netMix.eval()
files = os.listdir(test_img_path)
for f in files:
if f == 'ying.jpg':
t0 = time.time()
print('picture_name:',f)
im = getImage(os.path.join(test_img_path,f),bDel = True)
im = im.to(device)
mask = getImage('samples/ying_maskk_dilated.png',bDel = True)
mask = mask.to(device)
im1=im
im = torch.mul(im,mask)
fixnoise2 = torch.FloatTensor(1,nz, im.shape[2] // 2 ** nDep, im.shape[3] // 2 ** nDep)
fixnoise2 = fixnoise2.to(device)
fixnoise2=setNoise(fixnoise2)
'''
else:
if False:
print('false')
drift=(fixnoise2*1.0).uniform_(-1, 1)
fixnoise2[:, zGL:fixnoise2.shape[1]-zPeriodic]+=0.05*drift[:, zGL:fixnoise2.shape[1]-zPeriodic]
'''
fakebig =ganGeneration(im, fixnoise2)
print(fakebig.shape)
fakebig = fakebig
print(f,"test image size", im.shape,"inference time", -t0+time.time())
content = content.to(device)
if epoch==0 and i==0:
print ("template size",templatePatch.shape)
# train with real
netD.zero_grad()
text, _ = data
batch_size = content.size(0)##if we use texture and content of diff size may have issue -- just trim
text=text.to(device)
output = netD(text)##used to find correct size for label
errD_real = criterion(output, output.detach()*0+real_label)
errD_real.backward()
D_x = output.mean()
# train with fake
noise=setNoise(noise)
fake, alpha, A, mixedI = famosGeneration(content, noise, templatePatch, True)
output = netD(fake.detach())#???why detach
errD_fake = criterion(output, output.detach()*0+fake_label)
errD_fake.backward()
if opt.fAdvM > 0:
loss_adv_mixed = criterion(netD(mixedI.detach()), output.detach() * 0 + fake_label)
loss_adv_mixed.backward()
D_G_z1 = output.mean()
errD = errD_real + errD_fake
if opt.WGAN:
gradient_penalty = calc_gradient_penalty(netD, text, fake[:text.shape[0]])##for case fewer text images
gradient_penalty.backward()