def __init__(self, options):
super(Stage2Trainer, self).__init__(options, subfolders = ['samples', 'reconstructions'], copy_keys = copy_keys)
transforms = []
if options.crop_size is not None:
transforms.append(T.CenterCrop(options.crop_size))
transforms.append(T.Resize(options.image_size))
transforms.append(T.ToTensor())
transforms.append(T.Normalize((0.5, 0.5, 0.5, 0), (0.5, 0.5, 0.5, 1)))
image_set = ImageFolder(options.data_root, transform = T.Compose(transforms))
enc_codes = torch.load(os.path.join(options.enc_path, 'codes', '{0}_codes.pt'.format(options.enc_iter)))
code_set = torch.utils.data.TensorDataset(enc_codes[:, 0], enc_codes[:, 1])
self.dataset = ParallelDataset(image_set, code_set)
self.dataloader = torch.utils.data.DataLoader(self.dataset, batch_size = options.batch_size, shuffle = True, drop_last = True, num_workers = options.nloader)
self.data_iter = iter(self.dataloader)
enc_stats = torch.load(os.path.join(options.enc_path, 'codes', '{0}_stats.pt'.format(options.enc_iter)))
self.con_full_mean = enc_stats['full_mean']
self.con_full_std = enc_stats['full_std']
self.con_eigval = enc_stats['eigval']
self.con_eigvec = enc_stats['eigvec']
self.dim_weight = enc_stats['dim_weight']
if self.con_weight > 0:
self.enc = models.Encoder(options.image_size, options.image_size, options.enc_features, options.enc_blocks, options.enc_adain_features, options.enc_adain_blocks, options.content_size)
self.enc.to(self.device)
self.enc.load_state_dict(torch.load(os.path.join(options.enc_path, 'models', '{0}_enc.pt'.format(options.enc_iter)), map_location = self.device))
self.gen = models.TwoPartNestedDropoutGenerator(options.image_size, options.image_size, options.gen_features, options.gen_blocks, options.gen_adain_features, options.gen_adain_blocks, options.content_size, options.style_size)
self.gen.to(self.device)
if (self.load_path is None) and not options.reset_gen:
self.gen.load_state_dict(torch.load(os.path.join(options.enc_path, 'models', '{0}_gen.pt'.format(options.enc_iter)), map_location = self.device))
self.gen_optim = optim.RMSprop(self.gen.parameters(), lr = self.lr, eps = 1e-4)
self.add_model('gen', self.gen, self.gen_optim)
self.cla = models.ClassifierOrDiscriminator(options.image_size, options.image_size, options.cla_features, options.cla_blocks, options.cla_adain_features, options.cla_adain_blocks, self.nclass)
self.cla.to(self.device)
if (self.load_path is None) and (options.cla_path is not None) and not options.reset_cla:
self.cla.load_state_dict(torch.load(os.path.join(options.cla_path, 'models', '{0}_cla.pt'.format(options.cla_iter)), map_location = self.device))
self.cla_optim = optim.RMSprop(self.cla.parameters(), lr = self.lr, eps = 1e-4)
self.add_model('cla', self.cla, self.cla_optim)
self.dis = models.ClassifierOrDiscriminator(options.image_size, options.image_size, options.dis_features, options.dis_blocks, options.dis_adain_features, options.dis_adain_blocks, self.nclass)
self.dis.to(self.device)
if (self.load_path is None) and (options.cla_path is not None) and not options.reset_dis:
self.dis.load_state_dict(torch.load(os.path.join(options.cla_path, 'models', '{0}_cla.pt'.format(options.cla_iter)), map_location = self.device))
self.dis.convert()
self.dis_optim = optim.RMSprop(self.dis.parameters(), lr = self.lr, eps = 1e-4)
self.add_model('dis', self.dis, self.dis_optim)
self.sty = models.NormalizedStyleBank(self.nclass, options.style_size, image_set.get_class_freq())
self.sty.to(self.device)
