def __init__(
self,
num_cutouts = 128,
loss_coef = 100,
image_size = 512,
bilinear = False,
max_classes = None,
class_temperature = 2.,
experimental_resample = False,
ema_decay = 0.99,
center_bias = False,
larger_clip = False
):
super().__init__()
self.loss_coef = loss_coef
self.image_size = image_size
self.num_cutouts = num_cutouts
self.experimental_resample = experimental_resample
self.center_bias = center_bias
self.interpolation_settings = {'mode': 'bilinear', 'align_corners': False} if bilinear else {'mode': 'nearest'}
model_name = 'ViT-B/32' if not larger_clip else 'ViT-L/14'
self.perceptor, self.normalize_image = load(model_name, jit = False)
self.model = Model(
image_size = image_size,
max_classes = max_classes,
class_temperature = class_temperature,
ema_decay = ema_decay
)
offset_y = int(random.gauss(mu=center, sigma=std))
# resample uniformly if over boundaries
offset_x = random.randint(min_offset, max_offset) if (
offset_x > max_offset or offset_x < min_offset) else offset_x
offset_y = random.randint(min_offset, max_offset) if (
offset_y > max_offset or offset_y < min_offset) else offset_y
else:
offset_x = random.randint(min_offset, max_offset)
offset_y = random.randint(min_offset, max_offset)
cutout = image[:, :, offset_x:offset_x + size, offset_y:offset_y + size]
return cutout
# load clip
perceptor, normalize_image = load('ViT-B/32', jit=False)
# load biggan
class Latents(torch.nn.Module):
def __init__(self,
num_latents=15,
num_classes=1000,
z_dim=128,
max_classes=None,
class_temperature=2.):
super().__init__()
self.normu = torch.nn.Parameter(
torch.zeros(num_latents, z_dim).normal_(std=1))
self.cls = torch.nn.Parameter(
topk_tensors = []
for i in range(k):
is_last = i == (k - 1)
values, indices = (x / temperature).softmax(dim=-1).topk(1, dim=-1)
topks = torch.zeros_like(x).scatter_(-1, indices, values)
topk_tensors.append(topks)
if not is_last:
x = x.scatter(-1, indices, float('-inf'))
topks = torch.cat(topk_tensors, dim=-1)
return topks.reshape(n, k, dim).sum(dim = 1)
# load clip
perceptor, preprocess = load()
# load biggan
class Latents(torch.nn.Module):
def __init__(
self,
num_latents = 32,
max_classes = None,
class_temperature = 2.
):
super().__init__()
self.normu = torch.nn.Parameter(torch.zeros(num_latents, 128).normal_(std = 1))
self.cls = torch.nn.Parameter(torch.zeros(num_latents, 1000).normal_(mean = -3.9, std = .3))
self.register_buffer('thresh_lat', torch.tensor(1))
offset_x = (random.randint(min_offset, max_offset) if
(offset_x > max_offset
or offset_x < min_offset) else offset_x)
offset_y = (random.randint(min_offset, max_offset) if
(offset_y > max_offset
or offset_y < min_offset) else offset_y)
else:
offset_x = random.randint(min_offset, max_offset)
offset_y = random.randint(min_offset, max_offset)
cutout = image[:, :, offset_x:offset_x + size, offset_y:offset_y + size]
return cutout
# load clip
perceptor, normalize_image = load("ViT-B/32", jit=False)
# load biggan
class Latents(torch.nn.Module):
def __init__(
self,
num_latents=15,
num_classes=1000,
z_dim=128,
max_classes=None,
class_temperature=2.0,
):
super().__init__()
self.normu = torch.nn.Parameter(
