def generate_samples(self):
total_imgs = len(self.list_imgs)
print('Total ' + self.mode + ' data to generate samples:', total_imgs)
print('Mode: ' + self.mode + ' 2d sub grids samples to generate: ',
self.samples, ' Input images: ', total_imgs)
for j in range(total_imgs):
input_path = self.list_imgs[j]
label_path = self.list_labels[j]
img_numpy = img_loader.load_2d_image(input_path, type="RGB")
label_numpy = img_loader.load_2d_image(label_path, type='LA')
for i in range(self.per_image_sample):
h_crop, w_crop = self.generate_patch(img_numpy)
img_cropped = img_numpy[h_crop:(h_crop + self.crop_dim[0]),
w_crop:(w_crop + self.crop_dim[1]), :]
label_cropped = label_numpy[h_crop:(h_crop + self.crop_dim[0]),
w_crop:(w_crop + self.crop_dim[1])]
img_tensor = torch.from_numpy(img_cropped).float()
label_tensor = torch.from_numpy(label_cropped)
img_tensor = img_tensor.permute(2, 0, 1)
img_tensor = self.norm_img(img_tensor)
if self.save:
filename = self.sub_vol_path + 'id_' + str(
j) + '_s_' + str(i) + '_'
f_1 = filename + 'input.npy'
f_2 = filename + 'label.npy'
np.save(f_1, img_tensor)
np.save(f_2, label_tensor)
self.sample_list.append(tuple((f_1, f_2)))
else:
self.sample_list.append(tuple((img_tensor, label_tensor)))
def load_image(self, img_path, resize_dim):
if not os.path.exists(img_path):
print("IMAGE DOES NOT EXIST {}".format(img_path))
image = img_loader.load_2d_image(img_path, resize_dim)
t = transforms.ToTensor()
norm = transforms.Normalize(mean=[0.5, 0.5, 0.5], std=[1, 1, 1])
return norm(t(image))
def check_path_in_list(path, list):
"""
Checks a path if exist in the other list
"""
key = path.split('/')[-1]
for full_path in list:
path_id = full_path.split('/')[-1]
if path_id == key:
image_numpy = img_loader.load_2d_image(full_path, type="LA")
return image_numpy
return None
def preprocess_labels(maps_tuple):
"""
Majority labeling vote to produce ground truth labels
"""
label_list = []
m1, m2, m3, m4, m5, m6 = maps_tuple
for j in range(len(m5)):
path = m5[j] # as a reference annotation
key = path.split('/')[-1]
image_list = []
# voter 5
image_list.append(img_loader.load_2d_image(path, type="LA"))
# voter 1
image = check_path_in_list(path, m1)
if image is not None:
image_list.append(image)
# voter 2
image = check_path_in_list(path, m2)
if image is not None:
image_list.append(image)
# voter 3
image = check_path_in_list(path, m3)
if image is not None:
image_list.append(image)
# voter 4
image = check_path_in_list(path, m4)
if image is not None:
image_list.append(image)
# voter 6
image = check_path_in_list(path, m6)
if image is not None:
image_list.append(image)
stacked_labels = np.stack(image_list, axis=0)
label = vote(stacked_labels)
