def build_transforms():
def update_transforms(p):
crop_size = tuple(
[round(size // 2 + (size - size // 2) * p) for size in IMAGE_SIZE])
random_crop.reset(crop_size)
center_crop.reset(crop_size)
print('update transforms p: {:.2f}, crop_size: {}'.format(
p, crop_size))
random_crop = Resettable(T.RandomCrop)
center_crop = Resettable(T.CenterCrop)
to_tensor_and_norm = T.Compose([
T.ToTensor(),
T.Normalize(np.mean((0.485, 0.456, 0.406), keepdims=True),
np.mean((0.229, 0.224, 0.225), keepdims=True)),
])
train_transform = T.Compose([
Invert(),
T.RandomAffine(degrees=15,
scale=(0.8, 1 / 0.8),
resample=Image.BILINEAR),
random_crop,
to_tensor_and_norm,
# T.RandomErasing(value=0),
])
eval_transform = T.Compose([
Invert(),
center_crop,
to_tensor_and_norm,
])
return train_transform, eval_transform, update_transforms
def build_transform():
def update_transform(image_size):
print("image_size {}".format(image_size))
resize.reset(image_size)
crop.reset(image_size)
def to_rgb(input):
return input.convert("RGB")
resize = Resettable(T.Resize)
crop = Resettable(T.CenterCrop)
transform = T.Compose([
T.Lambda(to_rgb),
resize,
crop,
T.ToTensor(),
T.Normalize(mean=[0.5], std=[0.5]),
])
return transform, update_transform
size = round(np.random.uniform(self.min_size, self.max_size))
input = Resize(size)(input)
return input
def dist_sharpen(dist, temp):
assert dist.dim() == 2
dist = dist**(1 / temp)
dist = dist / dist.sum(1, keepdim=True)
return dist
random_crop = Resettable(RandomCrop)
center_crop = Resettable(CenterCrop)
infer_image_transform = Resettable(lambda tta: test_image_transform
if tta else eval_image_transform)
to_tensor = ToTensor()
if config.normalize is None:
normalize = T.Compose([])
elif config.normalize == 'experiment':
normalize = NormalizeByExperimentStats(
torch.load('./experiment_stats.pth'))
elif config.normalize == 'plate':
normalize = NormalizeByPlateStats(torch.load('./plate_stats.pth'))
else:
raise AssertionError('invalid normalization {}'.format(config.normalize))
assert c.mode == mode
s1 = [np.array(c) for c in s1]
s2 = [np.array(c) for c in s2]
for c1, c2 in zip(s1, s2):
c1[x1:x2, y1:y2] = c2[x1:x2, y1:y2]
s1 = [Image.fromarray(c) for c in s1]
assert s1[0].mode == mode
return s1
random_resize = Resettable(RandomResize)
random_crop = Resettable(RandomCrop)
center_crop = Resettable(CenterCrop)
to_tensor = ToTensor()
if config.normalize is None:
normalize = T.Compose([])
elif config.normalize == 'experiment':
normalize = NormalizeByExperimentStats(
torch.load('./experiment_stats.pth')) # TODO: needs realtime computation on private
elif config.normalize == 'plate':
normalize = NormalizeByPlateStats(
torch.load('./plate_stats.pth')) # TODO: needs realtime computation on private
else:
raise AssertionError('invalide normalization {}'.format(config.normalize))