def augment(img_in, img_tar, img_bic, flip_h=True, flip_v=True, rot=True):
info_aug = {'flip_h': False, 'flip_v': False, 'trans': False}
if random.random() < 0.5 and flip_h:
for img in [img_in, img_tar, img_bic]:
for idx_c in range(0):
img[:, :, idx_c] = np.fliplr(img)
info_aug['flip_h'] = True
if random.random() < 0.5 and flip_v:
for img in [img_in, img_tar, img_bic]:
for idx_c in range(0):
img[:, :, idx_c] = np.flipup(img)
info_aug['flip_v'] = True
if rot:
cnt_rot = int(random.random() // 0.25)
for img in [img_in, img_tar, img_bic]:
for idx_c in range(0):
img[:, :, idx_c] = np.rot90(img, cnt_rot)
info_aug['trans'] = True
return img_in, img_tar, img_bic, info_aug
def flip(self, axis, img):
if axis == 0:
img = np.fliplr(img)
if axis == 1:
img = np.flipup(img)
return img
def _pi_flipud(board):
'''
Flips the pi boards up to down according to our conventions
:return: the flipped board
'''
return np.roll(np.flipup(board),-1,0)
def reshape(self, bottom, top):
if self.split == 'train':
# self.train_timing = (self.train_timing + self.batch_size) % self.train_images_num
self.data = np.zeros(
[self.batch_size, 3, self.resize_size[0], self.resize_size[1]])
# self.sobel_0 = np.zeros([self.batch_size, 1, self.resize_size[0], self.resize_size[1]])
# self.sobel_1 = np.zeros([self.batch_size, 1, self.resize_size[0]/2, self.resize_size[1]/2])
# self.sobel_2 = np.zeros([self.batch_size, 1, self.resize_size[0]/4, self.resize_size[1]/4])
# self.sobel_3 = np.zeros([self.batch_size, 1, self.resize_size[0]/8, self.resize_size[1]/8])
# self.sobel_4 = np.zeros([self.batch_size, 1, self.resize_size[0]/16, self.resize_size[1]/16])
# self.sobel_5 = np.zeros([self.batch_size, 1, self.resize_size[0]/32, self.resize_size[1]/32])
self.labels = np.zeros(
[self.batch_size, 1, self.resize_size[0], self.resize_size[1]])
for i in range(self.batch_size):
self.train_timing = (self.train_timing +
1) % self.train_images_num
orignial_image_data = mping.imread(
osp.join(self.data_dir, self.train_data_dir,
self.train_images[self.train_timing]))
# orignial_image_sobel = mping.imread(osp.join(self.data_dir, self.train_sobel_dir, self.train_images[self.train_timing]))
orignial_image_labels = mping.imread(
osp.join(
self.data_dir, self.train_labels_dir,
self.train_images[self.train_timing].split('.jpg')[0] +
'.png'))
# print osp.join(self.data_dir, self.train_labels_dir, self.train_images[self.train_timing].split('.jpg')[0] + '.png')
GV.data_name = self.train_images[self.train_timing].split(
'.jpg')[0] + '.png'
if orignial_image_data.shape[:
2] != orignial_image_labels.shape[:
2]:
raise Exception('image and labels must be same size')
height, width = orignial_image_data.shape[:2]
if height < self.resize_size[0]:
height_add_flag = 1
else:
height_add_flag = 0
if width < self.resize_size[1]:
width_add_flag = 1
else:
width_add_flag = 0
if height_add_flag or width_add_flag:
add_image = np.zeros([
height * (1 + height_add_flag),
width * (1 + width_add_flag), 3
])
add_label = np.zeros([
height * (1 + height_add_flag),
width * (1 + width_add_flag), 3
])
if height_add_flag and not width_add_flag:
add_image[:height, :width] = np.flipup(
orignial_image_data)
add_image[height:2 *
height, :width] = orignial_image_data
add_label[:height, :width] = np.flipup(
orignial_image_labels)
add_label[height:2 *
height, :width] = orignial_image_labels
elif not height_add_flag and width_add_flag:
add_image[:height, :width] = np.fliplr(
orignial_image_data)
add_image[:height:width,
2 * width] = orignial_image_data
add_label[:height, :width] = np.fliplr(
orignial_image_labels)
add_label[:height,
width:2 * width] = orignial_image_labels
elif height_add_flag and width_add_flag:
add_image[height:2 * height, :width] = np.fliplr(
orignial_image_data)
add_image[height:2 * height, width,
2 * width] = orignial_image_data
add_image[:height, width,
2 * width] = np.flipud(orignial_image_data)
add_image[:height, :width] = np.fliplr(
np.flipud(orignial_image_data))
add_label[height:2 * height, :width] = np.fliplr(
orignial_image_labels)
add_label[height:2 * height, width,
2 * width] = orignial_image_labels
add_label[:height, width,
2 * width] = np.flipud(orignial_image_labels)
add_label[:height, :width] = np.fliplr(
np.flipud(orignial_image_labels))
orignial_image_data = add_image
orignial_image_labels = add_label
height, width = orignial_image_data.shape[:2]
start_x = random.randint(0, height - self.resize_size[0])
start_y = random.randint(0, width - self.resize_size[1])
image_data = orignial_image_data[start_x:start_x +
self.resize_size[0],
start_y:start_y +
self.resize_size[1]]
# image_sobel = orignial_image_sobel[end_x - int(height * tmp_crop_ratio) : end_x, end_y - int(width * tmp_crop_ratio) : end_y]
image_labels = orignial_image_labels[start_x:start_x +
