTypeToUse + emotionsTest])
dataname = "{}_{}".format(database, emoTest)
dataDf = utils.load_IEMOCAP(dataPath, actTypeToUse, emotionsTest)
# confusion matrix
cmV = np.zeros((num_classes, num_classes), dtype=int)
cmT = np.zeros((num_classes, num_classes), dtype=int)
for sp_test in range(10):
print("test speaker:", sp_test)
(x_train, y_train, y_train_raw, x_val, y_val, y_val_raw, x_test,
y_test, y_test_raw) = utils.Preprocess(sp_test, emotionsTest, dataDf,
tMax, fMax)
# set training parameters
mc_cb = keras.callbacks.ModelCheckpoint('BestModel_CNN.h5',
monitor='val_acc',
verbose=1,
save_best_only=True,
save_weights_only=True)
mc_es = keras.callbacks.EarlyStopping(monitor="acc",
patience=10,
verbose=1)
# class weight
cw = class_weight.compute_class_weight('balanced',
np.unique(y_train_raw[:, 0]),
y_train_raw[:, 0])
cw2 = {0: cw[0], 1: cw[1], 2: cw[2], 3: cw[3]}
model = cnn_model(num_classes=num_classes,
input_shape=input_shape,
dr=dr,
lr=lr)
epochs = 16
model.fit(x_train,
y_train,
class_weight=cw2,
batch_size=batch_size,
epochs=epochs,
verbose=1,
validation_data=(x_val, y_val))
epochs = 50
model.fit(x_train,
y_train,
class_weight=cw2,
batch_size=batch_size,
epochs=epochs,
verbose=1,
callbacks=[mc_cb, mc_es],
validation_data=(x_val, y_val))
del mc_cb
model.load_weights("BestModel_CNN.h5")
y_val_p = model.predict(x_val, batch_size=batch_size, verbose=1)
y_val_p_C = utils.y_raw_transform(y_val_p, y_val_raw)
cmV += confusion_matrix(y_val_raw[:, 0], y_val_p_C)
y_test_p = model.predict(x_test, batch_size=batch_size, verbose=1)
y_test_p_C = utils.y_raw_transform(y_test_p, y_test_raw)
cmT += confusion_matrix(y_test_raw[:, 0], y_test_p_C)
print("speaker: " + str(sp_test))
print(cmV)
print(cmT)
model.save("models/CNN/{}.h5".format(sp_test))
del model
keras.backend.clear_session()
# Evaluation
waV, uaV = np.around(utils.waua(cmV), decimals=4)
waT, uaT = np.around(utils.waua(cmT), decimals=4)
cmpV = cmV / np.reshape(np.sum(cmV, 1), (4, 1))
cmpT = cmT / np.reshape(np.sum(cmT, 1), (4, 1))
imageName = "results/CNN/Train_{}_CNN_V.png".format(dataname)
title = "wa={}, ua={}".format(waV, uaV)
utils.plot_wauacm(title, cmpV, emotionsTest, imageName)
imageName = "results/CNN/Train_{}_CNN_T.png".format(dataname)
title = "wa={}, ua={}".format(waT, uaT)
utils.plot_wauacm(title, cmpT, emotionsTest, imageName)
print("waV: " + str(waV) + ", uaV: " + str(uaV))
print(cmpV)
print("waT: " + str(waT) + ", uaT: " + str(uaT))
print(cmpT)
def DNN_Train(features, emotionsTest, config):
dr = 0.4
lr = 0.0002
batch_size = 128
epochs = 100
nSpeaker = len(features.index.unique('speaker'))
cmV = np.zeros((len(emotionsTest), len(emotionsTest)), dtype=int)
cmT = np.zeros((len(emotionsTest), len(emotionsTest)), dtype=int)
for sp in range(nSpeaker):
(x_train_scaled, y_train_C, x_val_scaled, y_val_C, x_test_scaled,
y_test_C) = preprocess(sp, sp - 1, features, emotionsTest)
num_classes = len(emotionsTest)
y_train = keras.utils.to_categorical(y_train_C, num_classes)
# y_test = keras.utils.to_categorical(y_test_C, num_classes)
y_val = keras.utils.to_categorical(y_val_C, num_classes)
# DNN model
input_shape = (features.shape[1], )
mc_cb = keras.callbacks.ModelCheckpoint('BestModel_DNN.h5',
monitor='val_acc',
verbose=1,
save_best_only=True,
save_weights_only=True)
es_cb = keras.callbacks.EarlyStopping(monitor="acc",
patience=10,
verbose=1)
# class weight
cw = class_weight.compute_class_weight('balanced',
np.unique(y_train_C), y_train_C)
cw2 = {0: cw[0], 1: cw[1], 2: cw[2], 3: cw[3]}
tf.reset_default_graph()
graph = tf.Graph()
with graph.as_default() as g:
with tf.Session(graph=g):
model = make_model(num_classes=num_classes,
input_shape=input_shape,
dr=dr,
