def initialize(self, opt):
super(PoseParsingModel, self).initialize(opt)
###################################
# create model
###################################
if opt.which_model_PP == 'resnet':
self.netPP = networks.ResnetGenerator(
input_nc = self.get_data_dim(opt.pp_input_type),
output_nc = self.get_data_dim(opt.pp_pose_type),
ngf = opt.pp_nf,
norm_layer = networks.get_norm_layer(opt.norm),
activation = nn.ReLU,
use_dropout = False,
n_blocks = opt.pp_nblocks,
gpu_ids = opt.gpu_ids,
output_tanh = False,
)
elif opt.which_model_PP == 'unet':
self.netPP = networks.UnetGenerator_v2(
input_nc = self.get_data_dim(opt.pp_input_type),
output_nc = self.get_data_dim(opt.pp_pose_type),
num_downs = 8,
ngf = opt.pp_nf,
max_nf = opt.pp_nf*(2**3),
norm_layer = networks.get_norm_layer(opt.norm),
use_dropout = False,
gpu_ids = opt.gpu_ids,
output_tanh = False,
)
else:
raise NotImplementedError()
if opt.gpu_ids:
self.netPP.cuda()
###################################
# init/load model
###################################
if self.is_train and (not opt.continue_train):
networks.init_weights(self.netPP, init_type=opt.init_type)
else:
self.load_network(self.netPP, 'netPP', opt.which_epoch)
###################################
# optimizers and schedulers
###################################
if self.is_train:
self.optim = torch.optim.Adam(self.netPP.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2))
self.optimizers = [self.optim]
self.schedulers = []
for optim in self.optimizers:
self.schedulers.append(networks.get_scheduler(optim, opt))
def initialize(self, opt):
super(DomainTransferModel, self).initialize()
###################################
# define Enc_A and Dec_B (VUnet)
###################################
self.netA = networks.VariationalUnet(
input_nc_dec=self.get_pose_dim(opt.pose_type),
input_nc_enc=3,
output_nc=3,
nf=opt.vunet_nf,
max_nf=opt.vunet_max_nf,
input_size=opt.fine_size,
n_latent_scales=opt.vunet_n_latent_scales,
bottleneck_factor=opt.vunet_bottleneck_factor,
box_factor=opt.vunet_box_factor,
n_residual_blocks=2,
norm_layer=networks.get_norm_layer(opt.norm),
activation=nn.ReLU(False),
use_dropout=False,
gpu_ids=opt.gpu_ids,
)
if opt.gpu_ids:
self.netA.cuda()
networks.init_weights(self.netA, init_type=opt.init_type)
###################################
# define Enc_B and Dec_B (VAE)
###################################
self.netB = networks.VariationalAutoEncoder(
input_nc=3,
output_nc=3,
nf=opt.vae_nf,
max_nf=opt.vae_max_nf,
latent_nf=opt.vae_latent_nf,
input_size=opt.fine_size,
bottleneck_factor=opt.vae_bottleneck_factor,
n_residual_blocks=2,
norm_layer=networks.get_norm_layer(opt.norm),
activation=nn.ReLU(False),
use_dropout=False,
gpu_ids=opt.gpu_ids,
)
if opt.gpu_ids:
self.netB.cuda()
network.init_weights(self.netB, init_type=opt.init_type)
###################################
# define feature transfer network
###################################
self.netFT = networks.VUnetLatentTransformer()
def __init__(self, gpu_ids=[0],load_model=None):
self.gpu = gpu_ids
norm_layer = networks.get_norm_layer('instance')
self.net =networks.UnetGenerator(3,1,8,64,use_dropout=False,norm_layer = norm_layer,gpu_ids = gpu_ids)
if load_model is not None:
self.net.load_state_dict(torch.load(load_model))
self.net.cuda(0)
def _create_stage_1_net(self, opt):
'''
stage-1 network should be a pretrained pose transfer model.
