def predict_Unet(data, ids, log=True):
warnings.filterwarnings('ignore', category=UserWarning, module='skimage')
model = load_model('model-dsbowl2018-1(0.304).h5', custom_objects={'mean_iou': mean_iou})
preds_test_upsampled = []
print("Predict masks")
for i, d in tqdm(enumerate(data), total=len(data)):
img = d
h_img = img.shape[0]
w_img = img.shape[1]
curimg = np.array(
[pic[:dp.IMG_HEIGHT, :dp.IMG_WIDTH].reshape(128, 128, 1) for pic in da.cut_image(img)])
curimg = curimg * 255
curimg = curimg.astype(np.uint8)
f = model.predict(curimg, verbose=1)
pred_img = da.glue_image(f, h_img, w_img, func_merg='max').reshape(h_img, w_img)
if log:
imsave("../data/OUTP/true_" + str(i) + ".png", pred_img)
preds_test_upsampled.append(pred_img)
new_test_ids = []
rles = []
res_encod = []
for n, id_ in enumerate(ids):
rle = list(prob_to_rles(preds_test_upsampled[n]))
rles.extend(rle)
res_encod.append(rle)
new_test_ids.extend([id_] * len(rle))
# Create submission DataFrame
sub = pd.DataFrame()
sub['ImageId'] = new_test_ids
sub['EncodedPixels'] = pd.Series(rles).apply(lambda x: ' '.join(str(y) for y in x))
if log:
sub.to_csv('sub-dsbowl2018-1.csv', index=False)
return sub, new_test_ids, rles, res_encod
def main_merge_with_NMS_optimizer():
warnings.filterwarnings('ignore', category=UserWarning, module='skimage')
new_test_ids = []
rles = []
detector = None
detector_second = None
with CustomObjectScope({
'relu6':
applications.mobilenet.relu6,
'DepthwiseConv2D':
applications.mobilenet.DepthwiseConv2D
}):
detector = load_model('detector_MobileNet.h5')
detector_second = load_model('detector_MobileNet_Second.h5')
u_net = load_model("U-net/Unet(32x32).h5",
custom_objects={'mean_iou': mean_iou})
dir_ = "../../data/stage1_test/"
#dir_ = "../../data/1/"
# Считывание данных
test_data = []
print("Read data")
ids = next(os.walk(dir_))[1]
for n, id_ in tqdm(enumerate(ids), total=len(ids)):
path = dir_ + id_ + "/images/"
test_data.append(imread(path + id_ + ".png", as_grey=True))
# Подготовка данных
sizes_nucl = load_obj("../sizes_nuclears")
for i in range(len(test_data)):
median = sizes_nucl[ids[i]]
coef = median / DIAG_IMAGE
if coef < 1:
test_data[i] = resize(
test_data[i],
(int(test_data[i].shape[0] / (median / DIAG_IMAGE)),
int(test_data[i].shape[1] / (median / DIAG_IMAGE))),
mode='constant',
preserve_range=True)
ids = next(os.walk(dir_))[1]
for n, id_ in tqdm(enumerate(ids), total=len(ids)):
try:
image = test_data[n]
images = []
masks = []
bounding_boxs = []
bounding_boxs_nms = []
# Создаем разные размеры исходной картинки
cur_size = (int(image.shape[0] * MAX_RESIZE),
int(image.shape[1] * MAX_RESIZE))
min_size = (image.shape[0] // MIN_RESIZE,
image.shape[1] // MIN_RESIZE)
images.append(image)
masks.append(np.zeros(image.shape, dtype=np.float64))
while True:
images.append(
resize(image,
cur_size,
mode='constant',
preserve_range=True))
masks.append(np.zeros(cur_size, dtype=np.float64))
cur_size = (int(cur_size[0] // COEF_RES),
int(cur_size[1] // COEF_RES))
if cur_size[0] <= min_size[0] or cur_size[1] <= min_size[1]:
break
# Обрабатываем каждую получившуюся картинку
for i in range(len(images)):
im = images[i]
# Вырезаем из картинки квадратики для детектора
square_for_detector = []
bounds = []
bbs = []
for h in range(SIZE_IMAGE[0], im.shape[0] + STRIDES[0],
STRIDES[0]):
lower_bound = min(h, im.shape[0])
for w in range(SIZE_IMAGE[1], im.shape[1] + STRIDES[1],
