"_ngf"+str(ngf)+"_N"+str(N)+"_dep"+str(nDep)
if opt.WGAN:
desc +='_WGAN'
if opt.LS:
desc += '_LS'
if bMirror:
desc += '_mirror'
if opt.contentScale !=1 or opt.textureScale !=1:
desc +="_scale"+str(opt.contentScale)+";"+str(opt.textureScale)
desc += '_cLoss'+str(opt.cLoss)
if not opt.coordCopy:
desc += "no coord copy"
targetMosaic,templates=getTemplates(opt,N,vis=True,path=opt.outputFolder+desc)
fixnoise = torch.FloatTensor(1, nz, targetMosaic.shape[2]//2**nDep,targetMosaic.shape[3]//2**nDep)
print("fixed variables",fixnoise.data.shape,targetMosaic.data.shape)
netD = Discriminator(ndf, opt.nDepD, bSigm=not opt.LS and not opt.WGAN)
##################################
if opt.multiScale:
netMix = NetU_MultiScale(ngf, nDep, nz, bSkip=opt.skipConnections, nc=N + 5, ncIn=5, bTanh=False, bCopyIn=opt.coordCopy, Ubottleneck=opt.Ubottleneck)
else:
netMix =NetUskip(ngf, nDep, nz, bSkip=opt.skipConnections, nc=N + 5, ncIn=5, bTanh=False, bCopyIn=opt.coordCopy, Ubottleneck=opt.Ubottleneck)##copy coords more often
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")
print ("device",device)
Gnets=[netMix]
if opt.refine:
ndf = int(opt.ndf)
desc = "fc" + str(opt.fContent) + "_ngf" + str(ngf) + "_ndf" + str(
ndf) + "_dep" + str(nDep) + "-" + str(opt.nDepD)
if opt.WGAN:
desc += '_WGAN'
if opt.LS:
desc += '_LS'
if bMirror:
desc += '_mirror'
if opt.contentScale != 1 or opt.textureScale != 1:
desc += "_scale" + str(opt.contentScale) + ";" + str(opt.textureScale)
desc += '_cLoss' + str(opt.cLoss)
targetMosaic = getTemplates(opt, N)
fixnoise = torch.FloatTensor(1, nz, targetMosaic.shape[2] // 2**nDep,
targetMosaic.shape[3] // 2**nDep)
print("fixed variables", fixnoise.data.shape, targetMosaic.data.shape)
netD = Discriminator(ndf, opt.nDepD, bSigm=not opt.LS and not opt.WGAN)
##################################
netMix = NetUskip(ngf,
nDep,
nz,
bSkip=opt.skipConnections,
nc=3,
ncIn=3,
bTanh=True,
Ubottleneck=opt.Ubottleneck)
device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu")