def main():
print opt
# prefixModelName = 'Regressor_1112_'
# prefixPredictedFN = 'preSub1_1112_'
# showTrainLossEvery = 100
# lr = 1e-4
# showTestPerformanceEvery = 2000
# saveModelEvery = 2000
# decLREvery = 40000
# numofIters = 200000
# how2normalize = 0
netD = Discriminator()
netD.apply(weights_init)
netD.cuda()
optimizerD = optim.Adam(netD.parameters(),lr=1e-3)
criterion_bce=nn.BCELoss()
criterion_bce.cuda()
#net=UNet()
if opt.whichNet==1:
net = UNet(in_channel=opt.numOfChannel_allSource, n_classes=1)
elif opt.whichNet==2:
net = ResUNet(in_channel=opt.numOfChannel_allSource, n_classes=1)
elif opt.whichNet==3:
net = UNet_LRes(in_channel=opt.numOfChannel_allSource, n_classes=1)
elif opt.whichNet==4:
net = ResUNet_LRes(in_channel=opt.numOfChannel_allSource, n_classes=1, dp_prob = opt.dropout_rate)
#net.apply(weights_init)
net.cuda()
params = list(net.parameters())
print('len of params is ')
print(len(params))
print('size of params is ')
print(params[0].size())
optimizer = optim.Adam(net.parameters(),lr=opt.lr)
criterion_L2 = nn.MSELoss()
criterion_L1 = nn.L1Loss()
criterion_RTL1 = RelativeThreshold_RegLoss(opt.RT_th)
criterion_gdl = gdl_loss(opt.gdlNorm)
#criterion = nn.CrossEntropyLoss()
# criterion = nn.NLLLoss2d()
given_weight = torch.cuda.FloatTensor([1,4,4,2])
criterion_3d = CrossEntropy3d(weight=given_weight)
criterion_3d = criterion_3d.cuda()
criterion_L2 = criterion_L2.cuda()
criterion_L1 = criterion_L1.cuda()
criterion_RTL1 = criterion_RTL1.cuda()
criterion_gdl = criterion_gdl.cuda()
#inputs=Variable(torch.randn(1000,1,32,32)) #here should be tensor instead of variable
#targets=Variable(torch.randn(1000,10,1,1)) #here should be tensor instead of variable
# trainset=data_utils.TensorDataset(inputs, targets)
# trainloader = data_utils.DataLoader(trainset, batch_size=4, shuffle=True, num_workers=2)
# inputs=torch.randn(1000,1,32,32)
# targets=torch.LongTensor(1000)
path_test ='/home/niedong/DataCT/data_niigz/'
path_patients_h5 = '/home/niedong/DataCT/h5Data_snorm/trainBatch2D_H5'
path_patients_h5_test ='/home/niedong/DataCT/h5Data_snorm/valBatch2D_H5'
# batch_size=10
#data_generator = Generator_2D_slices(path_patients_h5,opt.batchSize,inputKey='data3T',outputKey='data7T')
#data_generator_test = Generator_2D_slices(path_patients_h5_test,opt.batchSize,inputKey='data3T',outputKey='data7T')
data_generator = Generator_2D_slicesV1(path_patients_h5,opt.batchSize, inputKey='dataLPET', segKey='dataCT', contourKey='dataHPET')
data_generator_test = Generator_2D_slicesV1(path_patients_h5_test,opt.batchSize, inputKey='dataLPET', segKey='dataCT', contourKey='dataHPET')
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
net.load_state_dict(checkpoint['model'])
opt.start_epoch = 100000
opt.start_epoch = checkpoint["epoch"] + 1
# net.load_state_dict(checkpoint["model"].state_dict())
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
########### We'd better use dataloader to load a lot of data,and we also should train several epoches###############
########### We'd better use dataloader to load a lot of data,and we also should train several epoches###############
running_loss = 0.0
start = time.time()
for iter in range(opt.start_epoch, opt.numofIters+1):
#print('iter %d'%iter)
inputs, exinputs, labels = data_generator.next()
# xx = np.transpose(inputs,(5,64,64))
# inputs = np.transpose(inputs,(0,3,1,2))
inputs = np.squeeze(inputs) #5x64x64
# exinputs = np.transpose(exinputs,(0,3,1,2))
exinputs = np.squeeze(exinputs) #5x64x64
# print 'shape is ....',inputs.shape
labels = np.squeeze(labels) #64x64
# labels = labels.astype(int)
inputs = inputs.astype(float)
inputs = torch.from_numpy(inputs)
inputs = inputs.float()
exinputs = exinputs.astype(float)
exinputs = torch.from_numpy(exinputs)
exinputs = exinputs.float()
labels = labels.astype(float)
labels = torch.from_numpy(labels)
labels = labels.float()
#print type(inputs), type(exinputs)
if opt.isMultiSource:
source = torch.cat((inputs, exinputs),dim=1)
else:
source = inputs
#source = inputs
mid_slice = opt.numOfChannel_singleSource//2
residual_source = inputs[:, mid_slice, ...]