assume it is a vunet for now
'''
# load options
opt_s1 = argparse.Namespace()
dict_opt_s1 = io.load_json(
os.path.join('checkpoints', opt.which_model_stage_1,
'train_opt.json'))
opt_s1.__dict__.update(dict_opt_s1)
self.opt_s1 = opt_s1
# create model
if opt_s1.which_model_T == 'vunet':
self.netT_s1 = networks.VariationalUnet(
input_nc_dec=self.get_pose_dim(opt_s1.pose_type),
input_nc_enc=self.get_appearance_dim(opt_s1.appearance_type),
output_nc=self.get_output_dim(opt_s1.output_type),
nf=opt_s1.vunet_nf,
max_nf=opt_s1.vunet_max_nf,
input_size=opt_s1.fine_size,
n_latent_scales=opt_s1.vunet_n_latent_scales,
bottleneck_factor=opt_s1.vunet_bottleneck_factor,
box_factor=opt_s1.vunet_box_factor,
n_residual_blocks=2,
norm_layer=networks.get_norm_layer(opt_s1.norm),
activation=nn.ReLU(False),
use_dropout=False,
gpu_ids=opt.gpu_ids,
output_tanh=False,
)
if opt.gpu_ids:
self.netT_s1.cuda()
else:
raise NotImplementedError()
def __init__(self, opt):
"""Initialize the pix2pix class.
Parameters:
opt (Option class)-- stores all the experiment flags; needs to be a subclass of BaseOptions
"""
BaseModel.__init__(self, opt)
# specify the training losses you want to print out. The training/test scripts will call <BaseModel.get_current_losses>
self.loss_names = ['G_GAN', 'G_L1', 'D_real', 'D_fake']
# specify the images you want to save/display. The training/test scripts will call <BaseModel.get_current_visuals>
self.visual_names = ['real_A', 'fake_B', 'real_B']
# specify the models you want to save to the disk. The training/test scripts will call <BaseModel.save_networks> and <BaseModel.load_networks>
if self.isTrain:
self.model_names = ['G', 'D']
else: # during test time, only load G
self.model_names = ['G']
# define norm layer
norm_layer = networks.get_norm_layer(opt.norm)
# define networks (both generator and discriminator)
self.netG = self.define_net(
UnetGenerator(opt.input_nc, opt.output_nc, 8, opt.ngf, norm_layer,
not opt.no_dropout))
if self.isTrain: # define a discriminator; conditional GANs need to take both input and output images; Therefore, #channels for D is input_nc + output_nc
self.netD = self.define_net(
NLayerDiscriminator(opt.input_nc + opt.output_nc,
opt.ndf,
norm_layer=norm_layer))
if self.isTrain:
# define loss functions
self.criterionGAN = GANLoss(opt.gan_mode).to(self.device)
self.criterionL1 = torch.nn.L1Loss()