self.resize_size[0],
start_y:start_y +
self.resize_size[1]]
# end_x = random.randint(int(height * self.crop_ratio), height)
# end_y = random.randint(int(width * self.crop_ratio), width)
# image_data = orignial_image_data[:end_x, :end_y]
# image_sobel = orignial_image_sobel[:end_x, :end_y]
# image_labels = orignial_image_labels[:end_x, :end_y]
# plt.figure(1)
# plt.subplot(221)
# plt.imshow(orignial_image_data)
# plt.subplot(222)
# plt.imshow(orignial_image_labels)
# plt.subplot(223)
# plt.imshow(image_sobel)
# plt.subplot(224)
# plt.imshow(image_labels)
# GV.data_name = self.test_images[self.train_timing].split('.')[0]
# print self.data_dir, self.train_data_dir, self.train_images[self.train_timing]
# image_data = imresize(np.array(image_data, dtype = np.uint8), [self.resize_size[0], self.resize_size[1], 3])
# image_labels = imresize(np.array(image_labels, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
flip = np.random.randint(0, 2)
if len(image_data.shape) == 3:
a = 1
# image_data = imresize(np.array(image_data, dtype = np.uint8), [self.resize_size[0], self.resize_size[1], 3])
# image_sobel_0 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
# image_sobel_1 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/2, self.resize_size[1]/2])
# image_sobel_2 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/4, self.resize_size[1]/4])
# image_sobel_3 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/8, self.resize_size[1]/8])
# image_sobel_4 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/16, self.resize_size[1]/16])
# image_sobel_5 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/32, self.resize_size[1]/32])
else:
# image_data = imresize(np.array(image_data, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
image_data = np.tile(image_data[:, :, np.newaxis],
[1, 1, 3])
# image_sobel_0 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
# image_sobel_1 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/2, self.resize_size[1]/2])
# image_sobel_2 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/4, self.resize_size[1]/4])
# image_sobel_3 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/8, self.resize_size[1]/8])
# image_sobel_4 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/16, self.resize_size[1]/16])
# image_sobel_5 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/32, self.resize_size[1]/32])
GV.abnormal_files.append(GV.data_name)
# image_data = np.concatenate((image_data, image_sobel[:,:,np.newaxis]), axis = 2)
if len(image_labels.shape) == 3:
image_labels = image_labels.mean(axis=2)
# GV.abnormal_files.append(GV.data_name)
else:
pass
image_labels[np.where(image_labels > 0.1)] = 1
image_labels[np.where(image_labels < 0.1)] = 0
if flip == 1:
image_data = np.fliplr(image_data)
image_labels = np.fliplr(image_labels)
# flip = np.random.randint(0, 2)
# if flip == 1:
# image_data = np.flipud(image_data)
# image_labels = np.flipud(image_labels)
GV, image = image_data
GV.image_labels = image_labels
self.data[i] = image_data.transpose(2, 0, 1)
# self.sobel_0[i, 0] = image_sobel_0
# self.sobel_1[i, 0] = image_sobel_1
# self.sobel_2[i, 0] = image_sobel_2
# self.sobel_3[i, 0] = image_sobel_3
# self.sobel_4[i, 0] = image_sobel_4
# self.sobel_5[i, 0] = image_sobel_5
self.labels[i, 0] = image_labels
# self.data = np.array(self.data, dtype = np.float32)
# self.labels = np.array(self.labels, dtype = np.float32)
elif self.split == 'test':
# self.resize_size[0] = 64
# self.resize_size[1] = 64
self.data = np.zeros([
self.test_batch_size, 3, self.resize_size[0],
self.resize_size[1]
])
# self.sobel_0 = np.zeros([self.test_batch_size, 1, self.resize_size[0], self.resize_size[1]])
# self.sobel_1 = np.zeros([self.test_batch_size, 1, self.resize_size[0]/2, self.resize_size[1]/2])
# self.sobel_2 = np.zeros([self.test_batch_size, 1, self.resize_size[0]/4, self.resize_size[1]/4])
# self.sobel_3 = np.zeros([self.test_batch_size, 1, self.resize_size[0]/8, self.resize_size[1]/8])
# self.sobel_4 = np.zeros([self.test_batch_size, 1, self.resize_size[0]/16, self.resize_size[1]/16])
# self.sobel_5 = np.zeros([self.test_batch_size, 1, self.resize_size[0]/32, self.resize_size[1]/32])
self.labels = np.zeros([
self.test_batch_size, 1, self.resize_size[0],
self.resize_size[1]
])
for i in range(self.test_batch_size):
# print i
self.test_timing = (self.test_timing +
1) % self.test_images_num
suffix = osp.join(
self.data_dir, self.test_data_dir,
self.test_images[self.test_timing]).split('.')[-1]
if suffix == 'png':
image_data = mping.imread(
osp.join(self.data_dir, self.test_data_dir,
self.test_images[self.test_timing])) * 255
