emb_weights = getGloveEmbeddings('../embeddings', cache, t2i)
emb_weights = torch.tensor(emb_weights, device=device)
if args['--encoder'] == 'simple':
emb_size = 50
text_encoder = models.SimpleTextEncoder(emb_weights)
elif args['--encoder'] == 'mere':
emb_size = 64
text_encoder = models.TextEncoder(emb_weights, proj=True)
elif args['--encoder'] == 'joint':
pass
else:
raise TypeError('Invalid encoder type')
dim_Z = 50
generator = models.VideoGenerator(dim_Z, emb_size)
image_discriminator = models.ImageDiscriminator(
cond_size=emb_size, noise=args['--noise'],
sigma=float(args['--sigma']))
video_discriminator = models.VideoDiscriminator(
cond_size=emb_size, noise=args['--noise'],
sigma=float(args['--sigma']))
generator.to(device)
text_encoder.to(device)
image_discriminator.to(device)
video_discriminator.to(device)
dis_dict = {'image': image_discriminator,
'video': video_discriminator}
# Dataset of all
dataset = data.VideoFolderDataset(args['<dataset>'], cache=os.path.join(args['<dataset>'], 'local.db'))
# Object to get images from dataset VideoFolderDataset above
image_dataset = data.ImageDataset(dataset, image_transforms)
# Dataloader to load images
image_loader = DataLoader(image_dataset, batch_size=image_batch, drop_last=True, num_workers=2, shuffle=True)
# Object to get videos from dataset VideoFolderDataset above
video_dataset = data.VideoDataset(dataset, video_length=16, every_nth=2, transform=video_transforms)
# Dataloader to load videos
video_loader = DataLoader(video_dataset, batch_size=video_batch, drop_last=True, num_workers=2, shuffle=True)
# Create object of VideoGenerator
generator = models.VideoGenerator(n_channels, dim_z_content=dim_z_content, dim_z_view=dim_z_view,
dim_z_motion=dim_z_motion, dim_z_category=dim_z_category,
video_length=video_length)
# Create object of ImageDiscriminator
image_discriminator = build_discriminator(args['--image_discriminator'], n_channels=n_channels,
dim_z_view=dim_z_view, use_noise=args['--use_noise'],
noise_sigma=float(args['--noise_sigma']))
# Create object of VideoDiscriminator
video_discriminator = build_discriminator(args['--video_discriminator'], n_channels=n_channels,
dim_z_view=dim_z_view, dim_z_category=dim_z_category,
use_noise=args['--use_noise'],
noise_sigma=float(args['--noise_sigma']))
# Push all networks to GPU
print("torch.cuda.is_available() = ", torch.cuda.is_available())
image_dataset = data.ImageDataset(dataset, image_transforms)
image_loader = DataLoader(image_dataset,
batch_size=image_batch,
drop_last=True,
num_workers=2,
shuffle=True)
video_dataset = data.VideoDataset(dataset, 16, 2, video_transforms)
video_loader = DataLoader(video_dataset,
batch_size=video_batch,
drop_last=True,
num_workers=2,
shuffle=True)
generator = models.VideoGenerator(n_channels, dim_z_content,
dim_z_category, dim_z_motion,
video_length)
image_discriminator = build_discriminator(args['--image_discriminator'],
n_channels=n_channels,
use_noise=args['--use_noise'],
noise_sigma=float(
args['--noise_sigma']))
video_discriminator = build_discriminator(args['--video_discriminator'],
dim_categorical=dim_z_category,
n_channels=n_channels,
use_noise=args['--use_noise'],
noise_sigma=float(
args['--noise_sigma']))
image_batch = int(args['--image_batch'])
video_batch = int(args['--video_batch'])
dim_z_content = int(args['--dim_z_content'])
dim_z_motion = int(args['--dim_z_motion'])
dim_z_category = int(args['--dim_z_category'])
dataset = data.VideoFolderDataset(args['<dataset>'], cache=os.path.join(args['<dataset>'], 'local.db'))
image_dataset = data.ImageDataset(dataset, image_transforms)
image_loader = DataLoader(image_dataset, batch_size=image_batch, drop_last=True, num_workers=2, shuffle=True)
content_image_loader = DataLoader(image_dataset, batch_size=image_batch, drop_last=True, num_workers=2, shuffle=True)
video_dataset = data.VideoDataset(dataset, 16, 2, video_transforms)
video_loader = DataLoader(video_dataset, batch_size=video_batch, drop_last=True, num_workers=2, shuffle=True)
generator = models.VideoGenerator(n_channels, dim_z_content, dim_z_category, dim_z_motion, video_length, content_image_sampler=content_image_loader, use_cuda=use_cuda)