#inputs = inputs.cuda()
#exinputs = exinputs.cuda()
source = source.cuda()
residual_source = residual_source.cuda()
labels = labels.cuda()
#we should consider different data to train
#wrap them into Variable
source, residual_source, labels = Variable(source),Variable(residual_source), Variable(labels)
#inputs, exinputs, labels = Variable(inputs),Variable(exinputs), Variable(labels)
## (1) update D network: maximize log(D(x)) + log(1 - D(G(z)))
if opt.isAdLoss:
outputG = net(source,residual_source) #5x64x64->1*64x64
if len(labels.size())==3:
labels = labels.unsqueeze(1)
outputD_real = netD(labels)
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
netD.zero_grad()
batch_size = inputs.size(0)
real_label = torch.ones(batch_size,1)
real_label = real_label.cuda()
#print(real_label.size())
real_label = Variable(real_label)
#print(outputD_real.size())
loss_real = criterion_bce(outputD_real,real_label)
loss_real.backward()
#train with fake data
fake_label = torch.zeros(batch_size,1)
# fake_label = torch.FloatTensor(batch_size)
# fake_label.data.resize_(batch_size).fill_(0)
fake_label = fake_label.cuda()
fake_label = Variable(fake_label)
loss_fake = criterion_bce(outputD_fake,fake_label)
loss_fake.backward()
lossD = loss_real + loss_fake
# print 'loss_real is ',loss_real.data[0],'loss_fake is ',loss_fake.data[0],'outputD_real is',outputD_real.data[0]
# print('loss for discriminator is %f'%lossD.data[0])
#update network parameters
optimizerD.step()
if opt.isWDist:
one = torch.FloatTensor([1])
mone = one * -1
one = one.cuda()
mone = mone.cuda()
netD.zero_grad()
outputG = net(source,residual_source) #5x64x64->1*64x64
if len(labels.size())==3:
labels = labels.unsqueeze(1)
outputD_real = netD(labels)
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
batch_size = inputs.size(0)
D_real = outputD_real.mean()
# print D_real
D_real.backward(mone)
D_fake = outputD_fake.mean()
D_fake.backward(one)
gradient_penalty = opt.lambda_D_WGAN_GP*calc_gradient_penalty(netD, labels.data, outputG.data)
gradient_penalty.backward()
D_cost = D_fake - D_real + gradient_penalty
Wasserstein_D = D_real - D_fake
optimizerD.step()
## (2) update G network: minimize the L1/L2 loss, maximize the D(G(x))
# print inputs.data.shape
#outputG = net(source) #here I am not sure whether we should use twice or not
outputG = net(source,residual_source) #5x64x64->1*64x64
net.zero_grad()
if opt.whichLoss==1:
lossG_G = criterion_L1(torch.squeeze(outputG), torch.squeeze(labels))
elif opt.whichLoss==2:
lossG_G = criterion_RTL1(torch.squeeze(outputG), torch.squeeze(labels))
else:
lossG_G = criterion_L2(torch.squeeze(outputG), torch.squeeze(labels))
lossG_G = opt.lossBase * lossG_G
lossG_G.backward() #compute gradients
if opt.isGDL:
lossG_gdl = opt.lambda_gdl * criterion_gdl(outputG,torch.unsqueeze(labels,1))
lossG_gdl.backward() #compute gradients
if opt.isAdLoss:
#we want to fool the discriminator, thus we pretend the label here to be real. Actually, we can explain from the
#angel of equation (note the max and min difference for generator and discriminator)
#outputG = net(inputs)
outputG = net(source,residual_source) #5x64x64->1*64x64
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD = netD(outputG)
lossG_D = opt.lambda_AD*criterion_bce(outputD,real_label) #note, for generator, the label for outputG is real, because the G wants to confuse D
lossG_D.backward()
if opt.isWDist:
#we want to fool the discriminator, thus we pretend the label here to be real. Actually, we can explain from the
#angel of equation (note the max and min difference for generator and discriminator)
#outputG = net(inputs)
outputG = net(source,residual_source) #5x64x64->1*64x64
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
outputD_fake = outputD_fake.mean()
lossG_D = opt.lambda_AD*outputD_fake.mean() #note, for generator, the label for outputG is real, because the G wants to confuse D
lossG_D.backward(mone)
#for other losses, we can define the loss function following the pytorch tutorial
optimizer.step() #update network parameters
#print('loss for generator is %f'%lossG.data[0])
#print statistics
running_loss = running_loss + lossG_G.data[0]
if iter%opt.showTrainLossEvery==0: #print every 2000 mini-batches
print '************************************************'
print 'time now is: ' + time.asctime(time.localtime(time.time()))
# print 'running loss is ',running_loss
print 'average running loss for generator between iter [%d, %d] is: %.5f'%(iter - 100 + 1,iter,running_loss/100)
print 'lossG_G is %.5f respectively.'%(lossG_G.data[0])
if opt.isGDL:
print('loss for GDL loss is %f'%lossG_gdl.data[0])
if opt.isAdLoss:
print 'loss_real is ',loss_real.data[0],'loss_fake is ',loss_fake.data[0],'outputD_real is',outputD_real.data[0]
print('loss for discriminator is %f'%lossD.data[0])
if opt.isWDist:
print 'loss_real is ',D_real.data[0],'loss_fake is ',D_fake.data[0]
print('loss for discriminator is %f'%Wasserstein_D.data[0], ' D cost is %f'%D_cost)
print 'cost time for iter [%d, %d] is %.2f'%(iter - 100 + 1,iter, time.time()-start)
print '************************************************'
running_loss = 0.0
start = time.time()
if iter%opt.saveModelEvery==0: #save the model
state = {
'epoch': iter+1,
'model': net.state_dict()
}
torch.save(state, opt.prefixModelName+'%d.pt'%iter)
print 'save model: '+opt.prefixModelName+'%d.pt'%iter
if opt.isAdLoss:
torch.save(netD.state_dict(), opt.prefixModelName+'_net_D%d.pt'%iter)
if iter%opt.decLREvery==0:
opt.lr = opt.lr*0.5
adjust_learning_rate(optimizer, opt.lr)
if iter%opt.showValPerformanceEvery==0: #test one subject
# to test on the validation dataset in the format of h5
inputs,exinputs,labels = data_generator_test.next()
# inputs = np.transpose(inputs,(0,3,1,2))
inputs = np.squeeze(inputs)
# exinputs = np.transpose(exinputs, (0, 3, 1, 2))
exinputs = np.squeeze(exinputs) # 5x64x64
labels = np.squeeze(labels)
inputs = torch.from_numpy(inputs)
inputs = inputs.float()
exinputs = torch.from_numpy(exinputs)
exinputs = exinputs.float()
labels = torch.from_numpy(labels)
labels = labels.float()
mid_slice = opt.numOfChannel_singleSource // 2
residual_source = inputs[:, mid_slice, ...]
