def main():
dataset_add = '/home/bardia/Documents/University/Seventh Semester/Principles of Computational Intelligence/Digits Dataset/classes/'
splitRatio = 0.33
sample_per_class = 20
radius = 6000
train_set, test_set = splitDataset(dataset_add, sample_per_class, splitRatio)
predictions = getPredictions(train_set, test_set, radius)
print('-------------')
accuracy = getAccuracy(predictions)
print('Accuracy is :' + str(accuracy) + '%')
def main():
dataset_add = '/home/bardia/Documents/University/Seventh Semester/Principles of Computational Intelligence/Digits Dataset/classes/'
splitRatio = 0.33
sample_per_class = 5
train_set, test_set = splitDataset(dataset_add, sample_per_class, splitRatio)
print('train=%d and test=%d' % (len(train_set), len(test_set)))
# prepare model
separated_by_class = separateByClass(train_set)
summaries = summarizeByClass(separated_by_class)
# test model
predictions = getPredictions(summaries, test_set)
accuracy = getAccuracy(predictions)
print('Accuracy: ' + str(accuracy) + '%')
S = norm_float2int(S)
N = norm_float2int(N)
# mix noise and speech, trim 10 seconds if TOY, trim 2min 30seconds otherwise
X, S = dataMix(S, N, fs, mix, TOY = args.TOY)
X_int = norm_float2int(X)
wavfile.write("X_train.wav", fs, X_int)
S_int = norm_float2int(S)
wavfile.write("S_train.wav", fs, S_int)
S_mat = seq2mat(S, Frame_len, shift)
X_mat = seq2mat(X, Frame_len, shift)
S_train, S_dev, S_test = splitDataset(S_mat, seed = 1)
X_train, X_dev, X_test = splitDataset(X_mat, seed = 1)
FS_train = time2CompSpectro(S_train)
FX_train = time2CompSpectro(X_train)
FS_dev = time2CompSpectro(S_dev)
FX_dev = time2CompSpectro(X_dev)
F2M = freq2mel(n_Mel, fs, FX_train)
# get M2F matrix by transpose
M2F = F2M.T
# normalize M2F matrix, plus a very small number to avoid underflow
M2F = M2F / (np.sum(M2F, axis = 0, keepdims = True) + 0.00000001)
MEL_S_train = np.dot(np.abs(FS_train), F2M) # output of NN (MEL_S/MEL_X)
MEL_X_train = np.dot(np.abs(FX_train), F2M) # input of NN (MEL_X)
end_t = 1500
n_epochs = 1000 # number of complete pass in the dataset for training
patience = 3 # patience for the early stopping
l1_reg_const = 0.001 # const for l1 regularization penalty
l2_reg_const = 0.001 # const for l2 regularization penalty
######################################
if section == "4.3.1.1":
_, output = generateDataset(start_t, end_t)
t = [i for i in range(start_t, end_t)]
plotTimeSeries(t, output)
if section == "4.3.1.2":
input, output = generateDataset(start_t, end_t)
x_train, x_val, x_test, y_train, y_val, y_test = splitDataset(input, output)
sizes = [8, 16, 32, 64, 128, 256, 512] # models of NN to be trained
nn2error = defaultdict(list)
# Model
for size in tqdm(sizes): # for each model size
for i in range(5): # train 5 times for statistical significance
model = tf.keras.Sequential([
layers.Dense(size, activation='relu', kernel_regularizer=regularizers.l1_l2(l1_reg_const, l2_reg_const),
input_shape=[len(x_train[0])]),
layers.Dense(1)
])
model.compile(optimizer='adam', loss='mse', metrics=['mse'])