File size: 11,849 Bytes
a4d9f29
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
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
import torch
from torch.utils.data import Dataset,DataLoader
import torch.nn as nn
import torch.nn.functional as F
import torch.optim as optim
import numpy as np
import datetime

class AugmentationParams():
    def __init__(self,tshift_max=0, tshift_prob=0,
                      tmask_min=0, tmask_max=0, tmask_prob=0,
                      fshift_max=0, fshift_prob=0,
                      fmask_min=0, fmask_max=0, fmask_prob=0,
                      scale_min=1, scale_max=1, scale_prob=0, 
                      fgmix_weight_min=0,fgmix_weight_max=0, fgmix_prob=0,
                      bgmix_weight_min=0,bgmix_weight_max=0, bgmix_prob=0,
                      pixelnoise_rate_min=0, pixelnoise_rate_max=0, pixelnoise_intensity_min=0, pixelnoise_intensity_max=0, pixelnoise_prob=0):
        self.tshift_max = tshift_max
        self.tshift_prob = tshift_prob
        self.tmask_min = tmask_min
        self.tmask_max = tmask_max	
        self.tmask_prob = tmask_prob
        self.fshift_max = fshift_max	
        self.fshift_prob = fshift_prob
        self.fmask_min = fmask_min
        self.fmask_max = fmask_max        
        self.fmask_prob = fmask_prob
        self.scale_min = scale_min
        self.scale_max = scale_max
        self.scale_prob = scale_prob
        self.fgmix_weight_max = fgmix_weight_max
        self.fgmix_weight_min = fgmix_weight_min
        self.fgmix_prob = fgmix_prob
        self.bgmix_weight_max = bgmix_weight_max
        self.bgmix_weight_min = bgmix_weight_min
        self.bgmix_prob = bgmix_prob
        self.pixelnoise_rate_min = pixelnoise_rate_min
        self.pixelnoise_rate_max = pixelnoise_rate_max
        self.pixelnoise_intensity_min = pixelnoise_intensity_min
        self.pixelnoise_intensity_max = pixelnoise_intensity_max        
        self.pixelnoise_prob = pixelnoise_prob


def augment(img1, ap):
    """data augmentation
       img1: original image
       ap: AugmentationParams
    """

    img = np.copy(img1)
    ntime, nfreq = img.shape
    
    if ap.tshift_prob > np.random.uniform():
        tshift = np.random.randint(low=0, high=ap.tshift_max+1)
        img = np.concatenate((img[tshift:,:], img[:tshift,:]),axis=0)
	
    if ap.fshift_prob > np.random.uniform():
        fshift = np.random.randint(low=-ap.fshift_max, high=ap.fshift_max+1)
        if fshift < 0:
	    # shift down
            fshift2 = -fshift
            img = np.concatenate((img[:,fshift2:], np.repeat(img[:,nfreq-1],fshift2).reshape((ntime,fshift2))),axis=1)
        elif fshift > 0:
	    # shift up
            img = np.concatenate((np.repeat(img[:,0],fshift).reshape((ntime,fshift)), img[:,:-fshift]),axis=1)
	    
    if ap.scale_prob > np.random.uniform():
        # scale max between scale_min and scale_max
        origmax = np.max(img)
        w = np.random.uniform(low=ap.scale_min, high=ap.scale_max)
        img *= w/origmax
            
    if ap.tmask_prob > np.random.uniform():
        tpos = np.random.randint(low=0, high=ntime)
        twidth = np.random.randint(low=ap.tmask_min, high=ap.tmask_max)
        tpos2=min(tpos+twidth, ntime)
        img[tpos:tpos2,:] = 0
       	
    if ap.fmask_prob > np.random.uniform():
        fpos = np.random.randint(low=0, high=nfreq)
        fwidth = np.random.randint(low=ap.fmask_min, high=ap.fmask_max)
        fpos2=min(fpos+fwidth, nfreq)
        img[:,fpos:fpos2] = 0

    if ap.pixelnoise_prob > np.random.uniform():
        r = np.random.uniform(low=ap.pixelnoise_rate_min, high=ap.pixelnoise_rate_max)
        nn = np.int(ntime * nfreq * r)
        ii = np.random.randint(0,ntime, size=nn)
        jj = np.random.randint(0,nfreq, size=nn)
        img[ii,jj] += np.random.uniform(low=ap.pixelnoise_intensity_min, high=ap.pixelnoise_intensity_max, size=nn)

