-
Notifications
You must be signed in to change notification settings - Fork 2
/
utils.py
210 lines (171 loc) · 8.6 KB
/
utils.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
'''
pytorch implementation Adrien Bitton
link: https://github.com/adrienchaton/PerceptualAudio_pytorch
paper codes Pranay Manocha
link: https://github.com/pranaymanocha/PerceptualAudio
'''
import torch
import numpy as np
import matplotlib
matplotlib.rcParams['agg.path.chunksize'] = 10000
matplotlib.use('Agg') # for the server
from matplotlib import pyplot as plt
from sklearn.metrics import classification_report,accuracy_score
###############################################################################
### misc.
def print_time(s_duration):
m,s = divmod(s_duration,60)
h, m = divmod(m, 60)
print('elapsed time = '+"%d:%02d:%02d" % (h, m, s))
###############################################################################
### data import
def import_data(data_path,subsets,Lsize,batch_size,train_ratio=0.8,rgains=False,sr=22050):
train_y0 = []
train_y1 = []
train_labels = []
test_y0 = []
test_y1 = []
test_labels = []
for subset in subsets:
print('loading '+subset)
data_dic = np.load(data_path+subset+'_data.npy',allow_pickle=True).item()
# one numpy dic per pre-processed subset of audio distortion
# each dic entry is [first signal, second signal, human label]
fcount = 0
for fid in data_dic:
y0 = data_dic[fid][0] # first signal
y1 = data_dic[fid][1] # second signal
lab = data_dic[fid][2] # human label
min_len = np.min([y0.shape[0],y1.shape[0]])
N = min_len//Lsize
if N>0:
if np.random.rand()>train_ratio:
test_y0.append(y0[:N*Lsize].reshape(N,Lsize))
test_y1.append(y1[:N*Lsize].reshape(N,Lsize))
test_labels.append(np.zeros((N,1),dtype='int')+lab)
else:
train_y0.append(y0[:N*Lsize].reshape(N,Lsize))
train_y1.append(y1[:N*Lsize].reshape(N,Lsize))
train_labels.append(np.zeros((N,1),dtype='int')+lab)
fcount+=1
print('paired files amount to ',fcount)
train_y0 = torch.from_numpy(np.concatenate(train_y0)).float()
train_y1 = torch.from_numpy(np.concatenate(train_y1)).float()
train_labels = torch.from_numpy(np.concatenate(train_labels)).long()
test_y0 = torch.from_numpy(np.concatenate(test_y0)).float()
test_y1 = torch.from_numpy(np.concatenate(test_y1)).float()
test_labels = torch.from_numpy(np.concatenate(test_labels)).long()
train_ones = float(torch.sum(train_labels).item())
test_ones = float(torch.sum(test_labels).item())
print('train/test Lsize pairs amount to ',train_y0.shape[0],test_y0.shape[0])
print('train/test labels == one ("different") are ',int(train_ones),int(test_ones))
print('train/test ratio of labels == one ("different") are ',train_ones/train_y0.shape[0],test_ones/test_y0.shape[0])
train_dataset = torch.utils.data.TensorDataset(train_y0,train_y1,train_labels)
train_loader = torch.utils.data.DataLoader(train_dataset,batch_size=batch_size,shuffle=True,drop_last=True)
# if applying random gains, train_loader is scaled at every forward
# but train_refloader should have a fixed pre-scaling that is consistent, as for test data
if rgains is False:
train_refloader = torch.utils.data.DataLoader(train_dataset,batch_size=batch_size,shuffle=False,drop_last=False)
test_dataset = torch.utils.data.TensorDataset(test_y0,test_y1,test_labels)
test_loader = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=True,drop_last=True)
test_refloader = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=False,drop_last=False)
else:
print('reference trainset and test data have fixed random pre-scaling')
g = torch.zeros(train_y0.shape[0])
torch.nn.init.uniform_(g,rgains[0],rgains[1])
train_y0_scaled = train_y0*g.unsqueeze(1)
train_y1_scaled = train_y1*g.unsqueeze(1)
train_dataset_scaled = torch.utils.data.TensorDataset(train_y0_scaled,train_y1_scaled,train_labels)
