-
Notifications
You must be signed in to change notification settings - Fork 8
/
classify.py
361 lines (254 loc) · 10 KB
/
classify.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
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
import librosa
import librosa.display
import os, sys, cv2, matplotlib.pyplot as plt, numpy as np, shutil, itertools, pickle, pandas as pd, seaborn as sn, math, time
from random import seed, random, randint
from scipy.spatial import distance
import random
from sklearn.metrics import confusion_matrix
from sklearn.preprocessing import normalize
from mpl_toolkits.mplot3d import Axes3D
import soundfile as sf
from tqdm import tqdm
from keras import backend as K
from keras.models import Model, load_model, Sequential
from keras.callbacks import ModelCheckpoint
from sklearn.decomposition import PCA
from sklearn.model_selection import train_test_split
from sklearn.metrics import confusion_matrix
from generate_triple_combinations import generate_pairs
from model import get_model
from scipy import spatial
# Read metadata file
metadata = pd.read_csv(os.getcwd()+"/BritishBirdSongDataset/birdsong_metadata.csv")
header = list(metadata.head())
# Get bird names
bird_name = metadata['english_cname'].values
u, f = np.unique(bird_name, return_counts=True)
# Take 5 unique bird names for this project
uniq_birds = list(u[4:10]) + list(u[12:15])
data_train = []
data_test = []
y_train = []
y_test = []
bird_name_dict = {}
# Get file_id corresponding to bird names
for i in range(len(uniq_birds)) :
df = metadata[metadata['english_cname'] == uniq_birds[i]]
df = df['file_id'].values
df = df.tolist()
data_train.append(df[0])
y_train.append(i)
bird_name_dict[i] = uniq_birds[i]
data_test += df[1:]
y_test += [i] * (len(df) - 1)
# Read audio files using librosa. Divide each audio clip into frames of 2 sec duration. They are more than 40 sec long clips. So, using frames
# we will get a lot of data.
# Read training data and split into frames
frames_train = []
frames_test = []
frame_len = 22050*2 # equivalent of 2 seconds
y_frames_train = []
y_frames_test = []
for i in tqdm(range(len(data_train))) :
# Read audio
curr = data_train[i]
curr = os.getcwd() + "/BritishBirdSongDataset/songs/songs/xc" + str(curr) + ".flac"
y, sr = librosa.load(curr)
# Normalize time series
y = ((y-np.amin(y))*2)/(np.amax(y) - np.amin(y)) - 1
# Remove silence from the audio
org_len = len(y)
intervals = librosa.effects.split(y, top_db= 15, ref= np.max)
intervals = intervals.tolist()
y = (y.flatten()).tolist()
nonsilent_y = []
for p,q in intervals :
nonsilent_y = nonsilent_y + y[p:q+1]
y = np.array(nonsilent_y)
final_len = len(y)
sil = org_len - final_len
# Divide audio into frames
start = 0
end = frame_len
for j in range(0, len(y), int(frame_len*0.5)) :
frame = y[j:j+frame_len]
if len(frame) < frame_len :
frame = frame.tolist() + [0]* (frame_len-len(frame))
frame = np.array(frame)
S = np.abs(librosa.stft(frame, n_fft=512))
freqs = librosa.fft_frequencies(sr=sr, n_fft=512)
upper = ([x for x in range(len(freqs)) if freqs[x] >= 8000])[0]
lower = ([x for x in range(len(freqs)) if freqs[x] <= 1000])[-1]