STRIDES[1]):
right_bound = min(w, im.shape[1])
bounds.append((right_bound, lower_bound))
square_for_detector.append(im[lower_bound - SIZE_IMAGE[0]: lower_bound,
right_bound - SIZE_IMAGE[1]: right_bound]\
.reshape((SIZE_IMAGE[0], SIZE_IMAGE[1], 1)))
square_for_detector = np.array(square_for_detector)
# Говорит что в квадрате есть клетки
is_nucl = detector.predict(square_for_detector)
square_for_detector = np.array([
square_for_detector[i]
for i in range(len(square_for_detector))
if is_nucl[i][0] > TRESHHOLD_FIRST_DETECTOR
])
bounds = [
bounds[i] for i in range(len(bounds))
if is_nucl[i][0] > TRESHHOLD_FIRST_DETECTOR
]
# Говорит, что в квадрате только одна клетка
if len(square_for_detector) > 0:
is_one_nucl = detector_second.predict(square_for_detector)
bounds = [
bounds[i] for i in range(len(bounds))
if is_one_nucl[i][0] > TRESHHOLD_SECOND_DETECTOR
]
for box in bounds:
bbs.append((box[0] - SIZE_IMAGE[0],
box[1] - SIZE_IMAGE[1], box[0], box[1]))
bounding_boxs.append(bbs)
bounding_boxs_nms.append(
non_max_suppression_fast(np.array(bbs), 0.6))
result_mask = np.zeros(image.shape, dtype=np.float64)
result_bbs = []
# Мержим все боксы
for i in range(len(images)):
h_coef = images[i].shape[0] / image.shape[0]
w_coef = images[i].shape[1] / image.shape[1]
for bb in bounding_boxs_nms[i]:
box = (bb[0] // w_coef, bb[1] // h_coef, bb[2] // w_coef,
bb[3] // h_coef)
result_bbs.append(box)
result_nms = non_max_suppression_fast(np.array(result_bbs), 0.5)
# Предсказание по NMS
for box in result_nms:
im_to_unet = np.ndarray((1, SIZE_IMAGE[0], SIZE_IMAGE[1], 1),
dtype=np.float64)
# Приводим размеры квадрата к размерам входа Юнет
im_to_unet[0] = resize(image[box[1]:box[3], box[0]:box[2]],
SIZE_IMAGE,
mode='constant',
preserve_range=True).reshape(
(SIZE_IMAGE[0], SIZE_IMAGE[1], 1))
# Предсказывание
pred_mask = u_net.predict(im_to_unet)[0]
# Возвращаем начальный размер и вставляем в итоговую маску
copy_arr_to_arr(
result_mask[box[1]:box[3], box[0]:box[2]],
resize(pred_mask, (box[3] - box[1], box[2] - box[0]),
mode='constant',
preserve_range=True), "mean_except_zero")
# Кодирование
mask_to_encde = result_mask
rle = list(prob_to_rles(mask_to_encde, cutoff=0.5))
rles.extend(rle)
new_test_ids.extend([id_] * len(rle))
if len(rle) == 0:
rles.extend([[1, 1]])
new_test_ids.extend([id_])
save_result("../../data/detector_unet_pred/{}.png".format(id_),
image, mask_to_encde, images, bounding_boxs,
bounding_boxs_nms, result_bbs, result_nms)
except StopIteration:
print("Exception id: " + id_)
# Create submission DataFrame
sub = pd.DataFrame()
sub['ImageId'] = new_test_ids
sub['EncodedPixels'] = pd.Series(rles).apply(
lambda x: ' '.join(str(y) for y in x))
# Чистка от мусора
def f(x):
res = 0
for i in range(1, len(x), 2):
res += i
return res
sub['CountPix'] = pd.Series(rles).apply(f)
sub = sub[sub['CountPix'] > 10][['ImageId', 'EncodedPixels']]
sub.to_csv('detector_Unet.csv', index=False)
def main_ROI_NMS():
warnings.filterwarnings('ignore', category=UserWarning, module='skimage')
new_test_ids = []
rles = []
detector = None
with CustomObjectScope({
'relu6':
applications.mobilenet.relu6,
'DepthwiseConv2D':
applications.mobilenet.DepthwiseConv2D
}):
detector = load_model('detector_MobileNet.h5')
u_net = load_model("U-net/Unet(32x32).h5",
custom_objects={'mean_iou': mean_iou})
dir_ = "../../data/stage1_test/"
#dir_ = "../../data/1/"