train_refloader = torch.utils.data.DataLoader(train_dataset_scaled,batch_size=batch_size,shuffle=False,drop_last=False)
g = torch.zeros(test_y0.shape[0])
torch.nn.init.uniform_(g,rgains[0],rgains[1])
test_y0_scaled = test_y0*g.unsqueeze(1)
test_y1_scaled = test_y1*g.unsqueeze(1)
test_dataset = torch.utils.data.TensorDataset(test_y0_scaled,test_y1_scaled,test_labels)
test_loader = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=True,drop_last=True)
test_refloader = torch.utils.data.DataLoader(test_dataset,batch_size=batch_size,shuffle=False,drop_last=False)
return train_loader,test_loader,train_refloader,test_refloader
###############################################################################
### evaluation functions
def loss_plot(plot_name,loss_log):
plt.figure(figsize=(12,8))
plt.suptitle('loss log, rows=train/test')
plt.subplot(2,1,1)
plt.plot(loss_log[:,0])
plt.subplot(2,1,2)
plt.plot(loss_log[:,1])
plt.savefig(plot_name+'.png',format='png')
plt.close()
def acc_plot(plot_name,epoch_log,train_acc_log,test_acc_log):
plt.figure(figsize=(12,8))
plt.suptitle('accuracy log, rows=train/test')
plt.subplot(2,1,1)
plt.plot(epoch_log,train_acc_log)
plt.subplot(2,1,2)
plt.plot(epoch_log,test_acc_log)
plt.savefig(plot_name+'.png',format='png')
plt.close()
def eval_scores(model,train_refloader,test_refloader,device,report=True):
train_pred = []
train_labels = []
train_dist = []
test_pred = []
test_labels = []
test_dist = []
with torch.no_grad():
train_loss = 0
for _,minibatch in enumerate(train_refloader):
try:
xref = minibatch[0].to(device)
xper = minibatch[1].to(device)
labels = minibatch[2].to(device)
loss,dist,class_pred,class_prob = model.forward(xref,xper,labels)
labels = labels.squeeze()
train_pred.append(class_pred.cpu().numpy())
train_labels.append(labels.cpu().numpy())
train_loss += loss.item()
train_dist.append(dist.squeeze().cpu().numpy())
except:
print('passing one minibatch of evaluate train_refloader')
train_loss /= len(train_pred)
# loss is averaged in the minibatch "(reduction='mean')", then divided by the number of minibatches
test_loss = 0
for _,minibatch in enumerate(test_refloader):
try:
xref = minibatch[0].to(device)
xper = minibatch[1].to(device)
labels = minibatch[2].to(device)
loss,dist,class_pred,class_prob = model.forward(xref,xper,labels)
labels = labels.squeeze()
test_pred.append(class_pred.cpu().numpy())
test_labels.append(labels.cpu().numpy())
test_loss += loss.item()
test_dist.append(dist.squeeze().cpu().numpy())
except:
print('passing one minibatch of evaluate test_refloader')
test_loss /= len(test_pred)
train_pred = np.concatenate(train_pred)
train_labels = np.concatenate(train_labels)
test_pred = np.concatenate(test_pred)
test_labels = np.concatenate(test_labels)
train_dist = np.concatenate(train_dist)
test_dist = np.concatenate(test_dist)
train_dist_0 = np.mean(train_dist[np.where(train_labels==0)])
train_dist_1 = np.mean(train_dist[np.where(train_labels==1)])
test_dist_0 = np.mean(test_dist[np.where(test_labels==0)])
test_dist_1 = np.mean(test_dist[np.where(test_labels==1)])
if report is True:
print('TRAINING SET')
print('average training loss = ',train_loss)
print(classification_report(train_labels, train_pred, labels=[0,1], target_names=['same','different']))
train_acc = accuracy_score(train_labels, train_pred)
print('average training accuracy = ',train_acc)
print('average distance for train groudtruth 0,1 = ',train_dist_0,train_dist_1)
if report is True:
print('TEST SET')
print('average test loss = ',test_loss)
print(classification_report(test_labels, test_pred, labels=[0,1], target_names=['same','different']))
test_acc = accuracy_score(test_labels, test_pred)
print('average test accuracy = ',test_acc)
print('average distance for test groudtruth 0,1 = ',test_dist_0,test_dist_1)
return train_acc,test_acc,train_loss,test_loss