K = 1
else:
optimizer = optim.Adam(resnet.parameters(), lr=lr, betas=(0.5, 0.999))
K = 4
epoch = 0
niter = 5
loss_iters = np.zeros(niter)
while epoch < 2:
epoch += 1
# for idx, (rdf_inputs, rdfs, masks, weights, wGs, D) in enumerate(trainLoader): # for real rdf
for idx, (qsms, rdf_inputs, rdfs, wG, weights, masks_csf,
D) in enumerate(trainLoader): # for simulated rdf
if epoch == 1:
unet3d.eval(), resnet.eval(), unet3d_p.eval(), unet3d_.eval(
), unet3d_pdi.eval()
unet3d.to(device), unet3d_p.to(device), unet3d_.to(
device), unet3d_pdi.to(device)
rdf_inputs = rdf_inputs.to(device, dtype=torch.float)
with torch.no_grad():
qsm_inputs = unet3d(rdf_inputs).cpu().detach()
QSMnet_p = unet3d_p(rdf_inputs).cpu().detach()
QSMnet_ = unet3d_(rdf_inputs).cpu().detach()
PDI = unet3d_pdi((rdf_inputs + 0.15) * 3).cpu().detach()
# QSMnet = np.squeeze(np.asarray(qsm_inputs))
QSMnet = np.squeeze(np.asarray(QSMnet_))
QSMnet_p = np.squeeze(np.asarray(QSMnet_p))
PDI = np.squeeze(np.asarray(PDI[0, 0, ...])) / 3
else:
epoch = 0
loss_iters = np.zeros(niter)
while epoch < niter:
epoch += 1
# training phase
for idx, (rdfs, masks, weights, wGs) in enumerate(trainLoader):
rdfs = (rdfs.to(device, dtype=torch.float) + trans) * scale
masks = masks.to(device, dtype=torch.float)
weights = weights.to(device, dtype=torch.float)
wGs = wGs.to(device, dtype=torch.float)
if epoch == 1:
unet3d.eval()
means = unet3d(rdfs)[:, 0, ...]
stds = unet3d(rdfs)[:, 1, ...]
QSM = np.squeeze(np.asarray(means.cpu().detach()))
STD = np.squeeze(np.asarray(stds.cpu().detach()))
print('Saving initial results')
adict = {}
adict['QSM'] = QSM
sio.savemat(rootDir + '/QSM_0.mat', adict)
adict = {}
adict['STD'] = STD
sio.savemat(rootDir + '/STD_0.mat', adict)
# network
model = Unet(input_channels=1,
output_channels=2,
num_filters=[2**i for i in range(5, 10)],
bilateral=1,
use_deconv=1,
use_deconv2=1,
renorm=0,
flag_r_train=0,
flag_UTFI=1)
model.to(device)
model.load_state_dict(
torch.load(
rootDir +
'/weights/weight_pdf={0}_tv={1}.pt'.format(lambda_pdf, lambda_tv)))
model.eval()
chi_bs, chi_ls = [], []
with torch.no_grad():
for idx, (ifreqs, masks, data_weights, wGs) in enumerate(testLoader):
ifreqs = ifreqs.to(device, dtype=torch.float)
masks = masks.to(device, dtype=torch.float)
outputs = model(ifreqs)
chi_b, chi_l = outputs[:, 0:1, ...] * (
1 - masks), outputs[:, 1:2, ...] * masks
chi_bs.append(chi_b[0, ...].cpu().detach())
chi_ls.append(chi_l[0, ...].cpu().detach())
chi_bs = np.concatenate(chi_bs, axis=0)
chi_ls = np.concatenate(chi_ls, axis=0)
loss1 = loss_QSMnet(outputs1, qsms, masks, D)
loss2 = loss_QSMnet(outputs2, qsms, masks, D)
loss = loss1 + loss2
optimizer.zero_grad()
loss.backward()
optimizer.step()
loss1_sum += loss1
loss2_sum += loss2
gen_iterations += 1
scheduler.step(epoch)
# validation phase
resnet.eval(), unet3d.eval()
loss_total = 0
idx = 0
with torch.no_grad(): # to solve memory exploration issue
for idx, (rdfs, masks, qsms) in enumerate(valLoader):
idx += 1
rdfs = (rdfs.to(device, dtype=torch.float) + trans) * scale
qsms = (qsms.to(device, dtype=torch.float) + trans) * scale
masks = masks.to(device, dtype=torch.float)
outputs1 = unet3d(rdfs)
outputs2 = resnet(torch.cat((rdfs, outputs1), 1))
loss1 = loss_QSMnet(outputs1, qsms, masks, D)
loss2 = loss_QSMnet(outputs2, qsms, masks, D)
loss = loss1 + loss2
loss_total += loss
'images',
None,
False,
0,
transform=transform_test)
dataloader = torch.utils.data.DataLoader(dataset,
batch_size=1,
shuffle=False,
num_workers=8)
print('model initalize')
net = Unet(1, 1, 16).cuda()
net = nn.DataParallel(net)
print('model load')
net.load_state_dict(torch.load('./ckpt/unet.pth'))
net.eval()
with open('output/result.csv', 'w') as f:
f.write('id,rle_mask\n')
for batch_image, batch_name in tqdm(dataloader):
outputs = net(batch_image)
outputs = F.softmax(outputs, dim=1)[:, 1, :, :]
outputs = outputs > 0.50
# pdb.set_trace()
for k, v in zip(batch_name, outputs):
run = rle_encode(np.array(v))
f.write('{},{}\n'.format(k[:-4], run))
pass