-
Notifications
You must be signed in to change notification settings - Fork 1
/
tf_0.py
164 lines (129 loc) · 5.21 KB
/
tf_0.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
import tensorflow as tf
from tensorflow import keras
from keras import callbacks
from generate_dataframe import generate_dataframe
import librosa # for audio processing
import numpy as np
# NOTE: Since we are not displaying any graphs here we do not need librosa.display
# Sources: https://www.pythonpool.com/spectrogram-python/
# https://www.codespeedy.com/determine-input-shape-in-keras-tensorflow/
# Generate dataframe of audio file locations with database_maker function
df = generate_dataframe()
# Retrieve the mix and bass for training
mix_df = df[['Track', 'Mixture']]
bass_df = df[['Track', 'Bass']]
print('Dataframes Generated')
# Create arrays of locations of files from the database
mix_location_arr = [[]]
for i, j in mix_df.iterrows():
if i > 0:
mix_location_arr.append(str(df.loc[i][0] + '/' + df.loc[i][1]))
bass_location_arr = [[]]
for i, j in bass_df.iterrows():
if i > 0:
bass_location_arr.append(str(df.loc[i][0] + '/' + df.loc[i][1]))
# Load audio files into an array
mix_audio_arr = []
for i in range(len(mix_location_arr)):
if i > 0:
temp, _ = librosa.load(mix_location_arr[i])
mix_audio_arr.append(temp)
bass_audio_arr = []
for i in range(len(bass_location_arr)):
if i > 0:
temp, _ = librosa.load(bass_location_arr[i])
bass_audio_arr.append(temp)
print('Files loaded')
# Find STFT of given audio
mix_stft_arr = []
for i in range(len(mix_audio_arr)):
mix_stft_arr.append(librosa.stft(mix_audio_arr[i]))
bass_stft_arr = []
for i in range(len(bass_audio_arr)):
bass_stft_arr.append(librosa.stft(bass_audio_arr[i]))
print('STFT done')
# Convert frequency in STFT to dB
mix_arr = []
for i in range(len(mix_stft_arr)):
mix_arr.append(librosa.amplitude_to_db(abs(mix_stft_arr[i])))
bass_arr = []
for i in range(len(bass_stft_arr)):
bass_arr.append(librosa.amplitude_to_db(abs(bass_stft_arr[i])))
print('Conversion to dB done')
# Split data into test and train sets
mix_arr_test = [[]]
mix_arr_train = [[]]
for i in range(len(mix_arr)):
temp_test, temp_train = np.array_split(mix_arr[i], 2)
mix_arr_test.append(temp_test)
mix_arr_train.append(temp_train)
bass_arr_test = [[]]
bass_arr_train = [[]]
for i in range(len(bass_arr)):
temp_test, temp_train = np.array_split(bass_arr[i], 2)
bass_arr_test.append(temp_test)
bass_arr_train.append(temp_train)
print('Data split into test and train sets')
'''
# Normalize data
mix_arr_test = tf.keras.utils.to_categorical(mix_arr_test/(np.linalg.norm(mix_arr_test)))
mix_arr_train = tf.keras.utils.to_categorical(mix_arr_train/(np.linalg.norm(mix_arr_train)))
bass_arr_test = bass_arr_test/(np.linalg.norm(bass_arr_test))
bass_arr_train = bass_arr_train/(np.linalg.norm(bass_arr_train))
print('Data normalized')
print('Test:', mix_arr_test.shape)
print('Train:', mix_arr_train.shape)
'''
for i in range(len(mix_arr)):
mix_arr_test[i] = mix_arr_test[i]/(np.linalg.norm(mix_arr_test[i]))
mix_arr_train[i] = mix_arr_train[i]/(np.linalg.norm(mix_arr_train[i]))
for i in range(len(bass_arr)):
bass_arr_test[i] = bass_arr_test[i]/(np.linalg.norm(bass_arr_test[i]))
bass_arr_train[i] = bass_arr_train[i]/(np.linalg.norm(bass_arr_train[i]))
print('Data normalized')
# Set up data as np arrays
mix_arr_test = np.asarray(mix_arr_test, dtype=object)
mix_arr_train = np.asarray(mix_arr_train, dtype=object)
bass_arr_test = np.asarray(bass_arr_test, dtype=object)
bass_arr_train = np.asarray(bass_arr_train, dtype=object)
print('Data converted to Numpy arrays\n\n\n')
print('Test Shape:', mix_arr_test[1].shape)
print('Train Shape:', mix_arr_train[1].shape)
print(mix_arr_test)
# Set early stopping
earlystopping = callbacks.EarlyStopping(monitor ="accuracy",
mode ="min", patience = 5,
restore_best_weights = True)
print("Creating Model")
model = tf.keras.Sequential([
tf.keras.layers.Conv2D(4, (3,3), input_shape=[512,1292,1]),
tf.keras.layers.MaxPool2D(pool_size=(2,2)),
tf.keras.layers.Conv2D(4,(3,3), input_shape=[255,645,4]),
tf.keras.layers.MaxPool2D(pool_size=(2,2)),
tf.keras.layers.Conv2D(4,(3,3), input_shape=[126,321,4]),
tf.keras.layers.MaxPool2D(pool_size=(2,2)),
tf.keras.layers.Flatten(),
tf.keras.layers.Dense(1, activation=tf.nn.sigmoid)
])
model.summary()
# Compile model with stocastic gradient descent
model.compile(optimizer="Adam", loss="binary_crossentropy", metrics=["mae", "accuracy"])
print(mix_arr_train[1].shape)
print(mix_arr_train[2].shape)
print(mix_arr_train[3].shape)
print(mix_arr_train[4].shape)
tf.config.run_functions_eagerly(True)
hist = model.fit(np.ndarray.tolist(mix_arr_train)[1:], np.ndarray.tolist(bass_arr_train)[1:],
batch_size = 1,
epochs = 100,
verbose = 1,
validation_data = (np.ndarray.tolist(mix_arr_test)[1:], np.ndarray.tolist(bass_arr_test)[1:]),
callbacks=[earlystopping]
)
# Printing the accuracy
model_test = model.evaluate(mix_arr_test[1], mix_arr_test[1], verbose=2)
print(f" Model mse, mae and accuracy: {model_test}")
TrackPredictionMask=model.predict(mix_arr_test[1])
print("Pred", TrackPredictionMask)
print("Og", mix_arr_test[1])
print("Masked Test", (mix_arr_test[1]*TrackPredictionMask))