if opt.isMultiSource:
source = torch.cat((inputs, exinputs), dim=1)
else:
source = inputs
source = source.cuda()
residual_source = residual_source.cuda()
labels = labels.cuda()
source,residual_source,labels = Variable(source),Variable(residual_source), Variable(labels)
# source = inputs
#outputG = net(inputs)
outputG = net(source,residual_source) #5x64x64->1*64x64
if opt.whichLoss == 1:
lossG_G = criterion_L1(torch.squeeze(outputG), torch.squeeze(labels))
elif opt.whichLoss == 2:
lossG_G = criterion_RTL1(torch.squeeze(outputG), torch.squeeze(labels))
else:
lossG_G = criterion_L2(torch.squeeze(outputG), torch.squeeze(labels))
lossG_G = opt.lossBase * lossG_G
print '.......come to validation stage: iter {}'.format(iter),'........'
print 'lossG_G is %.5f.'%(lossG_G.data[0])
if opt.isGDL:
lossG_gdl = criterion_gdl(outputG, labels)
print('loss for GDL loss is %f'%lossG_gdl.data[0])
if iter % opt.showTestPerformanceEvery == 0: # test one subject
mr_test_itk=sitk.ReadImage(os.path.join(path_test,'sub1_sourceCT.nii.gz'))
ct_test_itk=sitk.ReadImage(os.path.join(path_test,'sub1_extraCT.nii.gz'))
hpet_test_itk = sitk.ReadImage(os.path.join(path_test, 'sub1_targetCT.nii.gz'))
spacing = hpet_test_itk.GetSpacing()
origin = hpet_test_itk.GetOrigin()
direction = hpet_test_itk.GetDirection()
mrnp=sitk.GetArrayFromImage(mr_test_itk)
ctnp=sitk.GetArrayFromImage(ct_test_itk)
hpetnp=sitk.GetArrayFromImage(hpet_test_itk)
##### specific normalization #####
# mu = np.mean(mrnp)
# maxV, minV = np.percentile(mrnp, [99 ,25])
# #mrimg=mrimg
# mrnp = (mrnp-minV)/(maxV-minV)
#for training data in pelvicSeg
if opt.how2normalize == 1:
maxV, minV = np.percentile(mrnp, [99 ,1])
print 'maxV,',maxV,' minV, ',minV
mrnp = (mrnp-mu)/(maxV-minV)
print 'unique value: ',np.unique(ctnp)
#for training data in pelvicSeg
if opt.how2normalize == 2:
maxV, minV = np.percentile(mrnp, [99 ,1])
print 'maxV,',maxV,' minV, ',minV
mrnp = (mrnp-mu)/(maxV-minV)
print 'unique value: ',np.unique(ctnp)
#for training data in pelvicSegRegH5
if opt.how2normalize== 3:
std = np.std(mrnp)
mrnp = (mrnp - mu)/std
print 'maxV,',np.ndarray.max(mrnp),' minV, ',np.ndarray.min(mrnp)
if opt.how2normalize == 4:
maxLPET = 149.366742
maxPercentLPET = 7.76
minLPET = 0.00055037
meanLPET = 0.27593288
stdLPET = 0.75747500
# for rsCT
maxCT = 27279
maxPercentCT = 1320
minCT = -1023
meanCT = -601.1929
stdCT = 475.034
# for s-pet
maxSPET = 156.675962
maxPercentSPET = 7.79
minSPET = 0.00055037
meanSPET = 0.284224789
stdSPET = 0.7642257
#matLPET = (mrnp - meanLPET) / (stdLPET)
matLPET = (mrnp - minLPET) / (maxPercentLPET - minLPET)
matCT = (ctnp - meanCT) / stdCT
matSPET = (hpetnp - minSPET) / (maxPercentSPET - minSPET)
if opt.how2normalize == 5:
# for rsCT
maxCT = 27279
maxPercentCT = 1320
minCT = -1023
meanCT = -601.1929
stdCT = 475.034
print
'ct, max: ', np.amax(ctnp), ' ct, min: ', np.amin(ctnp)
# matLPET = (mrnp - meanLPET) / (stdLPET)
matLPET = mrnp
matCT = (ctnp - meanCT) / stdCT
matSPET = hpetnp
if opt.how2normalize == 6:
maxPercentPET, minPercentPET = np.percentile(mrnp, [99.5, 0])
maxPercentCT, minPercentCT = np.percentile(ctnp, [99.5, 0])
print 'maxPercentPET: ', maxPercentPET, ' minPercentPET: ', minPercentPET, ' maxPercentCT: ', maxPercentCT, 'minPercentCT: ', minPercentCT
matLPET = (mrnp - minPercentPET) / (maxPercentPET - minPercentPET)
matSPET = (hpetnp - minPercentPET) / (maxPercentPET - minPercentPET)
matCT = (ctnp - minPercentCT) / (maxPercentCT - minPercentCT)
if not opt.isMultiSource:
matFA = matLPET
matGT = hpetnp
print 'matFA shape: ',matFA.shape, ' matGT shape: ', matGT.shape
matOut = testOneSubject_aver_res(matFA,matGT,[5,64,64],[1,64,64],[1,32,32],net,opt.prefixModelName+'%d.pt'%iter)
print 'matOut shape: ',matOut.shape
if opt.how2normalize==6:
ct_estimated = matOut * (maxPercentPET - minPercentPET) + minPercentPET
else:
ct_estimated = matOut
itspsnr = psnr(ct_estimated, matGT)
print 'pred: ',ct_estimated.dtype, ' shape: ',ct_estimated.shape
print 'gt: ',ctnp.dtype,' shape: ',ct_estimated.shape
print 'psnr = ',itspsnr
volout = sitk.GetImageFromArray(ct_estimated)
volout.SetSpacing(spacing)
volout.SetOrigin(origin)
volout.SetDirection(direction)
sitk.WriteImage(volout,opt.prefixPredictedFN+'{}'.format(iter)+'.nii.gz')
else:
matFA = matLPET
matGT = hpetnp
print 'matFA shape: ', matFA.shape, ' matGT shape: ', matGT.shape