    return img

def normalize(data):
    return (data-np.mean(data))/np.std(data)

class Dataset(Dataset):
    """ data and labels in input and numpy arrays, they are converted into tensors 
        n1: number of fg samples, n2: number of bg samples to train clean bg class, n3: number of bg samples only to be mixed with fg
    """
    def __init__(self, data, labels, n1,n2=0,n3=0, ntime=512, ap=None, eps=0.01, nclasses=22):
        self.data = data
        # label is float because BCEWithLogitsLoss supports labels that are probabilities
        self.labels = torch.nn.functional.one_hot(torch.from_numpy(labels),num_classes=nclasses).float()
        self.labels = torch.clamp(self.labels, min=eps, max=1.0-eps)
        self.ap = ap
        self.n1 = n1;
        self.n2 = n2;
        self.n3 = n3;
        self.ntime=ntime
        
    def __len__(self):
        return self.n1+self.n2+self.n3

    def __getitem__(self, idx):
        img = self.data[idx]
        lab = self.labels[idx]

        if self.ap:
            img = augment(img, self.ap)
            # sample 2nd image only when it is needed...
            if idx < self.n1 and self.ap.fgmix_prob > np.random.uniform():
                # mix only foreground species
                idx2 = np.random.randint(low=0, high=self.n1)
                img2 = self.data[idx2]
                img2 = augment(img2, self.ap)
                lab2 = self.labels[idx2]    
                #w = np.random.uniform(low=self.ap.fgmix_weight_min, high=self.ap.fgmix_weight_max)
                #img = (1-w) * img + w * img2
                #img = np.maximum(img, w*img2)
                img = np.maximum(img, img2)
                lab = torch.maximum(lab, lab2)
                
            if idx < self.n1 and self.ap.bgmix_prob > np.random.uniform():
                # mix with background
                idx2 = np.random.randint(low=self.n1, high=self.n1+self.n2)
                img2 = self.data[idx2]
                img2 = augment(img2, self.ap)
                lab2 = self.labels[idx2]    
                #w = np.random.uniform(low=self.ap.bgmix_weight_min, high=self.ap.bgmix_weight_max)
                #img = (1-w) * img + w * img2
                #img = np.maximum(img, w*img2)
                img = np.maximum(img, img2)
                lab = torch.maximum(lab, lab2)
            #img=normalize(img)
            
        img = torch.from_numpy(img[:self.ntime]).unsqueeze(0)
        return img, lab


class Net(nn.Module):

    def __init__(self, ntime=512, nfreq=128, nclasses=22):
        super(Net, self).__init__()
        # 1 input 100x64 image channel, 32 output channels, 3x3 square convolution kernel
        self.conv1 = nn.Conv2d(  1,  32, 3, padding=1)
        self.conv2 = nn.Conv2d( 32,  64, 3, padding=1)
        self.conv3 = nn.Conv2d( 64,  128, 3, padding=1)
        self.conv4 = nn.Conv2d( 128, 256, 3, padding=1)
        self.conv5 = nn.Conv2d( 256, 512, 3, padding=1)
        self.conv6 = nn.Conv2d( 512, 512, 3, padding=1)
        
        # image dimension after maxpool layers        
        n_maxpool = 6
        nt = ntime
        nr = nfreq
        for i in range(n_maxpool):
            nt //= 2
            nr //= 2
        nr *= 4    
            
        self.fc1 = nn.Linear(512 * nt * nr, 512)
        self.fc2 = nn.Linear(512, 128)
        self.fc3 = nn.Linear(128, nclasses)
        
    def forward(self, x):
        # Max pooling over a (2, 2) window
        x = F.max_pool2d(F.relu(self.conv1(x)), (2, 2))
        # If the size is a square, you can specify with a single number, default stride is kernel size
        x = F.max_pool2d(F.relu(self.conv2(x)), 2)
        x = F.max_pool2d(F.relu(self.conv3(x)), 2)
        # x = F.max_pool2d(F.relu(self.conv4(x)), (8,2))
        x = F.max_pool2d(F.relu(self.conv4(x)), (2,2))
        x = F.max_pool2d(F.relu(self.conv5(x)), (2,1))
        x = F.max_pool2d(F.relu(self.conv6(x)), (2,1))
        
