-
Notifications
You must be signed in to change notification settings - Fork 9
/
Copy pathmain.py
206 lines (164 loc) · 9.22 KB
/
main.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
import os
import auxil
import argparse
import numpy as np
import dense_suppli as spp
import dense_net3D_SA as nt
from keras.optimizers import Adam
from keras.utils.np_utils import to_categorical
from keras.layers import Input
from keras.models import Model
def load_hyper(args):
if args.dataset not in ["UH", "DIP", "DUP"]:
acrstate = None if args.random_state == None else args.random_state+args.idtest
pixelsO, labelsO, numclass = auxil.loadData(args.dataset, num_components=args.components, rand_state=acrstate)
pixelsO, labelsO = auxil.createImageCubes(pixelsO, labelsO, windowSize=args.spatialsize, removeZeroLabels = False)
pixels = pixelsO[labelsO!=0]
labels = labelsO[labelsO!=0] - 1
bands = pixels.shape[-1]; numberofclass = len(np.unique(labels))
pixels = pixels.reshape((pixels.shape[0], args.spatialsize, args.spatialsize, bands, 1))
pixelsO = pixelsO.reshape((pixelsO.shape[0], args.spatialsize, args.spatialsize, bands, 1))
x_train, x_test, y_train, y_test = auxil.split_data(pixels, labels, args.tr_percent, rand_state=acrstate)
del pixels, labels
else:
x_train, x_test, y_train, y_test, pixelsO, labelsO, bands, numberofclass = auxil.loadDataFIX(args)
return (x_train, y_train), (x_test, y_test), (pixelsO, labelsO), numberofclass, bands
def main():
parser = argparse.ArgumentParser(description='PyTorch DCNNs Training')
parser.add_argument('--max_step', default=8000, type=int, help='number of total epochs to run')
parser.add_argument('--idtest', default=0, type=int, help='id of experiment')
parser.add_argument('--lr', '--learning-rate', default=0.1, type=float, help='initial learning rate')
parser.add_argument('--components', default=None, type=int, help='dimensionality reduction')
parser.add_argument('--dataset', default='IP', type=str, help='dataset (options: IP, UP, SV, KSC)')
parser.add_argument('--tr_percent', default=0.10, type=float, help='samples of train set')
parser.add_argument('--tr_bsize', default=100, type=int, help='mini-batch train size (default: 100)')
parser.add_argument('--inplanes', dest='inplanes', default=16, type=int, help='bands before blocks')
parser.add_argument('--no-bottleneck', dest='bottleneck', action='store_false', help='to use basicblock (default: bottleneck)')
parser.add_argument('--spatialsize', dest='spatialsize', default=7, type=int, help='spatial-spectral patch dimension')
parser.add_argument('--momentum', default=0.9, type=float, help='momentum')
parser.add_argument('--weight-decay', '--wd', default=1e-4, type=float, help='weight decay (default: 1e-4)')
parser.add_argument('--random_state', default=None, type=int, help='random seed')
args = parser.parse_args()
state = {k: v for k, v in args._get_kwargs()}
#args.components = 35 if args.dataset == 'IP' else 15
(trainS, labelTr), (testS, labelTs), (pixels, labels), numberofclass, bands = load_hyper(args)
n, m = trainS.shape[0], testS.shape[0]
labelTr, labelTs, c, pInClass, _, newOrder = spp.relabel(labelTr, labelTs)
imbalancedCls, toBalance, imbClsNum, ir=spp.irFind(pInClass, c)
labelsCat = to_categorical(labelTr)
shuffleIndex=np.random.choice(np.arange(n), size=(n,), replace=False)
trainS=trainS[shuffleIndex]
labelTr=labelTr[shuffleIndex]
labelsCat=labelsCat[shuffleIndex]
classMap=list()
for i in range(c):
classMap.append(np.where(labelTr==i)[0])
adamOpt=Adam(0.0002, 0.5)
latDim, modelSamplePd, resSamplePd=100, 8000, 500
# model initialization
mlp=nt.denseMlpCreate(sh=(args.spatialsize, args.spatialsize, bands, 1, ), num_class=numberofclass)
mlp.compile(loss='mean_squared_error', optimizer=adamOpt)
mlp.trainable=False
dis=nt.denseDisCreate((args.spatialsize, args.spatialsize, bands, 1, ), num_class=numberofclass)
dis.compile(loss='mean_squared_error', optimizer=adamOpt)
dis.trainable=False
gen=nt.denseGamoGenCreate(latDim, numberofclass)
gen_processed, genP_mlp, genP_dis=list(), list(), list()
for i in range(imbClsNum):
dataMinor=trainS[classMap[i], :]