matOut = testOneSubject_aver_res_multiModal(matFA, matCT, matGT, [5, 64, 64], [1, 64, 64], [1, 32, 32], net,
opt.prefixModelName + '%d.pt' % iter)
print 'matOut shape: ', matOut.shape
if opt.how2normalize==6:
ct_estimated = matOut * (maxPercentPET - minPercentPET) + minPercentPET
else:
ct_estimated = matOut
itspsnr = psnr(ct_estimated, matGT)
print 'pred: ', ct_estimated.dtype, ' shape: ', ct_estimated.shape
print 'gt: ', ctnp.dtype, ' shape: ', ct_estimated.shape
print 'psnr = ', itspsnr
volout = sitk.GetImageFromArray(ct_estimated)
volout.SetSpacing(spacing)
volout.SetOrigin(origin)
volout.SetDirection(direction)
sitk.WriteImage(volout, opt.prefixPredictedFN + '{}'.format(iter) + '.nii.gz')
print('Finished Training')
def main():
print opt
# load/define networks
netG = Generator(opt.finalSize*opt.finalSize, opt.ngf, gpu_ids=opt.gpuID)
# netG = define_G(opt.fineSize*opt.fineSize, opt.ngf,opt.init_type, opt.gpuID)
optimizerG = optim.Adam(netG.parameters(),lr=opt.lr_G)
netG.apply(weights_init)
netG.cuda()
if opt.isAdLoss:
netD = Discriminator(opt.ndf, opt.gpuID)
netD.apply(weights_init)
netD.cuda()
optimizerD = optim.Adam(netD.parameters(),lr=opt.lr_D)
if opt.isWDist:
netD = Discriminator(opt.ndf, opt.gpuID)
netD.apply(weights_init)
netD.cuda()
optimizerD = optim.Adam(netD.parameters(),lr=opt.lr_D)
criterion_bce=nn.BCELoss()
criterion_bce.cuda()
params = list(netG.parameters())
print('len of params is ')
print(len(params))
print('size of params is ')
print(params[0].size())
path_test ='/shenlab/lab_stor5/dongnie/3T7T/3t7tHistData/'
path_patients_h5 = '/shenlab/lab_stor5/dongnie/3T7T/histH5Data'
path_patients_h5_test ='/shenlab/lab_stor5/dongnie/3T7T/histH5Data'
data_generator = Generator_2D_slices_oneKey(path_patients_h5,opt.batchSize, inputKey='data7T')
data_generator_test = Generator_2D_slices_oneKey(path_patients_h5_test,opt.batchSize, inputKey='data7T')
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
netG.load_state_dict(checkpoint['model'])
opt.start_epoch = 100000
opt.start_epoch = checkpoint["epoch"] + 1
# net.load_state_dict(checkpoint["model"].state_dict())
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
########### We'd better use dataloader to load a lot of data,and we also should train several epoches###############
########### We'd better use dataloader to load a lot of data,and we also should train several epoches###############
running_loss = 0.0
start = time.time()
for iter in range(opt.start_epoch, opt.numofIters+1):
#print('iter %d'%iter)
# we use a 128-dim vector as the noise of the input
noise = torch.randn(opt.batchSize, 128)
if opt.gpuID!=None:
noise = noise.cuda()
labels = data_generator.next() # labels means real images
labels = np.squeeze(labels) #64x64
labels = np.resize(labels, [opt.batchSize, opt.outputSizeOfG,opt.outputSizeOfG]) # here, we take 32 as output size
# labels = labels.astype(int)
inputs = noise
labels = labels.astype(float)
labels = torch.from_numpy(labels)
labels = labels.float()
source = inputs
source = source.cuda()
# residual_source = residual_source.cuda()
labels = labels.cuda()
#we should consider different data to train
source, labels = Variable(source), Variable(labels)
## (1) update D network: maximize log(D(x)) + log(1 - D(G(z)))
if opt.isAdLoss:
# outputG = net(source,residual_source) #5x64x64->1*64x64
if len(labels.size())==3:
labels = labels.unsqueeze(1)
# print 'labels.shape: ',labels.shape
outputD_real = netD(labels)
outputD_real = F.sigmoid(outputD_real)
outputG = netG(source) #1x64x64->1*64x64
# print 'outputG.shape: ',outputG.shape
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
outputD_fake = F.sigmoid(outputD_fake)
## update D network
netD.zero_grad()
batch_size = inputs.size(0)
real_label = torch.ones(batch_size,1)
real_label = real_label.cuda()
#print(real_label.size())
real_label = Variable(real_label)
# print 'outputD_real.shape: ',outputD_real.shape,' outputD_fake.size(): ',outputD_fake.shape
loss_real = criterion_bce(outputD_real,real_label)
loss_real.backward()
#train with fake data
fake_label = torch.zeros(batch_size,1)
fake_label = fake_label.cuda()
fake_label = Variable(fake_label)
loss_fake = criterion_bce(outputD_fake,fake_label)
loss_fake.backward()
lossD = loss_real + loss_fake
optimizerD.step()
if opt.isWDist:
one = torch.FloatTensor([1])
mone = one * -1
one = one.cuda()
mone = mone.cuda()
netD.zero_grad()
outputG = netG(source) #5x64x64->1*64x64
if len(labels.size())==3:
labels = labels.unsqueeze(1)
outputD_real = netD(labels)
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
batch_size = inputs.size(0)
D_real = outputD_real.mean()
# print D_real
D_real.backward(mone)
D_fake = outputD_fake.mean()
D_fake.backward(one)
gradient_penalty = opt.lambda_D_WGAN_GP*calc_gradient_penalty(netD, labels.data, outputG.data)
gradient_penalty.backward()
D_cost = D_fake - D_real + gradient_penalty
Wasserstein_D = D_real - D_fake
optimizerD.step()
## (2) update G network: maximize the D(G(x))
netG.zero_grad()
if opt.isAdLoss:
#we want to fool the discriminator, thus we pretend the label here to be real. Actually, we can explain from the
#angel of equation (note the max and min difference for generator and discriminator)
outputG = netG(source) #5x64x64->1*64x64
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD = netD(outputG)
outputD = F.sigmoid(outputD)
lossG_D = opt.lambda_AD*criterion_bce(outputD,real_label) #note, for generator, the label for outputG is real, because the G wants to confuse D
lossG_D.backward()
if opt.isWDist:
#we want to fool the discriminator, thus we pretend the label here to be real. Actually, we can explain from the
#angel of equation (note the max and min difference for generator and discriminator)
#outputG = net(inputs)
outputG = netG(source) #5x64x64->1*64x64
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
outputD_fake = outputD_fake.mean()
lossG_D = opt.lambda_AD*outputD_fake.mean() #note, for generator, the label for outputG is real, because the G wants to confuse D
lossG_D.backward(mone)
#for other losses, we can define the loss function following the pytorch tutorial
optimizerG.step() #update network parameters
if iter%opt.showTrainLossEvery==0: #print every 2000 mini-batches
print '************************************************'
print 'time now is: ' + time.asctime(time.localtime(time.time()))
# print 'running loss is ',running_loss
# print 'average running loss for generator between iter [%d, %d] is: %.5f'%(iter - 100 + 1,iter,running_loss/100)
if opt.isAdLoss:
print 'loss_real is ',loss_real.data[0],'loss_fake is ',loss_fake.data[0],'outputD_real is',torch.mean(outputD_real).data[0], 'outputD_fake is',torch.mean(outputD_fake).data[0]
print 'lossG_D is ', lossG_D.data[0]
print 'loss for discriminator is %f'%lossD.data[0]
if opt.isWDist:
print 'loss_real is ',D_real.data[0],'loss_fake is ',D_fake.data[0], 'outputD_real is',torch.mean(outputD_real).data[0], 'outputD_fake is',torch.mean(outputD_fake).data[0]
print 'lossG_D is ', lossG_D.data[0]
print 'loss for discriminator is %f'%Wasserstein_D.data[0], ' D cost is %f'%D_cost
print 'cost time for iter [%d, %d] is %.2f'%(iter - 100 + 1,iter, time.time()-start)
print '************************************************'
running_loss = 0.0
start = time.time()
if iter%opt.saveModelEvery==0: #save the model
state = {
'epoch': iter+1,
'model': netG.state_dict()
}
torch.save(state, opt.prefixModelName+'%d.pt'%iter)
print 'save model: '+opt.prefixModelName+'%d.pt'%iter
if opt.isAdLoss or opt.isWDist:
torch.save(netD.state_dict(), opt.prefixModelName+'netD_%d.pt'%iter)
if iter%opt.decLREvery==0:
opt.lr_G = opt.lr_G*0.5
adjust_learning_rate(optimizerG, opt.lr_G)
if opt.isAdLoss or opt.isWDist:
opt.lr_D = opt.lr_D*0.2
adjust_learning_rate(optimizerD, opt.lr_D)
if iter%opt.showValPerformanceEvery==0: #test one subject
# to test on the validation dataset in the format of h5
# we use a 128-dim vector as the noise of the input
noise = torch.randn(opt.batchSize, 128)
if opt.gpuID!=None:
noise = noise.cuda()
inputs = noise
# inputs,exinputs,labels = data_generator_test.next()
labels = data_generator_test.next()
labels = np.squeeze(labels)
labels = torch.from_numpy(labels)
labels = labels.float()
source = inputs
source = source.cuda()
source = Variable(source)
# residual_source = residual_source.cuda()
labels = labels.cuda()
labels = Variable(labels)
if len(labels.size())==3:
labels = labels.unsqueeze(1)
outputD_real = netD(labels)
outputG = netG(source) #noise -> img
if len(outputG.size())==3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
print 'outputD_real is ', torch.mean(outputD_real).data[0], ' outputD_real is ', torch.mean(outputD_real).data[0]