# initialize optimizers; schedulers will be automatically created by function <BaseModel.setup>.
self.optimizer_G = torch.optim.Adam(self.netG.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
self.optimizer_D = torch.optim.Adam(self.netD.parameters(),
lr=opt.lr,
betas=(opt.beta1, 0.999))
self.optimizers.append(self.optimizer_G)
self.optimizers.append(self.optimizer_D)
def initialize(self, opt):
super(SupervisedPoseTransferModel, self).initialize(opt)
###################################
# define transformer
###################################
if opt.which_model_T == 'resnet':
self.netT = networks.ResnetGenerator(
input_nc=3 + self.get_pose_dim(opt.pose_type),
output_nc=3,
ngf=opt.T_nf,
norm_layer=networks.get_norm_layer(opt.norm),
use_dropout=not opt.no_dropout,
n_blocks=9,
gpu_ids=opt.gpu_ids)
elif opt.which_model_T == 'unet':
self.netT = networks.UnetGenerator_v2(
input_nc=3 + self.get_pose_dim(opt.pose_type),
output_nc=3,
num_downs=8,
ngf=opt.T_nf,
norm_layer=networks.get_norm_layer(opt.norm),
use_dropout=not opt.no_dropout,
gpu_ids=opt.gpu_ids)
else:
raise NotImplementedError()
if opt.gpu_ids:
self.netT.cuda()
networks.init_weights(self.netT, init_type=opt.init_type)
###################################
# define discriminator
###################################
self.use_GAN = self.is_train and opt.loss_weight_gan > 0
if self.use_GAN > 0:
self.netD = networks.define_D_from_params(
input_nc=3 +
self.get_pose_dim(opt.pose_type) if opt.D_cond else 3,
ndf=opt.D_nf,
which_model_netD='n_layers',
n_layers_D=3,
norm=opt.norm,
which_gan=opt.which_gan,
init_type=opt.init_type,
gpu_ids=opt.gpu_ids)
else:
self.netD = None
###################################
# loss functions
###################################
if self.is_train:
self.loss_functions = []
self.schedulers = []
self.optimizers = []
self.crit_L1 = nn.L1Loss()
self.crit_vgg = networks.VGGLoss_v2(self.gpu_ids)
# self.crit_vgg_old = networks.VGGLoss(self.gpu_ids)
self.crit_psnr = networks.PSNR()
self.crit_ssim = networks.SSIM()
self.loss_functions += [self.crit_L1, self.crit_vgg]
self.optim = torch.optim.Adam(self.netT.parameters(),
lr=opt.lr,
betas=(opt.beta1, opt.beta2))
self.optimizers += [self.optim]
if self.use_GAN:
self.crit_GAN = networks.GANLoss(
use_lsgan=opt.which_gan == 'lsgan', tensor=self.Tensor)
self.optim_D = torch.optim.Adam(self.netD.parameters(),
lr=opt.lr_D,
betas=(opt.beta1, opt.beta2))
self.loss_functions.append(self.use_GAN)
self.optimizers.append(self.optim_D)
# todo: add pose loss
for optim in self.optimizers:
self.schedulers.append(networks.get_scheduler(optim, opt))
self.fake_pool = ImagePool(opt.pool_size)
###################################
# load trained model
###################################
if not self.is_train:
self.load_network(self.netT, 'netT', opt.which_model)
def initialize(self, opt):
super(VUnetPoseTransferModel, self).initialize(opt)
###################################
# define transformer
###################################
self.netT = networks.VariationalUnet(
input_nc_dec = self.get_pose_dim(opt.pose_type),
input_nc_enc = self.get_appearance_dim(opt.appearance_type),
output_nc = self.get_output_dim(opt.output_type),
nf = opt.vunet_nf,
max_nf = opt.vunet_max_nf,
input_size = opt.fine_size,
n_latent_scales = opt.vunet_n_latent_scales,
bottleneck_factor = opt.vunet_bottleneck_factor,
box_factor = opt.vunet_box_factor,
n_residual_blocks = 2,
norm_layer = networks.get_norm_layer(opt.norm),
activation = nn.ReLU(False),
use_dropout = False,