numMinor=dataMinor.shape[0]
gen_processed.append(nt.denseGenProcessCreate(numMinor, dataMinor,sh = \
(args.spatialsize, args.spatialsize, bands, 1),mul = args.spatialsize * args.spatialsize * bands))
ip1=Input(shape=(latDim,))
ip2=Input(shape=(c,))
op1=gen([ip1, ip2])
op2=gen_processed[i](op1)
op3=mlp(op2)
genP_mlp.append(Model(inputs=[ip1, ip2], outputs=op3))
genP_mlp[i].compile(loss='mean_squared_error', optimizer=adamOpt)
ip1=Input(shape=(latDim,))
ip2=Input(shape=(c,))
ip3=Input(shape=(c,))
op1=gen([ip1, ip2])
op2=gen_processed[i](op1)
op3=dis([op2, ip3])
genP_dis.append(Model(inputs=[ip1, ip2, ip3], outputs=op3))
genP_dis[i].compile(loss='mean_squared_error', optimizer=adamOpt)
batchDiv, numBatches, bSStore = spp.batchDivision(n, args.tr_bsize)
genClassPoints=int(np.ceil(args.tr_bsize / c))
fileStart = './SavedModel/'
savePath = './SavedModel/'
fileEnd = '_Model.h5'
if not os.path.exists(fileStart):
os.makedirs(fileStart)
picPath=savePath+'Pictures'
if not os.path.exists(picPath):
os.makedirs(picPath)
iter=np.int(np.ceil(args.max_step/resSamplePd)+1)
acsaSaveTr, gmSaveTr, accSaveTr=np.zeros((iter,)), np.zeros((iter,)), np.zeros((iter,))
acsaSaveTs, gmSaveTs, accSaveTs=np.zeros((iter,)), np.zeros((iter,)), np.zeros((iter,))
confMatSaveTr, confMatSaveTs=np.zeros((iter, c, c)), np.zeros((iter, c, c))
tprSaveTr, tprSaveTs=np.zeros((iter, c)), np.zeros((iter, c))
step=0
bestacc = -1
while step < args.max_step:
for j in range(numBatches):
x1, x2=batchDiv[j, 0], batchDiv[j+1, 0]
validR=np.ones((bSStore[j, 0],1))-np.random.uniform(0,0.1, size=(bSStore[j, 0], 1))
mlp.train_on_batch(trainS[x1:x2], labelsCat[x1:x2])
dis.train_on_batch([trainS[x1:x2], labelsCat[x1:x2]], validR)
invalid=np.zeros((bSStore[j, 0], 1))+np.random.uniform(0, 0.1, size=(bSStore[j, 0], 1))
randNoise=np.random.normal(0, 1, (bSStore[j, 0], latDim))
fakeLabel=spp.randomLabelGen(toBalance, bSStore[j, 0], c)
rLPerClass=spp.rearrange(fakeLabel, imbClsNum)
fakePoints=np.zeros((bSStore[j, 0],args.spatialsize, args.spatialsize, bands, 1))
genFinal=gen.predict([randNoise, fakeLabel])
for i1 in range(imbClsNum):
if rLPerClass[i1].shape[0]!=0:
temp=genFinal[rLPerClass[i1]]
fakePoints[rLPerClass[i1]]=gen_processed[i1].predict(temp)
mlp.train_on_batch(fakePoints, fakeLabel)
dis.train_on_batch([fakePoints, fakeLabel], invalid)
for i1 in range(imbClsNum):
validA=np.ones((genClassPoints, 1))
randomLabel=np.zeros((genClassPoints, c))
randomLabel[:, i1]=1
randNoise=np.random.normal(0, 1, (genClassPoints, latDim))
oppositeLabel=np.ones((genClassPoints, c))-randomLabel
genP_mlp[i1].train_on_batch([randNoise, randomLabel], oppositeLabel)
genP_dis[i1].train_on_batch([randNoise, randomLabel, randomLabel], validA)
if step%resSamplePd==0:
saveStep=int(step//resSamplePd)
pLabel=np.argmax(mlp.predict(trainS), axis=1)
acsa, gm, tpr, confMat, acc=spp.indices(pLabel, labelTr)
print('Train: Step: ', step, 'ACSA: ', np.round(acsa, 4), 'GM: ', np.round(gm, 4))
print('TPR: ', np.round(tpr, 2))
acsaSaveTr[saveStep], gmSaveTr[saveStep], accSaveTr[saveStep]=acsa, gm, acc
confMatSaveTr[saveStep]=confMat
tprSaveTr[saveStep]=tpr
pLabel=np.argmax(mlp.predict(testS), axis=1)
acsa, gm, tpr, confMat, acc=spp.indices(pLabel, labelTs)
print('Test: Step: ', step, 'ACSA: ', np.round(acsa, 4), 'GM: ', np.round(gm, 4))
print('TPR: ', np.round(tpr, 2))
acsaSaveTs[saveStep], gmSaveTs[saveStep], accSaveTs[saveStep]=acsa, gm, acc
confMatSaveTs[saveStep]=confMat
tprSaveTs[saveStep]=tpr
results = auxil.reports(pLabel, labelTs)[2]
if bestacc <= results[0]:
bestacc = auxil.reports(pLabel, labelTs)[2][0]
resultsF = auxil.reports(pLabel, labelTs)[2]
if step%modelSamplePd==0 and step!=0:
direcPath=savePath+'gamo_models_'+str(step)
if not os.path.exists(direcPath):
os.makedirs(direcPath)
gen.save(direcPath+'/GEN_'+str(step)+fileEnd)
mlp.save(direcPath+'/MLP_'+str(step)+fileEnd)
dis.save(direcPath+'/DIS_'+str(step)+fileEnd)
for i in range(imbClsNum):
gen_processed[i].save(direcPath+'/GenP_'+str(i)+'_'+str(step)+fileEnd)
step=step+2
if step>=args.max_step: break
print(newOrder)
print(resultsF)
if __name__ == '__main__':
main()