/
mlp.py
155 lines (106 loc) · 3.65 KB
/
mlp.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
from tools.autoencoder_model import *
from keras.models import Sequential
from keras.layers.core import Dense, Activation, Dropout
from keras.optimizers import Adam, SGD
# from IPython.display import clear_output
from tqdm import *
import sys
# sys.path.append('../python')
from tools.data import Corpus, History
from sklearn.metrics import log_loss
train=Corpus('data/TIMIT_train.hdf5',load_normalized=True)
dev=Corpus('data/TIMIT_dev.hdf5',load_normalized=True)
# test=Corpus('../data/TIMIT_test.hdf5',load_normalized=True)
enc='mfcc'
if enc=='mfcc':
tr_in,tr_out_dec=train.get()
dev_in,dev_out_dec=dev.get()
else:
enc='encoded'
tr_in,tr_out_dec=train.get_enc()
dev_in,dev_out_dec=dev.get_enc()
# tst_in,tst_out_dec=test.get()
print tr_in.shape
print tr_in[0].shape
print tr_out_dec.shape
print tr_out_dec[0].shape
for u in range(tr_in.shape[0]):
tr_in[u]=tr_in[u][:,:26]
for u in range(dev_in.shape[0]):
dev_in[u]=dev_in[u][:,:26]
# for u in range(tst_in.shape[0]):
# tst_in[u]=tst_in[u][:,:26]
input_dim=tr_in[0].shape[1]
output_dim=61
hidden_num=512
epoch_num=3
def dec2onehot(dec):
ret=[]
for u in dec:
assert np.all(u<output_dim)
num=u.shape[0]
r=np.zeros((num,output_dim))
r[range(0,num),u]=1
ret.append(r)
return np.array(ret)
tr_out=dec2onehot(tr_out_dec)
dev_out=dec2onehot(dev_out_dec)
# tst_out=dec2onehot(tst_out_dec)
model = Sequential()
model.add(Dense(input_dim=input_dim,output_dim=hidden_num))
model.add(Activation('tanh'))
model.add(Dropout(0.3))
model.add(Dense(output_dim=hidden_num))
model.add(Activation('tanh'))
model.add(Dense(output_dim=output_dim))
model.add(Activation('softmax'))
# optimizer= SGD(lr=0.003,momentum=0.9,nesterov=True)
optimizer = Adam()
loss='categorical_crossentropy'
metrics=['accuracy']
model.compile(loss=loss, optimizer=optimizer, metrics=metrics)
from random import shuffle
tr_hist=History('Train')
dev_hist=History('Dev')
tst_hist=History('Test')
tr_it=range(tr_in.shape[0])
for e in range(epoch_num):
print 'Epoch #{}/{}'.format(e+1,epoch_num)
sys.stdout.flush()
shuffle(tr_it)
for u in tqdm(tr_it):
l,a=model.train_on_batch(tr_in[u],tr_out[u])
tr_hist.r.addLA(l,a,tr_out[u].shape[0])
# clear_output()
tr_hist.log()
for u in range(dev_in.shape[0]):
l,a=model.test_on_batch(dev_in[u],dev_out[u])
dev_hist.r.addLA(l,a,dev_out[u].shape[0])
dev_hist.log()
# for u in range(tst_in.shape[0]):
# l,a=model.test_on_batch(tst_in[u],tst_out[u])
# tst_hist.r.addLA(l,a,tst_out[u].shape[0])
# tst_hist.log()
pickle.dump(model, open('models/classifier_enc.pkl','wb'))
pickle.dump(dev_hist, open('models/testHist_enc.pkl','wb'))
pickle.dump(tr_hist, open('models/trainHist_enc.pkl','wb'))
import matplotlib.pyplot as P
fig,ax=P.subplots(2,sharex=True,figsize=(12,10))
ax[0].set_title('Loss')
ax[0].plot(tr_hist.loss,label='Train')
ax[0].plot(dev_hist.loss,label='Dev')
# ax[0].plot(tst_hist.loss,label='Test')
ax[0].legend()
ax[0].set_ylim((1.4,2.0))
ax[1].set_title('PER %')
ax[1].plot(100*(1-np.array(tr_hist.acc)),label='Train')
ax[1].plot(100*(1-np.array(dev_hist.acc)),label='Dev')
# ax[1].plot(100*(1-np.array(tst_hist.acc)),label='Test')
ax[1].legend()
ax[1].set_ylim((45,55))
fig.savefig('train_%s_.png' % (enc))
print 'Min train PER: {:%}'.format(1-np.max(tr_hist.acc))
# print 'Min test PER: {:%}'.format(1-np.max(tst_hist.acc))
print 'Min dev PER epoch: #{}'.format((np.argmax(dev_hist.acc)+1))
# print 'Test PER on min dev: {:%}'.format(1-tst_hist.acc[np.argmax(dev_hist.acc)])
######################################################################################