freqs = freqs[lower:upper]
S = S[lower:upper,:]
if S.shape != (163, 345) :
print(S.shape)
assert S.shape == (163, 345)
frames_train.append(S)
y_frames_train.append(y_train[i])
# Read testing data and split into frames
for i in tqdm(range(len(data_test))) :
# Read audio
curr = data_test[i]
curr = os.getcwd() + "/BritishBirdSongDataset/songs/songs/xc" + str(curr) + ".flac"
y, sr = librosa.load(curr)
dur = librosa.get_duration(y, sr)
# Normalize time series
y = ((y-np.amin(y))*2)/(np.amax(y) - np.amin(y)) - 1
# Remove silence from the audio
org_len = len(y)
intervals = librosa.effects.split(y, top_db= 15, ref= np.max)
intervals = intervals.tolist()
y = (y.flatten()).tolist()
nonsilent_y = []
for p,q in intervals :
nonsilent_y = nonsilent_y + y[p:q+1]
y = np.array(nonsilent_y)
final_len = len(y)
sil = org_len - final_len
dur = librosa.get_duration(y, sr)
start = 0
end = frame_len
for j in range(0, len(y), int(frame_len*0.5)) :
frame = y[j:j+frame_len]
if len(frame) < frame_len :
frame = frame.tolist() + [0]* (frame_len-len(frame))
frame = np.array(frame)
S = np.abs(librosa.stft(frame, n_fft=512))
freqs = librosa.fft_frequencies(sr=sr, n_fft=512)
upper = ([x for x in range(len(freqs)) if freqs[x] >= 8000])[0]
lower = ([x for x in range(len(freqs)) if freqs[x] <= 1000])[-1]
freqs = freqs[lower:upper]
S = S[lower:upper,:]
assert S.shape == (163, 345)
frames_test.append(S)
y_frames_test.append(y_test[i])
print("Training samples : ", len(frames_train), len(y_frames_train), np.unique(y_frames_train, return_counts= True))
print("Testing samples : ", len(frames_test), len(y_frames_test), np.unique(y_frames_test, return_counts= True))
# Convert all data to nd array
y_frames_train = np.array(y_frames_train)
y_frames_test = np.array(y_frames_test)
r,c = frames_train[0].shape
frames_train = np.array(frames_train)
frames_train = frames_train.reshape((len(frames_train), r, c))
frames_test = np.array(frames_test)
frames_test = frames_test.reshape((len(frames_test), r, c))
dataX = np.concatenate((frames_train, frames_test), axis=0)
datay = np.concatenate((y_frames_train, y_frames_test), axis=0)
frames_train, frames_test, y_frames_train, y_frames_test = train_test_split(dataX, datay, test_size=0.6)
# Write training and testing data into a pickle file
f = open(os.getcwd() + "/training_frames1.pkl", 'wb')
pickle.dump([frames_train, y_frames_train], f)
f.close()
f = open(os.getcwd() + "/testing_frames1.pkl", 'wb')
pickle.dump([frames_test, y_frames_test], f)
f.close()
# Read training and testing data from the pickle file
f = open(os.getcwd() + "/training_frames1.pkl", 'rb')
frames_train, y_frames_train = pickle.load(f)
f.close()
f = open(os.getcwd() + "/testing_frames1.pkl", 'rb')
frames_test, y_frames_test = pickle.load(f)
f.close()
# Standardize the data
mu = frames_train.mean()
std = frames_train.std()
frames_train = (frames_train-mu)/std
frames_test = (frames_test-mu)/std
print("Training samples : ", frames_train.shape, len(y_frames_train), np.unique(y_frames_train, return_counts= True))
print("Testing samples : ", frames_test.shape, len(y_frames_test), np.unique(y_frames_test, return_counts= True))