if iter % opt.showTestPerformanceEvery == 0: # test one subject
mr_test_itk=sitk.ReadImage(os.path.join(path_test,'S10to1_3t.nii.gz'))
spacing = mr_test_itk.GetSpacing()
origin = mr_test_itk.GetOrigin()
direction = mr_test_itk.GetDirection()
mrnp=sitk.GetArrayFromImage(mr_test_itk)
##### specific normalization #####
mu = np.mean(mrnp)
#for training data in pelvicSeg
if opt.how2normalize == 1:
maxV, minV = np.percentile(mrnp, [99 ,1])
print 'maxV,',maxV,' minV, ',minV
mrnp = (mrnp-mu)/(maxV-minV)
print 'unique value: ',np.unique(mrnp)
#for training data in pelvicSeg
if opt.how2normalize == 2:
maxV, minV = np.percentile(mrnp, [99 ,1])
print 'maxV,',maxV,' minV, ',minV
mrnp = (mrnp-mu)/(maxV-minV)
print 'unique value: ',np.unique(mrnp)
#for training data in pelvicSegRegH5
if opt.how2normalize== 3:
std = np.std(mrnp)
mrnp = (mrnp - mu)/std
print 'maxV,',np.ndarray.max(mrnp),' minV, ',np.ndarray.min(mrnp)
if opt.how2normalize == 4:
maxLPET = 149.366742
maxPercentLPET = 7.76
minLPET = 0.00055037
meanLPET = 0.27593288
stdLPET = 0.75747500
matLPET = (mrnp - minLPET) / (maxPercentLPET - minLPET)
if opt.how2normalize == 5:
# for rsCT
maxCT = 27279
maxPercentCT = 1320
minCT = -1023
meanCT = -601.1929
stdCT = 475.034
print 'ct, max: ', np.amax(mrnp), ' ct, min: ', np.amin(mrnp)
matLPET = mrnp
if opt.how2normalize == 6:
maxPercentPET, minPercentPET = np.percentile(mrnp, [99.5, 0])
matLPET = (mrnp - minPercentPET) / (maxPercentPET - minPercentPET)
matFA = matLPET
# matGT = hpetnp
print 'matFA shape: ',matFA.shape
pred = testOneSubject4Cla(matFA,[1,64,64],[1,32,32],netD, opt.prefixModelName+'%d.pt'%iter)
print 'predicted result is ', pred
print('Finished Training')
def main():
print(opt)
netD = Discriminator()
netD.apply(weights_init)
netD.cuda()
optimizerD = optim.Adam(netD.parameters(), lr=opt.lr_netD)
criterion_bce = nn.BCELoss()
criterion_bce.cuda()
if opt.whichNet == 1:
net = UNet(in_channel=opt.numOfChannel_allSource, n_classes=1)
elif opt.whichNet == 2:
net = ResUNet(in_channel=opt.numOfChannel_allSource, n_classes=1)
elif opt.whichNet == 3:
net = UNet_LRes(in_channel=opt.numOfChannel_allSource, n_classes=1)
elif opt.whichNet == 4:
net = ResUNet_LRes(in_channel=opt.numOfChannel_allSource,
n_classes=1,
dp_prob=opt.dropout_rate)
net.cuda()
params = list(net.parameters())
print('len of params is ')
print(len(params))
print('size of params is ')
print(params[0].size())
optimizer = optim.Adam(net.parameters(), lr=opt.lr)
criterion_L2 = nn.MSELoss()
criterion_L1 = nn.L1Loss()
criterion_RTL1 = RelativeThreshold_RegLoss(opt.RT_th)
criterion_gdl = gdl_loss(opt.gdlNorm)
given_weight = torch.cuda.FloatTensor([1, 4, 4, 2])
criterion_3d = CrossEntropy3d(weight=given_weight)
criterion_3d = criterion_3d.cuda()
criterion_L2 = criterion_L2.cuda()
criterion_L1 = criterion_L1.cuda()
criterion_RTL1 = criterion_RTL1.cuda()
criterion_gdl = criterion_gdl.cuda()
path_test = opt.path_test
path_train = opt.path_train
path_dev = opt.path_dev
data_generator = Generator_2D_slices(path_train,
opt.batchSize,
inputKey='noisy',
outputKey='clear')
data_generator_dev = Generator_2D_slices(path_dev,
opt.batchSize,
inputKey='noisy',
outputKey='noisy')
if opt.resume:
if os.path.isfile(opt.resume):
print("=> loading checkpoint '{}'".format(opt.resume))
checkpoint = torch.load(opt.resume)
net.load_state_dict(checkpoint['model'])
opt.start_epoch = checkpoint["epoch"] + 1
else:
print("=> no checkpoint found at '{}'".format(opt.resume))
########### We'd better use dataloader to load a lot of data,and we also should train several epoches###############
########### We'd better use dataloader to load a lot of data,and we also should train several epoches###############
running_loss = 0.0
start = time.time()
for iter in range(opt.start_epoch, opt.numofIters + 1):
inputs, labels = next(data_generator)
exinputs = inputs
inputs = np.squeeze(inputs) #5x64x64
exinputs = np.squeeze(exinputs) #5x64x64
labels = np.squeeze(labels) #64x64
inputs = inputs.astype(float)
inputs = torch.from_numpy(inputs)
inputs = inputs.float()
exinputs = exinputs.astype(float)
exinputs = torch.from_numpy(exinputs)
exinputs = exinputs.float()
labels = labels.astype(float)
labels = torch.from_numpy(labels)
labels = labels.float()
source = inputs
mid_slice = opt.numOfChannel_singleSource // 2
residual_source = inputs[:, mid_slice, ...]