        x = torch.flatten(x, 1) # flatten all dimensions except the batch dimension, input now is 8*8 image (64*512 filter outputs)
        x = F.relu(self.fc1(x))
        x = F.relu(self.fc2(x))
        x = self.fc3(x)
        return x

def train(net, loss_fn, optimizer, train_dataloader, validation_dataloader, device, nepochs=10, info=0, model_outfile=None):
    """ returns accuracy for train and validation data from each training epoch 
        if model_outfile defined, save the best model
    """
    acc_train=np.zeros(nepochs)
    acc_valid=np.zeros(nepochs)
    bestval=0
    for epoch in range(nepochs):
        net.train()
        running_loss = 0.0
        for i, data in enumerate(train_dataloader, 0):
            inputs, labels = data
            inputs = inputs.to(device=device)
            labels = labels.to(device=device)
            outputs = net(inputs)
            loss = loss_fn(outputs, labels)
            
            optimizer.zero_grad()
            loss.backward()
            optimizer.step()

            if info>1:
                running_loss += loss.item()
                if i % 100 == 99:
                    print('[%d, %5d] train_loss: %.3f' % (epoch + 1, i + 1, running_loss / 100))
                    running_loss = 0.0

        net.eval()        
        train_accuracy = validate(net, train_dataloader, device)
        validation_accuracy = validate(net, validation_dataloader, device)
        acc_train[epoch] = train_accuracy
        acc_valid[epoch] = validation_accuracy
        if info:
            print(f'{datetime.datetime.now()} epoch {epoch}, train_accuracy: {train_accuracy:.3f} validation_accuracy: {validation_accuracy:.3f}')

        if validation_accuracy > bestval:
            bestval = validation_accuracy
            if model_outfile:
                torch.save(net.state_dict(), model_outfile)
    return acc_train, acc_valid

    
def validate1(net, dataloader, device):
    """ assumes only one given label """
    correct=0
    total=0
    net.eval()
    with torch.no_grad():
        for inputs, labels in dataloader:
            inputs = inputs.to(device=device)
            labels = labels.to(device=device)
            outputs = net(inputs)
            predicted = torch.argmax(outputs, 1)
            given_labels = torch.argmax(labels, 1)
            total += labels.shape[0]
            correct += int((predicted == given_labels).sum())

    return correct/total


def validate(net, dataloader, device): 
    """ allows multi-labeling """
    correct = 0.0
    total = 0
    net.eval()
    with torch.no_grad():
        for inputs, labels in dataloader:
            inputs = inputs.to(device=device)
            labels = labels.to(device=device)
            outputs = net(inputs)
            for i,lab in enumerate(labels):
                target = torch.where(lab > 0.5, 1, 0)
                ntarget = target.sum()
                out, ind = torch.sort(outputs[i], descending=True)
                correct += target[ind[:ntarget]].sum() / ntarget
            total += labels.shape[0]

    return correct/total


def classify(net, dataloader, device, nclasses=22):
    """ compute logits using dataloader """
    n=len(dataloader.dataset)
    out=np.zeros((n,nclasses))
    i1=0
    net.eval()
    with torch.no_grad():
        for inputs, labels in dataloader:
            inputs = inputs.to(device=device)
            outputs = net(inputs)
            i2=i1+len(outputs)
            out[i1:i2] = outputs.detach().numpy()
            i1=i2
    return out

def classify1_cpu(dat, net, nclasses=22):
    """ compute logits for data matrix using for loop (cpu only) """
    net.eval()
    n=len(dat)
    out=np.zeros((n,nclasses))
    for i in range(n):
       out[i]=net(torch.unsqueeze(torch.unsqueeze(torch.from_numpy(dat[i]),0),0)).detach().numpy()
    return out

def classify_cpu(dat, net):
    """ compute logits for entire data matrix (cpu only) """
    net.eval()
    # add dimension for number of channels (1) so that tensor is [num_segments num_channels ntime nfreq]
    out = net(torch.unsqueeze(torch.from_numpy(dat),1)).detach().numpy()
    return out