if (self.load_path is None) and not options.reset_sty:
self.sty.load_state_dict(torch.load(os.path.join(options.enc_path, 'models', '{0}_sty.pt'.format(options.enc_iter)), map_location = self.device))
self.sty_optim = optim.Adam(self.sty.parameters(), lr = self.sty_lr, eps = 1e-8)
self.add_model('sty', self.sty, self.sty_optim)
if self.load_path is not None:
rec_images = torch.load(os.path.join(self.load_path, 'reconstructions', 'images.pt'), map_location = self.device)
self.rec_codes = torch.load(os.path.join(self.load_path, 'reconstructions', 'codes.pt'), map_location = self.device)
self.rec_labels = torch.load(os.path.join(self.load_path, 'reconstructions', 'labels.pt'), map_location = self.device)
self.vis_codes = torch.load(os.path.join(self.load_path, 'samples', 'codes.pt'), map_location = self.device)
self.vis_style_noise = torch.load(os.path.join(self.load_path, 'samples', 'style_noise.pt'), map_location = self.device)
self.load(options.load_iter)
else:
rec_images = []
rec_codes = []
rec_labels = []
rec_index = random.sample(range(len(self.dataset)), options.vis_row * options.vis_col)
for k in rec_index:
image, label, code, _ = self.dataset[k]
rec_images.append(image)
rec_codes.append(code)
rec_labels.append(label)
rec_images = torch.stack(rec_images, dim = 0)
self.rec_codes = torch.stack(rec_codes, dim = 0).to(self.device)
self.rec_labels = one_hot(torch.tensor(rec_labels, dtype = torch.int32), self.nclass).to(self.device)
self.vis_codes = self.noise_to_con_code(gaussian_noise(options.vis_row * options.vis_col, options.content_size)).to(self.device)
self.vis_style_noise = gaussian_noise(options.vis_row * options.vis_col, options.style_size).to(self.device)
self.state.dis_total_batches = 0
if self.save_path != self.load_path:
torch.save(rec_images, os.path.join(self.save_path, 'reconstructions', 'images.pt'))
torch.save(self.rec_codes, os.path.join(self.save_path, 'reconstructions', 'codes.pt'))
torch.save(self.rec_labels, os.path.join(self.save_path, 'reconstructions', 'labels.pt'))
torch.save(self.vis_codes, os.path.join(self.save_path, 'samples', 'codes.pt'))
torch.save(self.vis_style_noise, os.path.join(self.save_path, 'samples', 'style_noise.pt'))
save_image(rec_images.add(1).div(2), os.path.join(self.save_path, 'reconstructions', 'target.png'), self.vis_col)
self.add_periodic_func(self.visualize_fixed, options.visualize_iter)
self.visualize_fixed()
self.loss_avg_factor = 0.9
self.sty_drop_prob = torch.Tensor(options.style_size)
for i in range(options.style_size):
self.sty_drop_prob[i] = options.style_dropout ** i
def __init__(self, options):
super(Stage1Trainer, self).__init__(options,
subfolders=['reconstructions'],
copy_keys=copy_keys)
transforms = []
if options.crop_size is not None:
transforms.append(T.CenterCrop(options.crop_size))
transforms.append(T.Resize(options.image_size))
transforms.append(T.ToTensor())
image_transforms = transforms + [
T.Normalize((0.5, 0.5, 0.5, 0), (0.5, 0.5, 0.5, 1))
]
image_set = ImageFolder(options.data_root,
transform=T.Compose(image_transforms))
if options.weight_root is not None:
self.has_weight = True
weight_transforms = transforms + [lambda x: x[0]]
weight_set = ImageFolder(options.weight_root,
transform=T.Compose(weight_transforms))
self.dataset = ParallelDataset(image_set, weight_set)
else:
self.has_weight = False
self.dataset = image_set
self.dataloader = torch.utils.data.DataLoader(
self.dataset,
batch_size=options.batch_size,
shuffle=True,