source = source.cuda()
residual_source = residual_source.cuda()
labels = labels.cuda()
#wrap them into Variable
source, residual_source, labels = Variable(source), Variable(
residual_source), Variable(labels)
## (1) update D network: maximize log(D(x)) + log(1 - D(G(z)))
if opt.isAdLoss:
if opt.whichNet == 3 or opt.whichNet == 4:
outputG = net(source, residual_source) # 5x64x64->1*64x64
else:
outputG = net(source) # 5x64x64->1*64x64
if len(labels.size()) == 3:
labels = labels.unsqueeze(1)
outputD_real = netD(labels)
outputD_real = torch.sigmoid(outputD_real)
if len(outputG.size()) == 3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
outputD_fake = torch.sigmoid(outputD_fake)
netD.zero_grad()
batch_size = inputs.size(0)
real_label = torch.ones(batch_size, 1)
real_label = real_label.cuda()
real_label = Variable(real_label)
loss_real = criterion_bce(outputD_real, real_label)
loss_real.backward()
#train with fake data
fake_label = torch.zeros(batch_size, 1)
fake_label = fake_label.cuda()
fake_label = Variable(fake_label)
loss_fake = criterion_bce(outputD_fake, fake_label)
loss_fake.backward()
lossD = loss_real + loss_fake
#update network parameters
optimizerD.step()
if opt.isWDist:
one = torch.FloatTensor([1])
mone = one * -1
one = one.cuda()
mone = mone.cuda()
netD.zero_grad()
if opt.whichNet == 3 or opt.whichNet == 4:
outputG = net(source, residual_source) # 5x64x64->1*64x64
else:
outputG = net(source) # 5x64x64->1*64x64
if len(labels.size()) == 3:
labels = labels.unsqueeze(1)
outputD_real = netD(labels)
if len(outputG.size()) == 3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
batch_size = inputs.size(0)
D_real = outputD_real.mean()
D_real.backward(mone)
D_fake = outputD_fake.mean()
D_fake.backward(one)
gradient_penalty = opt.lambda_D_WGAN_GP * calc_gradient_penalty(
netD, labels.data, outputG.data)
gradient_penalty.backward()
D_cost = D_fake - D_real + gradient_penalty
Wasserstein_D = D_real - D_fake
optimizerD.step()
## (2) update G network: minimize the L1/L2 loss, maximize the D(G(x))
if opt.whichNet == 3 or opt.whichNet == 4:
outputG = net(source, residual_source) # 5x64x64->1*64x64
else:
outputG = net(source) # 5x64x64->1*64x64
net.zero_grad()
if opt.whichLoss == 1:
lossG_G = criterion_L1(torch.squeeze(outputG),
torch.squeeze(labels))
elif opt.whichLoss == 2:
lossG_G = criterion_RTL1(torch.squeeze(outputG),
torch.squeeze(labels))
else:
lossG_G = criterion_L2(torch.squeeze(outputG),
torch.squeeze(labels))
lossG_G = opt.lossBase * lossG_G
lossG_G.backward(retain_graph=True) #compute gradients
if opt.isGDL:
lossG_gdl = opt.lambda_gdl * criterion_gdl(
outputG, torch.unsqueeze(torch.squeeze(labels, 1), 1))
lossG_gdl.backward() #compute gradients
if opt.isAdLoss:
#we want to fool the discriminator, thus we pretend the label here to be real. Actually, we can explain from the
#angel of equation (note the max and min difference for generator and discriminator)
if opt.whichNet == 3 or opt.whichNet == 4:
outputG = net(source, residual_source) # 5x64x64->1*64x64
else:
outputG = net(source) # 5x64x64->1*64x64
if len(outputG.size()) == 3:
outputG = outputG.unsqueeze(1)
outputD = netD(outputG)
outputD = torch.sigmoid(outputD)
lossG_D = opt.lambda_AD * criterion_bce(
outputD, real_label
) #note, for generator, the label for outputG is real, because the G wants to confuse D
lossG_D.backward()
if opt.isWDist:
#we want to fool the discriminator, thus we pretend the label here to be real. Actually, we can explain from the
#angel of equation (note the max and min difference for generator and discriminator)
if opt.whichNet == 3 or opt.whichNet == 4:
outputG = net(source, residual_source) # 5x64x64->1*64x64
else:
outputG = net(source) # 5x64x64->1*64x64
if len(outputG.size()) == 3:
outputG = outputG.unsqueeze(1)
outputD_fake = netD(outputG)
outputD_fake = outputD_fake.mean()
lossG_D = opt.lambda_AD * outputD_fake.mean(
) #note, for generator, the label for outputG is real, because the G wants to confuse D
lossG_D.backward(mone)
#for other losses, we can define the loss function following the pytorch tutorial
optimizer.step() #update network parameters
running_loss = running_loss + lossG_G.data.item()
if iter % opt.showTrainLossEvery == 0: #print every 2000 mini-batches
print('************************************************')
print('time now is: ' + time.asctime(time.localtime(time.time())))
print(
'average running loss for generator between iter [%d, %d] is: %.5f'
% (iter - 100 + 1, iter, running_loss / 100))
print('lossG_G is %.5f respectively.' % (lossG_G.data.item()))
if opt.isGDL:
print('loss for GDL loss is %f' % lossG_gdl.data.item())
if opt.isAdLoss:
print('loss for discriminator is %f' % lossD.data.item())
print('lossG_D for discriminator is %f' % lossG_D.data.item())
if opt.isWDist:
print('loss_real is ',
torch.mean(D_real).data.item(), 'loss_fake is ',
torch.mean(D_fake).data.item())
print(
'loss for discriminator is %f' % Wasserstein_D.data.item(),
' D cost is %f' % D_cost)
print('lossG_D for discriminator is %f' % lossG_D.data.item())
print('cost time for iter [%d, %d] is %.2f' %
(iter - 100 + 1, iter, time.time() - start))
print('************************************************')
running_loss = 0.0
start = time.time()
if iter % opt.saveModelEvery == 0: #save the model
state = {'epoch': iter + 1, 'model': net.state_dict()}
torch.save(state, opt.prefixModelName + '%d.pt' % iter)
print('save model: ' + opt.prefixModelName + '%d.pt' % iter)
if opt.isAdLoss or opt.isWDist:
torch.save(netD.state_dict(),
opt.prefixModelName + '_net_D%d.pt' % iter)
if iter % opt.decLREvery == 0:
opt.lr = opt.lr * opt.lrDecRate
adjust_learning_rate(optimizer, opt.lr)
if opt.isAdLoss or opt.isWDist:
opt.lr_netD = opt.lr_netD * opt.lrDecRate_netD
adjust_learning_rate(optimizerD, opt.lr_netD)
if iter % opt.showValPerformanceEvery == 0: #test one subject
# to test on the validation dataset in the format of h5
inputs, labels = next(data_generator_dev)
exinputs = inputs
inputs = np.squeeze(inputs)
exinputs = np.squeeze(exinputs) # 5x64x64
labels = np.squeeze(labels)
inputs = torch.from_numpy(inputs)
inputs = inputs.float()
exinputs = torch.from_numpy(exinputs)
exinputs = exinputs.float()
labels = torch.from_numpy(labels)
labels = labels.float()
mid_slice = opt.numOfChannel_singleSource // 2
residual_source = inputs[:, mid_slice, ...]
