def test_sequential(self):
f = nn.Sequential(
RandomHorizontalFlip(1.0, return_transform=True),
RandomHorizontalFlip(1.0, return_transform=True),
)
f1 = nn.Sequential(
RandomHorizontalFlip(1.0, return_transform=True),
RandomHorizontalFlip(1.0),
)
input = torch.tensor([[[[0., 0., 0.],
[0., 0., 0.],
[0., 1., 1.]]]]) # 1 x 1 x 3 x 3
expected_transform = torch.tensor([[[-1., 0., 3.],
[0., 1., 0.],
[0., 0., 1.]]]) # 1 x 3 x 3
expected_transform_1 = expected_transform @ expected_transform
assert(f(input)[0] == input).all()
assert(f(input)[1] == expected_transform_1).all()
assert(f1(input)[0] == input).all()
assert(f1(input)[1] == expected_transform).all()
def test_random_hflip(self):
f = RandomHorizontalFlip(p=1.0, return_transform=True)
f1 = RandomHorizontalFlip(p=0., return_transform=True)
f2 = RandomHorizontalFlip(p=1.)
f3 = RandomHorizontalFlip(p=0.)
input = torch.tensor([[0., 0., 0.],
[0., 0., 0.],
[0., 1., 1.]]) # 3 x 3
expected = torch.tensor([[0., 0., 0.],
[0., 0., 0.],
[1., 1., 0.]]) # 3 x 3
expected_transform = torch.tensor([[-1., 0., 3.],
[0., 1., 0.],
[0., 0., 1.]]) # 3 x 3
identity = torch.tensor([[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]]) # 3 x 3
assert (f(input)[0] == expected).all()
assert (f(input)[1] == expected_transform).all()
assert (f1(input)[0] == input).all()
assert (f1(input)[1] == identity).all()
assert (f2(input) == expected).all()
assert (f3(input) == input).all()
def test_batch_random_hflip(self):
f = RandomHorizontalFlip(p=1.0, return_transform=True)
f1 = RandomHorizontalFlip(p=0.0, return_transform=True)
input = torch.tensor([[[[0., 0., 0.],
[0., 0., 0.],
[0., 1., 1.]]]]) # 1 x 1 x 3 x 3
expected = torch.tensor([[[[0., 0., 0.],
[0., 0., 0.],
[1., 1., 0.]]]]) # 1 x 1 x 3 x 3
expected_transform = torch.tensor([[[-1., 0., 3.],
[0., 1., 0.],
[0., 0., 1.]]]) # 1 x 3 x 3
identity = torch.tensor([[[1., 0., 0.],
[0., 1., 0.],
[0., 0., 1.]]]) # 1 x 3 x 3
input = input.repeat(5, 3, 1, 1) # 5 x 3 x 3 x 3
expected = expected.repeat(5, 3, 1, 1) # 5 x 3 x 3 x 3
expected_transform = expected_transform.repeat(5, 1, 1) # 5 x 3 x 3
identity = identity.repeat(5, 1, 1) # 5 x 3 x 3
assert (f(input)[0] == expected).all()
assert (f(input)[1] == expected_transform).all()
assert (f1(input)[0] == input).all()
assert (f1(input)[1] == identity).all()
def test_gradcheck(self):
input = torch.rand((3, 3)).double() # 3 x 3
input = utils.tensor_to_gradcheck_var(input) # to var
assert gradcheck(RandomHorizontalFlip(p=1.), (input, ), raise_exception=True)
def monochrome_model(gaussian_p, solarize_p):
return nn.Sequential(
RandomResizedCrop((image_size, image_size),
interpolation="BICUBIC"), RandomHorizontalFlip(),
RandomApply(GaussianBlur2d(get_kernel_size(image_size),
(0.1, 2.0)),
p=gaussian_p), RandomSolarize(0, 0, p=solarize_p))
def color_model(gaussian_p, solarize_p):
return nn.Sequential(
RandomResizedCrop((image_size, image_size),
interpolation="BICUBIC"), RandomHorizontalFlip(),
ColorJitter(0.4, 0.4, 0.2, 0.1, p=0.8), RandomGrayscale(p=0.2),
RandomApply(GaussianBlur2d(get_kernel_size(image_size),
(0.1, 2.0)),
p=gaussian_p), RandomSolarize(0, 0, p=solarize_p))
def train_transforms(image_size,
train_img_scale=(0.35, 1),
normalize: bool = True,
mean=torch.tensor([0.485, 0.456, 0.406]),
std=torch.tensor([0.229, 0.224, 0.225])):
"""Transforms for train augmentation with Kornia."""
transforms = [
AccimageImageToTensorNN(),
RandomResizedCrop((image_size, image_size),
train_img_scale,
keepdim=True),
RandomHorizontalFlip(keepdim=True)
]
if normalize:
transforms.append(Normalize(mean=std, std=std, keepdim=True))
return torch.nn.Sequential(*transforms)
def __init__(self, input_shape, s=1.0, apply_transforms=None):
assert len(input_shape) == 3, "input_shape should be (H, W, C)"
self.input_shape = input_shape
self.H, self.W, self.C = input_shape[0], input_shape[1], input_shape[2]
self.s = s
self.apply_transforms = apply_transforms
if self.apply_transforms is None:
kernel_size = int(0.1 * self.H)
sigma = self._get_sigma()
self.apply_transforms = KorniaCompose([
RandomResizedCrop(size=(self.H, self.W), scale=(0.08, 1.0)),
RandomHorizontalFlip(p=0.5),
ColorJitter(0.8 * self.s, 0.8 * self.s, 0.8 * self.s, 0.2 * self.s),
RandomGrayscale(p=0.2),
GaussianBlur2d(kernel_size=(kernel_size, kernel_size),
sigma=(sigma, sigma))
])
def random_flip(self, p: float = 0.5) -> TransformType:
return RandomHorizontalFlip(p=p)
def smoke_test(self):
f = RandomHorizontalFlip(0.5)
repr = "RandomHorizontalFlip(p=0.5, return_transform=False)"
assert str(f) == repr
from itertools import chain
import torch
import torch.nn as nn
import torch.optim as optim
from kornia.augmentation import RandomHorizontalFlip, RandomCrop
from kornia.geometry.transform import resize
from fastcore.all import *
from .data_anime_heads import Datasets as AnimeHeadsDatasets, DataLoaders as AnimeHeadsDataLoaders
from .data_birds import Datasets as BirdsDatasets, DataLoaders as BirdsDataLoaders
from .model import RnnEncoder, CnnEncoder
from .torch_utils import *
# Internal Cell
ce_loss = nn.CrossEntropyLoss()
random_hflip = RandomHorizontalFlip().requires_grad_(False).eval()
random_crop = RandomCrop((229, 229)).requires_grad_(False).eval()
# Internal Cell
def cosine_similarity(x1, x2, dim=1, eps=1e-8):
"""Returns cosine similarity between x1 and x2, computed along dim.
"""
w12 = torch.sum(x1 * x2, dim)
w1 = torch.norm(x1, 2, dim)
w2 = torch.norm(x2, 2, dim)
return (w12 / (w1 * w2).clamp(min=eps)).squeeze()
def func_attention(query, context, gamma1):
"""
