return features
sF_loss = get_features(styleV, resnet)
cF_loss = get_features(contentV, resnet)
tF = get_features(transfer, resnet)
loss, styleLoss, contentLoss = criterion(tF, sF_loss, cF_loss)
# backward & optimization
loss.backward()
optimizer.step()
print(
'Iteration: [%d/%d] Loss: %.4f contentLoss: %.4f styleLoss: %.4f Learng Rate is %.6f'
% (opt.niter, iteration, loss, contentLoss, styleLoss,
optimizer.param_groups[0]['lr']))
adjust_learning_rate(optimizer, iteration)
if ((iteration) % opt.log_interval == 0):
transfer = transfer.clamp(0, 1)
concat = torch.cat((content, style, transfer.cpu()), dim=0)
vutils.save_image(concat,
'%s/%d.png' % (opt.outf, iteration),
normalize=True,
scale_each=True,
nrow=opt.batchSize)
if (iteration > 0 and (iteration) % opt.save_interval == 0):
torch.save(matrix.state_dict(),
'%s/%s_resnet_softmax_imagenet.pth' % (opt.outf, opt.layer))
else:
feature, transmatrix = matrix(cF, sF)
transfer = dec(feature)
sF_loss = vgg5(styleV)
cF_loss = vgg5(contentV)
tF = vgg5(transfer)
loss, styleLoss, contentLoss = criterion(tF, sF_loss, cF_loss)
# backward & optimization
loss.backward()
optimizer.step()
print(
'Iteration: [%d/%d] Loss: %.4f contentLoss: %.4f styleLoss: %.4f Learng Rate is %.6f'
% (opt.niter, iteration, loss, contentLoss, styleLoss,
optimizer.param_groups[0]['lr']))
adjust_learning_rate(optimizer, iteration)
if ((iteration) % opt.log_interval == 0):
transfer = transfer.clamp(0, 1)
concat = torch.cat((content, style, transfer.data.cpu()), dim=0)
vutils.save_image(concat,
'%s/%d.png' % (opt.outf, iteration),
normalize=True,
scale_each=True,
nrow=opt.batchSize)
if (iteration > 0 and (iteration) % opt.save_interval == 0):
torch.save(matrix.state_dict(), '%s/%s.pth' % (opt.outf, opt.layer))
if str(name) in layers3:
features[layers3[str(name)]] = x
for name, layer in enumerate(model.layer4):
x = layer(x)
if str(name) in layers4:
features[layers4[str(name)]] = x
return features
sF_loss = get_features(styleV, resnet)
cF_loss = get_features(contentV, resnet)
tF = get_features(transfer, resnet)
loss,styleLoss,contentLoss = criterion(tF,sF_loss,cF_loss)
# backward & optimization
loss.backward()
optimizer.step()
print('Iteration: [%d/%d] Loss: %.4f contentLoss: %.4f styleLoss: %.4f Learng Rate is %.6f'%
(opt.niter,iteration,loss,contentLoss,styleLoss,optimizer.param_groups[0]['lr']))
adjust_learning_rate(optimizer,iteration)
if((iteration) % opt.log_interval == 0):
transfer = transfer.clamp(0,1)
concat = torch.cat((content,style,transfer.cpu()),dim=0)
vutils.save_image(concat,'%s/%d.png'%(opt.outf,iteration),normalize=True,scale_each=True,nrow=opt.batchSize)
if(iteration > 0 and (iteration) % opt.save_interval == 0):
torch.save(matrix.state_dict(), '%s/%s_resnet_random.pth' % (opt.outf,opt.layer))