def generateGridLabel(x_path, y, img_size):
# scale_file = open(scale_path, 'rb')
# label_min = pickle.load(scale_file)
# label_max = pickle.load(scale_file)
# scale_file.close()
if len(img_size) != len(y):
print('The length of size_info and y is not same.')
os._exit(0)
if Check_Resize2:
for i in range(len(y)):
x_path[i] = '.' + x_path[i]
x, y = sr.sampleResize(x_path=x_path,
y=y,
im_size=img_size,
short_side=ShortSide,
long_side=LongSide,
resizeh=ResizeH,
resizew=ResizeW,
resizeX=True,
resizeY=True)
sp.show_bbox_in_one_image(x, y)
input('Stop')
y = sr.sampleResize(y=y,
im_size=img_size,
short_side=ShortSide,
long_side=LongSide,
resizeh=ResizeH,
resizew=ResizeW,
resizeY=True)
label_c, label_location = hs.relabel_2D_location(y, ResizeH, ResizeW,
DSR_x, DSR_y, CLASS,
CrossRatio)
# label_c = np_utils.to_categorical(label_c, num_classes=CLASS + 1)
# label_bbox = mms.vtMinMaxScaler(label_bbox, label_min, label_max)
return label_c, label_location
def checkResize(x_path, y, im_size):
for i in range(len(y)):
x_path[i] = '.' + x_path[i]
sx, sy = sr.sampleResize(x_path=x_path[i:i + 1],
y=y[i:i + 1],
im_size=im_size[i:i + 1],
short_side=ShortSide,
long_side=LongSide,
resizeh=ResizeH,
resizew=ResizeW,
resizeX=True,
resizeY=True)
# print(len(sx[0]),len(sy[0]))
print(x_path[i])
sp.show_bbox_in_one_image(sx, sy)
def getBatch(X_path, batch_size, label_path, val=False):
FMAP_w = ResizeW // DSR_x
FMAP_h = ResizeH // DSR_y
X_path = np.array(X_path)
while 1:
# 打乱数据非常重要
permutation = np.random.permutation(X_path.shape[0])
X_path = X_path[permutation]
# label_c = label_c[permutation, :, :, :]
# label_bbox = label_bbox[permutation, :, :, :]
# X, label_c, label_bbox = shuffleTogether(X_path, label_c, label_bbox)
print(X_path[0]) # 用于确认是否打乱
if val:
Epoch_len = X_path.shape[0]
else:
Epoch_len = X_path.shape[0] # Todo:如果要每次用部分训练样本，设置这里
for i in range(0, Epoch_len, batch_size):
x = sr.sampleResize(x_path=X_path[i:i + batch_size],
short_side=ShortSide,
long_side=LongSide,
resizeh=ResizeH,
resizew=ResizeW,
resizeX=True)
Y_number = len(X_path[i:i + batch_size])
cls = np.zeros((Y_number, FMAP_h, FMAP_w, 1))
loc = np.zeros((Y_number, FMAP_h, FMAP_w, 1))
for j in range(Y_number):
row = 0
Y_path = label_path + str(
os.path.basename(X_path[i + j]).split('.')[0]) + '.txt'
with open(Y_path, 'r') as f:
for line in f:
line_split = line.strip().split(' ')
cls[j][row // FMAP_w][row % FMAP_w][0] = int(
float(line_split[0]))
loc[j][row // FMAP_w][row % FMAP_w][0] = int(
float(line_split[1]))
row += 1
# cls = label_c[i:i + batch_size, :, :, :]
# bbox = label_bbox[i:i + batch_size, :, :, :]
# 最重要的就是这个yield，它代表返回，返回以后循环还是会继续，然后再返回。
# 就比如有一个机器一直在作累加运算，但是会把每次累加中间结果告诉你一样，直到把所有数加完
yield ({
'input': x
}, {
'cls_out': cls,
'loc_out': np_utils.to_categorical(loc, num_classes=5)
})
def addTest_ResizeInfo(test_img):
# start_time = time.time()
X_path, Y, size = getInformation(test_img)
# print(len(Y))
# x = sr.sampleResize(x_path=X_path, short_side=ShortSide, long_side=LongSide,
# resizeh=ResizeH, resizew=ResizeW, resizeX=True)
size_info, pad_info = sr.sampleResize(y=Y,
im_size=size,
short_side=ShortSide,
long_side=LongSide,
resizeh=ResizeH,
resizew=ResizeW,
resizeY=True,
sizeinfo=True)
for i in range(len(Y)):
# test_img[i]['pixel'] = x[i]
test_img[i]['resizeinfo'] = size_info[i]
test_img[i]['padinfo'] = pad_info[i]
# print(test_img[i]['filepath'])
# print(test_img[i]['resizeinfo'])
# print(test_img[i]['padinfo'])
# input('stop')
# print('The average time of reading an image is {} seconds.'.format((time.time() - start_time) / len(test_img)))
return test_img
def testFast(test_img, md, name_model):
# (height, width, channel) = test_img[0]['pixel'].shape
# test_x = np.zeros((len(test_img), height, widt h, channel))
predict_cls = np.zeros((len(test_img), ResizeH // DSR_y, ResizeW // DSR_x, 1))
predict_loc = np.zeros((len(test_img), ResizeH // DSR_y, ResizeW // DSR_x, 5))
# scale_file = open(scale_path, 'rb')
# label_min = pickle.load(scale_file)
# label_max = pickle.load(scale_file)
# scale_file.close()
X_path, Y, size = getInformation(test_img)
x = sr.sampleResize(x_path=X_path, short_side=ShortSide, long_side=LongSide,
resizeh=ResizeH, resizew=ResizeW, resizeX=True)
bbox = []
start = time.time()
for i in range(len(test_img)):
