-
Notifications
You must be signed in to change notification settings - Fork 33
/
metric_use.py
181 lines (126 loc) · 6.2 KB
/
metric_use.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
import numpy as np
import numpy
import os
from tqdm import tqdm
import numpy as np
from tqdm import tqdm
from scipy.io import wavfile
import os, csv
import tensorflow as tf
import pickle
import numpy as np
import argparse
from network_model import *
from helper import *
import argparse
import librosa
def argument_parser():
"""
Get an argument parser for a training script.
"""
parser = argparse.ArgumentParser(formatter_class=argparse.ArgumentDefaultsHelpFormatter)
parser.add_argument('--saved_name', help='name of the file saved as a pickle file',default='m_example',type=str)
parser.add_argument('--modfolder', help='path of the saved model to be used for inference', default='../pre-model/scratch_loss',type=str)
parser.add_argument('--type', help='pretrained/linear/finetune/scratch', default='scratch',type=str)
return parser
args = argument_parser().parse_args()
#sample loading the audio files in a dictionary. '4.wav is perceptually farther than 2.ref from reference.'
def load_full_data_list(datafolder='dataset'): # check change path names
dataset={}
print("Loading Files....")
dataset['all']={}
dataset['all']['inname']=[]
dataset['all']['outname']=[]
dataset['all']['inname'].append('../sample_audio/ref.wav')
dataset['all']['outname'].append('../sample_audio/4.wav')
dataset['all']['inname'].append('../sample_audio/ref.wav')
dataset['all']['outname'].append('../sample_audio/2.wav')
return dataset
def load_full_data(dataset):
dataset['all']['inaudio'] = [None]*len(dataset['all']['inname'])
dataset['all']['outaudio'] = [None]*len(dataset['all']['outname'])
for id in tqdm(range(len(dataset['all']['inname']))):
if dataset['all']['inaudio'][id] is None:
inputData, sr = librosa.load(dataset['all']['inname'][id],sr=22050)
outputData, sr = librosa.load(dataset['all']['outname'][id],sr=22050)
## convert to 16 bit floating point
inputData = np.round(inputData.astype(np.float)*32768)
outputData = np.round(outputData.astype(np.float)*32768)
inputData_wav = np.reshape(inputData, [-1, 1])
outputData_wav = np.reshape(outputData, [-1, 1])
shape_wav = np.shape(inputData_wav)
shape_wav1 = np.shape(outputData_wav)
inputData_wav = np.reshape(inputData_wav, [1, 1,shape_wav[0], shape_wav[1]])
outputData_wav = np.reshape(outputData_wav, [1, 1,shape_wav1[0], shape_wav1[1]])
inputData_wav = np.float32(inputData_wav)
outputData_wav = np.float32(outputData_wav)
dataset['all']['inaudio'][id] = inputData_wav
dataset['all']['outaudio'][id] = outputData_wav
return dataset
######### Data loading
dataset=load_full_data_list()
dataset=load_full_data(dataset)
######### Parameters of the model
#MAKE SURE THAT YOU UPDATE THESE PARAMETERS IF YOU MAKE ANY CHANGES TO THE MODEL.
#################################
SE_LAYERS = 13 # NUMBER OF INTERNAL LAYERS
SE_CHANNELS = 64 # NUMBER OF FEATURE CHANNELS PER LAYER
SE_LOSS_LAYERS = 14 # NUMBER OF FEATURE LOSS LAYERS
SE_NORM = "NM" # TYPE OF LAYER NORMALIZATION (NM, SBN or None)
SE_LOSS_TYPE = "FL" # TYPE OF TRAINING LOSS (L1, L2 or FL)
# FEATURE LOSS NETWORK
LOSS_LAYERS = 14 # NUMBER OF INTERNAL LAYERS
LOSS_BASE_CHANNELS = 32 # NUMBER OF FEATURE CHANNELS PER LAYER IN FIRT LAYER
LOSS_BLK_CHANNELS = 5 # NUMBER OF LAYERS BETWEEN CHANNEL NUMBER UPDATES
LOSS_NORM = 'SBN' # TYPE OF LAYER NORMALIZATION (NM, SBN or None)
SET_WEIGHT_EPOCH = 10 # NUMBER OF EPOCHS BEFORE FEATURE LOSS BALANCE
SAVE_EPOCHS = 10 # NUMBER OF EPOCHS BETWEEN MODEL SAVES
FILTER_SIZE=3
def load_full_data(dataset,sets,id_value):
noisy=dataset[sets]['inaudio'][id_value]
clean=dataset['all']['outaudio'][id_value]
clean=np.reshape(clean,[clean.shape[2]])
noisy=np.reshape(noisy,[noisy.shape[2]])
shape1=clean.shape[0]
shape2=noisy.shape[0]
if shape1>shape2:
difference=shape1-shape2
a=(np.zeros(difference))
noisy=np.append(a,noisy,axis=0)
elif shape1<shape2:
difference=shape2-shape1
a=(np.zeros(difference))
clean=np.append(a,clean,axis=0)
clean=np.reshape(clean,[1,1,clean.shape[0],1])
noisy=np.reshape(noisy,[1,1,noisy.shape[0],1])
return [clean,noisy]
def model_run():
modfolder= args.modfolder
tf.reset_default_graph()
with tf.variable_scope(tf.get_variable_scope()):
input1_wav=tf.placeholder(tf.float32,shape=[None, None, None,1])
clean1_wav=tf.placeholder(tf.float32,shape=[None, None, None,1])
keep_prob = tf.placeholder_with_default(1.0, shape=())
if args.type!='pretrained':
others,loss_sum = featureloss(input1_wav,clean1_wav,keep_prob,loss_layers=SE_LOSS_LAYERS,n_layers=LOSS_LAYERS, norm_type=LOSS_NORM, base_channels=LOSS_BASE_CHANNELS,blk_channels=LOSS_BLK_CHANNELS,ksz=FILTER_SIZE)
else:
others,loss_sum = featureloss_pretrained(input1_wav,clean1_wav,keep_prob,loss_layers=SE_LOSS_LAYERS,n_layers=LOSS_LAYERS, norm_type=LOSS_NORM, base_channels=LOSS_BASE_CHANNELS,blk_channels=LOSS_BLK_CHANNELS,ksz=FILTER_SIZE)
distance = loss_sum
distance_overall=[]
with tf.Session() as sess:
#sess.run(tf.global_variables_initializer())
loss_saver = tf.train.Saver([var for var in tf.trainable_variables()])
if args.type=='pretrained':
loss_saver.restore(sess, "%s/loss_model.ckpt" % modfolder)
else:
loss_saver.restore(sess, "%s/my_test_model" % modfolder)
for j in tqdm(range(len(dataset['all']['inname']))):
wav_in,wav_out=load_full_data(dataset,'all',j)
dist= sess.run([distance],feed_dict={input1_wav:wav_out, clean1_wav:wav_in})
distance_overall.append(dist)
return [distance_overall]
distance=[]
distance_overall=model_run()
distance.append(distance_overall)
with open('saved_distances/'+str(args.saved_name)+'.p', 'wb') as f:
pickle.dump(distance, f)