# image_sobel = mping.imread(osp.join(self.data_dir, self.test_sobel_dir, self.test_images[self.test_timing]))
else:
image_data = mping.imread(
osp.join(self.data_dir, self.test_data_dir,
self.test_images[self.test_timing]))
# image_sobel = mping.imread(osp.join(self.data_dir, self.test_sobel_dir, self.test_images[self.test_timing]))
# print self.test_images[self.test_timing], suffix
image_labels = mping.imread(
osp.join(
self.data_dir, self.test_labels_dir,
self.test_images[self.test_timing].split('.' +
suffix)[0] +
'.png'))
# print osp.join(self.data_dir, self.test_labels_dir, self.test_images[self.test_timing].split('.' + suffix)[0] + '.png')
image_labels[np.where(image_labels > 0.1)] = 1
# hegiht, width, _ = image_labels.shape
# max_height = 64
# if height > max_height:
# image_labels = imresize(image_labels, [max_height, max_height * width / height, 3])
# image_labels = imresize(image_labels, [max_height, max_height * width / height, 3])
GV.data_name = self.test_images[self.test_timing].split('.')[0]
GV.data_dir = osp.join(self.data_dir, self.test_data_dir)
print self.data_dir, self.test_labels_dir, self.test_images[
self.test_timing], GV.data_name
GV.a = self.data_dir
GV.b = self.test_labels_dir
GV.c = self.test_images[self.test_timing]
GV.d = GV.data_name
# print 'hello', GV.data_name
if len(image_data.shape) == 3:
if image_data.shape[2] != 3:
GV.image = image_data
image_data = imresize(
np.array(image_data[:, :, 0] * 255,
dtype=np.uint8),
[self.resize_size[0], self.resize_size[1]])
image_data = np.tile(image_data[:, :, np.newaxis],
[1, 1, 3])
# image_sobel_0 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
# image_sobel_1 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/2, self.resize_size[1]/2])
# image_sobel_2 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/4, self.resize_size[1]/4])
# image_sobel_3 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/8, self.resize_size[1]/8])
# image_sobel_4 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/16, self.resize_size[1]/16])
# image_sobel_5 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/32, self.resize_size[1]/32])
else:
image_data = imresize(
np.array(image_data, dtype=np.uint8),
[self.resize_size[0], self.resize_size[1], 3])
# image_sobel_0 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
# image_sobel_1 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/2, self.resize_size[1]/2])
# image_sobel_2 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/4, self.resize_size[1]/4])
# image_sobel_3 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/8, self.resize_size[1]/8])
# image_sobel_4 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/16, self.resize_size[1]/16])
# image_sobel_5 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/32, self.resize_size[1]/32])
# dasda
elif len(image_data.shape) == 2:
image_data = imresize(
np.array(image_data, dtype=np.uint8),
[self.resize_size[0], self.resize_size[1]])
image_data = np.tile(image_data[:, :, np.newaxis],
[1, 1, 3])
# image_sobel_0 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0], self.resize_size[1]])
# image_sobel_1 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/2, self.resize_size[1]/2])
# image_sobel_2 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/4, self.resize_size[1]/4])
# image_sobel_3 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/8, self.resize_size[1]/8])
# image_sobel_4 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/16, self.resize_size[1]/16])
# image_sobel_5 = imresize(np.array(image_sobel, dtype = np.uint8), [self.resize_size[0]/32, self.resize_size[1]/32])
if len(image_labels.shape) == 3:
image_labels = imresize(
np.array(image_labels[:, :, 0], dtype=np.uint8),
[self.resize_size[0], self.resize_size[1]])
else:
image_labels = imresize(
np.array(image_labels, dtype=np.uint8),
[self.resize_size[0], self.resize_size[1]])
# image_data = np.concatenate((image_data, image_sobel[:,:,np.newaxis]), axis = 2)
GV.image = image_data
GV.labels = image_labels
self.data[i] = image_data.transpose(2, 0, 1)
# self.sobel_0[i, 0] = image_sobel_0
# self.sobel_1[i, 0] = image_sobel_1
# self.sobel_2[i, 0] = image_sobel_2
# self.sobel_3[i, 0] = image_sobel_3
# self.sobel_4[i, 0] = image_sobel_4
# self.sobel_5[i, 0] = image_sobel_5
self.labels[i, 0] = image_labels
top[0].reshape(*self.data.shape)
# top[1].reshape(*self.sobel_0.shape)
# top[2].reshape(*self.sobel_1.shape)
# top[3].reshape(*self.sobel_2.shape)
# top[4].reshape(*self.sobel_3.shape)
# top[5].reshape(*self.sobel_4.shape)
# top[6].reshape(*self.sobel_5.shape)
top[1].reshape(*self.labels.shape)
def mirror_horizontal(batch):
return np.flipup(batch)