# predicting...
img = x[i]
test_x = img.reshape((1, img.shape[0], img.shape[1], img.shape[2]))
predict_cls[i], predict_loc[i] = md.predict(test_x)
box = pp.processLocation(img, copy.deepcopy(predict_cls[i]), copy.deepcopy(predict_loc[i]),
DSR_x, DSR_y, filterBox=FilterBBox)
bbox.append(box)
# inverse data standardizing
# predict_bbox = mms.vtMinMaxScaler_inverse(predict_bbox, label_min, label_max) # 有点疑问，是否要reshape
# handle bounding box
# 明天从这里写下去
# bbox = handleBBOX(predict_loc, predict_cls)
# current_model = ModelPath + name_model + '.hdf5'
# NMS
# for i in range(len(bbox)):
# bbox[i] = NMS.non_max_suppression_fast(bbox[i], overlap_thresh=0.5,
# max_boxes=20, current_model=name_model)
cost_time = time.time() - start
print('The cost time for an image is', str(cost_time / len(test_img)) + '.')
# if len(bbox) != len(test_img):
# print('The lengths of bboxes and images are inconsistent.')
# os._exit(0)
# 画小切块的类别
if GridImage:
ResultImage2_md = ResultImage + name_model + '_grid/'
if not os.path.exists(ResultImage2_md):
os.makedirs(ResultImage2_md)
for i in tqdm(range(x.shape[0])):
img_path = test_img[i]['filepath']
base_name = str(os.path.basename(img_path).split('.')[0])
img2 = x[i] * 255
img2 = sp.draw_grid_location(img2, predict_cls[i], predict_loc[i], CLASS)
result_image2_path = ResultImage2_md + base_name + '.png'
cv2.imwrite(result_image2_path, img2)
if LocationImage:
ResultImage3_md = ResultImage + name_model + '_location/'
if not os.path.exists(ResultImage3_md):
os.makedirs(ResultImage3_md)
for i in tqdm(range(x.shape[0])):
img_path = test_img[i]['filepath']
base_name = str(os.path.basename(img_path).split('.')[0])
img3 = x[i] * 255
img3 = pp.processLocation(img3, predict_cls[i], predict_loc[i], DSR_x, DSR_y, drawBBox=True,
filterBox=FilterBBox)
result_image3_path = ResultImage3_md + base_name + '.png'
cv2.imwrite(result_image3_path, img3)
# if ResizedImage_bbox:
# ResultImage3_md = ResultImage + name_model + '_bbox/'
# if not os.path.exists(ResultImage3_md):
# os.makedirs(ResultImage3_md)
# for i in tqdm(range(x.shape[0])):
# img_path = test_img[i]['filepath']
# base_name = str(os.path.basename(img_path).split('.')[0])
# img3 = x[i] * 255
# bboxes = {}
# for b in range(len(bbox[i])):
# class_name = int(bbox[i][b][5])
# class_prob = round(bbox[i][b][4], 2)
# x1 = round(bbox[i][b][0], 2)
# y1 = round(bbox[i][b][1], 2)
# x2 = round(bbox[i][b][2], 2)
# y2 = round(bbox[i][b][3], 2)
# if class_name not in bboxes:
# bboxes[class_name] = []
# bboxes[class_name].append([x1, y1, x2, y2, class_prob, class_name])
#
# img3 = sp.draw_bbox_in_one_image(img3, bboxes)
# # result_image_path = ResultImage_md + mark + base_name + '.png'
# result_image3_path = ResultImage3_md + base_name + '.png'
# cv2.imwrite(result_image3_path, img3)
#
for i in range(len(bbox)):
bbox[i] = sr.bboxesTransform_inverse(bbox[i], test_img[i]['resizeinfo'], test_img[i]['padinfo'], isWH=True)
return bbox