def __getitem__(self, index):
x_orig, classifier_target = self.dataset[index]
x_orig = x_orig.numpy()
classifier_target = classifier_target.numpy()
x_tf_0 = np.copy(x_orig)
x_tf_90 = np.rot90(x_orig.copy(), k=1).copy()
x_tf_180 = np.rot90(x_orig.copy(), k=2).copy()
x_tf_270 = np.rot90(x_orig.copy(), k=3).copy()
possible_translations = list(itertools.product([0, 8, -8], [0, 8, -8]))
num_possible_translations = len(possible_translations)
tx, ty = possible_translations[random.randint(
0, num_possible_translations - 1)]
tx_target = {0: 0, 8: 1, -8: 2}[tx]
ty_target = {0: 0, 8: 1, -8: 2}[ty]
x_tf_trans = cv2f.affine(np.asarray(x_orig).copy(),
0, (tx, ty),
1,
0,
interpolation=cv2.INTER_CUBIC,
mode=cv2.BORDER_REFLECT_101)
return \
normalize(trnF.to_tensor(x_tf_0)), \
normalize(trnF.to_tensor(x_tf_90)), \
normalize(trnF.to_tensor(x_tf_180)), \
normalize(trnF.to_tensor(x_tf_270)), \
normalize(trnF.to_tensor(x_tf_trans)), \
torch.tensor(tx_target), \
torch.tensor(ty_target), \
torch.tensor(classifier_target)
def __call__(self, img):
"""
img (numpy ndarray): Image to be transformed.
Returns:
numpy ndarray: Affine transformed image.
"""
ret = self.get_params(self.degrees, self.translate, self.scale,
self.shear, (img.shape[1], img.shape[0]))
return F.affine(img,
*ret,
interpolation=self.interpolation,
fillcolor=self.fillcolor)
def __call__(self, imgs):
"""
imgs (numpy ndarray): Image sequence (time*height*width*channel) to be transformed.
Returns:
numpy ndarray: Affine transformed image sequence.
"""
ret = self.get_params(self.degrees, self.translate, self.scale,
self.shear, (imgs.shape[2], imgs.shape[1]))
# Apply to all images
output_imgs = []
for I in imgs:
output_imgs.append(
F.affine(I,
*ret,
interpolation=self.interpolation,
fillcolor=self.fillcolor))
return np.array(output_imgs)
def __getitem__(self, index):
x, _ = self.dataset[index]
pert = pert_configs[self.pert_number]
x = np.asarray(resize_and_crop(x))
x = cv2f.affine(np.asarray(x),
0, (pert[0], pert[1]),
1,
0,
interpolation=cv2.INTER_LINEAR,
mode=cv2.BORDER_REFLECT_101)
x = np.rot90(x, pert[2])
return trnF.to_tensor(x.copy()), [
expanded_params[i].index(pert[i])
for i in range(len(expanded_params))
]
def __getitem__(self, index):
x, _ = self.dataset[index // num_perts]
pert = pert_configs[index % num_perts]
x = np.asarray(resize_and_crop(x))
if np.random.uniform() < 0.5:
x = x[:, ::-1]
x = cv2f.affine(np.asarray(x), 0, (pert[0], pert[1]), 1, 0,
interpolation=cv2.INTER_LINEAR, mode=cv2.BORDER_REFLECT_101)
label = [expanded_params[i].index(pert[i]) for i in range(len(expanded_params))]
label = np.vstack((label + [0], label + [1], label + [2], label + [3]))
x = trnF.to_tensor(x.copy()).unsqueeze(0).numpy()
x = np.concatenate((x, np.rot90(x, 1, axes=(2, 3)),
np.rot90(x, 2, axes=(2, 3)), np.rot90(x, 3, axes=(2, 3))), 0)
return torch.FloatTensor(x), label