lr=lr)
# if socket.getfqdn(socket.gethostname()) == "d8":
# model = multi_gpu_model(model, gpus=4)
model.fit(x_train_scaled,
y_train,
class_weight=cw2,
batch_size=batch_size,
epochs=epochs,
verbose=1,
callbacks=[mc_cb, es_cb],
validation_data=(x_val_scaled, y_val))
# evaluation
del mc_cb
model.load_weights('BestModel_DNN.h5')
y_val_p = model.predict(x_val_scaled,
batch_size=batch_size,
verbose=1)
y_val_p_C = [emoC.argmax() for emoC in y_val_p]
cmV += confusion_matrix(y_val_C,
y_val_p_C,
labels=list(range(num_classes)))
y_test_p = model.predict(x_test_scaled,
batch_size=batch_size,
verbose=1)
cmT += confusion_matrix(y_test_C,
[emoC.argmax() for emoC in y_test_p],
labels=list(range(num_classes)))
print("speaker: " + str(sp))
print(cmV)
print(cmT)
model.save("models/DNN/{}_{}.h5".format(config, sp))
waV, uaV = np.around(waua(cmV), decimals=4)
waT, uaT = np.around(waua(cmT), decimals=4)
cmpV = cmV / np.reshape(np.sum(cmV, 1), (4, 1))
cmpT = cmT / np.reshape(np.sum(cmT, 1), (4, 1))
imageName = "results/DNN/Train_{}_V.png".format(config)
title = "wa={}, ua={}".format(waV, uaV)
plot_wauacm(title, cmpV, emotionsTest, imageName)
imageName = "results/DNN/Train_{}_T.png".format(config)
title = "wa={}, ua={}".format(waT, uaT)
plot_wauacm(title, cmpT, emotionsTest, imageName)
print("waV: " + str(waV) + ", uaV: " + str(uaV))
print(cmpV)
print("waT: " + str(waT) + ", uaT: " + str(uaT))
print(cmpT)
def main():
if not os.path.exists("results"):
os.mkdir("results")
if not os.path.exists("models"):
os.mkdir("models")
specPath = "spectrograms/"
if socket.getfqdn(socket.gethostname()) == "d8":
os.environ["CUDA_VISIBLE_DEVICES"] = "2,3"
batch_size = 64 * 2
else:
batch_size = 64
# parameters
t_uplim = 200 # 2 s
f_uplim = 400 # 4 kHz
t_shift_train = 100 # 1 s
t_shift_test = 40 # 0.4 s
# load the DataFrame of IEMOCAP
dataDf = pd.read_csv(specPath + "IEMOCAP.csv", index_col=0)
speakers = list(dataDf.speaker.unique())
emotionsTest = list(dataDf.emotion.unique())
num_classes = len(emotionsTest)
# cross-validation
cmTest = np.zeros((num_classes, num_classes))
# %%
for sp_test in range(len(speakers)):
speakers_train = speakers.copy()
speaker_test = speakers_train.pop(sp_test)
print("Test speaker: {}".format(speaker_test))
# preprocessing
# train set
(x_train, y_train,
y_train_raw) = utils.prepro(specPath, dataDf, speakers_train,
emotionsTest, "train", t_shift_train,
f_uplim, t_uplim)
# test set
(x_test, y_test, y_test_raw) = utils.prepro(specPath, dataDf,
speaker_test, emotionsTest,
"test", t_shift_test,
f_uplim, t_uplim)
# class weight
cw = class_weight.compute_class_weight('balanced',
np.unique(y_train_raw[:, 0]),
y_train_raw[:, 0])
cw2 = {0: cw[0], 1: cw[1], 2: cw[2], 3: cw[3]}
K.clear_session()
model = cnn_model(f_uplim, t_uplim, num_classes)
lrs_cb = keras.callbacks.LearningRateScheduler(lrschedule, verbose=1)
epochs = 50
model.compile(optimizer=SGD(lr=0.05,
momentum=0.9,
decay=0.0001,
nesterov=True),
loss="categorical_crossentropy",
metrics=["accuracy"])
model.fit(x_train,
y_train,
batch_size=batch_size,
epochs=epochs,
class_weight=cw2,
callbacks=[lrs_cb],
validation_data=(x_test, y_test))
# %% test
y_test_p = model.predict(x_test, batch_size=batch_size, verbose=1)
cmTest += utils.y_raw_evaluate(y_test_p, y_test_raw)
model.save("models/Inception_{}.h5".format(sp_test))
del model
# %% print results
waTest, uaTest = utils.waua(cmTest)
# test set
print("Results on test set: wa={}, ua={}".format(waTest, uaTest))
print(cmTest)
title = "wa={}, ua={}".format(waTest, uaTest)
imageName = "results/Inception_Train_T.png"
cmTestP = cmTest / np.reshape(np.sum(cmTest, 1), (4, 1))
utils.plot_wauacm(title, cmTestP, emotionsTest, imageName, ".4f")