# There are imbalanced classes. Repeat the data so that all classes have same number of samples.
u, f = np.unique(y_frames_train, return_counts= True)
frames_train1 = []
y_frames_train1 = []
maximum = max(f)
count = 0
for i in u :
ind, = np.where(y_frames_train == i)
ind = ind.tolist()
while len(ind) < maximum :
ind = ind + ind
ind = ind[:maximum]
temp = frames_train[ind]
if count == 0 :
frames_train1 = temp
count += 1
else :
frames_train1 = np.concatenate((frames_train1, temp), axis= 0)
y_frames_train1 += [i] * maximum
y_frames_train1 = np.array(y_frames_train1)
# Make pairs for siamese network
anchor, pos, neg = generate_pairs(frames_train1, y_frames_train1, rand_samples= -1, pos_pair_size=1200)
anchor = anchor.astype(np.float16)
pos = pos.astype(np.float16)
neg = neg.astype(np.float16)
np.savez_compressed(os.getcwd()+'/training_siamese_frames', a=anchor, b= pos, c= neg)
data = np.load(os.getcwd()+'/training_siamese_frames.npz')
anchor = data['a']
pos = data['b']
neg = data['c']
print("Siamese pairs ", anchor.shape, pos.shape, neg.shape)
# Train the model
_,r,c = anchor.shape
encoder, model = get_model(c,r)
anchor = pos.transpose(0, 2, 1)
pos = pos.transpose(0, 2, 1)
neg = neg.transpose(0, 2, 1)
y = np.ones((len(anchor), 32*3))
# Fit model
mc = ModelCheckpoint(os.getcwd()+'/model_checkpoint.h5',
save_weights_only=False, save_freq=10)
# Train the model
# model.fit([anchor, pos, neg], y, epochs= 30, callbacks= [mc], batch_size= 256, verbose= 1)
# model.save(os.getcwd() + "/siamese.h5")
# encoder.save(os.getcwd() + "/encoder.h5")
# Load the model
model = load_model(os.getcwd() + "/siamese.h5", compile= False)
encoder = load_model(os.getcwd() + "/encoder.h5", compile= False)
# Create the confusion and similarity matrix
frames_test = frames_test.transpose(0,2,1)
frames_train1 = frames_train1.transpose(0,2,1)
sim_mat = np.zeros((len(frames_test), len(frames_test)))
ind = np.arange(0, len(frames_test), 1)
ind = [x for _,x in sorted(zip(y_frames_test, ind))]
frames_test = frames_test[ind]
y_frames_test = y_frames_test[ind]
encoded = encoder.predict(frames_test)
pred = []
for i in tqdm(range(len(frames_test))) :
curr = frames_test[i:i+1]
sim = encoder.predict(curr)
sim = [1 - spatial.distance.cosine(sim[0], encoded[x]) for x in range(len(encoded))]
sim = np.array(sim)
sim[sim<0] = 0
ind = np.copy(y_frames_test)
ind = [x for _,x in sorted(zip(sim, ind))]
ind = ind[::-1]
ind = ind[:21]
u, f = np.unique(ind, return_counts= True)
best = u[np.argmax(f)]
pred.append(best)
sim_mat[i] = sim
conf_mat = confusion_matrix(y_frames_test, pred, normalize= 'true')
print("Accuracy on testset : ", np.trace(conf_mat)/np.sum(conf_mat))
u = np.unique(y_frames_test)
# Plot confusion matrix
conf_mat = pd.DataFrame(conf_mat, columns= [bird_name_dict[x] for x in u], index= [bird_name_dict[x] for x in u])
plt.figure(figsize = (20,15))
sn.heatmap(conf_mat, annot=True, annot_kws={"size": 10}, cmap='jet')
plt.tick_params(labelsize=8)
plt.xticks(rotation= 60)
plt.title("Bird Song classification")
# plt.show()
plt.savefig(os.getcwd()+"/ConfusionMatrix_test.png")
plt.close()
# Plot similarity matrix
sim_mat = pd.DataFrame(sim_mat, columns= [bird_name_dict[x] for x in y_frames_test], index= [bird_name_dict[x] for x in y_frames_test])
plt.figure(figsize = (20,15))
sn.heatmap(sim_mat, annot=False, annot_kws={"size": 12}, cmap='viridis')
plt.tick_params(labelsize=10)
plt.xticks(rotation= 90)
plt.title("Bird Song classification")
# plt.show()
plt.savefig(os.getcwd()+"/SimilarityMatrix_test.png")
plt.close()