gpu_ids = opt.gpu_ids,
output_tanh = False,
)
if opt.gpu_ids:
self.netT.cuda()
networks.init_weights(self.netT, init_type=opt.init_type)
###################################
# define discriminator
###################################
self.use_GAN = self.is_train and opt.loss_weight_gan > 0
if self.use_GAN:
self.netD = networks.define_D_from_params(
input_nc=3+self.get_pose_dim(opt.pose_type) if opt.D_cond else 3,
ndf=opt.D_nf,
which_model_netD='n_layers',
n_layers_D=opt.D_n_layer,
norm=opt.norm,
which_gan=opt.which_gan,
init_type=opt.init_type,
gpu_ids=opt.gpu_ids)
else:
self.netD = None
###################################
# loss functions
###################################
self.crit_psnr = networks.PSNR()
self.crit_ssim = networks.SSIM()
if self.is_train:
self.optimizers =[]
self.crit_vgg = networks.VGGLoss_v2(self.gpu_ids, opt.content_layer_weight, opt.style_layer_weight, opt.shifted_style)
# self.crit_vgg_old = networks.VGGLoss(self.gpu_ids)
self.optim = torch.optim.Adam(self.netT.parameters(), lr=opt.lr, betas=(opt.beta1, opt.beta2), weight_decay=opt.weight_decay)
self.optimizers += [self.optim]
if self.use_GAN:
self.crit_GAN = networks.GANLoss(use_lsgan=opt.which_gan=='lsgan', tensor=self.Tensor)
self.optim_D = torch.optim.Adam(self.netD.parameters(), lr=opt.lr_D, betas=(opt.beta1, opt.beta2))
self.optimizers.append(self.optim_D)
# todo: add pose loss
self.fake_pool = ImagePool(opt.pool_size)
###################################
# load trained model
###################################
if not self.is_train:
self.load_network(self.netT, 'netT', opt.which_epoch)
elif opt.continue_train:
self.load_network(self.netT, 'netT', opt.which_epoch)
self.load_optim(self.optim, 'optim', opt.which_epoch)
if self.use_GAN:
self.load_network(self.netD, 'netD', opt.which_epoch)
self.load_optim(self.optim_D, 'optim_D', opt.which_epoch)
###################################
# schedulers
###################################
if self.is_train:
self.schedulers = []
for optim in self.optimizers:
self.schedulers.append(networks.get_scheduler(optim, opt))
def initialize(self, opt):
super(TwoStagePoseTransferModel, self).initialize(opt)
###################################
# load pretrained stage-1 (coarse) network
###################################
self._create_stage_1_net(opt)
###################################
# define stage-2 (refine) network
###################################
# local patch encoder
if opt.which_model_s2e == 'patch_embed':
self.netT_s2e = networks.LocalPatchEncoder(
n_patch=len(opt.patch_indices),
input_nc=3,
output_nc=opt.s2e_nof,
nf=opt.s2e_nf,
max_nf=opt.s2e_max_nf,
input_size=opt.patch_size,
bottleneck_factor=opt.s2e_bottleneck_factor,
n_residual_blocks=2,
norm_layer=networks.get_norm_layer(opt.norm),
activation=nn.ReLU(False),
use_dropout=False,
gpu_ids=opt.gpu_ids,
)
s2e_nof = opt.s2e_nof
elif opt.which_model_s2e == 'patch':
self.netT_s2e = networks.LocalPatchRearranger(
n_patch=len(opt.patch_indices),
image_size=opt.fine_size,
)
s2e_nof = 3
elif opt.which_model_s2e == 'seg_embed':
self.netT_s2e = networks.SegmentRegionEncoder(
seg_nc=self.opt.seg_nc,
input_nc=3,
output_nc=opt.s2e_nof,
nf=opt.s2d_nf,
input_size=opt.fine_size,
n_blocks=3,
norm_layer=networks.get_norm_layer(opt.norm),
activation=nn.ReLU,
use_dropout=False,
grid_level=opt.s2e_grid_level,
gpu_ids=opt.gpu_ids,
)
s2e_nof = opt.s2e_nof + opt.s2e_grid_level
else:
raise NotImplementedError()
if opt.gpu_ids:
self.netT_s2e.cuda()
# decoder
if self.opt.which_model_s2d == 'resnet':