drop_last=True,
num_workers=options.nloader)
self.data_iter = iter(self.dataloader)
self.enc = models.Encoder(options.image_size, options.image_size,
options.enc_features, options.enc_blocks,
options.enc_adain_features,
options.enc_adain_blocks,
options.content_size)
self.enc.to(self.device)
self.enc_optim = optim.Adam(self.enc.parameters(),
lr=self.lr,
eps=1e-4)
self.add_model('enc', self.enc, self.enc_optim)
self.gen = models.TwoPartNestedDropoutGenerator(
options.image_size, options.image_size, options.gen_features,
options.gen_blocks, options.gen_adain_features,
options.gen_adain_blocks, options.content_size, options.style_size)
self.gen.to(self.device)
self.gen_optim = optim.Adam(self.gen.parameters(),
lr=self.lr,
eps=1e-4)
self.add_model('gen', self.gen, self.gen_optim)
if self.mlp:
self.cla = models.MLPClassifier(options.content_size,
options.mlp_features,
options.mlp_layers, self.nclass)
else:
self.cla = models.ClassifierOrDiscriminator(
options.image_size, options.image_size, options.cla_features,
options.cla_blocks, options.cla_adain_features,
options.cla_adain_blocks, self.nclass)
self.cla.to(self.device)
self.cla_optim = optim.Adam(self.cla.parameters(),
lr=self.lr,
eps=1e-4)
self.add_model('cla', self.cla, self.cla_optim)
self.sty = models.NormalizedStyleBank(self.nclass, options.style_size,
image_set.get_class_freq())
self.sty.to(self.device)
self.sty_optim = optim.Adam(self.sty.parameters(),
lr=self.sty_lr,
eps=1e-8)
self.add_model('sty', self.sty, self.sty_optim)
if self.load_path is not None:
self.vis_images = torch.load(os.path.join(self.load_path,
'reconstructions',
'images.pt'),
map_location=self.device)
self.vis_labels = torch.load(os.path.join(self.load_path,
'reconstructions',
'labels.pt'),
map_location=self.device)
if self.has_weight:
self.vis_weights = torch.load(os.path.join(
self.load_path, 'reconstructions', 'weights.pt'),
map_location=self.device)
self.load(options.load_iter)
else:
vis_images = []
vis_labels = []
if self.has_weight:
vis_weights = []
vis_index = random.sample(range(len(image_set)),
options.vis_row * options.vis_col)
for k in vis_index:
image, label = image_set[k]
vis_images.append(image)
vis_labels.append(label)
if self.has_weight:
weight, _ = weight_set[k]
vis_weights.append(weight)
self.vis_images = torch.stack(vis_images, dim=0).to(self.device)
self.vis_labels = one_hot(
torch.tensor(vis_labels, dtype=torch.int32),
self.nclass).to(self.device)
if self.has_weight:
self.vis_weights = torch.stack(vis_weights,
dim=0).to(self.device)
if self.save_path != self.load_path:
torch.save(
self.vis_images,
os.path.join(self.save_path, 'reconstructions', 'images.pt'))
torch.save(
self.vis_labels,
os.path.join(self.save_path, 'reconstructions', 'labels.pt'))
save_image(
self.vis_images.add(1).div(2),
os.path.join(self.save_path, 'reconstructions', 'target.png'),
self.vis_col)
if self.has_weight:
torch.save(
self.vis_weights,
os.path.join(self.save_path, 'reconstructions',
'weights.pt'))
save_image(
self.vis_weights.unsqueeze(1),
os.path.join(self.save_path, 'reconstructions',
'weight.png'), self.vis_col)
self.add_periodic_func(self.visualize_fixed, options.visualize_iter)
self.visualize_fixed()
self.con_drop_prob = torch.Tensor(options.content_size)
for i in range(options.content_size):
self.con_drop_prob[i] = options.content_dropout**i
self.sty_drop_prob = torch.Tensor(options.style_size)
for i in range(options.style_size):
self.sty_drop_prob[i] = options.style_dropout**i