def __init__(self,
text,
*,
lr=1e-5,
batch_size=4,
gradient_accumulate_every=4,
save_every=100,
image_width=512,
num_layers=16,
epochs=20,
iterations=1050,
save_progress=False,
seed=None,
open_folder=True):
super().__init__()
if exists(seed):
print(f'setting seed: {seed}')
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
self.epochs = epochs
self.iterations = iterations
total_batches = epochs * iterations * batch_size * gradient_accumulate_every
model = DeepDaze(total_batches=total_batches,
batch_size=batch_size,
image_width=image_width,
num_layers=num_layers).cuda()
self.model = model
self.scaler = GradScaler()
self.optimizer = Adam(model.parameters(), lr)
self.gradient_accumulate_every = gradient_accumulate_every
self.save_every = save_every
self.text = text
textpath = self.text.replace(' ', '_')
self.textpath = textpath
self.filename = Path(f'./{textpath}.png')
self.save_progress = save_progress
self.encoded_text = tokenize(text).cuda()
self.open_folder = open_folder
def create_text_encoding(self, text):
tokenized_text = tokenize(text).cuda()
with torch.no_grad():
text_encoding = perceptor.encode_text(tokenized_text).detach()
return text_encoding
def create_text_encoding(text):
tokenized_text = tokenize(text).cuda()
text_encoding = perceptor.encode_text(tokenized_text).detach()
return text_encoding
def __init__(
self,
text,
*,
lr=1e-5,
batch_size=4,
gradient_accumulate_every=4,
save_every=100,
image_width=512,
num_layers=16,
epochs=20,
iterations=1050,
save_progress=False,
seed=None,
open_folder=True,
save_date_time=False,
start_image_path=None,
start_image_train_iters=10,
start_image_lr=3e-4,
theta_initial=None,
theta_hidden=None,
):
super().__init__()
if exists(seed):
tqdm.write(f'setting seed: {seed}')
torch.manual_seed(seed)
torch.cuda.manual_seed(seed)
random.seed(seed)
torch.backends.cudnn.deterministic = True
self.epochs = epochs
self.iterations = iterations
total_batches = epochs * iterations * batch_size * gradient_accumulate_every
model = DeepDaze(total_batches=total_batches,
batch_size=batch_size,
image_width=image_width,
num_layers=num_layers,
theta_initial=theta_initial,
theta_hidden=theta_hidden).cuda()
self.model = model
self.scaler = GradScaler()
self.optimizer = AdamP(model.parameters(), lr)
self.gradient_accumulate_every = gradient_accumulate_every
self.save_every = save_every
self.save_date_time = save_date_time
self.open_folder = open_folder
self.save_progress = save_progress
self.text = text
self.textpath = text.replace(" ", "_")
self.filename = self.image_output_path()
self.encoded_text = tokenize(text).cuda()
self.start_image = None
self.start_image_train_iters = start_image_train_iters
self.start_image_lr = start_image_lr
if exists(start_image_path):
file = Path(start_image_path)
assert file.exists(
), f'file does not exist at given starting image path {self.start_image_path}'
image = Image.open(str(file))
transform = T.Compose([
T.Resize(image_width),
T.CenterCrop((image_width, image_width)),
T.ToTensor(),
T.Normalize(0.5, 0.5)
])
image_tensor = transform(image)[None, ...].cuda()
self.start_image = image_tensor