source = inputs
source = source.cuda()
residual_source = residual_source.cuda()
labels = labels.cuda()
source, residual_source, labels = Variable(source), Variable(
residual_source), Variable(labels)
if opt.whichNet == 3 or opt.whichNet == 4:
outputG = net(source, residual_source) # 5x64x64->1*64x64
else:
outputG = net(source) # 5x64x64->1*64x64
if opt.whichLoss == 1:
lossG_G = criterion_L1(torch.squeeze(outputG),
torch.squeeze(labels))
elif opt.whichLoss == 2:
lossG_G = criterion_RTL1(torch.squeeze(outputG),
torch.squeeze(labels))
else:
lossG_G = criterion_L2(torch.squeeze(outputG),
torch.squeeze(labels))
lossG_G = opt.lossBase * lossG_G
print('.......come to validation stage: iter {}'.format(iter),
'........')
print('lossG_G is %.5f.' % (lossG_G.data.item()))
if opt.isGDL:
lossG_gdl = criterion_gdl(
outputG, torch.unsqueeze(torch.squeeze(labels, 1), 1))
print('loss for GDL loss is %f' % lossG_gdl.data.item())
if iter % opt.showTestPerformanceEvery == 0: # test one subject
noisy_np = np.load(
os.path.join(path_test, opt.test_input_file_name))
noisy_np = noisy_np[0]
noisy_np = noisy_np.reshape(1, noisy_np.shape[0],
noisy_np.shape[1])
clear_np = np.load(
os.path.join(path_test, opt.test_label_file_name))
clear_np = clear_np[0]
clear_np = clear_np.reshape(1, clear_np.shape[0],
clear_np.shape[1])
hpetnp = clear_np
#for training data in pelvicSeg
if opt.how2normalize == 1:
maxV, minV = np.percentile(noisy_np, [99, 1])
print('maxV,', maxV, ' minV, ', minV)
noisy_np = (noisy_np - mu) / (maxV - minV)
print('unique value: ', np.unique(clear_np))
#for training data in pelvicSeg
if opt.how2normalize == 2:
maxV, minV = np.percentile(noisy_np, [99, 1])
print('maxV,', maxV, ' minV, ', minV)
noisy_np = (noisy_np - mu) / (maxV - minV)
print('unique value: ', np.unique(clear_np))
#for training data in pelvicSegRegH5
if opt.how2normalize == 3:
std = np.std(noisy_np)
noisy_np = (noisy_np - mu) / std
print('maxV,', np.ndarray.max(noisy_np), ' minV, ',
np.ndarray.min(noisy_np))
if opt.how2normalize == 4:
maxLPET = 149.366742
maxPercentLPET = 7.76
minLPET = 0.00055037
meanLPET = 0.27593288
stdLPET = 0.75747500
# for rsCT
maxCT = 27279
maxPercentCT = 1320
minCT = -1023
meanCT = -601.1929
stdCT = 475.034
# for s-pet
maxSPET = 156.675962
maxPercentSPET = 7.79
minSPET = 0.00055037
meanSPET = 0.284224789
stdSPET = 0.7642257
matLPET = (noisy_np - minLPET) / (maxPercentLPET - minLPET)
matCT = (clear_np - meanCT) / stdCT
matSPET = (hpetnp - minSPET) / (maxPercentSPET - minSPET)
if opt.how2normalize == 5:
# for rsCT
maxCT = 27279
maxPercentCT = 1320
minCT = -1023
meanCT = -601.1929
stdCT = 475.034
print('ct, max: ', np.amax(clear_np), ' ct, min: ',
np.amin(clear_np))
matLPET = noisy_np
matCT = (clear_np - meanCT) / stdCT
matSPET = hpetnp
if opt.how2normalize == 6:
maxPercentPET, minPercentPET = np.percentile(
noisy_np, [99.5, 0])
maxPercentCT, minPercentCT = np.percentile(clear_np, [99.5, 0])
print('maxPercentPET: ', maxPercentPET, ' minPercentPET: ',
minPercentPET, ' maxPercentCT: ', maxPercentCT,
'minPercentCT: ', minPercentCT)
matLPET = (noisy_np - minPercentPET) / (maxPercentPET -
minPercentPET)
matSPET = (hpetnp - minPercentPET) / (maxPercentPET -
minPercentPET)
matCT = (clear_np - minPercentCT) / (maxPercentCT -
minPercentCT)
matFA = matLPET
matGT = hpetnp
print('matFA shape: ', matFA.shape, ' matGT shape: ', matGT.shape)
matOut = testOneSubject_aver_res(
matFA, matGT, [2, 64, 64], [1, 64, 64], [1, 8, 8], net,
opt.prefixModelName + '%d.pt' % iter)
print('matOut shape: ', matOut.shape)
if opt.how2normalize == 6:
clear_estimated = matOut * (maxPercentPET -
minPercentPET) + minPercentPET
else:
clear_estimated = matOut
itspsnr = psnr(clear_estimated, matGT)
clear_estimated = clear_estimated.reshape(clear_estimated.shape[1],
clear_estimated.shape[2])
print('pred: ', clear_estimated.dtype, ' shape: ',
clear_estimated.shape)
print('gt: ', clear_np.dtype, ' shape: ', clear_estimated.shape)
print('psnr = ', itspsnr)
volout = sitk.GetImageFromArray(clear_estimated)
volout = sitk.Cast(
sitk.RescaleIntensity(volout,
outputMinimum=0,
outputMaximum=65535), sitk.sitkUInt16)
sitk.WriteImage(
volout, opt.prefixPredictedFN + '{}'.format(iter) + '.tiff')
np.save(opt.prefixPredictedFN + '{}'.format(iter) + '.npy',
clear_estimated)
print('Finished Training')