ids_test = os.walk(dir_)
ids = next(ids_test)[1]
for n, id_ in tqdm(enumerate(ids), total=len(ids)):
try:
path = dir_ + id_ + "/images/"
image = imread(path + id_ + ".png", as_grey=True)
images = []
masks = []
bounding_boxs = []
bounding_boxs_nms = []
# Уменьшаем размер картинки
cur_size = (image.shape[0], image.shape[1])
min_size = (image.shape[0] // MIN_RESIZE,
image.shape[1] // MIN_RESIZE)
images.append(image)
masks.append(np.zeros(cur_size, dtype=np.float64))
while SIZE_IMAGE[0] <= cur_size[0] >= min_size[0] and SIZE_IMAGE[
1] <= cur_size[1] >= min_size[1]:
cur_size = (int(cur_size[0] // COEF_RES),
int(cur_size[1] // COEF_RES))
if SIZE_IMAGE[0] > cur_size[0] or SIZE_IMAGE[1] > cur_size[1]:
break
images.append(
resize(image,
cur_size,
mode='constant',
preserve_range=True))
masks.append(np.zeros(cur_size, dtype=np.float64))
# masks[-1].fill(0.45)
# Предсказание
for i in range(len(images)):
im = images[i]
# Проход детектора
bbs = []
for h in range(SIZE_IMAGE[0], im.shape[0] + STRIDES[0],
STRIDES[0]):
lower_bound = min(h, im.shape[0])
for w in range(SIZE_IMAGE[1], im.shape[1] + STRIDES[1],
STRIDES[1]):
right_bound = min(w, im.shape[1])
part_image_analis = np.ndarray(
(1, SIZE_IMAGE[0], SIZE_IMAGE[1], 1),
dtype=np.float64)
part_image_analis[0] = im[lower_bound - SIZE_IMAGE[0]: lower_bound,
right_bound - SIZE_IMAGE[1]: right_bound] \
.reshape((SIZE_IMAGE[0], SIZE_IMAGE[1], 1))
is_nucl = detector.predict(part_image_analis)[0][0]
if is_nucl > 0.99:
bbs.append((right_bound - SIZE_IMAGE[0],
lower_bound - SIZE_IMAGE[1],
right_bound, lower_bound))
bounding_boxs.append(bbs)
bounding_boxs_nms.append(
non_max_suppression_fast(np.array(bbs), 0.6))
# Предсказание по NMS
for box in bounding_boxs_nms[-1]:
im_to_unet = np.ndarray(
(1, SIZE_IMAGE[0], SIZE_IMAGE[1], 1), dtype=np.float64)
im_to_unet[0] = im[box[1]:box[3], box[0]:box[2]].reshape(
(SIZE_IMAGE[0], SIZE_IMAGE[1], 1))
pred_mask = u_net.predict(im_to_unet)[0]
copy_arr_to_arr(masks[i][box[1]:box[3], box[0]:box[2]],
pred_mask.reshape(SIZE_IMAGE),
"mean_except_zero")
masks_resize = np.zeros(
(len(masks), image.shape[0], image.shape[1]), dtype=np.float64)
masks_resize[0] = masks[0]
for j in range(1, len(masks)):
masks_resize[j] = resize(masks[j],
(image.shape[0], image.shape[1]),
mode='constant',
preserve_range=True)
copy_arr_to_arr(masks_resize[0],
masks_resize[j],
mode="mean_except_zero")
# Кодирование
mask_to_encde = masks_resize[0]
rle = list(prob_to_rles(mask_to_encde, cutoff=0.5))
rles.extend(rle)
new_test_ids.extend([id_] * len(rle))
if len(rle) == 0:
rles.extend([[1, 1]])
new_test_ids.extend([id_])
save_result("../../data/detector_unet_pred/{}.png".format(id_),
image, mask_to_encde, masks, bounding_boxs,
bounding_boxs_nms)
except StopIteration:
print("Exception id: " + id_)
# Create submission DataFrame
sub = pd.DataFrame()
sub['ImageId'] = new_test_ids
sub['EncodedPixels'] = pd.Series(rles).apply(
lambda x: ' '.join(str(y) for y in x))
# Чистка от мусора
def f(x):
res = 0
for i in range(1, len(x), 2):
res += i
return res
sub['CountPix'] = pd.Series(rles).apply(f)
sub = sub[sub['CountPix'] > 10][['ImageId', 'EncodedPixels']]
sub.to_csv('detector_Unet.csv', index=False)
def main_defolt():
warnings.filterwarnings('ignore', category=UserWarning, module='skimage')
new_test_ids = []
rles = []
detector = None
with CustomObjectScope({
'relu6':
applications.mobilenet.relu6,