image_discriminator = build_discriminator(args['--image_discriminator'], n_channels=n_channels,
use_noise=args['--use_noise'], noise_sigma=float(args['--noise_sigma']))
video_discriminator = build_discriminator(args['--video_discriminator'], dim_categorical=dim_z_category,
n_channels=n_channels, use_noise=args['--use_noise'],
noise_sigma=float(args['--noise_sigma']))
if torch.cuda.is_available():
generator.cuda()
image_discriminator.cuda()
video_discriminator.cuda()
trainer = Trainer(image_loader, video_loader,
int(args['--print_every']),
video_transforms = functools.partial(video_transform,
image_transform=image_transforms)
factor = int(args["--factor"])
dim_z_view = int(args["--dim_z_view"])
dim_z_category = int(args["--dim_z_category"])
num_class = dim_z_view * dim_z_category
video_length = int(args["--video_length"])
input_path = str(args["--input_path"])
# Create object of VideoGenerator
generator = models.VideoGenerator(n_channels=3,
dim_z_content=50,
dim_z_view=dim_z_view,
dim_z_motion=10,
dim_z_category=dim_z_category,
dim_z_object=2,
video_length=video_length,
video_transforms=video_transforms,
image_transforms=image_transforms,
ngf=64)
print(generator)
print("torch.cuda.is_available() = ", torch.cuda.is_available())
if torch.cuda.is_available():
print("torch.cuda.is_available() = ", torch.cuda.is_available())
generator.cuda()
generator.load_state_dict(checkpoint['generator'])
generator.eval()
output_folder = "./test_encode_module"
if not os.path.exists(output_folder):
dim_z_category = int(args["--dim_z_category"])
num_class = dim_z_view * dim_z_category
video_length = int(args["--video_length"])
output_folder = str(args["--output_folder"])
performer_name = str(args["--performer_name"])
is_6_dau = args["--is_6_dau"]
offset = 0
if (is_6_dau == False):
offset = 6
# Create object of VideoGenerator
generator = models.VideoGenerator(n_channels=3,
dim_z_content=50,
dim_z_view=dim_z_view,
dim_z_motion=10,
dim_z_category=dim_z_category,
video_length=video_length,
ngf=64)
print(generator)
print("torch.cuda.is_available() = ", torch.cuda.is_available())
if torch.cuda.is_available():
print("torch.cuda.is_available() = ", torch.cuda.is_available())
generator.cuda()
generator.load_state_dict(checkpoint['generator'])
generator.eval()
# Create output folder
if not os.path.exists(output_folder):
os.makedirs(output_folder)
num_workers=2,
shuffle=True)
video_dataset = data.VideoDataset(dataset, 16, 2, video_transforms)
video_loader = DataLoader(video_dataset,
batch_size=video_batch,
drop_last=True,
num_workers=2,
shuffle=True)
generator = models.VideoGenerator(
n_channels,
dim_z_content,
dim_z_category,
dim_z_motion,
video_length,
use_cgan_proj_discr=use_cgan_proj_discr,
n_categories=n_categories,
n_content_categories=n_content_categories,
resnet=resnet,
content_only_style=content_only_style)
print('generator')
print(generator)
if not use_cgan_proj_discr:
image_discriminator = build_discriminator(
args['--image_discriminator'],
n_channels=n_channels,
use_noise=args['--use_noise'],
noise_sigma=float(args['--noise_sigma']))
else:
# Object to get images from dataset VideoFolderDataset above
encode_image_dataset = data.ImageDataset(dataset, image_transforms)
# Dataloader to load images
encode_image_loader = DataLoader(encode_image_dataset,
batch_size=image_batch,
drop_last=True,
num_workers=2,
shuffle=True)
# Create object of VideoGenerator
generator = models.VideoGenerator(n_channels,
dim_z_content=dim_z_content,
dim_z_view=dim_z_view,
dim_z_motion=dim_z_motion,
dim_z_category=dim_z_category,
dim_z_object=dim_z_object,
video_length=video_length,
encode_video_loader=encode_video_loader,
encode_image_loader=encode_image_loader,
video_transforms=video_transforms,
image_transforms=image_transforms)
# Create object of ImageDiscriminator
image_discriminator = build_discriminator(args['--image_discriminator'],
n_channels=n_channels,
dim_z_view=dim_z_view,
dim_z_object=dim_z_object,
use_noise=args['--use_noise'],
noise_sigma=float(
args['--noise_sigma']))