self.netT_s2d = networks.ResnetGenerator(
input_nc=3 + s2e_nof,
output_nc=3,
ngf=opt.s2d_nf,
norm_layer=networks.get_norm_layer(opt.norm),
activation=nn.ReLU,
use_dropout=False,
n_blocks=opt.s2d_nblocks,
gpu_ids=opt.gpu_ids,
output_tanh=False,
)
elif self.opt.which_model_s2d == 'unet':
self.netT_s2d = networks.UnetGenerator_v2(
input_nc=3 + s2e_nof,
output_nc=3,
num_downs=8,
ngf=opt.s2d_nf,
max_nf=opt.s2d_nf * 2**3,
norm_layer=networks.get_norm_layer(opt.norm),
use_dropout=False,
gpu_ids=opt.gpu_ids,
output_tanh=False,
)
elif self.opt.which_model_s2d == 'rpresnet':
self.netT_s2d = networks.RegionPropagationResnetGenerator(
input_nc=3 + s2e_nof,
output_nc=3,
ngf=opt.s2d_nf,
norm_layer=networks.get_norm_layer(opt.norm),
activation=nn.ReLU,
use_dropout=False,
nblocks=opt.s2d_nblocks,
gpu_ids=opt.gpu_ids,
output_tanh=False)
else:
raise NotImplementedError()
if opt.gpu_ids:
self.netT_s2d.cuda()
###################################
# define discriminator
###################################
self.use_GAN = self.is_train and opt.loss_weight_gan > 0
if self.use_GAN:
self.netD = networks.define_D_from_params(
input_nc=3 +
self.get_pose_dim(opt.pose_type) if opt.D_cond else 3,
ndf=opt.D_nf,
which_model_netD='n_layers',
n_layers_D=opt.D_n_layer,
norm=opt.norm,
which_gan=opt.which_gan,
init_type=opt.init_type,
gpu_ids=opt.gpu_ids)
else:
self.netD = None
###################################
# loss functions
###################################
self.crit_psnr = networks.PSNR()
self.crit_ssim = networks.SSIM()
if self.is_train:
self.optimizers = []
self.crit_vgg = networks.VGGLoss_v2(self.gpu_ids,
opt.content_layer_weight,
opt.style_layer_weight,
opt.shifted_style)
self.optim = torch.optim.Adam([{
'params': self.netT_s2e.parameters()
}, {
'params': self.netT_s2d.parameters()
}],
lr=opt.lr,
betas=(opt.beta1, opt.beta2))
self.optimizers.append(self.optim)
if opt.train_s1:
self.optim_s1 = torch.optim.Adam(self.netT_s1.parameters(),
lr=opt.lr_s1,
betas=(opt.beta1, opt.beta2))
self.optimizers.append(self.optim_s1)
if self.use_GAN:
self.crit_GAN = networks.GANLoss(
use_lsgan=opt.which_gan == 'lsgan', tensor=self.Tensor)
self.optim_D = torch.optim.Adam(self.netD.parameters(),
lr=opt.lr_D,
betas=(opt.beta1, opt.beta2))
self.optimizers.append(self.optim_D)
self.fake_pool = ImagePool(opt.pool_size)
###################################
# init/load model
###################################
if self.is_train:
if not opt.continue_train:
self.load_network(self.netT_s1, 'netT', 'latest',
self.opt_s1.id)
networks.init_weights(self.netT_s2e, init_type=opt.init_type)
networks.init_weights(self.netT_s2d, init_type=opt.init_type)
if self.use_GAN:
networks.init_weights(self.netD, init_type=opt.init_type)
else:
self.load_network(self.netT_s1, 'netT_s1', opt.which_epoch)
self.load_network(self.netT_s2e, 'netT_s2e', opt.which_epoch)
self.load_network(self.netT_s2d, 'netT_s2d', opt.which_epoch)
self.load_optim(self.optim, 'optim', opt.which_epoch)
if self.use_GAN:
self.load_network(self.netD, 'netD', opt.which_epoch)
self.load_optim(self.optim_D, 'optim_D', opt.which_epoch)
else:
self.load_network(self.netT_s1, 'netT_s1', opt.which_epoch)
self.load_network(self.netT_s2e, 'netT_s2e', opt.which_epoch)
self.load_network(self.netT_s2d, 'netT_s2d', opt.which_epoch)
###################################
# schedulers
###################################
if self.is_train:
self.schedulers = []
for optim in self.optimizers:
self.schedulers.append(networks.get_scheduler(optim, opt))