'DepthwiseConv2D':
applications.mobilenet.DepthwiseConv2D
}):
detector = load_model('detector_MobileNet.h5')
u_net = load_model("U-net/Unet(32x32).h5",
custom_objects={'mean_iou': mean_iou})
dir_ = "../../data/stage1_test/"
ids_test = os.walk(dir_)
ids = next(ids_test)[1]
for n, id_ in tqdm(enumerate(ids), total=len(ids)):
try:
path = dir_ + id_ + "/images/"
image = imread(path + id_ + ".png", as_grey=True)
images = []
masks = []
# Уменьшаем размер картинки
cur_size = (image.shape[0], image.shape[1])
min_size = (image.shape[0] // MIN_RESIZE,
image.shape[1] // MIN_RESIZE)
images.append(image)
masks.append(np.zeros(cur_size, dtype=np.float64))
while SIZE_IMAGE[0] <= cur_size[0] >= min_size[0] and SIZE_IMAGE[
1] <= cur_size[1] >= min_size[1]:
cur_size = (int(cur_size[0] // COEF_RES),
int(cur_size[1] // COEF_RES))
if SIZE_IMAGE[0] > cur_size[0] or SIZE_IMAGE[1] > cur_size[1]:
break
images.append(
resize(image,
cur_size,
mode='constant',
preserve_range=True))
masks.append(np.zeros(cur_size, dtype=np.float64))
# masks[-1].fill(0.45)
# Предсказание
for i in range(len(images)):
im = images[i]
# Проход детектора
for h in range(SIZE_IMAGE[0], im.shape[0] + STRIDES[0],
STRIDES[0]):
lower_bound = min(h, im.shape[0] - 1)
for w in range(SIZE_IMAGE[1], im.shape[1] + STRIDES[1],
STRIDES[1]):
right_bound = min(w, im.shape[1] - 1)
part_image_analis = np.ndarray(
(1, SIZE_IMAGE[0], SIZE_IMAGE[1], 1),
dtype=np.float64)
part_image_analis[0] = im[lower_bound - SIZE_IMAGE[0]: lower_bound,
right_bound - SIZE_IMAGE[1]: right_bound] \
.reshape((SIZE_IMAGE[0], SIZE_IMAGE[1], 1))
is_nucl = detector.predict(part_image_analis)[0][0]
if is_nucl > 0.99:
pred_mask = u_net.predict(part_image_analis)[0]
copy_arr_to_arr(
masks[i][lower_bound -
SIZE_IMAGE[0]:lower_bound,
right_bound -
SIZE_IMAGE[1]:right_bound],
pred_mask.reshape(SIZE_IMAGE),
"mean_except_zero")
masks_resize = np.zeros(
(len(masks), image.shape[0], image.shape[1]), dtype=np.float64)
masks_resize[0] = masks[0]
for j in range(1, len(masks)):
masks_resize[j] = resize(masks[j],
(image.shape[0], image.shape[1]),
mode='constant',
preserve_range=True)
copy_arr_to_arr(masks_resize[0],
masks_resize[j],
mode="mean_except_zero")
# Кодирование
mask_to_encde = masks_resize[0]
rle = list(prob_to_rles(mask_to_encde, cutoff=0.5))
rles.extend(rle)
new_test_ids.extend([id_] * len(rle))
if len(rle) == 0:
rles.extend([[1, 1]])
new_test_ids.extend([id_])
# Вывод
asdd = np.array(masks_resize[0] > 0.5, dtype=np.float64)
if not os.path.isdir("../../data/detect/" + id_ + "/"):
os.makedirs("../../data/detect/" + id_ + "/")
imsave("../../data/detect/" + id_ + "/" + "source.png", image)
for i in range(len(masks)):
imsave("../../data/detect/" + id_ + "/" + str(i) + ".png",
masks[i])
fig = plt.figure(figsize=(8, 4))
fig.add_subplot(1, 3, 1)
plt.imshow(image)
plt.title("Исходное")
fig.add_subplot(1, 3, 2)
plt.imshow(mask_to_encde)
plt.title("Выход сети")
fig.add_subplot(1, 3, 3)
plt.imshow(np.array(mask_to_encde > 0.5, dtype=np.float64))
plt.title("Маска для сабмита")
# plt.show(block=True)
plt.savefig("../../data/detector_unet_pred/{}.png".format(id_))
plt.close()
except:
print("Exception id: " + id_)
# Create submission DataFrame
sub = pd.DataFrame()
sub['ImageId'] = new_test_ids
sub['EncodedPixels'] = pd.Series(rles).apply(
lambda x: ' '.join(str(y) for y in x))
sub.to_csv('detector_Unet.csv', index=False)