def __getitem__(self, index):
random.seed()
# choose whether to pull labeled or unlabaled
labeled = 1 if random.random(
) > self.unlabeled_ratio else 0 # print(imgs.size())
if labeled == 1:
img = img_as_float(
imread(self.labeled_img_paths +
self.labeled_ids[index]))[:, :, :3]
msk = imread(self.mask_paths + self.labeled_ids[index]).astype(
np.bool)
msk = np.expand_dims(msk, axis=-1)
else:
img = img_as_float(
imread(self.unlabeled_img_paths +
self.unlabeled_ids[self.unlabeled_idx]))[:, :, :3]
msk = np.ones((101, 101, 1)) * -1.
self.unlabeled_idx += 1
# if we go through all the labeled images, shuffle them and start counter over
if self.unlabeled_idx >= len(self.unlabeled_ids):
self.unlabeled_ids = shuffle(self.unlabeled_ids)
self.unlabeled_idx = 0
# the geometric augmentions have to be the same
img, msk = augment_img([img, msk], imsize=self.imsize, mt=True)
# brightness, gamma, and gaussian noise can be different
img_a = mt_noise(img)
img_b = mt_noise(img)
msk_tch = torch.from_numpy(msk.astype(np.float32))
msk_half_tch = torch.from_numpy(
resize(msk.astype(np.float32), (64, 64), preserve_range=True))
msk_qrtr_tch = torch.from_numpy(
resize(msk.astype(np.float32), (32, 32), preserve_range=True))
out_dict = {
'img_a':
self.normalize(torch.from_numpy(img_a.astype(np.float32))),
'img_b':
self.normalize(torch.from_numpy(img_b.astype(np.float32))),
'msk': msk_tch,
'msk_half': msk_half_tch,
'msk_qrtr': msk_qrtr_tch,
'has_msk': msk_tch.sum() > 0,
'is_labeled': torch.tensor(labeled).long()
}
return out_dict
def __getitem__(self, index):
random.seed()
# super sample small masks
#if random.random() > 0.25 or self.valid:
img = img_as_float(imread(self.img_paths +
self.img_ids[index]))[:, :, :3]
#img[:,:,2] = 1. - fltrs.laplace(img[:,:,1])
msk = imread(self.mask_paths + self.img_ids[index]).astype(np.bool)
msk = np.expand_dims(msk, axis=-1)
boundry = find_boundaries(msk) * 1
if not self.valid:
img_np, msk_np, boundry_np = augment_img([img, msk, boundry],
imsize=self.imsize)
else:
#img_np = resize(np.asarray(img), (self.imsize, self.imsize),
# preserve_range=True, mode='reflect')
#msk_np = resize(msk, (self.imsize, self.imsize),
# preserve_range=True, mode='reflect')
img_np = reflect_pad(img, int((self.imsize - img.shape[0]) / 2))
msk_np = reflect_pad(msk, int((self.imsize - msk.shape[0]) / 2))
boundry_np = reflect_pad(boundry,
int((self.imsize - boundry.shape[0]) / 2))
img_np = img_np.transpose((2, 0, 1)).astype(np.float32)
msk_np = msk_np.transpose((2, 0, 1)).astype(np.float32)
boundry_np = boundry_np.transpose((2, 0, 1)).astype(np.float32)
#print(img_np.shape, msk_np.shape)
# get image ready for torch
img_tch = self.normalize(torch.from_numpy(img_np.astype(np.float32)))
msk_tch = torch.from_numpy(msk_np.astype(np.float32))
boundry_tch = torch.from_numpy(boundry_np.astype(np.float32))
img_tch = add_depth_channels(img_tch)
out_dict = {
'img': img_tch,
'msk': msk_tch,
'has_msk': msk_tch.sum() > 0,
'boundry': boundry_tch,
'id': self.img_ids[index].replace('.png', '')
}
return out_dict
def __getitem__(self, index):
random.seed()
# super sample small masks
#if random.random() > 0.25 or self.valid:
img = img_as_float(imread(self.img_paths +
self.img_ids[index]))[:, :, :3]
#img[:,:,2] = 1. - fltrs.laplace(img[:,:,0])
msk = imread(self.mask_paths + self.img_ids[index]).astype(np.bool)
msk = np.expand_dims(msk, axis=-1)
#else:
# small_idx = random.randint(0, len(self.small_msk_ids))
# img = img_as_float(imread('../data/train/small_masks/images/' + self.small_msk_ids[small_idx]))[:,:,:3]
# msk = imread('../data/train/small_masks/masks/' + self.small_msk_ids[small_idx]).astype(np.bool)
# msk = np.expand_dims(msk, axis=-1)
if not self.valid:
img_np, msk_np = augment_img([img, msk], imsize=self.imsize)
else:
#img_np = resize(np.asarray(img), (self.imsize, self.imsize),
# preserve_range=True, mode='reflect')
#msk_np = resize(msk, (self.imsize, self.imsize),
# preserve_range=True, mode='reflect')
img_np = reflect_pad(img, int((self.imsize - img.shape[0]) / 2))
msk_np = reflect_pad(msk, int((self.imsize - msk.shape[0]) / 2))
img_np = img_np.transpose((2, 0, 1)).astype(np.float32)
msk_np = msk_np.transpose((2, 0, 1)).astype(np.float32)
#print(img_np.shape, msk_np.shape)
# get image ready for torch
img_tch = self.normalize(torch.from_numpy(img_np.astype(np.float32)))
msk_tch = torch.from_numpy(msk_np.astype(np.float32))
msk_half_np = np.expand_dims(resize(msk_np[0].astype(np.float32),
(64, 64),
preserve_range=True),
axis=-1).transpose(2, 0, 1)
msk_half_tch = torch.from_numpy(msk_half_np.astype(np.float32))
msk_qrtr_np = np.expand_dims(resize(msk_np[0].astype(np.float32),
(32, 32),
preserve_range=True),
axis=-1).transpose(2, 0, 1)
msk_qrtr_tch = torch.from_numpy(msk_qrtr_np.astype(np.float32))
msk_eigt_np = np.expand_dims(resize(msk_np[0].astype(np.float32),
(16, 16),
preserve_range=True),
axis=-1).transpose(2, 0, 1)
msk_eigt_tch = torch.from_numpy(msk_eigt_np.astype(np.float32))
msk_sixteen_np = np.expand_dims(resize(msk_np[0].astype(np.float32),
(8, 8),
preserve_range=True),
axis=-1).transpose(2, 0, 1)
msk_sixteen_tch = torch.from_numpy(msk_sixteen_np.astype(np.float32))
#img_tch = add_depth_channels(img_tch)
out_dict = {
'img':
img_tch,
'msk': [
msk_tch, msk_half_tch, msk_qrtr_tch, msk_eigt_tch,
msk_sixteen_tch
],
#'msk_dwnsample': [msk_half_tch, msk_qrtr_tch, msk_eigt_tch, msk_sixteen_tch],
#'msk_qrtr': msk_qrtr_tch,
#'msk_eigt': msk_eigt_tch,
'has_msk':
msk_tch.sum() > 0,
'id':
self.img_ids[index].replace('.png', '')
}
return out_dict